lora retrieval

lora/dataset.py

@@ -0,0 +1,54 @@
+import torch, os
+import pandas as pd
+from PIL import Image
+from torchvision.transforms import v2
+from diffsynth.data.video import crop_and_resize
+
+
+class LoraDataset(torch.utils.data.Dataset):
+    def __init__(self, base_path, metadata_path, steps_per_epoch=1000, loras_per_item=1):
+        self.base_path = base_path
+        data_df = pd.read_csv(metadata_path)
+        self.model_file = data_df["model_file"].tolist()
+        self.image_file = data_df["image_file"].tolist()
+        self.text = data_df["text"].tolist()
+        self.max_resolution = 1920 * 1080
+        self.steps_per_epoch = steps_per_epoch
+        self.loras_per_item = loras_per_item
+        
+        
+    def read_image(self, image_file):
+        image = Image.open(image_file).convert("RGB")
+        width, height = image.size
+        if width * height > self.max_resolution:
+            scale = (width * height / self.max_resolution) ** 0.5
+            image = image.resize((int(width / scale), int(height / scale)))
+            width, height = image.size
+        if width % 16 != 0 or height % 16 != 0:
+            image = crop_and_resize(image, height // 16 * 16, width // 16 * 16)
+        image = v2.functional.to_image(image)
+        image = v2.functional.to_dtype(image, dtype=torch.float32, scale=True)
+        image = v2.functional.normalize(image, [0.5], [0.5])
+        return image
+    
+    
+    def get_data(self, data_id):
+        data = {
+            "model_file": os.path.join(self.base_path, self.model_file[data_id]),
+            "image": self.read_image(os.path.join(self.base_path, self.image_file[data_id])),
+            "text": self.text[data_id]
+        }
+        return data
+
+
+    def __getitem__(self, index):
+        data = []
+        while len(data) < self.loras_per_item:
+            data_id = torch.randint(0, len(self.model_file), (1,))[0]
+            data_id = (data_id + index) % len(self.model_file) # For fixed seed.
+            data.append(self.get_data(data_id))
+        return data
+
+
+    def __len__(self):
+        return self.steps_per_epoch

lora/merger.py

@@ -0,0 +1,61 @@
+import torch
+
+
+class LoraMerger(torch.nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.weight_base = torch.nn.Parameter(torch.randn((dim,)))
+        self.weight_lora = torch.nn.Parameter(torch.randn((dim,)))
+        self.weight_cross = torch.nn.Parameter(torch.randn((dim,)))
+        self.weight_out = torch.nn.Parameter(torch.ones((dim,)))
+        self.bias = torch.nn.Parameter(torch.randn((dim,)))
+        self.activation = torch.nn.Sigmoid()
+        self.norm_base = torch.nn.LayerNorm(dim, eps=1e-5)
+        self.norm_lora = torch.nn.LayerNorm(dim, eps=1e-5)
+        
+    def forward(self, base_output, lora_outputs):
+        norm_base_output = self.norm_base(base_output)
+        norm_lora_outputs = self.norm_lora(lora_outputs)
+        gate = self.activation(
+            norm_base_output * self.weight_base \
+            + norm_lora_outputs * self.weight_lora \
+            + norm_base_output * norm_lora_outputs * self.weight_cross + self.bias
+        )
+        output = base_output + (self.weight_out * gate * lora_outputs).sum(dim=0)
+        return output
+
+
+class LoraPatcher(torch.nn.Module):
+    def __init__(self, lora_patterns=None):
+        super().__init__()
+        if lora_patterns is None:
+            lora_patterns = self.default_lora_patterns()
+        model_dict = {}
+        for lora_pattern in lora_patterns:
+            name, dim = lora_pattern["name"], lora_pattern["dim"]
+            model_dict[name.replace(".", "___")] = LoraMerger(dim)
+        self.model_dict = torch.nn.ModuleDict(model_dict)
+        
+    def default_lora_patterns(self):
+        lora_patterns = []
+        lora_dict = {
+            "attn.a_to_qkv": 9216, "attn.a_to_out": 3072, "ff_a.0": 12288, "ff_a.2": 3072, "norm1_a.linear": 18432,
+            "attn.b_to_qkv": 9216, "attn.b_to_out": 3072, "ff_b.0": 12288, "ff_b.2": 3072, "norm1_b.linear": 18432,
+        }
+        for i in range(19):
+            for suffix in lora_dict:
+                lora_patterns.append({
+                    "name": f"blocks.{i}.{suffix}",
+                    "dim": lora_dict[suffix]
+                })
+        lora_dict = {"to_qkv_mlp": 21504, "proj_out": 3072, "norm.linear": 9216}
+        for i in range(38):
+            for suffix in lora_dict:
+                lora_patterns.append({
+                    "name": f"single_blocks.{i}.{suffix}",
+                    "dim": lora_dict[suffix]
+                })
+        return lora_patterns
+        
+    def forward(self, base_output, lora_outputs, name):
+        return self.model_dict[name.replace(".", "___")](base_output, lora_outputs)

