File size: 10,100 Bytes
fb4fac3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
from diffsynth import ModelManager, SD3ImagePipeline
from peft import LoraConfig, inject_adapter_in_model
from torchvision import transforms
from PIL import Image
import lightning as pl
import pandas as pd
import torch, os, argparse
os.environ["TOKENIZERS_PARALLELISM"] = "True"



class TextImageDataset(torch.utils.data.Dataset):
    def __init__(self, dataset_path, steps_per_epoch=10000, height=1024, width=1024, center_crop=True, random_flip=False):
        self.steps_per_epoch = steps_per_epoch
        metadata = pd.read_csv(os.path.join(dataset_path, "train/metadata.csv"))
        self.path = [os.path.join(dataset_path, "train", file_name) for file_name in metadata["file_name"]]
        self.text = metadata["text"].to_list()
        self.image_processor = transforms.Compose(
            [
                transforms.Resize(max(height, width), interpolation=transforms.InterpolationMode.BILINEAR),
                transforms.CenterCrop((height, width)) if center_crop else transforms.RandomCrop((height, width)),
                transforms.RandomHorizontalFlip() if random_flip else transforms.Lambda(lambda x: x),
                transforms.ToTensor(),
                transforms.Normalize([0.5], [0.5]),
            ]
        )


    def __getitem__(self, index):
        data_id = torch.randint(0, len(self.path), (1,))[0]
        data_id = (data_id + index) % len(self.path) # For fixed seed.
        text = self.text[data_id]
        image = Image.open(self.path[data_id]).convert("RGB")
        image = self.image_processor(image)
        return {"text": text, "image": image}
    

    def __len__(self):
        return self.steps_per_epoch
    


class LightningModel(pl.LightningModule):
    def __init__(self, torch_dtype=torch.float16, learning_rate=1e-4, pretrained_weights=[], lora_rank=4, lora_alpha=4, use_gradient_checkpointing=True):
        super().__init__()

        # Load models
        model_manager = ModelManager(torch_dtype=torch_dtype, device=self.device)
        model_manager.load_models(pretrained_weights)
        self.pipe = SD3ImagePipeline.from_model_manager(model_manager)

        # Freeze parameters
        self.pipe.text_encoder_1.requires_grad_(False)
        self.pipe.text_encoder_2.requires_grad_(False)
        if self.pipe.text_encoder_3 is not None:
            self.pipe.text_encoder_3.requires_grad_(False)
        self.pipe.dit.requires_grad_(False)
        self.pipe.vae_decoder.requires_grad_(False)
        self.pipe.vae_encoder.requires_grad_(False)
        self.pipe.text_encoder_1.eval()
        self.pipe.text_encoder_2.eval()
        if self.pipe.text_encoder_3 is not None:
            self.pipe.text_encoder_3.eval()
        self.pipe.dit.train()
        self.pipe.vae_decoder.eval()
        self.pipe.vae_encoder.eval()

        # Add LoRA to DiT
        lora_config = LoraConfig(
            r=lora_rank,
            lora_alpha=lora_alpha,
            init_lora_weights="gaussian",
            target_modules=["a_to_qkv", "b_to_qkv"],
        )
        self.pipe.dit = inject_adapter_in_model(lora_config, self.pipe.dit)
        for param in self.pipe.dit.parameters():
            # Upcast LoRA parameters into fp32
            if param.requires_grad:
                param.data = param.to(torch.float32)

        # Set other parameters
        self.learning_rate = learning_rate
        self.use_gradient_checkpointing = use_gradient_checkpointing
        self.pipe.scheduler.set_timesteps(1000)
    

    def training_step(self, batch, batch_idx):
        # Data
        text, image = batch["text"], batch["image"]

        # Prepare input parameters
        self.pipe.device = self.device
        prompt_emb, pooled_prompt_emb = self.pipe.prompter.encode_prompt(
            self.pipe.text_encoder_1, self.pipe.text_encoder_2, self.pipe.text_encoder_3,
            text, device=self.device
        )
        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, 1000, (1,))
        timestep = self.pipe.scheduler.timesteps[timestep_id].to(self.device)
        noisy_latents = self.pipe.scheduler.add_noise(latents, noise, self.pipe.scheduler.timesteps[timestep_id])
        training_target = self.pipe.scheduler.training_target(latents, noise, timestep)

