mirror of
https://github.com/modelscope/DiffSynth-Studio.git
synced 2026-03-18 22:08:13 +00:00
svd
This commit is contained in:
Artiprocher
118
diffsynth/extensions/ESRGAN/__init__.py
Normal file
118
diffsynth/extensions/ESRGAN/__init__.py
Normal file
@@ -0,0 +1,118 @@
|
||||
import torch
|
||||
from einops import repeat
|
||||
from PIL import Image
|
||||
import numpy as np
|
||||
|
||||
|
||||
class ResidualDenseBlock(torch.nn.Module):
|
||||
|
||||
def __init__(self, num_feat=64, num_grow_ch=32):
|
||||
super(ResidualDenseBlock, self).__init__()
|
||||
self.conv1 = torch.nn.Conv2d(num_feat, num_grow_ch, 3, 1, 1)
|
||||
self.conv2 = torch.nn.Conv2d(num_feat + num_grow_ch, num_grow_ch, 3, 1, 1)
|
||||
self.conv3 = torch.nn.Conv2d(num_feat + 2 * num_grow_ch, num_grow_ch, 3, 1, 1)
|
||||
self.conv4 = torch.nn.Conv2d(num_feat + 3 * num_grow_ch, num_grow_ch, 3, 1, 1)
|
||||
self.conv5 = torch.nn.Conv2d(num_feat + 4 * num_grow_ch, num_feat, 3, 1, 1)
|
||||
self.lrelu = torch.nn.LeakyReLU(negative_slope=0.2, inplace=True)
|
||||
|
||||
def forward(self, x):
|
||||
x1 = self.lrelu(self.conv1(x))
|
||||
x2 = self.lrelu(self.conv2(torch.cat((x, x1), 1)))
|
||||
x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), 1)))
|
||||
x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), 1)))
|
||||
x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1))
|
||||
return x5 * 0.2 + x
|
||||
|
||||
|
||||
class RRDB(torch.nn.Module):
|
||||
|
||||
def __init__(self, num_feat, num_grow_ch=32):
|
||||
super(RRDB, self).__init__()
|
||||
self.rdb1 = ResidualDenseBlock(num_feat, num_grow_ch)
|
||||
self.rdb2 = ResidualDenseBlock(num_feat, num_grow_ch)
|
||||
self.rdb3 = ResidualDenseBlock(num_feat, num_grow_ch)
|
||||
|
||||
def forward(self, x):
|
||||
out = self.rdb1(x)
|
||||
out = self.rdb2(out)
|
||||
out = self.rdb3(out)
|
||||
return out * 0.2 + x
|
||||
|
||||
|
||||
class RRDBNet(torch.nn.Module):
|
||||
|
||||
def __init__(self, num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32):
|
||||
super(RRDBNet, self).__init__()
|
||||
self.conv_first = torch.nn.Conv2d(num_in_ch, num_feat, 3, 1, 1)
|
||||
self.body = torch.torch.nn.Sequential(*[RRDB(num_feat=num_feat, num_grow_ch=num_grow_ch) for _ in range(num_block)])
|
||||
self.conv_body = torch.nn.Conv2d(num_feat, num_feat, 3, 1, 1)
|
||||
# upsample
|
||||
self.conv_up1 = torch.nn.Conv2d(num_feat, num_feat, 3, 1, 1)
|
||||
self.conv_up2 = torch.nn.Conv2d(num_feat, num_feat, 3, 1, 1)
|
||||
self.conv_hr = torch.nn.Conv2d(num_feat, num_feat, 3, 1, 1)
|
||||
self.conv_last = torch.nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
|
||||
self.lrelu = torch.nn.LeakyReLU(negative_slope=0.2, inplace=True)
|
||||
|
||||
def forward(self, x):
|
||||
feat = x
|
||||
feat = self.conv_first(feat)
|
||||
body_feat = self.conv_body(self.body(feat))
|
||||
feat = feat + body_feat
|
||||
# upsample
|
||||
feat = repeat(feat, "B C H W -> B C (H 2) (W 2)")
|
||||
feat = self.lrelu(self.conv_up1(feat))
|
||||
feat = repeat(feat, "B C H W -> B C (H 2) (W 2)")
|
||||
feat = self.lrelu(self.conv_up2(feat))
|
||||
out = self.conv_last(self.lrelu(self.conv_hr(feat)))
|
||||
return out
|
||||
|
||||
|
||||
class ESRGAN(torch.nn.Module):
|
||||
def __init__(self, model):
|
||||
super().__init__()
|
||||
self.model = model
|
||||
|
||||
@staticmethod
|
||||
def from_pretrained(model_path):
|
||||
model = RRDBNet()
|
||||
state_dict = torch.load(model_path, map_location="cpu")["params_ema"]
|
||||
model.load_state_dict(state_dict)
|
||||
model.eval()
|
||||
return ESRGAN(model)
|
||||
|
||||
def process_image(self, image):
|
||||
image = torch.Tensor(np.array(image, dtype=np.float32) / 255).permute(2, 0, 1)
|
||||
return image
|
||||
|
||||
def process_images(self, images):
|
||||
images = [self.process_image(image) for image in images]
|
||||
images = torch.stack(images)
|
||||
return images
|
||||
|
||||
def decode_images(self, images):
|
||||
images = (images.permute(0, 2, 3, 1) * 255).clip(0, 255).numpy().astype(np.uint8)
|
||||
images = [Image.fromarray(image) for image in images]
|
||||
return images
|
||||
|
||||
@torch.no_grad()
|
||||
def upscale(self, images, batch_size=4, progress_bar=lambda x:x):
|
||||
# Preprocess
|
||||
input_tensor = self.process_images(images)
|
||||
|
||||
# Interpolate
|
||||
output_tensor = []
|
||||
for batch_id in progress_bar(range(0, input_tensor.shape[0], batch_size)):
|
||||
batch_id_ = min(batch_id + batch_size, input_tensor.shape[0])
|
||||
batch_input_tensor = input_tensor[batch_id: batch_id_]
|
||||
batch_input_tensor = batch_input_tensor.to(
|
||||
device=self.model.conv_first.weight.device,
|
||||
dtype=self.model.conv_first.weight.dtype)
|
||||
batch_output_tensor = self.model(batch_input_tensor)
|
||||
output_tensor.append(batch_output_tensor.cpu())
|
||||
|
||||
# Output
|
||||
output_tensor = torch.concat(output_tensor, dim=0)
|
||||
|
||||
# To images
|
||||
output_images = self.decode_images(output_tensor)
|
||||
return output_images
|
||||
@@ -167,7 +167,7 @@ class RIFEInterpolater:
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def interpolate(self, images, scale=1.0, batch_size=4, num_iter=1):
|
||||
def interpolate(self, images, scale=1.0, batch_size=4, num_iter=1, progress_bar=lambda x:x):
|
||||
# Preprocess
|
||||
processed_images = self.process_images(images)
|
||||
|
||||
@@ -177,7 +177,7 @@ class RIFEInterpolater:
|
||||
|
||||
# Interpolate
|
||||
output_tensor = []
|
||||
for batch_id in range(0, input_tensor.shape[0], batch_size):
|
||||
for batch_id in progress_bar(range(0, input_tensor.shape[0], batch_size)):
|
||||
batch_id_ = min(batch_id + batch_size, input_tensor.shape[0])
|
||||
batch_input_tensor = input_tensor[batch_id: batch_id_]
|
||||
batch_input_tensor = batch_input_tensor.to(device=self.device, dtype=self.torch_dtype)
|
||||
|
||||
@@ -16,6 +16,11 @@ from .sd_controlnet import SDControlNet
|
||||
|
||||
from .sd_motion import SDMotionModel
|
||||
|
||||
from .svd_image_encoder import SVDImageEncoder
|
||||
from .svd_unet import SVDUNet
|
||||
from .svd_vae_decoder import SVDVAEDecoder
|
||||
from .svd_vae_encoder import SVDVAEEncoder
|
||||
|
||||
|
||||
class ModelManager:
|
||||
def __init__(self, torch_dtype=torch.float16, device="cuda"):
|
||||
@@ -25,6 +30,10 @@ class ModelManager:
|
||||
self.model_path = {}
|
||||
self.textual_inversion_dict = {}
|
||||
|
||||
def is_stable_video_diffusion(self, state_dict):
|
||||
param_name = "model.diffusion_model.output_blocks.9.1.time_stack.0.norm_in.weight"
|
||||
return param_name in state_dict
|
||||
|
||||
def is_RIFE(self, state_dict):
|
||||
param_name = "block_tea.convblock3.0.1.weight"
|
||||
return param_name in state_dict or ("module." + param_name) in state_dict
|
||||
@@ -60,6 +69,21 @@ class ModelManager:
|
||||
param_name = "model.encoder.layers.5.self_attn_layer_norm.weight"
|
||||
return param_name in state_dict and len(state_dict) == 254
|
||||
|
||||
def load_stable_video_diffusion(self, state_dict, components=None, file_path=""):
|
||||
component_dict = {
|
||||
"image_encoder": SVDImageEncoder,
|
||||
"unet": SVDUNet,
|
||||
"vae_decoder": SVDVAEDecoder,
|
||||
"vae_encoder": SVDVAEEncoder,
|
||||
}
|
||||
if components is None:
|
||||
components = ["image_encoder", "unet", "vae_decoder", "vae_encoder"]
|
||||
for component in components:
|
||||
self.model[component] = component_dict[component]()
|
||||
self.model[component].load_state_dict(self.model[component].state_dict_converter().from_civitai(state_dict))
|
||||
self.model[component].to(self.torch_dtype).to(self.device)
|
||||
self.model_path[component] = file_path
|
||||
|
||||
def load_stable_diffusion(self, state_dict, components=None, file_path=""):
|
||||
component_dict = {
|
||||
"text_encoder": SDTextEncoder,
|
||||
@@ -190,7 +214,9 @@ class ModelManager:
|
||||
|
||||
def load_model(self, file_path, components=None, lora_alphas=[]):
|
||||
state_dict = load_state_dict(file_path, torch_dtype=self.torch_dtype)
|
||||
if self.is_animatediff(state_dict):
|
||||
if self.is_stable_video_diffusion(state_dict):
|
||||
self.load_stable_video_diffusion(state_dict, file_path=file_path)
|
||||
elif self.is_animatediff(state_dict):
|
||||
self.load_animatediff(state_dict, file_path=file_path)
|
||||
elif self.is_controlnet(state_dict):
|
||||
self.load_controlnet(state_dict, file_path=file_path)
|
||||
|
||||
@@ -165,6 +165,8 @@ class AttentionBlock(torch.nn.Module):
|
||||
encoder_hidden_states = text_emb.mean(dim=0, keepdim=True)
|
||||
else:
|
||||
encoder_hidden_states = text_emb
|
||||
if encoder_hidden_states.shape[0] != hidden_states.shape[0]:
|
||||
encoder_hidden_states = encoder_hidden_states.repeat(hidden_states.shape[0], 1, 1)
|
||||
|
||||
if tiled:
|
||||
tile_size = min(tile_size, min(height, width))
|
||||
|
||||
@@ -87,3 +87,416 @@ class SVDImageEncoderStateDictConverter:
|
||||
name_ = ".".join(["encoders", layer_id, attn_rename_dict[layer_type], tail])
|
||||
state_dict_[name_] = param
|
||||
return state_dict_
|
||||
|
||||
def from_civitai(self, state_dict):
|
||||
rename_dict = {
|
||||
"conditioner.embedders.0.open_clip.model.visual.class_embedding": "embeddings.class_embedding",
|
||||
"conditioner.embedders.0.open_clip.model.visual.conv1.weight": "embeddings.patch_embedding.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.ln_post.bias": "post_layernorm.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.ln_post.weight": "post_layernorm.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.ln_pre.bias": "pre_layernorm.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.ln_pre.weight": "pre_layernorm.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.positional_embedding": "embeddings.position_embeds",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.0.attn.in_proj_bias": ['encoders.0.attn.to_q.bias', 'encoders.0.attn.to_k.bias', 'encoders.0.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.0.attn.in_proj_weight": ['encoders.0.attn.to_q.weight', 'encoders.0.attn.to_k.weight', 'encoders.0.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.0.attn.out_proj.bias": "encoders.0.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.0.attn.out_proj.weight": "encoders.0.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.0.ln_1.bias": "encoders.0.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.0.ln_1.weight": "encoders.0.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.0.ln_2.bias": "encoders.0.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.0.ln_2.weight": "encoders.0.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.0.mlp.c_fc.bias": "encoders.0.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.0.mlp.c_fc.weight": "encoders.0.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.0.mlp.c_proj.bias": "encoders.0.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.0.mlp.c_proj.weight": "encoders.0.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.1.attn.in_proj_bias": ['encoders.1.attn.to_q.bias', 'encoders.1.attn.to_k.bias', 'encoders.1.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.1.attn.in_proj_weight": ['encoders.1.attn.to_q.weight', 'encoders.1.attn.to_k.weight', 'encoders.1.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.1.attn.out_proj.bias": "encoders.1.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.1.attn.out_proj.weight": "encoders.1.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.1.ln_1.bias": "encoders.1.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.1.ln_1.weight": "encoders.1.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.1.ln_2.bias": "encoders.1.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.1.ln_2.weight": "encoders.1.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.1.mlp.c_fc.bias": "encoders.1.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.1.mlp.c_fc.weight": "encoders.1.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.1.mlp.c_proj.bias": "encoders.1.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.1.mlp.c_proj.weight": "encoders.1.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.10.attn.in_proj_bias": ['encoders.10.attn.to_q.bias', 'encoders.10.attn.to_k.bias', 'encoders.10.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.10.attn.in_proj_weight": ['encoders.10.attn.to_q.weight', 'encoders.10.attn.to_k.weight', 'encoders.10.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.10.attn.out_proj.bias": "encoders.10.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.10.attn.out_proj.weight": "encoders.10.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.10.ln_1.bias": "encoders.10.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.10.ln_1.weight": "encoders.10.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.10.ln_2.bias": "encoders.10.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.10.ln_2.weight": "encoders.10.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.10.mlp.c_fc.bias": "encoders.10.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.10.mlp.c_fc.weight": "encoders.10.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.10.mlp.c_proj.bias": "encoders.10.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.10.mlp.c_proj.weight": "encoders.10.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.11.attn.in_proj_bias": ['encoders.11.attn.to_q.bias', 'encoders.11.attn.to_k.bias', 'encoders.11.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.11.attn.in_proj_weight": ['encoders.11.attn.to_q.weight', 'encoders.11.attn.to_k.weight', 'encoders.11.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.11.attn.out_proj.bias": "encoders.11.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.11.attn.out_proj.weight": "encoders.11.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.11.ln_1.bias": "encoders.11.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.11.ln_1.weight": "encoders.11.