lora/retriever.py

@@ -0,0 +1,149 @@
+import torch
+from diffsynth import SDTextEncoder
+from diffsynth.models.sd3_text_encoder import SD3TextEncoder1StateDictConverter
+from diffsynth.models.sd_text_encoder import CLIPEncoderLayer
+
+
+class LoRALayerBlock(torch.nn.Module):
+    def __init__(self, L, dim_in):
+        super().__init__()
+        self.x = torch.nn.Parameter(torch.randn(1, L, dim_in))
+
+    def forward(self, lora_A, lora_B):
+        out = self.x @ lora_A.T @ lora_B.T
+        return out
+    
+
+class LoRAEmbedder(torch.nn.Module):
+    def __init__(self, lora_patterns=None, L=1, out_dim=2048):
+        super().__init__()
+        if lora_patterns is None:
+            lora_patterns = self.default_lora_patterns()
+            
+        model_dict = {}
+        for lora_pattern in lora_patterns:
+            name, dim = lora_pattern["name"], lora_pattern["dim"][0]
+            model_dict[name.replace(".", "___")] = LoRALayerBlock(L, dim)
+        self.model_dict = torch.nn.ModuleDict(model_dict)
+        
+        proj_dict = {}
+        for lora_pattern in lora_patterns:
+            layer_type, dim = lora_pattern["type"], lora_pattern["dim"][1]
+            if layer_type not in proj_dict:
+                proj_dict[layer_type.replace(".", "___")] = torch.nn.Linear(dim, out_dim)
+        self.proj_dict = torch.nn.ModuleDict(proj_dict)
+        
+        self.lora_patterns = lora_patterns
+        
+        
+    def default_lora_patterns(self):
+        lora_patterns = []
+        lora_dict = {
+            "attn.a_to_qkv": (3072, 9216), "attn.a_to_out": (3072, 3072), "ff_a.0": (3072, 12288), "ff_a.2": (12288, 3072), "norm1_a.linear": (3072, 18432),
+            "attn.b_to_qkv": (3072, 9216), "attn.b_to_out": (3072, 3072), "ff_b.0": (3072, 12288), "ff_b.2": (12288, 3072), "norm1_b.linear": (3072, 18432),
+        }
+        for i in range(19):
+            for suffix in lora_dict:
+                lora_patterns.append({
+                    "name": f"blocks.{i}.{suffix}",
+                    "dim": lora_dict[suffix],
+                    "type": suffix,
+                })
+        lora_dict = {"to_qkv_mlp": (3072, 21504), "proj_out": (15360, 3072), "norm.linear": (3072, 9216)}
+        for i in range(38):
+            for suffix in lora_dict:
+                lora_patterns.append({
+                    "name": f"single_blocks.{i}.{suffix}",
+                    "dim": lora_dict[suffix],
+                    "type": suffix,
+                })
+        return lora_patterns
+        
+    def forward(self, lora):
+        lora_emb = []
+        for lora_pattern in self.lora_patterns:
+            name, layer_type = lora_pattern["name"], lora_pattern["type"]
+            lora_A = lora[name + ".lora_A.default.weight"]
+            lora_B = lora[name + ".lora_B.default.weight"]
+            lora_out = self.model_dict[name.replace(".", "___")](lora_A, lora_B)
+            lora_out = self.proj_dict[layer_type.replace(".", "___")](lora_out)
+            lora_emb.append(lora_out)
+        lora_emb = torch.concat(lora_emb, dim=1)