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.11.ln_2.bias": "encoders.11.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.11.ln_2.weight": "encoders.11.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.11.mlp.c_fc.bias": "encoders.11.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.11.mlp.c_fc.weight": "encoders.11.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.11.mlp.c_proj.bias": "encoders.11.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.11.mlp.c_proj.weight": "encoders.11.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.12.attn.in_proj_bias": ['encoders.12.attn.to_q.bias', 'encoders.12.attn.to_k.bias', 'encoders.12.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.12.attn.in_proj_weight": ['encoders.12.attn.to_q.weight', 'encoders.12.attn.to_k.weight', 'encoders.12.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.12.attn.out_proj.bias": "encoders.12.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.12.attn.out_proj.weight": "encoders.12.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.12.ln_1.bias": "encoders.12.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.12.ln_1.weight": "encoders.12.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.12.ln_2.bias": "encoders.12.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.12.ln_2.weight": "encoders.12.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.12.mlp.c_fc.bias": "encoders.12.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.12.mlp.c_fc.weight": "encoders.12.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.12.mlp.c_proj.bias": "encoders.12.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.12.mlp.c_proj.weight": "encoders.12.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.13.attn.in_proj_bias": ['encoders.13.attn.to_q.bias', 'encoders.13.attn.to_k.bias', 'encoders.13.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.13.attn.in_proj_weight": ['encoders.13.attn.to_q.weight', 'encoders.13.attn.to_k.weight', 'encoders.13.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.13.attn.out_proj.bias": "encoders.13.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.13.attn.out_proj.weight": "encoders.13.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.13.ln_1.bias": "encoders.13.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.13.ln_1.weight": "encoders.13.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.13.ln_2.bias": "encoders.13.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.13.ln_2.weight": "encoders.13.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.13.mlp.c_fc.bias": "encoders.13.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.13.mlp.c_fc.weight": "encoders.13.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.13.mlp.c_proj.bias": "encoders.13.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.13.mlp.c_proj.weight": "encoders.13.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.14.attn.in_proj_bias": ['encoders.14.attn.to_q.bias', 'encoders.14.attn.to_k.bias', 'encoders.14.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.14.attn.in_proj_weight": ['encoders.14.attn.to_q.weight', 'encoders.14.attn.to_k.weight', 'encoders.14.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.14.attn.out_proj.bias": "encoders.14.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.14.attn.out_proj.weight": "encoders.14.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.14.ln_1.bias": "encoders.14.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.14.ln_1.weight": "encoders.14.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.14.ln_2.bias": "encoders.14.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.14.ln_2.weight": "encoders.14.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.14.mlp.c_fc.bias": "encoders.14.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.14.mlp.c_fc.weight": "encoders.14.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.14.mlp.c_proj.bias": "encoders.14.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.14.mlp.c_proj.weight": "encoders.14.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.15.attn.in_proj_bias": ['encoders.15.attn.to_q.bias', 'encoders.15.attn.to_k.bias', 'encoders.15.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.15.attn.in_proj_weight": ['encoders.15.attn.to_q.weight', 'encoders.15.attn.to_k.weight', 'encoders.15.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.15.attn.out_proj.bias": "encoders.15.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.15.attn.out_proj.weight": "encoders.15.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.15.ln_1.bias": "encoders.15.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.15.ln_1.weight": "encoders.15.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.15.ln_2.bias": "encoders.15.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.15.ln_2.weight": "encoders.15.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.15.mlp.c_fc.bias": "encoders.15.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.15.mlp.c_fc.weight": "encoders.15.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.15.mlp.c_proj.bias": "encoders.15.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.15.mlp.c_proj.weight": "encoders.15.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.16.attn.in_proj_bias": ['encoders.16.attn.to_q.bias', 'encoders.16.attn.to_k.bias', 'encoders.16.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.16.attn.in_proj_weight": ['encoders.16.attn.to_q.weight', 'encoders.16.attn.to_k.weight', 'encoders.16.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.16.attn.out_proj.bias": "encoders.16.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.16.attn.out_proj.weight": "encoders.16.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.16.ln_1.bias": "encoders.16.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.16.ln_1.weight": "encoders.16.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.16.ln_2.bias": "encoders.16.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.16.ln_2.weight": "encoders.16.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.16.mlp.c_fc.bias": "encoders.16.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.16.mlp.c_fc.weight": "encoders.16.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.16.mlp.c_proj.bias": "encoders.16.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.16.mlp.c_proj.weight": "encoders.16.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.17.attn.in_proj_bias": ['encoders.17.attn.to_q.bias', 'encoders.17.attn.to_k.bias', 'encoders.17.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.17.attn.in_proj_weight": ['encoders.17.attn.to_q.weight', 'encoders.17.attn.to_k.weight', 'encoders.17.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.17.attn.out_proj.bias": "encoders.17.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.17.attn.out_proj.weight": "encoders.17.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.17.ln_1.bias": "encoders.17.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.17.ln_1.weight": "encoders.17.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.17.ln_2.bias": "encoders.17.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.17.ln_2.weight": "encoders.17.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.17.mlp.c_fc.bias": "encoders.17.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.17.mlp.c_fc.weight": "encoders.17.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.17.mlp.c_proj.bias": "encoders.17.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.17.mlp.c_proj.weight": "encoders.17.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.18.attn.in_proj_bias": ['encoders.18.attn.to_q.bias', 'encoders.18.attn.to_k.bias', 'encoders.18.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.18.attn.in_proj_weight": ['encoders.18.attn.to_q.weight', 'encoders.18.attn.to_k.weight', 'encoders.18.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.18.attn.out_proj.bias": "encoders.18.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.18.attn.out_proj.weight": "encoders.18.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.18.ln_1.bias": "encoders.18.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.18.ln_1.weight": "encoders.18.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.18.ln_2.bias": "encoders.18.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.18.ln_2.weight": "encoders.18.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.18.mlp.c_fc.bias": "encoders.18.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.18.mlp.c_fc.weight": "encoders.18.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.18.mlp.c_proj.bias": "encoders.18.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.18.mlp.c_proj.weight": "encoders.18.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.19.attn.in_proj_bias": ['encoders.19.attn.to_q.bias', 'encoders.19.attn.to_k.bias', 'encoders.19.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.19.attn.in_proj_weight": ['encoders.19.attn.to_q.weight', 'encoders.19.attn.to_k.weight', 'encoders.19.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.19.attn.out_proj.bias": "encoders.19.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.19.attn.out_proj.weight": "encoders.19.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.19.ln_1.bias": "encoders.19.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.19.ln_1.weight": "encoders.19.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.19.ln_2.bias": "encoders.19.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.19.ln_2.weight": "encoders.19.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.19.mlp.c_fc.bias": "encoders.19.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.19.mlp.c_fc.weight": "encoders.19.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.19.mlp.c_proj.bias": "encoders.19.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.19.mlp.c_proj.weight": "encoders.19.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.2.attn.in_proj_bias": ['encoders.2.attn.to_q.bias', 'encoders.2.attn.to_k.bias', 'encoders.2.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.2.attn.in_proj_weight": ['encoders.2.attn.to_q.weight', 'encoders.2.attn.to_k.weight', 'encoders.2.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.2.attn.out_proj.bias": "encoders.2.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.2.attn.out_proj.weight": "encoders.2.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.2.ln_1.bias": "encoders.2.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.2.ln_1.weight": "encoders.2.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.2.ln_2.bias": "encoders.2.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.2.ln_2.weight": "encoders.2.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.2.mlp.c_fc.bias": "encoders.2.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.2.mlp.c_fc.weight": "encoders.2.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.2.mlp.c_proj.bias": "encoders.2.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.2.mlp.c_proj.weight": "encoders.2.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.20.attn.in_proj_bias": ['encoders.20.attn.to_q.bias', 'encoders.20.attn.to_k.bias', 'encoders.20.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.20.attn.in_proj_weight": ['encoders.20.attn.to_q.weight', 'encoders.20.attn.to_k.weight', 'encoders.20.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.20.attn.out_proj.bias": "encoders.20.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.20.attn.out_proj.weight": "encoders.20.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.20.ln_1.bias": "encoders.20.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.20.ln_1.weight": "encoders.20.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.20.ln_2.bias": "encoders.20.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.20.ln_2.weight": "encoders.20.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.20.mlp.c_fc.bias": "encoders.20.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.20.mlp.c_fc.weight": "encoders.20.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.20.mlp.c_proj.bias": "encoders.20.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.20.mlp.c_proj.weight": "encoders.20.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.21.attn.in_proj_bias": ['encoders.21.attn.to_q.bias', 'encoders.21.attn.to_k.bias', 'encoders.21.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.21.attn.in_proj_weight": ['encoders.21.attn.to_q.weight', 'encoders.21.attn.to_k.weight', 'encoders.21.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.21.attn.out_proj.bias": "encoders.21.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.21.attn.out_proj.weight": "encoders.21.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.21.ln_1.bias": "encoders.21.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.21.ln_1.weight": "encoders.21.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.21.ln_2.bias": "encoders.21.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.21.ln_2.weight": "encoders.21.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.21.mlp.c_fc.bias": "encoders.21.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.21.mlp.c_fc.weight": "encoders.21.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.21.mlp.c_proj.bias": "encoders.21.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.21.mlp.c_proj.weight": "encoders.21.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.22.attn.in_proj_bias": ['encoders.22.attn.to_q.bias', 'encoders.22.attn.to_k.bias', 'encoders.22.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.22.attn.in_proj_weight": ['encoders.22.attn.to_q.weight', 'encoders.22.attn.to_k.weight', 'encoders.22.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.22.attn.out_proj.bias": "encoders.22.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.22.attn.out_proj.weight": "encoders.22.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.22.ln_1.bias": "encoders.22.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.22.ln_1.weight": "encoders.22.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.22.ln_2.bias": "encoders.22.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.22.ln_2.weight": "encoders.22.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.22.mlp.c_fc.bias": "encoders.22.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.22.mlp.c_fc.weight": "encoders.22.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.22.mlp.c_proj.bias": "encoders.22.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.22.mlp.c_proj.weight": "encoders.22.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.23.attn.in_proj_bias": ['encoders.23.attn.to_q.bias', 'encoders.23.attn.to_k.bias', 'encoders.23.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.23.attn.in_proj_weight": ['encoders.23.attn.to_q.weight', 'encoders.23.attn.to_k.weight', 'encoders.23.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.23.attn.out_proj.bias": "encoders.23.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.23.attn.out_proj.weight": "encoders.23.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.23.ln_1.bias": "encoders.23.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.23.ln_1.weight": "encoders.23.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.23.ln_2.bias": "encoders.23.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.23.ln_2.weight": "encoders.23.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.23.mlp.c_fc.bias": "encoders.23.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.23.mlp.c_fc.weight": "encoders.23.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.23.mlp.c_proj.bias": "encoders.23.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.23.mlp.c_proj.weight": "encoders.23.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.24.attn.in_proj_bias": ['encoders.24.attn.to_q.bias', 'encoders.24.attn.to_k.bias', 'encoders.24.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.24.attn.in_proj_weight": ['encoders.24.attn.to_q.weight', 'encoders.24.attn.to_k.weight', 'encoders.24.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.24.attn.out_proj.bias": "encoders.24.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.24.attn.out_proj.weight": "encoders.24.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.24.ln_1.bias": "encoders.24.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.24.ln_1.weight": "encoders.24.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.24.ln_2.bias": "encoders.24.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.24.ln_2.weight": "encoders.24.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.24.mlp.c_fc.bias": "encoders.24.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.24.mlp.c_fc.weight": "encoders.24.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.24.mlp.c_proj.bias": "encoders.24.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.24.mlp.c_proj.weight": "encoders.24.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.25.attn.in_proj_bias": ['encoders.25.attn.to_q.bias', 'encoders.25.attn.to_k.bias', 'encoders.25.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.25.attn.in_proj_weight": ['encoders.25.attn.to_q.weight', 'encoders.25.attn.to_k.weight', 'encoders.25.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.25.attn.out_proj.bias": "encoders.25.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.25.attn.out_proj.weight": "encoders.25.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.25.ln_1.bias": "encoders.25.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.25.ln_1.weight": "encoders.25.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.25.ln_2.bias": "encoders.25.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.25.ln_2.weight": "encoders.25.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.25.mlp.c_fc.bias": "encoders.25.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.25.mlp.c_fc.weight": "encoders.25.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.25.mlp.c_proj.bias": "encoders.25.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.25.mlp.c_proj.weight": "encoders.25.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.26.attn.in_proj_bias": ['encoders.26.attn.to_q.bias', 'encoders.26.attn.to_k.bias', 'encoders.26.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.26.attn.in_proj_weight": ['encoders.26.attn.to_q.weight', 'encoders.26.attn.to_k.weight', 'encoders.26.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.26.attn.out_proj.bias": "encoders.26.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.26.attn.out_proj.weight": "encoders.26.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.26.ln_1.bias": "encoders.26.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.26.ln_1.weight": "encoders.26.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.26.ln_2.bias": "encoders.26.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.26.ln_2.weight": "encoders.26.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.26.mlp.c_fc.bias": "encoders.26.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.26.mlp.c_fc.weight": "encoders.26.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.26.mlp.c_proj.bias": "encoders.26.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.26.mlp.c_proj.weight": "encoders.26.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.27.attn.in_proj_bias": ['encoders.27.attn.to_q.bias', 'encoders.27.attn.to_k.bias', 'encoders.27.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.27.attn.in_proj_weight": ['encoders.27.attn.to_q.weight', 'encoders.27.attn.to_k.weight', 'encoders.27.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.27.attn.out_proj.bias": "encoders.27.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.27.attn.out_proj.weight": "encoders.27.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.27.ln_1.bias": "encoders.27.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.27.ln_1.weight": "encoders.27.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.27.ln_2.bias": "encoders.27.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.27.ln_2.weight": "encoders.27.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.27.mlp.c_fc.bias": "encoders.27.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.27.mlp.c_fc.weight": "encoders.27.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.27.mlp.c_proj.bias": "encoders.27.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.27.mlp.c_proj.weight": "encoders.27.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.28.attn.in_proj_bias": ['encoders.28.attn.to_q.bias', 'encoders.28.attn.to_k.bias', 'encoders.28.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.28.attn.in_proj_weight": ['encoders.28.attn.to_q.weight', 'encoders.28.attn.to_k.weight', 'encoders.28.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.28.attn.out_proj.bias": "encoders.28.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.28.attn.out_proj.weight": "encoders.28.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.28.ln_1.bias": "encoders.28.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.28.ln_1.weight": "encoders.28.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.28.ln_2.bias": "encoders.28.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.28.ln_2.weight": "encoders.28.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.28.mlp.c_fc.bias": "encoders.28.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.28.mlp.c_fc.weight": "encoders.28.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.28.mlp.c_proj.bias": "encoders.28.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.28.mlp.c_proj.weight": "encoders.28.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.29.attn.in_proj_bias": ['encoders.29.attn.to_q.bias', 'encoders.29.attn.to_k.bias', 'encoders.29.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.29.attn.in_proj_weight": ['encoders.29.attn.to_q.weight', 'encoders.29.attn.to_k.weight', 'encoders.29.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.29.attn.out_proj.bias": "encoders.29.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.29.attn.out_proj.weight": "encoders.29.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.29.ln_1.bias": "encoders.29.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.29.ln_1.weight": "encoders.29.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.29.ln_2.bias": "encoders.29.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.29.ln_2.weight": "encoders.29.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.29.mlp.c_fc.bias": "encoders.29.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.29.mlp.c_fc.weight": "encoders.29.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.29.mlp.c_proj.bias": "encoders.29.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.29.mlp.c_proj.weight": "encoders.29.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.3.attn.in_proj_bias": ['encoders.3.attn.to_q.bias', 'encoders.3.attn.to_k.bias', 'encoders.3.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.3.attn.in_proj_weight": ['encoders.3.attn.to_q.weight', 'encoders.3.attn.to_k.weight', 'encoders.3.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.3.attn.out_proj.bias": "encoders.3.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.3.attn.out_proj.weight": "encoders.3.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.3.ln_1.bias": "encoders.3.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.3.ln_1.weight": "encoders.3.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.3.ln_2.bias": "encoders.3.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.3.ln_2.weight": "encoders.3.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.3.mlp.c_fc.bias": "encoders.3.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.3.mlp.c_fc.weight": "encoders.3.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.3.mlp.c_proj.bias": "encoders.3.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.3.mlp.c_proj.weight": "encoders.3.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.30.attn.in_proj_bias": ['encoders.30.attn.to_q.bias', 'encoders.30.attn.to_k.bias', 'encoders.30.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.30.attn.in_proj_weight": ['encoders.30.attn.to_q.weight', 'encoders.30.attn.to_k.weight', 'encoders.30.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.30.attn.out_proj.bias": "encoders.30.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.30.attn.out_proj.weight": "encoders.30.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.30.ln_1.bias": "encoders.30.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.30.ln_1.weight": "encoders.30.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.30.ln_2.bias": "encoders.30.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.30.ln_2.weight": "encoders.30.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.30.mlp.c_fc.bias": "encoders.30.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.30.mlp.c_fc.weight": "encoders.30.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.30.mlp.c_proj.bias": "encoders.30.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.30.mlp.c_proj.weight": "encoders.30.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.31.attn.in_proj_bias": ['encoders.31.attn.to_q.bias', 'encoders.31.attn.to_k.bias', 'encoders.31.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.31.attn.in_proj_weight": ['encoders.31.attn.to_q.weight', 'encoders.31.attn.to_k.weight', 'encoders.31.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.31.attn.out_proj.bias": "encoders.31.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.31.attn.out_proj.weight": "encoders.31.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.31.ln_1.bias": "encoders.31.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.31.ln_1.weight": "encoders.31.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.31.ln_2.bias": "encoders.31.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.31.ln_2.weight": "encoders.31.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.31.mlp.c_fc.bias": "encoders.31.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.31.mlp.c_fc.weight": "encoders.31.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.31.mlp.c_proj.bias": "encoders.31.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.31.mlp.c_proj.weight": "encoders.31.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.4.attn.in_proj_bias": ['encoders.4.attn.to_q.bias', 'encoders.4.attn.to_k.bias', 'encoders.4.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.4.attn.in_proj_weight": ['encoders.4.attn.to_q.weight', 'encoders.4.attn.to_k.weight', 'encoders.4.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.4.attn.out_proj.bias": "encoders.4.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.4.attn.out_proj.weight": "encoders.4.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.4.ln_1.bias": "encoders.4.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.4.ln_1.weight": "encoders.4.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.4.ln_2.bias": "encoders.4.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.4.ln_2.weight": "encoders.4.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.4.mlp.c_fc.bias": "encoders.4.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.4.mlp.c_fc.weight": "encoders.4.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.4.mlp.c_proj.bias": "encoders.4.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.4.mlp.c_proj.weight": "encoders.4.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.5.attn.in_proj_bias": ['encoders.5.attn.to_q.bias', 'encoders.5.attn.to_k.bias', 'encoders.5.