+        return lora_emb
+
+
+class TextEncoder(torch.nn.Module):
+    def __init__(self, embed_dim=768, vocab_size=49408, max_position_embeddings=77, num_encoder_layers=12, encoder_intermediate_size=3072):
+        super().__init__()
+
+        # token_embedding
+        self.token_embedding = torch.nn.Embedding(vocab_size, embed_dim)
+
+        # position_embeds (This is a fixed tensor)
+        self.position_embeds = torch.nn.Parameter(torch.zeros(1, max_position_embeddings, embed_dim))
+
+        # encoders
+        self.encoders = torch.nn.ModuleList([CLIPEncoderLayer(embed_dim, encoder_intermediate_size) for _ in range(num_encoder_layers)])
+
+        # attn_mask
+        self.attn_mask = self.attention_mask(max_position_embeddings)
+
+        # final_layer_norm
+        self.final_layer_norm = torch.nn.LayerNorm(embed_dim)
+
+    def attention_mask(self, length):
+        mask = torch.empty(length, length)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)
+        return mask
+
+    def forward(self, input_ids, clip_skip=1):
+        embeds = self.token_embedding(input_ids) + self.position_embeds
+        attn_mask = self.attn_mask.to(device=embeds.device, dtype=embeds.dtype)
+        for encoder_id, encoder in enumerate(self.encoders):
+            embeds = encoder(embeds, attn_mask=attn_mask)
+            if encoder_id + clip_skip == len(self.encoders):
+                break
+        embeds = self.final_layer_norm(embeds)
+        pooled_embeds = embeds[torch.arange(embeds.shape[0]), input_ids.to(dtype=torch.int).argmax(dim=-1)]
+        return pooled_embeds
+    
+    @staticmethod
+    def state_dict_converter():
+        return SD3TextEncoder1StateDictConverter()
+    
+    
+class LoRAEncoder(torch.nn.Module):
+    def __init__(self, embed_dim=768, max_position_embeddings=304, num_encoder_layers=2, encoder_intermediate_size=3072, L=1):
+        super().__init__()
+        max_position_embeddings *= L
+        
+        # Embedder
+        self.embedder = LoRAEmbedder(L=L, out_dim=embed_dim)
+
+        # position_embeds (This is a fixed tensor)
+        self.position_embeds = torch.nn.Parameter(torch.zeros(1, max_position_embeddings, embed_dim))
+
+        # encoders
+        self.encoders = torch.nn.ModuleList([CLIPEncoderLayer(embed_dim, encoder_intermediate_size) for _ in range(num_encoder_layers)])
+
+        # attn_mask
+        self.attn_mask = self.attention_mask(max_position_embeddings)
+
+        # final_layer_norm
+        self.final_layer_norm = torch.nn.LayerNorm(embed_dim)
+
+    def attention_mask(self, length):
+        mask = torch.empty(length, length)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)
+        return mask
+
+    def forward(self, lora):
+        embeds = self.embedder(lora) + self.position_embeds
+        attn_mask = self.attn_mask.to(device=embeds.device, dtype=embeds.dtype)
+        for encoder_id, encoder in enumerate(self.encoders):
+            embeds = encoder(embeds, attn_mask=attn_mask)
+        embeds = self.final_layer_norm(embeds)
+        embeds = embeds.mean(dim=1)
+        return embeds