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.5.attn.in_proj_weight": ['encoders.5.attn.to_q.weight', 'encoders.5.attn.to_k.weight', 'encoders.5.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.5.attn.out_proj.bias": "encoders.5.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.5.attn.out_proj.weight": "encoders.5.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.5.ln_1.bias": "encoders.5.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.5.ln_1.weight": "encoders.5.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.5.ln_2.bias": "encoders.5.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.5.ln_2.weight": "encoders.5.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.5.mlp.c_fc.bias": "encoders.5.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.5.mlp.c_fc.weight": "encoders.5.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.5.mlp.c_proj.bias": "encoders.5.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.5.mlp.c_proj.weight": "encoders.5.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.6.attn.in_proj_bias": ['encoders.6.attn.to_q.bias', 'encoders.6.attn.to_k.bias', 'encoders.6.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.6.attn.in_proj_weight": ['encoders.6.attn.to_q.weight', 'encoders.6.attn.to_k.weight', 'encoders.6.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.6.attn.out_proj.bias": "encoders.6.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.6.attn.out_proj.weight": "encoders.6.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.6.ln_1.bias": "encoders.6.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.6.ln_1.weight": "encoders.6.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.6.ln_2.bias": "encoders.6.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.6.ln_2.weight": "encoders.6.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.6.mlp.c_fc.bias": "encoders.6.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.6.mlp.c_fc.weight": "encoders.6.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.6.mlp.c_proj.bias": "encoders.6.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.6.mlp.c_proj.weight": "encoders.6.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.7.attn.in_proj_bias": ['encoders.7.attn.to_q.bias', 'encoders.7.attn.to_k.bias', 'encoders.7.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.7.attn.in_proj_weight": ['encoders.7.attn.to_q.weight', 'encoders.7.attn.to_k.weight', 'encoders.7.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.7.attn.out_proj.bias": "encoders.7.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.7.attn.out_proj.weight": "encoders.7.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.7.ln_1.bias": "encoders.7.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.7.ln_1.weight": "encoders.7.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.7.ln_2.bias": "encoders.7.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.7.ln_2.weight": "encoders.7.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.7.mlp.c_fc.bias": "encoders.7.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.7.mlp.c_fc.weight": "encoders.7.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.7.mlp.c_proj.bias": "encoders.7.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.7.mlp.c_proj.weight": "encoders.7.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.8.attn.in_proj_bias": ['encoders.8.attn.to_q.bias', 'encoders.8.attn.to_k.bias', 'encoders.8.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.8.attn.in_proj_weight": ['encoders.8.attn.to_q.weight', 'encoders.8.attn.to_k.weight', 'encoders.8.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.8.attn.out_proj.bias": "encoders.8.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.8.attn.out_proj.weight": "encoders.8.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.8.ln_1.bias": "encoders.8.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.8.ln_1.weight": "encoders.8.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.8.ln_2.bias": "encoders.8.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.8.ln_2.weight": "encoders.8.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.8.mlp.c_fc.bias": "encoders.8.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.8.mlp.c_fc.weight": "encoders.8.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.8.mlp.c_proj.bias": "encoders.8.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.8.mlp.c_proj.weight": "encoders.8.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.9.attn.in_proj_bias": ['encoders.9.attn.to_q.bias', 'encoders.9.attn.to_k.bias', 'encoders.9.attn.to_v.bias'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.9.attn.in_proj_weight": ['encoders.9.attn.to_q.weight', 'encoders.9.attn.to_k.weight', 'encoders.9.attn.to_v.weight'],
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.9.attn.out_proj.bias": "encoders.9.attn.to_out.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.9.attn.out_proj.weight": "encoders.9.attn.to_out.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.9.ln_1.bias": "encoders.9.layer_norm1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.9.ln_1.weight": "encoders.9.layer_norm1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.9.ln_2.bias": "encoders.9.layer_norm2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.9.ln_2.weight": "encoders.9.layer_norm2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.9.mlp.c_fc.bias": "encoders.9.fc1.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.9.mlp.c_fc.weight": "encoders.9.fc1.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.9.mlp.c_proj.bias": "encoders.9.fc2.bias",
|
||||
"conditioner.embedders.0.open_clip.model.visual.transformer.resblocks.9.mlp.c_proj.weight": "encoders.9.fc2.weight",
|
||||
"conditioner.embedders.0.open_clip.model.visual.proj": "visual_projection.weight",
|
||||
}
|
||||
state_dict_ = {}
|
||||
for name in state_dict:
|
||||
if name in rename_dict:
|
||||
param = state_dict[name]
|
||||
if name == "conditioner.embedders.0.open_clip.model.visual.class_embedding":
|
||||
param = param.reshape((1, 1, param.shape[0]))
|
||||
elif name == "conditioner.embedders.0.open_clip.model.visual.positional_embedding":
|
||||
param = param.reshape((1, param.shape[0], param.shape[1]))
|
||||
elif name == "conditioner.embedders.0.open_clip.model.visual.proj":
|
||||
param = param.T
|
||||
if isinstance(rename_dict[name], str):
|
||||
state_dict_[rename_dict[name]] = param
|
||||
else:
|
||||
length = param.shape[0] // 3
|
||||
for i, rename in enumerate(rename_dict[name]):
|
||||
state_dict_[rename] = param[i*length: i*length+length]
|
||||
return state_dict_
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
@@ -2,7 +2,7 @@ import torch
|
||||
from .attention import Attention
|
||||
from .sd_unet import ResnetBlock, UpSampler
|
||||
from .tiler import TileWorker
|
||||
from einops import rearrange
|
||||
from einops import rearrange, repeat
|
||||
|
||||
|
||||
class VAEAttentionBlock(torch.nn.Module):
|
||||
@@ -119,14 +119,13 @@ class SVDVAEDecoder(torch.nn.Module):
|
||||
self.conv_out = torch.nn.Conv2d(128, 3, kernel_size=3, padding=1)
|
||||
self.time_conv_out = torch.nn.Conv3d(3, 3, kernel_size=(3, 1, 1), padding=(1, 0, 0))
|
||||
|
||||
|
||||
def forward(self, sample):
|
||||
# 1. pre-process
|
||||
hidden_states = sample.flatten(0, 1)
|
||||
hidden_states = rearrange(sample, "C T H W -> T C H W")
|
||||
hidden_states = hidden_states / self.scaling_factor
|
||||
hidden_states = self.conv_in(hidden_states)
|
||||
time_emb = None
|
||||
text_emb = None
|
||||
res_stack = None
|
||||
time_emb, text_emb, res_stack = None, None, None
|
||||
|
||||
# 2. blocks
|
||||
for i, block in enumerate(self.blocks):
|
||||
@@ -136,11 +135,70 @@ class SVDVAEDecoder(torch.nn.Module):
|
||||
hidden_states = self.conv_norm_out(hidden_states)
|
||||
hidden_states = self.conv_act(hidden_states)
|
||||
hidden_states = self.conv_out(hidden_states)
|
||||
hidden_states = rearrange(hidden_states, "T C H W -> 1 C T H W")
|
||||
hidden_states = rearrange(hidden_states, "T C H W -> C T H W")
|
||||
hidden_states = self.time_conv_out(hidden_states)
|
||||
|
||||
return hidden_states
|
||||
|
||||
|
||||
def build_mask(self, data, is_bound):
|
||||
_, T, H, W = data.shape
|
||||
t = repeat(torch.arange(T), "T -> T H W", T=T, H=H, W=W)
|
||||
h = repeat(torch.arange(H), "H -> T H W", T=T, H=H, W=W)
|
||||
w = repeat(torch.arange(W), "W -> T H W", T=T, H=H, W=W)
|
||||
border_width = (T + H + W) // 6
|
||||
pad = torch.ones_like(t) * border_width
|
||||
mask = torch.stack([
|
||||
pad if is_bound[0] else t + 1,
|
||||
pad if is_bound[1] else T - t,
|
||||
pad if is_bound[2] else h + 1,
|
||||
pad if is_bound[3] else H - h,
|
||||
pad if is_bound[4] else w + 1,
|
||||
pad if is_bound[5] else W - w
|
||||
]).min(dim=0).values
|
||||
mask = mask.clip(1, border_width)
|
||||
mask = (mask / border_width).to(dtype=data.dtype, device=data.device)
|
||||
mask = rearrange(mask, "T H W -> 1 T H W")
|
||||
return mask
|
||||
|
||||
|
||||
def decode_video(
|
||||
self, sample,
|
||||
batch_time=8, batch_height=128, batch_width=128,
|
||||
stride_time=4, stride_height=32, stride_width=32,
|
||||
progress_bar=lambda x:x
|
||||
):
|
||||
sample = sample.permute(1, 0, 2, 3)
|
||||
data_device = sample.device
|
||||
computation_device = self.conv_in.weight.device
|
||||
torch_dtype = sample.dtype
|
||||
_, T, H, W = sample.shape
|
||||
|
||||
weight = torch.zeros((1, T, H*8, W*8), dtype=torch_dtype, device=data_device)
|
||||
values = torch.zeros((3, T, H*8, W*8), dtype=torch_dtype, device=data_device)
|
||||
|
||||
# Split tasks
|
||||
tasks = []
|
||||
for t in range(0, T, stride_time):
|
||||
for h in range(0, H, stride_height):
|
||||
for w in range(0, W, stride_width):
|
||||
if (t-stride_time >= 0 and t-stride_time+batch_time >= T)\
|
||||
or (h-stride_height >= 0 and h-stride_height+batch_height >= H)\
|
||||
or (w-stride_width >= 0 and w-stride_width+batch_width >= W):
|
||||
continue
|
||||
tasks.append((t, t+batch_time, h, h+batch_height, w, w+batch_width))
|
||||
|
||||
# Run
|
||||
for tl, tr, hl, hr, wl, wr in progress_bar(tasks):
|
||||
sample_batch = sample[:, tl:tr, hl:hr, wl:wr].to(computation_device)
|
||||
sample_batch = self.forward(sample_batch).to(data_device)
|
||||
mask = self.build_mask(sample_batch, is_bound=(tl==0, tr>=T, hl==0, hr>=H, wl==0, wr>=W))
|
||||
values[:, tl:tr, hl*8:hr*8, wl*8:wr*8] += sample_batch * mask
|
||||
weight[:, tl:tr, hl*8:hr*8, wl*8:wr*8] += mask
|
||||
values /= weight
|
||||
return values
|
||||
|
||||
|
||||
def state_dict_converter(self):
|
||||
return SVDVAEDecoderStateDictConverter()
|
||||
|
||||
@@ -238,3 +296,282 @@ class SVDVAEDecoderStateDictConverter:
|
||||
state_dict_[name_] = state_dict[name]
|
||||
|
||||
return state_dict_
|
||||
|
||||
|
||||
def from_civitai(self, state_dict):
|
||||
rename_dict = {
|
||||
"first_stage_model.decoder.conv_in.bias": "conv_in.bias",
|
||||
"first_stage_model.decoder.conv_in.weight": "conv_in.weight",
|
||||
"first_stage_model.decoder.conv_out.bias": "conv_out.bias",
|
||||
"first_stage_model.decoder.conv_out.time_mix_conv.bias": "time_conv_out.bias",
|
||||
"first_stage_model.decoder.conv_out.time_mix_conv.weight": "time_conv_out.weight",
|
||||
"first_stage_model.decoder.conv_out.weight": "conv_out.weight",
|
||||
"first_stage_model.decoder.mid.attn_1.k.bias": "blocks.2.transformer_blocks.0.to_k.bias",
|
||||
"first_stage_model.decoder.mid.attn_1.k.weight": "blocks.2.transformer_blocks.0.to_k.weight",
|
||||
"first_stage_model.decoder.mid.attn_1.norm.bias": "blocks.2.norm.bias",
|
||||
"first_stage_model.decoder.mid.attn_1.norm.weight": "blocks.2.norm.weight",
|
||||
"first_stage_model.decoder.mid.attn_1.proj_out.bias": "blocks.2.transformer_blocks.0.to_out.bias",
|
||||
"first_stage_model.decoder.mid.attn_1.proj_out.weight": "blocks.2.transformer_blocks.0.to_out.weight",
|
||||
"first_stage_model.decoder.mid.attn_1.q.bias": "blocks.2.transformer_blocks.0.to_q.bias",
|
||||
"first_stage_model.decoder.mid.attn_1.q.weight": "blocks.2.transformer_blocks.0.to_q.weight",
|
||||
"first_stage_model.decoder.mid.attn_1.v.bias": "blocks.2.transformer_blocks.0.to_v.bias",
|
||||
"first_stage_model.decoder.mid.attn_1.v.weight": "blocks.2.transformer_blocks.0.to_v.weight",
|
||||
"first_stage_model.decoder.mid.block_1.conv1.bias": "blocks.0.conv1.bias",
|
||||
"first_stage_model.decoder.mid.block_1.conv1.weight": "blocks.0.conv1.weight",
|
||||
"first_stage_model.decoder.mid.block_1.conv2.bias": "blocks.0.conv2.bias",
|
||||
"first_stage_model.decoder.mid.block_1.conv2.weight": "blocks.0.conv2.weight",
|
||||
"first_stage_model.decoder.mid.block_1.mix_factor": "blocks.1.mix_factor",
|
||||
"first_stage_model.decoder.mid.block_1.norm1.bias": "blocks.0.norm1.bias",
|
||||
"first_stage_model.decoder.mid.block_1.norm1.weight": "blocks.0.norm1.weight",
|
||||
"first_stage_model.decoder.mid.block_1.norm2.bias": "blocks.0.norm2.bias",
|
||||
"first_stage_model.decoder.mid.block_1.norm2.weight": "blocks.0.norm2.weight",
|
||||
"first_stage_model.decoder.mid.block_1.time_stack.in_layers.0.bias": "blocks.1.norm1.bias",
|
||||
"first_stage_model.decoder.mid.block_1.time_stack.in_layers.0.weight": "blocks.1.norm1.weight",
|
||||
"first_stage_model.decoder.mid.block_1.time_stack.in_layers.2.bias": "blocks.1.conv1.bias",
|
||||
"first_stage_model.decoder.mid.block_1.time_stack.in_layers.2.weight": "blocks.1.conv1.weight",
|
||||
"first_stage_model.decoder.mid.block_1.time_stack.out_layers.0.bias": "blocks.1.norm2.bias",
|
||||
"first_stage_model.decoder.mid.block_1.time_stack.out_layers.0.weight": "blocks.1.norm2.weight",
|
||||
"first_stage_model.decoder.mid.block_1.time_stack.out_layers.3.bias": "blocks.1.conv2.bias",
|
||||
"first_stage_model.decoder.mid.block_1.time_stack.out_layers.3.weight": "blocks.1.conv2.weight",