lora/test_merger.py

@@ -0,0 +1,46 @@
+from diffsynth import FluxImagePipeline, ModelManager, load_state_dict
+from diffsynth.models.lora import FluxLoRAConverter
+from diffsynth.pipelines.flux_image import lets_dance_flux
+from lora.dataset import LoraDataset
+from lora.merger import LoraPatcher
+from lora.utils import load_lora
+import torch, os
+from accelerate import Accelerator, DistributedDataParallelKwargs
+from tqdm import tqdm
+
+
+model_manager = ModelManager(torch_dtype=torch.bfloat16, device="cuda", model_id_list=["FLUX.1-dev"])
+pipe = FluxImagePipeline.from_model_manager(model_manager)
+pipe.enable_auto_lora()
+
+lora_patcher = LoraPatcher().to(dtype=torch.bfloat16, device="cuda")
+lora_patcher.load_state_dict(load_state_dict("models/lora_merger/epoch-3.safetensors"))
+
+dataset = LoraDataset("data/lora/models", "data/lora/lora_dataset_1000.csv", steps_per_epoch=800, loras_per_item=4)
+
+for seed in range(100):
+    batch = dataset[0]
+    num_lora = torch.randint(1, len(batch), (1,))[0]
+    lora_state_dicts = [
+        FluxLoRAConverter.align_to_diffsynth_format(load_lora(batch[i]["model_file"], device="cuda")) for i in range(num_lora)
+    ]
+    image = pipe(
+        prompt=batch[0]["text"],
+        seed=seed,
+    )
+    image.save(f"data/lora/lora_outputs/image_{seed}_nolora.jpg")
+    for i in range(num_lora):
+        image = pipe(
+            prompt=batch[0]["text"],
+            lora_state_dicts=[lora_state_dicts[i]], 
+            lora_patcher=lora_patcher,
+            seed=seed,
+        )
+        image.save(f"data/lora/lora_outputs/image_{seed}_{i}.jpg")
+    image = pipe(
+        prompt=batch[0]["text"],
+        lora_state_dicts=lora_state_dicts, 
+        lora_patcher=lora_patcher,
+        seed=seed,
+    )
+    image.save(f"data/lora/lora_outputs/image_{seed}_merger.jpg")

lora/test_retriever.py

@@ -0,0 +1,148 @@
+from diffsynth import FluxImagePipeline, ModelManager, load_state_dict
+from diffsynth.models.lora import FluxLoRAConverter
+from diffsynth.pipelines.flux_image import lets_dance_flux
+from lora.dataset import LoraDataset
+from lora.retriever import TextEncoder, LoRAEncoder
+from lora.merger import LoraPatcher
+from lora.utils import load_lora
+import torch, os
+from accelerate import Accelerator, DistributedDataParallelKwargs
+from tqdm import tqdm
+from transformers import CLIPTokenizer, CLIPModel
+import pandas as pd
+
+
+
+class LoRARetrieverTrainingModel(torch.nn.Module):
+    def __init__(self, pretrained_path):
+        super().__init__()
+        
+        self.text_encoder = TextEncoder().to(torch.bfloat16)
+        state_dict = load_state_dict("models/FLUX/FLUX.1-dev/text_encoder/model.safetensors")
+        self.text_encoder.load_state_dict(TextEncoder.state_dict_converter().from_civitai(state_dict))
+        self.text_encoder.requires_grad_(False)
+        self.text_encoder.eval()
+        
+        self.lora_encoder = LoRAEncoder().to(torch.bfloat16)
+        state_dict = load_state_dict(pretrained_path)
+        self.lora_encoder.load_state_dict(state_dict)
+        
+        self.tokenizer = CLIPTokenizer.from_pretrained("diffsynth/tokenizer_configs/stable_diffusion_3/tokenizer_1")
+        
+        
+    def to(self, *args, **kwargs):
+        device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(*args, **kwargs)
+        if device is not None:
+            self.device = device
+        if dtype is not None:
+            self.torch_dtype = dtype
+        super().to(*args, **kwargs)
+        return self
+
+
+    def forward(self, batch):
+        text = [data["text"] for data in batch]
+        input_ids = self.tokenizer(
+            text,
+            return_tensors="pt",
+            padding="max_length",
+            max_length=77,
+            truncation=True
+        ).input_ids.to(self.device)
+        text_emb = self.text_encoder(input_ids)
+        text_emb = text_emb / text_emb.norm()
+        
+        lora_emb = []
+        for data in batch:
+            lora = FluxLoRAConverter.align_to_diffsynth_format(load_lora(data["model_file"], device=self.device))
+            lora_emb.append(self.lora_encoder(lora))
+        lora_emb = torch.concat(lora_emb)
+        lora_emb = lora_emb / lora_emb.norm()
+        
+        similarity = text_emb @ lora_emb.T
+        print(similarity)
+        loss = -torch.log(torch.softmax(similarity, dim=0).diag()) - torch.log(torch.softmax(similarity, dim=1).diag())
+        loss = 10 * loss.mean()
+        return loss
+    
+    
+    def trainable_modules(self):
+        return self.lora_encoder.parameters()
+    
+    @torch.no_grad()
+    def process_lora_list(self, lora_list):
+        lora_emb = []
+        for lora in tqdm(lora_list):
+            lora = FluxLoRAConverter.align_to_diffsynth_format(load_lora(lora, device="cuda"))
+            lora_emb.append(self.lora_encoder(lora))
+        lora_emb = torch.concat(lora_emb)
+        lora_emb = lora_emb / lora_emb.norm()
+        self.lora_emb = lora_emb
+        self.lora_list = lora_list
+    
+    @torch.no_grad()
+    def retrieve(self, text, k=1):
+        input_ids = self.tokenizer(
+            text,
+            return_tensors="pt",
+            padding="max_length",
+            max_length=77,
+            truncation=True
+        ).input_ids.to(self.device)
+        text_emb = self.text_encoder(input_ids)
+        text_emb = text_emb / text_emb.norm()
+        
+        similarity = text_emb @ self.lora_emb.T
+        topk = torch.topk(similarity, k, dim=1).indices[0]
+        
+        lora_list = []
+        model_url_list = []
+        for lora_id in topk:
+            print(self.lora_list[lora_id])
+            lora = FluxLoRAConverter.align_to_diffsynth_format(load_lora(self.lora_list[lora_id], device="cuda"))
+            lora_list.append(lora)
+            model_id = self.lora_list[lora_id].split("/")[3:5]
+            model_url_list.append(f"https://www.modelscope.cn/models/{model_id[0]}/{model_id[1]}")
+        return lora_list, model_url_list
+
+
+
+model_manager = ModelManager(torch_dtype=torch.bfloat16, device="cuda", model_id_list=["FLUX.1-dev"])
+pipe = FluxImagePipeline.from_model_manager(model_manager)
+pipe.enable_auto_lora()
+
+lora_patcher = LoraPatcher().to(dtype=torch.bfloat16, device="cuda")
+lora_patcher.load_state_dict(load_state_dict("models/lora_merger/epoch-9.safetensors"))
+
+retriever = LoRARetrieverTrainingModel("models/lora_retriever/epoch-3.safetensors").to(dtype=torch.bfloat16, device="cuda")
+retriever.process_lora_list(list(set("data/lora/models/" + i for i in pd.read_csv("data/lora/lora_dataset_1000.csv")["model_file"])))
+
+dataset = LoraDataset("data/lora/models", "data/lora/lora_dataset_1000.csv", steps_per_epoch=800, loras_per_item=1)
+
+text_list = []
+model_url_list = []
+for seed in range(100):
+    text = dataset[0][0]["text"]
+    print(text)
+    loras, urls = retriever.retrieve(text, k=3)
+    print(urls)
+    image = pipe(
+        prompt=text,
+        seed=seed,
+    )
+    image.save(f"data/lora/lora_outputs/image_{seed}_top0.jpg")
+    for i in range(2, 3):
+        image = pipe(
+            prompt=text,
+            lora_state_dicts=loras[:i+1],
+            lora_patcher=lora_patcher,
+            seed=seed,
+        )
+        image.save(f"data/lora/lora_outputs/image_{seed}_top{i+1}.jpg")
+        
+    text_list.append(text)
+    model_url_list.append(urls)
+    df = pd.DataFrame()
+    df["text"] = text_list
+    df["models"] = [",".join(i) for i in model_url_list]
+    df.to_csv("data/lora/lora_outputs/metadata.csv", index=False, encoding="utf-8-sig")