|
||||
"first_stage_model.decoder.mid.block_2.conv1.bias": "blocks.3.conv1.bias",
|
||||
"first_stage_model.decoder.mid.block_2.conv1.weight": "blocks.3.conv1.weight",
|
||||
"first_stage_model.decoder.mid.block_2.conv2.bias": "blocks.3.conv2.bias",
|
||||
"first_stage_model.decoder.mid.block_2.conv2.weight": "blocks.3.conv2.weight",
|
||||
"first_stage_model.decoder.mid.block_2.mix_factor": "blocks.4.mix_factor",
|
||||
"first_stage_model.decoder.mid.block_2.norm1.bias": "blocks.3.norm1.bias",
|
||||
"first_stage_model.decoder.mid.block_2.norm1.weight": "blocks.3.norm1.weight",
|
||||
"first_stage_model.decoder.mid.block_2.norm2.bias": "blocks.3.norm2.bias",
|
||||
"first_stage_model.decoder.mid.block_2.norm2.weight": "blocks.3.norm2.weight",
|
||||
"first_stage_model.decoder.mid.block_2.time_stack.in_layers.0.bias": "blocks.4.norm1.bias",
|
||||
"first_stage_model.decoder.mid.block_2.time_stack.in_layers.0.weight": "blocks.4.norm1.weight",
|
||||
"first_stage_model.decoder.mid.block_2.time_stack.in_layers.2.bias": "blocks.4.conv1.bias",
|
||||
"first_stage_model.decoder.mid.block_2.time_stack.in_layers.2.weight": "blocks.4.conv1.weight",
|
||||
"first_stage_model.decoder.mid.block_2.time_stack.out_layers.0.bias": "blocks.4.norm2.bias",
|
||||
"first_stage_model.decoder.mid.block_2.time_stack.out_layers.0.weight": "blocks.4.norm2.weight",
|
||||
"first_stage_model.decoder.mid.block_2.time_stack.out_layers.3.bias": "blocks.4.conv2.bias",
|
||||
"first_stage_model.decoder.mid.block_2.time_stack.out_layers.3.weight": "blocks.4.conv2.weight",
|
||||
"first_stage_model.decoder.norm_out.bias": "conv_norm_out.bias",
|
||||
"first_stage_model.decoder.norm_out.weight": "conv_norm_out.weight",
|
||||
"first_stage_model.decoder.up.0.block.0.conv1.bias": "blocks.26.conv1.bias",
|
||||
"first_stage_model.decoder.up.0.block.0.conv1.weight": "blocks.26.conv1.weight",
|
||||
"first_stage_model.decoder.up.0.block.0.conv2.bias": "blocks.26.conv2.bias",
|
||||
"first_stage_model.decoder.up.0.block.0.conv2.weight": "blocks.26.conv2.weight",
|
||||
"first_stage_model.decoder.up.0.block.0.mix_factor": "blocks.27.mix_factor",
|
||||
"first_stage_model.decoder.up.0.block.0.nin_shortcut.bias": "blocks.26.conv_shortcut.bias",
|
||||
"first_stage_model.decoder.up.0.block.0.nin_shortcut.weight": "blocks.26.conv_shortcut.weight",
|
||||
"first_stage_model.decoder.up.0.block.0.norm1.bias": "blocks.26.norm1.bias",
|
||||
"first_stage_model.decoder.up.0.block.0.norm1.weight": "blocks.26.norm1.weight",
|
||||
"first_stage_model.decoder.up.0.block.0.norm2.bias": "blocks.26.norm2.bias",
|
||||
"first_stage_model.decoder.up.0.block.0.norm2.weight": "blocks.26.norm2.weight",
|
||||
"first_stage_model.decoder.up.0.block.0.time_stack.in_layers.0.bias": "blocks.27.norm1.bias",
|
||||
"first_stage_model.decoder.up.0.block.0.time_stack.in_layers.0.weight": "blocks.27.norm1.weight",
|
||||
"first_stage_model.decoder.up.0.block.0.time_stack.in_layers.2.bias": "blocks.27.conv1.bias",
|
||||
"first_stage_model.decoder.up.0.block.0.time_stack.in_layers.2.weight": "blocks.27.conv1.weight",
|
||||
"first_stage_model.decoder.up.0.block.0.time_stack.out_layers.0.bias": "blocks.27.norm2.bias",
|
||||
"first_stage_model.decoder.up.0.block.0.time_stack.out_layers.0.weight": "blocks.27.norm2.weight",
|
||||
"first_stage_model.decoder.up.0.block.0.time_stack.out_layers.3.bias": "blocks.27.conv2.bias",
|
||||
"first_stage_model.decoder.up.0.block.0.time_stack.out_layers.3.weight": "blocks.27.conv2.weight",
|
||||
"first_stage_model.decoder.up.0.block.1.conv1.bias": "blocks.28.conv1.bias",
|
||||
"first_stage_model.decoder.up.0.block.1.conv1.weight": "blocks.28.conv1.weight",
|
||||
"first_stage_model.decoder.up.0.block.1.conv2.bias": "blocks.28.conv2.bias",
|
||||
"first_stage_model.decoder.up.0.block.1.conv2.weight": "blocks.28.conv2.weight",
|
||||
"first_stage_model.decoder.up.0.block.1.mix_factor": "blocks.29.mix_factor",
|
||||
"first_stage_model.decoder.up.0.block.1.norm1.bias": "blocks.28.norm1.bias",
|
||||
"first_stage_model.decoder.up.0.block.1.norm1.weight": "blocks.28.norm1.weight",
|
||||
"first_stage_model.decoder.up.0.block.1.norm2.bias": "blocks.28.norm2.bias",
|
||||
"first_stage_model.decoder.up.0.block.1.norm2.weight": "blocks.28.norm2.weight",
|
||||
"first_stage_model.decoder.up.0.block.1.time_stack.in_layers.0.bias": "blocks.29.norm1.bias",
|
||||
"first_stage_model.decoder.up.0.block.1.time_stack.in_layers.0.weight": "blocks.29.norm1.weight",
|
||||
"first_stage_model.decoder.up.0.block.1.time_stack.in_layers.2.bias": "blocks.29.conv1.bias",
|
||||
"first_stage_model.decoder.up.0.block.1.time_stack.in_layers.2.weight": "blocks.29.conv1.weight",
|
||||
"first_stage_model.decoder.up.0.block.1.time_stack.out_layers.0.bias": "blocks.29.norm2.bias",
|
||||
"first_stage_model.decoder.up.0.block.1.time_stack.out_layers.0.weight": "blocks.29.norm2.weight",
|
||||
"first_stage_model.decoder.up.0.block.1.time_stack.out_layers.3.bias": "blocks.29.conv2.bias",
|
||||
"first_stage_model.decoder.up.0.block.1.time_stack.out_layers.3.weight": "blocks.29.conv2.weight",
|
||||
"first_stage_model.decoder.up.0.block.2.conv1.bias": "blocks.30.conv1.bias",
|
||||
"first_stage_model.decoder.up.0.block.2.conv1.weight": "blocks.30.conv1.weight",
|
||||
"first_stage_model.decoder.up.0.block.2.conv2.bias": "blocks.30.conv2.bias",
|
||||
"first_stage_model.decoder.up.0.block.2.conv2.weight": "blocks.30.conv2.weight",
|
||||
"first_stage_model.decoder.up.0.block.2.mix_factor": "blocks.31.mix_factor",
|
||||
"first_stage_model.decoder.up.0.block.2.norm1.bias": "blocks.30.norm1.bias",
|
||||
"first_stage_model.decoder.up.0.block.2.norm1.weight": "blocks.30.norm1.weight",
|
||||
"first_stage_model.decoder.up.0.block.2.norm2.bias": "blocks.30.norm2.bias",
|
||||
"first_stage_model.decoder.up.0.block.2.norm2.weight": "blocks.30.norm2.weight",
|
||||
"first_stage_model.decoder.up.0.block.2.time_stack.in_layers.0.bias": "blocks.31.norm1.bias",
|
||||
"first_stage_model.decoder.up.0.block.2.time_stack.in_layers.0.weight": "blocks.31.norm1.weight",
|
||||
"first_stage_model.decoder.up.0.block.2.time_stack.in_layers.2.bias": "blocks.31.conv1.bias",
|
||||
"first_stage_model.decoder.up.0.block.2.time_stack.in_layers.2.weight": "blocks.31.conv1.weight",
|
||||
"first_stage_model.decoder.up.0.block.2.time_stack.out_layers.0.bias": "blocks.31.norm2.bias",
|
||||
"first_stage_model.decoder.up.0.block.2.time_stack.out_layers.0.weight": "blocks.31.norm2.weight",
|
||||
"first_stage_model.decoder.up.0.block.2.time_stack.out_layers.3.bias": "blocks.31.conv2.bias",
|
||||
"first_stage_model.decoder.up.0.block.2.time_stack.out_layers.3.weight": "blocks.31.conv2.weight",
|
||||
"first_stage_model.decoder.up.1.block.0.conv1.bias": "blocks.19.conv1.bias",
|
||||
"first_stage_model.decoder.up.1.block.0.conv1.weight": "blocks.19.conv1.weight",
|
||||
"first_stage_model.decoder.up.1.block.0.conv2.bias": "blocks.19.conv2.bias",
|
||||
"first_stage_model.decoder.up.1.block.0.conv2.weight": "blocks.19.conv2.weight",
|
||||
"first_stage_model.decoder.up.1.block.0.mix_factor": "blocks.20.mix_factor",
|
||||
"first_stage_model.decoder.up.1.block.0.nin_shortcut.bias": "blocks.19.conv_shortcut.bias",
|
||||
"first_stage_model.decoder.up.1.block.0.nin_shortcut.weight": "blocks.19.conv_shortcut.weight",
|
||||
"first_stage_model.decoder.up.1.block.0.norm1.bias": "blocks.19.norm1.bias",
|
||||
"first_stage_model.decoder.up.1.block.0.norm1.weight": "blocks.19.norm1.weight",
|
||||
"first_stage_model.decoder.up.1.block.0.norm2.bias": "blocks.19.norm2.bias",
|
||||
"first_stage_model.decoder.up.1.block.0.norm2.weight": "blocks.19.norm2.weight",
|
||||
"first_stage_model.decoder.up.1.block.0.time_stack.in_layers.0.bias": "blocks.20.norm1.bias",
|
||||
"first_stage_model.decoder.up.1.block.0.time_stack.in_layers.0.weight": "blocks.20.norm1.weight",
|
||||
"first_stage_model.decoder.up.1.block.0.time_stack.in_layers.2.bias": "blocks.20.conv1.bias",
|
||||
"first_stage_model.decoder.up.1.block.0.time_stack.in_layers.2.weight": "blocks.20.conv1.weight",
|
||||
"first_stage_model.decoder.up.1.block.0.time_stack.out_layers.0.bias": "blocks.20.norm2.bias",
|
||||
"first_stage_model.decoder.up.1.block.0.time_stack.out_layers.0.weight": "blocks.20.norm2.weight",
|
||||
"first_stage_model.decoder.up.1.block.0.time_stack.out_layers.3.bias": "blocks.20.conv2.bias",
|
||||
"first_stage_model.decoder.up.1.block.0.time_stack.out_layers.3.weight": "blocks.20.conv2.weight",
|
||||
"first_stage_model.decoder.up.1.block.1.conv1.bias": "blocks.21.conv1.bias",
|
||||
"first_stage_model.decoder.up.1.block.1.conv1.weight": "blocks.21.conv1.weight",
|
||||
"first_stage_model.decoder.up.1.block.1.conv2.bias": "blocks.21.conv2.bias",
|
||||
"first_stage_model.decoder.up.1.block.1.conv2.weight": "blocks.21.conv2.weight",
|
||||
"first_stage_model.decoder.up.1.block.1.mix_factor": "blocks.22.mix_factor",
|
||||
"first_stage_model.decoder.up.1.block.1.norm1.bias": "blocks.21.norm1.bias",
|
||||
"first_stage_model.decoder.up.1.block.1.norm1.weight": "blocks.21.norm1.weight",
|
||||
"first_stage_model.decoder.up.1.block.1.norm2.bias": "blocks.21.norm2.bias",
|
||||
"first_stage_model.decoder.up.1.block.1.norm2.weight": "blocks.21.norm2.weight",
|
||||
"first_stage_model.decoder.up.1.block.1.time_stack.in_layers.0.bias": "blocks.22.norm1.bias",
|
||||
"first_stage_model.decoder.up.1.block.1.time_stack.in_layers.0.weight": "blocks.22.norm1.weight",
|
||||
"first_stage_model.decoder.up.1.block.1.time_stack.in_layers.2.bias": "blocks.22.conv1.bias",
|
||||
"first_stage_model.decoder.up.1.block.1.time_stack.in_layers.2.weight": "blocks.22.conv1.weight",
|
||||
"first_stage_model.decoder.up.1.block.1.time_stack.out_layers.0.bias": "blocks.22.norm2.bias",
|
||||
"first_stage_model.decoder.up.1.block.1.time_stack.out_layers.0.weight": "blocks.22.norm2.weight",
|
||||
"first_stage_model.decoder.up.1.block.1.time_stack.out_layers.3.bias": "blocks.22.conv2.bias",
|
||||
"first_stage_model.decoder.up.1.block.1.time_stack.out_layers.3.weight": "blocks.22.conv2.weight",
|
||||
"first_stage_model.decoder.up.1.block.2.conv1.bias": "blocks.23.conv1.bias",
|
||||
"first_stage_model.decoder.up.1.block.2.conv1.weight": "blocks.23.conv1.weight",
|
||||
"first_stage_model.decoder.up.1.block.2.conv2.bias": "blocks.23.conv2.bias",
|
||||
"first_stage_model.decoder.up.1.block.2.conv2.weight": "blocks.23.conv2.weight",
|
||||
"first_stage_model.decoder.up.1.block.2.mix_factor": "blocks.24.mix_factor",
|
||||
"first_stage_model.decoder.up.1.block.2.norm1.bias": "blocks.23.norm1.bias",
|
||||
"first_stage_model.decoder.up.1.block.2.norm1.weight": "blocks.23.norm1.weight",
|
||||
"first_stage_model.decoder.up.1.block.2.norm2.bias": "blocks.23.norm2.bias",
|
||||
"first_stage_model.decoder.up.1.block.2.norm2.weight": "blocks.23.norm2.weight",
|
||||
"first_stage_model.decoder.up.1.block.2.time_stack.in_layers.0.bias": "blocks.24.norm1.bias",
|
||||
"first_stage_model.decoder.up.1.block.2.time_stack.in_layers.0.weight": "blocks.24.norm1.weight",
|
||||
"first_stage_model.decoder.up.1.block.2.time_stack.in_layers.2.bias": "blocks.24.conv1.bias",
|
||||
"first_stage_model.decoder.up.1.block.2.time_stack.in_layers.2.weight": "blocks.24.conv1.weight",
|
||||
"first_stage_model.decoder.up.1.block.2.time_stack.out_layers.0.bias": "blocks.24.norm2.bias",
|
||||
"first_stage_model.decoder.up.1.block.2.time_stack.out_layers.0.weight": "blocks.24.norm2.weight",
|
||||
"first_stage_model.decoder.up.1.block.2.time_stack.out_layers.3.bias": "blocks.24.conv2.bias",
|
||||
"first_stage_model.decoder.up.1.block.2.time_stack.out_layers.3.weight": "blocks.24.conv2.weight",
|
||||
"first_stage_model.decoder.up.1.upsample.conv.bias": "blocks.25.conv.bias",
|
||||
"first_stage_model.decoder.up.1.upsample.conv.weight": "blocks.25.conv.weight",
|
||||
"first_stage_model.decoder.up.2.block.0.conv1.bias": "blocks.12.conv1.bias",
|
||||
"first_stage_model.decoder.up.2.block.0.conv1.weight": "blocks.12.conv1.weight",
|
||||
"first_stage_model.decoder.up.2.block.0.conv2.bias": "blocks.12.conv2.bias",
|
||||
"first_stage_model.decoder.up.2.block.0.conv2.weight": "blocks.12.conv2.weight",
|
||||
"first_stage_model.decoder.up.2.block.0.mix_factor": "blocks.13.mix_factor",
|
||||
"first_stage_model.decoder.up.2.block.0.norm1.bias": "blocks.12.norm1.bias",
|
||||
"first_stage_model.decoder.up.2.block.0.norm1.weight": "blocks.12.norm1.weight",
|
||||
"first_stage_model.decoder.up.2.block.0.norm2.bias": "blocks.12.norm2.bias",
|
||||
"first_stage_model.decoder.up.2.block.0.norm2.weight": "blocks.12.norm2.weight",
|
||||
"first_stage_model.decoder.up.2.block.0.time_stack.in_layers.0.bias": "blocks.13.norm1.bias",
|
||||
"first_stage_model.decoder.up.2.block.0.time_stack.in_layers.0.weight": "blocks.13.norm1.weight",
|
||||
"first_stage_model.decoder.up.2.block.0.time_stack.in_layers.2.bias": "blocks.13.conv1.bias",
|
||||
"first_stage_model.decoder.up.2.block.0.time_stack.in_layers.2.weight": "blocks.13.conv1.weight",
|
||||
"first_stage_model.decoder.up.2.block.0.time_stack.out_layers.0.bias": "blocks.13.norm2.bias",
|
||||
"first_stage_model.decoder.up.2.block.0.time_stack.out_layers.0.weight": "blocks.13.norm2.weight",
|
||||
"first_stage_model.decoder.up.2.block.0.time_stack.out_layers.3.bias": "blocks.13.conv2.bias",
|
||||
"first_stage_model.decoder.up.2.block.0.time_stack.out_layers.3.weight": "blocks.13.conv2.weight",
|
||||
"first_stage_model.decoder.up.2.block.1.conv1.bias": "blocks.14.conv1.bias",
|
||||
"first_stage_model.decoder.up.2.block.1.conv1.weight": "blocks.14.conv1.weight",
|
||||
"first_stage_model.decoder.up.2.block.1.conv2.bias": "blocks.14.conv2.bias",
|
||||
"first_stage_model.decoder.up.2.block.1.conv2.weight": "blocks.14.conv2.weight",