lora/train_merger.py

@@ -0,0 +1,119 @@
+from diffsynth import FluxImagePipeline, ModelManager
+from diffsynth.models.lora import FluxLoRAConverter
+from diffsynth.pipelines.flux_image import lets_dance_flux
+from lora.dataset import LoraDataset
+from lora.merger import LoraPatcher
+from lora.utils import load_lora
+import torch, os
+from accelerate import Accelerator, DistributedDataParallelKwargs
+from tqdm import tqdm
+
+
+
+class LoRAMergerTrainingModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        model_manager = ModelManager(torch_dtype=torch.bfloat16, device="cpu", model_id_list=["FLUX.1-dev"])
+        self.pipe = FluxImagePipeline.from_model_manager(model_manager)
+        self.lora_patcher = LoraPatcher()
+        self.pipe.enable_auto_lora()
+        self.freeze_parameters()
+        self.switch_to_training_mode()
+        self.use_gradient_checkpointing = True
+        self.state_dict_converter = FluxLoRAConverter.align_to_diffsynth_format
+        self.device = "cuda"
+        
+        
+    def to(self, *args, **kwargs):
+        device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(*args, **kwargs)
+        if device is not None:
+            self.device = device
+        if dtype is not None:
+            self.torch_dtype = dtype
+        super().to(*args, **kwargs)
+        return self
+        
+        
+    def switch_to_training_mode(self):
+        self.pipe.scheduler.set_timesteps(1000, training=True)
+
+
+    def freeze_parameters(self):
+        self.pipe.requires_grad_(False)
+        self.pipe.eval()
+        self.pipe.denoising_model().train()
+        self.lora_patcher.requires_grad_(True)
+
+
+    def forward(self, batch):
+        # Data
+        text, image = batch[0]["text"], batch[0]["image"].unsqueeze(0)
+        num_lora = torch.randint(1, len(batch), (1,))[0]
+        lora_state_dicts = [
+            self.state_dict_converter(load_lora(batch[i]["model_file"], device=self.device)) for i in range(num_lora)
+        ]
+        lora_alphas = None
+
+        # Prepare input parameters
+        self.pipe.device = self.device
+        prompt_emb = self.pipe.encode_prompt(text, positive=True)
+        latents = self.pipe.vae_encoder(image.to(dtype=self.pipe.torch_dtype, device=self.device))
+        noise = torch.randn_like(latents)
+        timestep_id = torch.randint(0, self.pipe.scheduler.num_train_timesteps, (1,))
+        timestep = self.pipe.scheduler.timesteps[timestep_id].to(self.device)
+        extra_input = self.pipe.prepare_extra_input(latents)
+        noisy_latents = self.pipe.scheduler.add_noise(latents, noise, timestep)
+        training_target = self.pipe.scheduler.training_target(latents, noise, timestep)
+
+        # Compute loss
+        noise_pred = lets_dance_flux(
+            self.pipe.dit,
+            hidden_states=noisy_latents, timestep=timestep, **prompt_emb, **extra_input,
+            lora_state_dicts=lora_state_dicts, lora_alphas=lora_alphas, lora_patcher=self.lora_patcher,
+            use_gradient_checkpointing=self.use_gradient_checkpointing
+        )
+        loss = torch.nn.functional.mse_loss(noise_pred.float(), training_target.float())
+        loss = loss * self.pipe.scheduler.training_weight(timestep)
+        return loss
+    
+    
+    def trainable_modules(self):
+        return self.lora_patcher.parameters()
+
+
+class ModelLogger:
+    def __init__(self, output_path, remove_prefix_in_ckpt=None):
+        self.output_path = output_path
+        self.remove_prefix_in_ckpt = remove_prefix_in_ckpt
+        
+    
+    def on_step_end(self, loss):
+        pass
+    
+    
+    def on_epoch_end(self, accelerator, model, epoch_id):
+        accelerator.wait_for_everyone()
+        if accelerator.is_main_process:
+            state_dict = accelerator.unwrap_model(model).lora_patcher.state_dict()
+            os.makedirs(self.output_path, exist_ok=True)
+            path = os.path.join(self.output_path, f"epoch-{epoch_id}.safetensors")
+            accelerator.save(state_dict, path, safe_serialization=True)
+
+
+if __name__ == '__main__':
+    model = LoRAMergerTrainingModel()
+    dataset = LoraDataset("data/lora/models/", "data/lora/lora_dataset_1000.csv", steps_per_epoch=800, loras_per_item=4)
+    dataloader = torch.utils.data.DataLoader(dataset, shuffle=True, batch_size=1, num_workers=1, collate_fn=lambda x: x[0])
+    optimizer = torch.optim.AdamW(model.trainable_modules(), lr=1e-4)
+    model_logger = ModelLogger("models/lora_merger")
+    accelerator = Accelerator(kwargs_handlers=[DistributedDataParallelKwargs(find_unused_parameters=True)])
+    model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
+    
+    for epoch_id in range(1000000):
+        for data in tqdm(dataloader):
+            with accelerator.accumulate(model):
+                optimizer.zero_grad()
+                loss = model(data)
+                accelerator.backward(loss)
+                optimizer.step()
+        model_logger.on_epoch_end(accelerator, model, epoch_id)