|
||||
"first_stage_model.decoder.up.2.block.1.mix_factor": "blocks.15.mix_factor",
|
||||
"first_stage_model.decoder.up.2.block.1.norm1.bias": "blocks.14.norm1.bias",
|
||||
"first_stage_model.decoder.up.2.block.1.norm1.weight": "blocks.14.norm1.weight",
|
||||
"first_stage_model.decoder.up.2.block.1.norm2.bias": "blocks.14.norm2.bias",
|
||||
"first_stage_model.decoder.up.2.block.1.norm2.weight": "blocks.14.norm2.weight",
|
||||
"first_stage_model.decoder.up.2.block.1.time_stack.in_layers.0.bias": "blocks.15.norm1.bias",
|
||||
"first_stage_model.decoder.up.2.block.1.time_stack.in_layers.0.weight": "blocks.15.norm1.weight",
|
||||
"first_stage_model.decoder.up.2.block.1.time_stack.in_layers.2.bias": "blocks.15.conv1.bias",
|
||||
"first_stage_model.decoder.up.2.block.1.time_stack.in_layers.2.weight": "blocks.15.conv1.weight",
|
||||
"first_stage_model.decoder.up.2.block.1.time_stack.out_layers.0.bias": "blocks.15.norm2.bias",
|
||||
"first_stage_model.decoder.up.2.block.1.time_stack.out_layers.0.weight": "blocks.15.norm2.weight",
|
||||
"first_stage_model.decoder.up.2.block.1.time_stack.out_layers.3.bias": "blocks.15.conv2.bias",
|
||||
"first_stage_model.decoder.up.2.block.1.time_stack.out_layers.3.weight": "blocks.15.conv2.weight",
|
||||
"first_stage_model.decoder.up.2.block.2.conv1.bias": "blocks.16.conv1.bias",
|
||||
"first_stage_model.decoder.up.2.block.2.conv1.weight": "blocks.16.conv1.weight",
|
||||
"first_stage_model.decoder.up.2.block.2.conv2.bias": "blocks.16.conv2.bias",
|
||||
"first_stage_model.decoder.up.2.block.2.conv2.weight": "blocks.16.conv2.weight",
|
||||
"first_stage_model.decoder.up.2.block.2.mix_factor": "blocks.17.mix_factor",
|
||||
"first_stage_model.decoder.up.2.block.2.norm1.bias": "blocks.16.norm1.bias",
|
||||
"first_stage_model.decoder.up.2.block.2.norm1.weight": "blocks.16.norm1.weight",
|
||||
"first_stage_model.decoder.up.2.block.2.norm2.bias": "blocks.16.norm2.bias",
|
||||
"first_stage_model.decoder.up.2.block.2.norm2.weight": "blocks.16.norm2.weight",
|
||||
"first_stage_model.decoder.up.2.block.2.time_stack.in_layers.0.bias": "blocks.17.norm1.bias",
|
||||
"first_stage_model.decoder.up.2.block.2.time_stack.in_layers.0.weight": "blocks.17.norm1.weight",
|
||||
"first_stage_model.decoder.up.2.block.2.time_stack.in_layers.2.bias": "blocks.17.conv1.bias",
|
||||
"first_stage_model.decoder.up.2.block.2.time_stack.in_layers.2.weight": "blocks.17.conv1.weight",
|
||||
"first_stage_model.decoder.up.2.block.2.time_stack.out_layers.0.bias": "blocks.17.norm2.bias",
|
||||
"first_stage_model.decoder.up.2.block.2.time_stack.out_layers.0.weight": "blocks.17.norm2.weight",
|
||||
"first_stage_model.decoder.up.2.block.2.time_stack.out_layers.3.bias": "blocks.17.conv2.bias",
|
||||
"first_stage_model.decoder.up.2.block.2.time_stack.out_layers.3.weight": "blocks.17.conv2.weight",
|
||||
"first_stage_model.decoder.up.2.upsample.conv.bias": "blocks.18.conv.bias",
|
||||
"first_stage_model.decoder.up.2.upsample.conv.weight": "blocks.18.conv.weight",
|
||||
"first_stage_model.decoder.up.3.block.0.conv1.bias": "blocks.5.conv1.bias",
|
||||
"first_stage_model.decoder.up.3.block.0.conv1.weight": "blocks.5.conv1.weight",
|
||||
"first_stage_model.decoder.up.3.block.0.conv2.bias": "blocks.5.conv2.bias",
|
||||
"first_stage_model.decoder.up.3.block.0.conv2.weight": "blocks.5.conv2.weight",
|
||||
"first_stage_model.decoder.up.3.block.0.mix_factor": "blocks.6.mix_factor",
|
||||
"first_stage_model.decoder.up.3.block.0.norm1.bias": "blocks.5.norm1.bias",
|
||||
"first_stage_model.decoder.up.3.block.0.norm1.weight": "blocks.5.norm1.weight",
|
||||
"first_stage_model.decoder.up.3.block.0.norm2.bias": "blocks.5.norm2.bias",
|
||||
"first_stage_model.decoder.up.3.block.0.norm2.weight": "blocks.5.norm2.weight",
|
||||
"first_stage_model.decoder.up.3.block.0.time_stack.in_layers.0.bias": "blocks.6.norm1.bias",
|
||||
"first_stage_model.decoder.up.3.block.0.time_stack.in_layers.0.weight": "blocks.6.norm1.weight",
|
||||
"first_stage_model.decoder.up.3.block.0.time_stack.in_layers.2.bias": "blocks.6.conv1.bias",
|
||||
"first_stage_model.decoder.up.3.block.0.time_stack.in_layers.2.weight": "blocks.6.conv1.weight",
|
||||
"first_stage_model.decoder.up.3.block.0.time_stack.out_layers.0.bias": "blocks.6.norm2.bias",
|
||||
"first_stage_model.decoder.up.3.block.0.time_stack.out_layers.0.weight": "blocks.6.norm2.weight",
|
||||
"first_stage_model.decoder.up.3.block.0.time_stack.out_layers.3.bias": "blocks.6.conv2.bias",
|
||||
"first_stage_model.decoder.up.3.block.0.time_stack.out_layers.3.weight": "blocks.6.conv2.weight",
|
||||
"first_stage_model.decoder.up.3.block.1.conv1.bias": "blocks.7.conv1.bias",
|
||||
"first_stage_model.decoder.up.3.block.1.conv1.weight": "blocks.7.conv1.weight",
|
||||
"first_stage_model.decoder.up.3.block.1.conv2.bias": "blocks.7.conv2.bias",
|
||||
"first_stage_model.decoder.up.3.block.1.conv2.weight": "blocks.7.conv2.weight",
|
||||
"first_stage_model.decoder.up.3.block.1.mix_factor": "blocks.8.mix_factor",
|
||||
"first_stage_model.decoder.up.3.block.1.norm1.bias": "blocks.7.norm1.bias",
|
||||
"first_stage_model.decoder.up.3.block.1.norm1.weight": "blocks.7.norm1.weight",
|
||||
"first_stage_model.decoder.up.3.block.1.norm2.bias": "blocks.7.norm2.bias",
|
||||
"first_stage_model.decoder.up.3.block.1.norm2.weight": "blocks.7.norm2.weight",
|
||||
"first_stage_model.decoder.up.3.block.1.time_stack.in_layers.0.bias": "blocks.8.norm1.bias",
|
||||
"first_stage_model.decoder.up.3.block.1.time_stack.in_layers.0.weight": "blocks.8.norm1.weight",
|
||||
"first_stage_model.decoder.up.3.block.1.time_stack.in_layers.2.bias": "blocks.8.conv1.bias",
|
||||
"first_stage_model.decoder.up.3.block.1.time_stack.in_layers.2.weight": "blocks.8.conv1.weight",
|
||||
"first_stage_model.decoder.up.3.block.1.time_stack.out_layers.0.bias": "blocks.8.norm2.bias",
|
||||
"first_stage_model.decoder.up.3.block.1.time_stack.out_layers.0.weight": "blocks.8.norm2.weight",
|
||||
"first_stage_model.decoder.up.3.block.1.time_stack.out_layers.3.bias": "blocks.8.conv2.bias",
|
||||
"first_stage_model.decoder.up.3.block.1.time_stack.out_layers.3.weight": "blocks.8.conv2.weight",
|
||||
"first_stage_model.decoder.up.3.block.2.conv1.bias": "blocks.9.conv1.bias",
|
||||
"first_stage_model.decoder.up.3.block.2.conv1.weight": "blocks.9.conv1.weight",
|
||||
"first_stage_model.decoder.up.3.block.2.conv2.bias": "blocks.9.conv2.bias",
|
||||
"first_stage_model.decoder.up.3.block.2.conv2.weight": "blocks.9.conv2.weight",
|
||||
"first_stage_model.decoder.up.3.block.2.mix_factor": "blocks.10.mix_factor",
|
||||
"first_stage_model.decoder.up.3.block.2.norm1.bias": "blocks.9.norm1.bias",
|
||||
"first_stage_model.decoder.up.3.block.2.norm1.weight": "blocks.9.norm1.weight",
|
||||
"first_stage_model.decoder.up.3.block.2.norm2.bias": "blocks.9.norm2.bias",
|
||||
"first_stage_model.decoder.up.3.block.2.norm2.weight": "blocks.9.norm2.weight",
|
||||
"first_stage_model.decoder.up.3.block.2.time_stack.in_layers.0.bias": "blocks.10.norm1.bias",
|
||||
"first_stage_model.decoder.up.3.block.2.time_stack.in_layers.0.weight": "blocks.10.norm1.weight",
|
||||
"first_stage_model.decoder.up.3.block.2.time_stack.in_layers.2.bias": "blocks.10.conv1.bias",
|
||||
"first_stage_model.decoder.up.3.block.2.time_stack.in_layers.2.weight": "blocks.10.conv1.weight",
|
||||
"first_stage_model.decoder.up.3.block.2.time_stack.out_layers.0.bias": "blocks.10.norm2.bias",
|
||||
"first_stage_model.decoder.up.3.block.2.time_stack.out_layers.0.weight": "blocks.10.norm2.weight",
|
||||
"first_stage_model.decoder.up.3.block.2.time_stack.out_layers.3.bias": "blocks.10.conv2.bias",
|
||||
"first_stage_model.decoder.up.3.block.2.time_stack.out_layers.3.weight": "blocks.10.conv2.weight",
|
||||
"first_stage_model.decoder.up.3.upsample.conv.bias": "blocks.11.conv.bias",
|
||||
"first_stage_model.decoder.up.3.upsample.conv.weight": "blocks.11.conv.weight",
|
||||
}
|
||||
state_dict_ = {}
|
||||
for name in state_dict:
|
||||
if name in rename_dict:
|
||||
param = state_dict[name]
|
||||
if "blocks.2.transformer_blocks.0" in rename_dict[name]:
|
||||
param = param.squeeze()
|
||||
state_dict_[rename_dict[name]] = param
|
||||
return state_dict_
|
||||
|
||||
138
diffsynth/models/svd_vae_encoder.py
Normal file
138
diffsynth/models/svd_vae_encoder.py
Normal file
@@ -0,0 +1,138 @@
|
||||
from .sd_vae_encoder import SDVAEEncoderStateDictConverter, SDVAEEncoder
|
||||
|
||||
|
||||
class SVDVAEEncoder(SDVAEEncoder):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.scaling_factor = 0.13025
|
||||
|
||||
def state_dict_converter(self):
|
||||
return SVDVAEEncoderStateDictConverter()
|
||||
|
||||
|
||||
class SVDVAEEncoderStateDictConverter(SDVAEEncoderStateDictConverter):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
|
||||
def from_diffusers(self, state_dict):
|
||||
return super().from_diffusers(state_dict)
|
||||
|
||||
def from_civitai(self, state_dict):
|
||||
rename_dict = {
|
||||
"conditioner.embedders.3.encoder.encoder.conv_in.bias": "conv_in.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.conv_in.weight": "conv_in.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.conv_out.bias": "conv_out.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.conv_out.weight": "conv_out.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.0.conv1.bias": "blocks.0.conv1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.0.conv1.weight": "blocks.0.conv1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.0.conv2.bias": "blocks.0.conv2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.0.conv2.weight": "blocks.0.conv2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.0.norm1.bias": "blocks.0.norm1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.0.norm1.weight": "blocks.0.norm1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.0.norm2.bias": "blocks.0.norm2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.0.norm2.weight": "blocks.0.norm2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.1.conv1.bias": "blocks.1.conv1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.1.conv1.weight": "blocks.1.conv1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.1.conv2.bias": "blocks.1.conv2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.1.conv2.weight": "blocks.1.conv2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.1.norm1.bias": "blocks.1.norm1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.1.norm1.weight": "blocks.1.norm1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.1.norm2.bias": "blocks.1.norm2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.block.1.norm2.weight": "blocks.1.norm2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.downsample.conv.bias": "blocks.2.conv.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.0.downsample.conv.weight": "blocks.2.conv.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.0.conv1.bias": "blocks.3.conv1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.0.conv1.weight": "blocks.3.conv1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.0.conv2.bias": "blocks.3.conv2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.0.conv2.weight": "blocks.3.conv2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.0.nin_shortcut.bias": "blocks.3.conv_shortcut.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.0.nin_shortcut.weight": "blocks.3.conv_shortcut.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.0.norm1.bias": "blocks.3.norm1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.0.norm1.weight": "blocks.3.norm1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.0.norm2.bias": "blocks.3.norm2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.0.norm2.weight": "blocks.3.norm2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.1.conv1.bias": "blocks.4.conv1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.1.conv1.weight": "blocks.4.conv1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.1.conv2.bias": "blocks.4.conv2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.1.conv2.weight": "blocks.4.conv2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.1.norm1.bias": "blocks.4.norm1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.1.norm1.weight": "blocks.4.norm1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.1.norm2.bias": "blocks.4.norm2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.block.1.norm2.weight": "blocks.4.norm2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.downsample.conv.bias": "blocks.5.conv.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.1.downsample.conv.weight": "blocks.5.conv.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.0.conv1.bias": "blocks.6.conv1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.0.conv1.weight": "blocks.6.conv1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.0.conv2.bias": "blocks.6.conv2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.0.conv2.weight": "blocks.6.conv2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.0.nin_shortcut.bias": "blocks.6.conv_shortcut.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.0.nin_shortcut.weight": "blocks.6.conv_shortcut.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.0.norm1.bias": "blocks.6.norm1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.0.norm1.weight": "blocks.6.norm1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.0.norm2.bias": "blocks.6.norm2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.0.norm2.weight": "blocks.6.norm2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.1.conv1.bias": "blocks.7.conv1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.1.conv1.weight": "blocks.7.conv1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.1.conv2.bias": "blocks.7.conv2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.1.conv2.weight": "blocks.7.conv2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.1.norm1.bias": "blocks.7.norm1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.1.norm1.weight": "blocks.7.norm1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.1.norm2.bias": "blocks.7.norm2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.block.1.norm2.weight": "blocks.7.norm2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.downsample.conv.bias": "blocks.8.conv.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.2.downsample.conv.weight": "blocks.8.conv.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.0.conv1.bias": "blocks.9.conv1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.0.conv1.weight": "blocks.9.conv1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.0.conv2.bias": "blocks.9.conv2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.0.conv2.weight": "blocks.9.conv2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.0.norm1.bias": "blocks.9.norm1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.0.norm1.weight": "blocks.9.norm1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.0.norm2.bias": "blocks.9.norm2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.0.norm2.weight": "blocks.9.norm2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.1.conv1.bias": "blocks.10.conv1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.1.conv1.weight": "blocks.10.conv1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.1.conv2.bias": "blocks.10.conv2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.1.conv2.weight": "blocks.10.conv2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.1.norm1.bias": "blocks.10.norm1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.1.norm1.weight": "blocks.10.norm1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.1.norm2.bias": "blocks.10.norm2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.down.3.block.1.norm2.weight": "blocks.10.norm2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.attn_1.k.bias": "blocks.12.transformer_blocks.0.to_k.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.attn_1.k.weight": "blocks.12.transformer_blocks.0.to_k.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.attn_1.norm.bias": "blocks.12.norm.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.attn_1.norm.weight": "blocks.12.norm.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.attn_1.proj_out.bias": "blocks.12.transformer_blocks.0.to_out.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.attn_1.proj_out.weight": "blocks.12.transformer_blocks.0.to_out.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.attn_1.q.bias": "blocks.12.transformer_blocks.0.to_q.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.attn_1.q.weight": "blocks.12.transformer_blocks.0.to_q.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.attn_1.v.bias": "blocks.12.transformer_blocks.0.to_v.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.attn_1.v.weight": "blocks.12.transformer_blocks.0.to_v.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_1.conv1.bias": "blocks.11.conv1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_1.conv1.weight": "blocks.11.conv1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_1.conv2.bias": "blocks.11.conv2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_1.conv2.weight": "blocks.11.conv2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_1.norm1.bias": "blocks.11.norm1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_1.norm1.weight": "blocks.11.norm1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_1.norm2.bias": "blocks.11.norm2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_1.norm2.weight": "blocks.11.norm2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_2.conv1.bias": "blocks.13.conv1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_2.conv1.weight": "blocks.13.conv1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_2.conv2.bias": "blocks.13.conv2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_2.conv2.weight": "blocks.13.conv2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_2.norm1.bias": "blocks.13.norm1.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_2.norm1.weight": "blocks.13.norm1.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_2.norm2.bias": "blocks.13.norm2.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.mid.block_2.norm2.weight": "blocks.13.norm2.weight",
|
||||
"conditioner.embedders.3.encoder.encoder.norm_out.bias": "conv_norm_out.bias",
|
||||
"conditioner.embedders.3.encoder.encoder.norm_out.weight": "conv_norm_out.weight",
|
||||
"conditioner.embedders.3.encoder.quant_conv.bias": "quant_conv.bias",
|
||||
"conditioner.embedders.3.encoder.quant_conv.weight": "quant_conv.weight",
|
||||
}
|
||||
state_dict_ = {}
|
||||
for name in state_dict:
|
||||
if name in rename_dict:
|
||||
param = state_dict[name]
|
||||
if "transformer_blocks" in rename_dict[name]:
|
||||
param = param.squeeze()
|
||||
state_dict_[rename_dict[name]] = param
|
||||
return state_dict_
|
||||
@@ -1,3 +1,4 @@
|
||||
from .stable_diffusion import SDImagePipeline
|
||||
from .stable_diffusion_xl import SDXLImagePipeline
|
||||
from .stable_diffusion_video import SDVideoPipeline, SDVideoPipelineRunner
|
||||
from .stable_video_diffusion import SVDVideoPipeline
|
||||
|
||||
289
diffsynth/pipelines/stable_video_diffusion.py
Normal file
289
diffsynth/pipelines/stable_video_diffusion.py
Normal file
@@ -0,0 +1,289 @@
|
||||
from ..models import ModelManager, SVDImageEncoder, SVDUNet, SVDVAEEncoder, SVDVAEDecoder
|
||||
from ..schedulers import ContinuousODEScheduler
|
||||
from ..data import save_video
|
||||
import torch
|
||||
from tqdm import tqdm
|
||||
from PIL import Image
|
||||
import numpy as np
|
||||
from einops import rearrange, repeat
|
||||
|
||||
|
||||
|
||||
class SVDVideoPipeline(torch.nn.Module):
|
||||
|
||||
def __init__(self, device="cuda", torch_dtype=torch.float16):
|
||||
super().__init__()
|
||||
self.scheduler = ContinuousODEScheduler()
|
||||
self.device = device
|
||||
self.torch_dtype = torch_dtype
|
||||
# models
|
||||
self.image_encoder: SVDImageEncoder = None
|
||||
self.unet: SVDUNet = None
|
||||
self.vae_encoder: SVDVAEEncoder = None
|
||||
self.vae_decoder: SVDVAEDecoder = None
|
||||
|
||||
|
||||
def fetch_main_models(self, model_manager: ModelManager):
|
||||
self.image_encoder = model_manager.image_encoder
|
||||
self.unet = model_manager.unet
|
||||
self.vae_encoder = model_manager.vae_encoder
|
||||
self.vae_decoder = model_manager.vae_decoder
|
||||
|
||||
|
||||
@staticmethod
|
||||
def from_model_manager(model_manager: ModelManager, **kwargs):
|
||||
pipe = SVDVideoPipeline(device=model_manager.device, torch_dtype=model_manager.torch_dtype)
|
||||
pipe.fetch_main_models(model_manager)
|
||||
return pipe
|
||||
|
||||
|
||||
def preprocess_image(self, image):
|
||||
image = torch.Tensor(np.array(image, dtype=np.float32) * (2 / 255) - 1).permute(2, 0, 1).unsqueeze(0)
|
||||
return image
|
||||
|
||||
|
||||
def decode_image(self, latent, tiled=False, tile_size=64, tile_stride=32):
|
||||
image = self.vae_decoder(latent.to(self.device), tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)[0]
|
||||
image = image.cpu().permute(1, 2, 0).numpy()
|
||||
image = Image.fromarray(((image / 2 + 0.5).clip(0, 1) * 255).astype("uint8"))
|
||||
return image
|
||||
|
||||
|
||||
def encode_image_with_clip(self, image):
|
||||
image = self.preprocess_image(image).to(device=self.device, dtype=self.torch_dtype)
|
||||
image = SVDCLIPImageProcessor().resize_with_antialiasing(image, (224, 224))
|
||||
image = (image + 1.0) / 2.0
|
||||
mean = torch.tensor([0.48145466, 0.4578275, 0.40821073]).reshape(1, 3, 1, 1).to(device=self.device, dtype=self.torch_dtype)
|
||||
std = torch.tensor([0.26862954, 0.26130258, 0.27577711]).reshape(1, 3, 1, 1).to(device=self.device, dtype=self.torch_dtype)
|
||||
image = (image - mean) / std
|
||||
image_emb = self.image_encoder(image)
|
||||
return image_emb
|
||||
|
||||
|
||||
def encode_image_with_vae(self, image, noise_aug_strength):
|
||||
image = self.preprocess_image(image).to(device=self.device, dtype=self.torch_dtype)
|
||||
noise = torch.randn(image.shape, device="cpu", dtype=self.torch_dtype).to(self.device)
|
||||
image = image + noise_aug_strength * noise
|
||||
image_emb = self.vae_encoder(image) / self.vae_encoder.scaling_factor
|
||||
return image_emb
|
||||
|
||||
|
||||
def encode_video_with_vae(self, video):
|
||||
video = torch.concat([self.preprocess_image(frame) for frame in video], dim=0)
|
||||
video = rearrange(video, "T C H W -> 1 C T H W")
|
||||
video = video.to(device=self.device, dtype=self.torch_dtype)
|
||||
latents = self.vae_encoder.encode_video(video)
|
||||
latents = rearrange(latents[0], "C T H W -> T C H W")
|
||||
return latents
|
||||
|
||||
|
||||
def tensor2video(self, frames):
|
||||
frames = rearrange(frames, "C T H W -> T H W C")
|
||||
frames = ((frames.float() + 1) * 127.5).clip(0, 255).cpu().numpy().astype(np.uint8)
|
||||
frames = [Image.fromarray(frame) for frame in frames]
|
||||
return frames
|
||||
|
||||
|
||||
def calculate_noise_pred(
|
||||
self,
|
||||
latents,
|
||||
timestep,
|
||||
add_time_id,
|
||||
cfg_scales,
|
||||
image_emb_vae_posi, image_emb_clip_posi,
|
||||
image_emb_vae_nega, image_emb_clip_nega
|
||||
):
|
||||
latents_input = self.scheduler.scale_model_input(latents, timestep)
|
||||
|
||||
# Positive side
|
||||
noise_pred_posi = self.unet(
|
||||
torch.cat([latents_input, image_emb_vae_posi], dim=1),
|
||||
timestep, image_emb_clip_posi, add_time_id
|
||||
)
|
||||
# Negative side
|
||||
noise_pred_nega = self.unet(
|
||||
torch.cat([latents_input, image_emb_vae_nega], dim=1),
|
||||
timestep, image_emb_clip_nega, add_time_id
|
||||
)
|
||||
|
||||
# Classifier-free guidance
|
||||
noise_pred = noise_pred_nega + cfg_scales * (noise_pred_posi - noise_pred_nega)
|
||||
|
||||
return noise_pred
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def __call__(
|
||||
self,
|
||||
input_image=None,
|
||||
input_video=None,
|
||||
min_cfg_scale=1.0,
|
||||
max_cfg_scale=3.0,
|
||||
denoising_strength=1.0,
|
||||
num_frames=25,
|
||||
height=576,
|
||||
width=1024,
|
||||
fps=7,
|
||||
motion_bucket_id=127,
|
||||
noise_aug_strength=0.02,
|
||||
num_inference_steps=20,
|
||||
progress_bar_cmd=tqdm,
|
||||
progress_bar_st=None,
|
||||
):
|
||||
# Prepare scheduler
|
||||
self.scheduler.set_timesteps(num_inference_steps, denoising_strength=denoising_strength)
|
||||
|
||||
# Prepare latent tensors
|
||||
noise = torch.randn((num_frames, 4, height//8, width//8), device="cpu", dtype=self.torch_dtype).to(self.device)
|
||||
if denoising_strength == 1.0:
|
||||
latents = noise * self.scheduler.init_noise_sigma
|
||||
else:
|
||||
latents = self.encode_video_with_vae(input_video)
|
||||
latents = self.scheduler.add_noise(latents, noise, self.scheduler.timesteps[0])
|
||||
|
||||
# Encode image
|
||||
image_emb_clip_posi = self.encode_image_with_clip(input_image)
|
||||
image_emb_clip_nega = torch.zeros_like(image_emb_clip_posi)
|
||||
image_emb_vae_posi = repeat(self.encode_image_with_vae(input_image, noise_aug_strength), "B C H W -> (B T) C H W", T=num_frames)
|
||||
image_emb_vae_nega = torch.zeros_like(image_emb_vae_posi)
|
||||
|
||||
# Prepare classifier-free guidance
|
||||
cfg_scales = torch.linspace(min_cfg_scale, max_cfg_scale, num_frames)
|
||||
cfg_scales = cfg_scales.reshape(num_frames, 1, 1, 1).to(device=self.device, dtype=self.torch_dtype)
|
||||
|
||||
# Prepare positional id
|
||||
add_time_id = torch.tensor([[fps-1, motion_bucket_id, noise_aug_strength]], device=self.device)
|
||||
|
||||
# Denoise
|
||||
for progress_id, timestep in enumerate(progress_bar_cmd(self.scheduler.timesteps)):
|
||||
|
||||
# Fetch model output
|
||||
noise_pred = self.calculate_noise_pred(
|
||||
latents, timestep, add_time_id, cfg_scales,
|
||||
image_emb_vae_posi, image_emb_clip_posi, image_emb_vae_nega, image_emb_clip_nega
|
||||
)
|
||||
|
||||
# Forward Euler
|
||||
latents = self.scheduler.step(noise_pred, timestep, latents)
|
||||
|
||||
# Update progress bar
|
||||
if progress_bar_st is not None:
|
||||
progress_bar_st.progress(progress_id / len(self.scheduler.timesteps))
|
||||
|
||||
# Decode image
|
||||
video = self.vae_decoder.decode_video(latents, progress_bar=progress_bar_cmd)
|
||||
video = self.tensor2video(video)
|
||||
|
||||
return video
|
||||
|
||||
|
||||
|
||||
class SVDCLIPImageProcessor:
|
||||
def __init__(self):
|
||||
pass
|
||||
|
||||
def resize_with_antialiasing(self, input, size, interpolation="bicubic", align_corners=True):
|
||||
h, w = input.shape[-2:]
|
||||
factors = (h / size[0], w / size[1])
|
||||
|
||||
# First, we have to determine sigma
|
||||
# Taken from skimage: https://github.com/scikit-image/scikit-image/blob/v0.19.2/skimage/transform/_warps.py#L171
|
||||
sigmas = (
|
||||
max((factors[0] - 1.0) / 2.0, 0.001),
|
||||
max((factors[1] - 1.0) / 2.0, 0.001),
|
||||
)
|
||||
|
||||
# Now kernel size. Good results are for 3 sigma, but that is kind of slow. Pillow uses 1 sigma
|
||||
# https://github.com/python-pillow/Pillow/blob/master/src/libImaging/Resample.c#L206
|
||||
# But they do it in the 2 passes, which gives better results. Let's try 2 sigmas for now
|
||||
ks = int(max(2.0 * 2 * sigmas[0], 3)), int(max(2.0 * 2 * sigmas[1], 3))
|
||||
|
||||
# Make sure it is odd
|
||||
if (ks[0] % 2) == 0:
|
||||
ks = ks[0] + 1, ks[1]
|
||||
|
||||
if (ks[1] % 2) == 0:
|
||||
ks = ks[0], ks[1] + 1
|
||||
|
||||
input = self._gaussian_blur2d(input, ks, sigmas)
|
||||
|
||||
output = torch.nn.functional.interpolate(input, size=size, mode=interpolation, align_corners=align_corners)
|
||||
return output
|
||||
|
||||
|
||||
def _compute_padding(self, kernel_size):
|
||||
"""Compute padding tuple."""