lora/train_retriever.py

@@ -0,0 +1,105 @@
+from diffsynth import FluxImagePipeline, ModelManager, load_state_dict
+from diffsynth.models.lora import FluxLoRAConverter
+from diffsynth.pipelines.flux_image import lets_dance_flux
+from lora.dataset import LoraDataset
+from lora.retriever import TextEncoder, LoRAEncoder
+from lora.utils import load_lora
+import torch, os
+from accelerate import Accelerator, DistributedDataParallelKwargs
+from tqdm import tqdm
+from transformers import CLIPTokenizer, CLIPModel
+
+
+
+class LoRARetrieverTrainingModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        
+        self.text_encoder = TextEncoder().to(torch.bfloat16)
+        state_dict = load_state_dict("models/FLUX/FLUX.1-dev/text_encoder/model.safetensors")
+        self.text_encoder.load_state_dict(TextEncoder.state_dict_converter().from_civitai(state_dict))
+        self.text_encoder.requires_grad_(False)
+        self.text_encoder.eval()
+        
+        self.lora_encoder = LoRAEncoder().to(torch.bfloat16)
+        
+        self.tokenizer = CLIPTokenizer.from_pretrained("diffsynth/tokenizer_configs/stable_diffusion_3/tokenizer_1")
+        
+        
+    def to(self, *args, **kwargs):
+        device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(*args, **kwargs)
+        if device is not None:
+            self.device = device
+        if dtype is not None:
+            self.torch_dtype = dtype
+        super().to(*args, **kwargs)
+        return self
+
+
+    def forward(self, batch):
+        text = [data["text"] for data in batch]
+        input_ids = self.tokenizer(
+            text,
+            return_tensors="pt",
+            padding="max_length",
+            max_length=77,
+            truncation=True
+        ).input_ids.to(self.device)
+        text_emb = self.text_encoder(input_ids)
+        text_emb = text_emb / text_emb.norm()
+        
+        lora_emb = []
+        for data in batch:
+            lora = FluxLoRAConverter.align_to_diffsynth_format(load_lora(data["model_file"], device=self.device))
+            lora_emb.append(self.lora_encoder(lora))
+        lora_emb = torch.concat(lora_emb)
+        lora_emb = lora_emb / lora_emb.norm()
+        
+        similarity = text_emb @ lora_emb.T
+        print(similarity)
+        loss = -torch.log(torch.softmax(similarity, dim=0).diag()) - torch.log(torch.softmax(similarity, dim=1).diag())
+        loss = 10 * loss.mean()
+        return loss
+    
+    
+    def trainable_modules(self):
+        return self.lora_encoder.parameters()
+
+
+class ModelLogger:
+    def __init__(self, output_path, remove_prefix_in_ckpt=None):
+        self.output_path = output_path
+        self.remove_prefix_in_ckpt = remove_prefix_in_ckpt
+        
+    
+    def on_step_end(self, loss):
+        pass
+    
+    
+    def on_epoch_end(self, accelerator, model, epoch_id):
+        accelerator.wait_for_everyone()
+        if accelerator.is_main_process:
+            state_dict = accelerator.unwrap_model(model).lora_encoder.state_dict()
+            os.makedirs(self.output_path, exist_ok=True)
+            path = os.path.join(self.output_path, f"epoch-{epoch_id}.safetensors")
+            accelerator.save(state_dict, path, safe_serialization=True)
+
+
+if __name__ == '__main__':
+    model = LoRARetrieverTrainingModel()
+    dataset = LoraDataset("data/lora/models/", "data/lora/lora_dataset_1000.csv", steps_per_epoch=100, loras_per_item=32)
+    dataloader = torch.utils.data.DataLoader(dataset, shuffle=True, batch_size=1, num_workers=1, collate_fn=lambda x: x[0])
+    optimizer = torch.optim.AdamW(model.trainable_modules(), lr=1e-4)
+    model_logger = ModelLogger("models/lora_retriever")
+    accelerator = Accelerator(kwargs_handlers=[DistributedDataParallelKwargs(find_unused_parameters=True)])
+    model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
+    
+    for epoch_id in range(1000000):
+        for data in tqdm(dataloader):
+            with accelerator.accumulate(model):
+                optimizer.zero_grad()
+                loss = model(data)
+                accelerator.backward(loss)
+                optimizer.step()
+                print(loss)
+        model_logger.on_epoch_end(accelerator, model, epoch_id)

lora/utils.py

@@ -0,0 +1,12 @@
+from diffsynth import load_state_dict
+import math, torch
+
+
+def load_lora(file_path, device):
+    sd = load_state_dict(file_path, torch_dtype=torch.bfloat16, device=device)
+    scale = math.sqrt(sd["lora_unet_single_blocks_9_modulation_lin.alpha"] / sd["lora_unet_single_blocks_9_modulation_lin.lora_down.weight"].shape[0])
+    if scale != 1:
+        sd = {i: sd[i] * scale for i in sd}
+    return sd
+
+