|
||||
# 4 or 6 ints: (padding_left, padding_right,padding_top,padding_bottom)
|
||||
# https://pytorch.org/docs/stable/nn.html#torch.nn.functional.pad
|
||||
if len(kernel_size) < 2:
|
||||
raise AssertionError(kernel_size)
|
||||
computed = [k - 1 for k in kernel_size]
|
||||
|
||||
# for even kernels we need to do asymmetric padding :(
|
||||
out_padding = 2 * len(kernel_size) * [0]
|
||||
|
||||
for i in range(len(kernel_size)):
|
||||
computed_tmp = computed[-(i + 1)]
|
||||
|
||||
pad_front = computed_tmp // 2
|
||||
pad_rear = computed_tmp - pad_front
|
||||
|
||||
out_padding[2 * i + 0] = pad_front
|
||||
out_padding[2 * i + 1] = pad_rear
|
||||
|
||||
return out_padding
|
||||
|
||||
|
||||
def _filter2d(self, input, kernel):
|
||||
# prepare kernel
|
||||
b, c, h, w = input.shape
|
||||
tmp_kernel = kernel[:, None, ...].to(device=input.device, dtype=input.dtype)
|
||||
|
||||
tmp_kernel = tmp_kernel.expand(-1, c, -1, -1)
|
||||
|
||||
height, width = tmp_kernel.shape[-2:]
|
||||
|
||||
padding_shape: list[int] = self._compute_padding([height, width])
|
||||
input = torch.nn.functional.pad(input, padding_shape, mode="reflect")
|
||||
|
||||
# kernel and input tensor reshape to align element-wise or batch-wise params
|
||||
tmp_kernel = tmp_kernel.reshape(-1, 1, height, width)
|
||||
input = input.view(-1, tmp_kernel.size(0), input.size(-2), input.size(-1))
|
||||
|
||||
# convolve the tensor with the kernel.
|
||||
output = torch.nn.functional.conv2d(input, tmp_kernel, groups=tmp_kernel.size(0), padding=0, stride=1)
|
||||
|
||||
out = output.view(b, c, h, w)
|
||||
return out
|
||||
|
||||
|
||||
def _gaussian(self, window_size: int, sigma):
|
||||
if isinstance(sigma, float):
|
||||
sigma = torch.tensor([[sigma]])
|
||||
|
||||
batch_size = sigma.shape[0]
|
||||
|
||||
x = (torch.arange(window_size, device=sigma.device, dtype=sigma.dtype) - window_size // 2).expand(batch_size, -1)
|
||||
|
||||
if window_size % 2 == 0:
|
||||
x = x + 0.5
|
||||
|
||||
gauss = torch.exp(-x.pow(2.0) / (2 * sigma.pow(2.0)))
|
||||
|
||||
return gauss / gauss.sum(-1, keepdim=True)
|
||||
|
||||
|
||||
def _gaussian_blur2d(self, input, kernel_size, sigma):
|
||||
if isinstance(sigma, tuple):
|
||||
sigma = torch.tensor([sigma], dtype=input.dtype)
|
||||
else:
|
||||
sigma = sigma.to(dtype=input.dtype)
|
||||
|
||||
ky, kx = int(kernel_size[0]), int(kernel_size[1])
|
||||
bs = sigma.shape[0]
|
||||
kernel_x = self._gaussian(kx, sigma[:, 1].view(bs, 1))
|
||||
kernel_y = self._gaussian(ky, sigma[:, 0].view(bs, 1))
|
||||
out_x = self._filter2d(input, kernel_x[..., None, :])
|
||||
out = self._filter2d(out_x, kernel_y[..., None])
|
||||
|
||||
return out
|
||||
@@ -1,65 +1,2 @@
|
||||
import torch, math
|
||||
|
||||
|
||||
class EnhancedDDIMScheduler():
|
||||
|
||||
def __init__(self, num_train_timesteps=1000, beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear"):
|
||||
self.num_train_timesteps = num_train_timesteps
|
||||
if beta_schedule == "scaled_linear":
|
||||
betas = torch.square(torch.linspace(math.sqrt(beta_start), math.sqrt(beta_end), num_train_timesteps, dtype=torch.float32))
|
||||
elif beta_schedule == "linear":
|
||||
betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
|
||||
else:
|
||||
raise NotImplementedError(f"{beta_schedule} is not implemented")
|
||||
self.alphas_cumprod = torch.cumprod(1.0 - betas, dim=0).tolist()
|
||||
self.set_timesteps(10)
|
||||
|
||||
|
||||
def set_timesteps(self, num_inference_steps, denoising_strength=1.0):
|
||||
# The timesteps are aligned to 999...0, which is different from other implementations,
|
||||
# but I think this implementation is more reasonable in theory.
|
||||
max_timestep = max(round(self.num_train_timesteps * denoising_strength) - 1, 0)
|
||||
num_inference_steps = min(num_inference_steps, max_timestep + 1)
|
||||
if num_inference_steps == 1:
|
||||
self.timesteps = [max_timestep]
|
||||
else:
|
||||
step_length = max_timestep / (num_inference_steps - 1)
|
||||
self.timesteps = [round(max_timestep - i*step_length) for i in range(num_inference_steps)]
|
||||
|
||||
|
||||
def denoise(self, model_output, sample, alpha_prod_t, alpha_prod_t_prev):
|
||||
weight_e = math.sqrt(1 - alpha_prod_t_prev) - math.sqrt(alpha_prod_t_prev * (1 - alpha_prod_t) / alpha_prod_t)
|
||||
weight_x = math.sqrt(alpha_prod_t_prev / alpha_prod_t)
|
||||
|
||||
prev_sample = sample * weight_x + model_output * weight_e
|
||||
|
||||
weight_e = -math.sqrt((1 - alpha_prod_t) / alpha_prod_t)
|
||||
weight_x = math.sqrt(1 / alpha_prod_t)
|
||||
|
||||
return prev_sample
|
||||
|
||||
|
||||
def step(self, model_output, timestep, sample, to_final=False):
|
||||
alpha_prod_t = self.alphas_cumprod[timestep]
|
||||
timestep_id = self.timesteps.index(timestep)
|
||||
if to_final or timestep_id + 1 >= len(self.timesteps):
|
||||
alpha_prod_t_prev = 1.0
|
||||
else:
|
||||
timestep_prev = self.timesteps[timestep_id + 1]
|
||||
alpha_prod_t_prev = self.alphas_cumprod[timestep_prev]
|
||||
|
||||
return self.denoise(model_output, sample, alpha_prod_t, alpha_prod_t_prev)
|
||||
|
||||
|
||||
def return_to_timestep(self, timestep, sample, sample_stablized):
|
||||
alpha_prod_t = self.alphas_cumprod[timestep]
|
||||
noise_pred = (sample - math.sqrt(alpha_prod_t) * sample_stablized) / math.sqrt(1 - alpha_prod_t)
|
||||
return noise_pred
|
||||
|
||||
|
||||
def add_noise(self, original_samples, noise, timestep):
|
||||
sqrt_alpha_prod = math.sqrt(self.alphas_cumprod[timestep])
|
||||
sqrt_one_minus_alpha_prod = math.sqrt(1 - self.alphas_cumprod[timestep])
|
||||
noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
|
||||
return noisy_samples
|
||||
|
||||
from .ddim import EnhancedDDIMScheduler
|
||||
from .continuous_ode import ContinuousODEScheduler
|
||||
|
||||
52
diffsynth/schedulers/continuous_ode.py
Normal file
52
diffsynth/schedulers/continuous_ode.py
Normal file
@@ -0,0 +1,52 @@
|
||||
import torch, math
|
||||
|
||||
|
||||
class ContinuousODEScheduler():
|
||||
|
||||
def __init__(self, num_inference_steps=100, sigma_max=700.0, sigma_min=0.002, rho=7.0):
|
||||
self.sigma_max = sigma_max
|
||||
self.sigma_min = sigma_min
|
||||
self.rho = rho
|
||||
self.init_noise_sigma = math.sqrt(sigma_max*sigma_max + 1)
|
||||
self.set_timesteps(num_inference_steps)
|
||||
|
||||
|
||||
def set_timesteps(self, num_inference_steps=100, denoising_strength=1.0):
|
||||
ramp = torch.linspace(0, denoising_strength, num_inference_steps)
|
||||
min_inv_rho = torch.pow(torch.tensor((self.sigma_min,)), (1 / self.rho))
|
||||
max_inv_rho = torch.pow(torch.tensor((self.sigma_max,)), (1 / self.rho))
|
||||
self.sigmas = torch.pow(max_inv_rho + ramp * (min_inv_rho - max_inv_rho), self.rho)
|
||||
self.timesteps = torch.log(self.sigmas) * 0.25
|
||||
|
||||
|
||||
def step(self, model_output, timestep, sample, to_final=False):
|
||||
timestep_id = torch.argmin((self.timesteps - timestep).abs())
|
||||
sigma = self.sigmas[timestep_id]
|
||||
estimated_sample = -sigma / (sigma*sigma + 1).sqrt() * model_output + 1 / (sigma*sigma + 1) * sample
|
||||
if to_final or timestep_id + 1 >= len(self.timesteps):
|
||||
prev_sample = estimated_sample
|
||||
else:
|
||||
dt = self.sigmas[timestep_id + 1] - sigma
|
||||
derivative = 1 / sigma * (sample - estimated_sample)
|
||||
prev_sample = sample + derivative * dt
|
||||
return prev_sample
|
||||
|
||||
|
||||
def scale_model_input(self, sample, timestep):
|
||||
timestep_id = torch.argmin((self.timesteps - timestep).abs())
|
||||
sigma = self.sigmas[timestep_id]
|
||||
sample = sample / (sigma*sigma + 1).sqrt()
|
||||
return sample
|
||||
|
||||
|
||||
def return_to_timestep(self, timestep, sample, sample_stablized):
|
||||
# This scheduler doesn't support this function.
|
||||
pass
|
||||
|
||||
|
||||
def add_noise(self, original_samples, noise, timestep):
|
||||
timestep_id = torch.argmin((self.timesteps - timestep).abs())
|
||||
sigma = self.sigmas[timestep_id]
|
||||
sample = original_samples + noise * sigma
|
||||
return sample
|
||||
|
||||
60
diffsynth/schedulers/ddim.py
Normal file
60
diffsynth/schedulers/ddim.py
Normal file
@@ -0,0 +1,60 @@
|
||||
import torch, math
|
||||
|
||||
|
||||
class EnhancedDDIMScheduler():
|
||||
|
||||
def __init__(self, num_train_timesteps=1000, beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear"):
|
||||
self.num_train_timesteps = num_train_timesteps
|
||||
if beta_schedule == "scaled_linear":
|
||||
betas = torch.square(torch.linspace(math.sqrt(beta_start), math.sqrt(beta_end), num_train_timesteps, dtype=torch.float32))
|
||||
elif beta_schedule == "linear":
|
||||
betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
|
||||
else:
|
||||
raise NotImplementedError(f"{beta_schedule} is not implemented")
|
||||
self.alphas_cumprod = torch.cumprod(1.0 - betas, dim=0).tolist()
|
||||
self.set_timesteps(10)
|
||||
|
||||
|
||||
def set_timesteps(self, num_inference_steps, denoising_strength=1.0):
|
||||
# The timesteps are aligned to 999...0, which is different from other implementations,
|
||||
# but I think this implementation is more reasonable in theory.
|
||||
max_timestep = max(round(self.num_train_timesteps * denoising_strength) - 1, 0)
|
||||
num_inference_steps = min(num_inference_steps, max_timestep + 1)
|
||||
if num_inference_steps == 1:
|
||||
self.timesteps = [max_timestep]
|
||||
else:
|
||||
step_length = max_timestep / (num_inference_steps - 1)
|
||||
self.timesteps = [round(max_timestep - i*step_length) for i in range(num_inference_steps)]
|
||||
|
||||
|
||||
def denoise(self, model_output, sample, alpha_prod_t, alpha_prod_t_prev):
|
||||
weight_e = math.sqrt(1 - alpha_prod_t_prev) - math.sqrt(alpha_prod_t_prev * (1 - alpha_prod_t) / alpha_prod_t)
|
||||
weight_x = math.sqrt(alpha_prod_t_prev / alpha_prod_t)
|
||||
prev_sample = sample * weight_x + model_output * weight_e
|
||||
return prev_sample
|
||||
|
||||
|
||||
def step(self, model_output, timestep, sample, to_final=False):
|
||||
alpha_prod_t = self.alphas_cumprod[timestep]
|
||||
timestep_id = self.timesteps.index(timestep)
|
||||
if to_final or timestep_id + 1 >= len(self.timesteps):
|
||||
alpha_prod_t_prev = 1.0
|
||||
else:
|
||||
timestep_prev = self.timesteps[timestep_id + 1]
|
||||
alpha_prod_t_prev = self.alphas_cumprod[timestep_prev]
|
||||
|
||||
return self.denoise(model_output, sample, alpha_prod_t, alpha_prod_t_prev)
|
||||
|
||||
|
||||
def return_to_timestep(self, timestep, sample, sample_stablized):
|
||||
alpha_prod_t = self.alphas_cumprod[timestep]
|
||||
noise_pred = (sample - math.sqrt(alpha_prod_t) * sample_stablized) / math.sqrt(1 - alpha_prod_t)
|
||||
return noise_pred
|
||||
|
||||
|
||||
def add_noise(self, original_samples, noise, timestep):
|
||||
sqrt_alpha_prod = math.sqrt(self.alphas_cumprod[timestep])
|
||||
sqrt_one_minus_alpha_prod = math.sqrt(1 - self.alphas_cumprod[timestep])
|
||||
noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
|
||||
return noisy_samples
|
||||
|
||||
37
examples/svd_text_to_video.py
Normal file
37
examples/svd_text_to_video.py
Normal file
@@ -0,0 +1,37 @@
|
||||
from diffsynth import save_video, SDXLImagePipeline, ModelManager, SVDVideoPipeline
|
||||
from diffsynth import ModelManager
|
||||
import torch
|
||||
|
||||
|
||||
# Download models
|
||||
# `models/stable_diffusion_xl/sd_xl_base_1.0.safetensors`: [link](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors)
|
||||
# `models/stable_video_diffusion/svd_xt.safetensors`: [link](https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt/resolve/main/svd_xt.safetensors)
|
||||
|
||||
|
||||
prompt = "cloud, wind"
|
||||
torch.manual_seed(0)
|
||||
|
||||
# 1. Text-to-image using SD-XL
|
||||
model_manager = ModelManager(torch_dtype=torch.float16, device="cuda")
|
||||
model_manager.load_models(["models/stable_diffusion_xl/sd_xl_base_1.0.safetensors"])
|
||||
pipe = SDXLImagePipeline.from_model_manager(model_manager)
|
||||
image = pipe(
|
||||
prompt=prompt,
|
||||
negative_prompt="",
|
||||
cfg_scale=6,
|
||||
height=1024, width=1024, num_inference_steps=50,
|
||||
)
|
||||
pipe.to("cpu")
|
||||
torch.cuda.empty_cache()
|
||||
|
||||
# 2. Image-to-video using SVD
|
||||
model_manager = ModelManager()
|
||||
model_manager.load_models(["models/stable_video_diffusion/svd_xt.safetensors"])
|
||||
pipe = SVDVideoPipeline.from_model_manager(model_manager)
|
||||
video = pipe(
|
||||
input_image=image,
|
||||
num_frames=25, fps=15, height=1024, width=1024,
|
||||
motion_bucket_id=127,
|
||||
num_inference_steps=50
|
||||
)
|
||||
save_video(video, "video.mp4", fps=15)
|
||||
Reference in New Issue
Block a user