        # Compute loss
        noise_pred = self.pipe.dit(noisy_latents, timestep, prompt_emb, pooled_prompt_emb, use_gradient_checkpointing=self.use_gradient_checkpointing)
        loss = torch.nn.functional.mse_loss(noise_pred, training_target)

        # Record log
        self.log("train_loss", loss, prog_bar=True)
        return loss


    def configure_optimizers(self):
        trainable_modules = filter(lambda p: p.requires_grad, self.pipe.dit.parameters())
        optimizer = torch.optim.AdamW(trainable_modules, lr=self.learning_rate)
        return optimizer
    

    def on_save_checkpoint(self, checkpoint):
        checkpoint.clear()
        trainable_param_names = list(filter(lambda named_param: named_param[1].requires_grad, self.pipe.dit.named_parameters()))
        trainable_param_names = set([named_param[0] for named_param in trainable_param_names])
        state_dict = self.pipe.dit.state_dict()
        for name, param in state_dict.items():
            if name in trainable_param_names:
                checkpoint[name] = param



def parse_args():
    parser = argparse.ArgumentParser(description="Simple example of a training script.")
    parser.add_argument(
        "--pretrained_path",
        type=str,
        default=None,
        required=True,
        help="Path to pretrained model. For example, `models/stable_diffusion_3/sd3_medium_incl_clips.safetensors` or `models/stable_diffusion_3/sd3_medium_incl_clips_t5xxlfp16.safetensors`.",
    )
    parser.add_argument(
        "--dataset_path",
        type=str,
        default=None,
        required=True,
        help="The path of the Dataset.",
    )
    parser.add_argument(
        "--output_path",
        type=str,
        default="./",
        help="Path to save the model.",
    )
    parser.add_argument(
        "--steps_per_epoch",
        type=int,
        default=500,
        help="Number of steps per epoch.",
    )
    parser.add_argument(
        "--height",
        type=int,
        default=1024,
        help="Image height.",
    )
    parser.add_argument(
        "--width",
        type=int,
        default=1024,
        help="Image width.",
    )
    parser.add_argument(
        "--center_crop",
        default=False,
        action="store_true",
        help=(
            "Whether to center crop the input images to the resolution. If not set, the images will be randomly"
            " cropped. The images will be resized to the resolution first before cropping."
        ),
    )
    parser.add_argument(
        "--random_flip",
        default=False,
        action="store_true",
        help="Whether to randomly flip images horizontally",
    )
    parser.add_argument(
        "--batch_size",
        type=int,
        default=1,
        help="Batch size (per device) for the training dataloader.",
    )
    parser.add_argument(
        "--dataloader_num_workers",
        type=int,
        default=0,
        help="Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.",
    )
    parser.add_argument(
        "--precision",
        type=str,
        default="16-mixed",
        choices=["32", "16", "16-mixed"],
        help="Training precision",
    )
    parser.add_argument(
        "--learning_rate",
        type=float,
        default=1e-4,
        help="Learning rate.",
    )
    parser.add_argument(
        "--lora_rank",
        type=int,
        default=4,
        help="The dimension of the LoRA update matrices.",
    )
    parser.add_argument(
        "--lora_alpha",
        type=float,
        default=4.0,
        help="The weight of the LoRA update matrices.",
    )
    parser.add_argument(
        "--use_gradient_checkpointing",
        default=False,
        action="store_true",
        help="Whether to use gradient checkpointing.",
    )
    parser.add_argument(
        "--accumulate_grad_batches",
        type=int,
        default=1,
        help="The number of batches in gradient accumulation.",
    )
    parser.add_argument(
        "--training_strategy",
        type=str,
        default="auto",
        choices=["auto", "deepspeed_stage_1", "deepspeed_stage_2", "deepspeed_stage_3"],
        help="Training strategy",
    )
    parser.add_argument(
        "--max_epochs",
        type=int,
        default=1,
        help="Number of epochs.",
    )
    args = parser.parse_args()
    return args



if __name__ == '__main__':
    # args
    args = parse_args()

    # dataset and data loader
    dataset = TextImageDataset(
        args.dataset_path,
        steps_per_epoch=args.steps_per_epoch * args.batch_size,
        height=args.height,
        width=args.width,
        center_crop=args.center_crop,
        random_flip=args.random_flip
    )
    train_loader = torch.utils.data.DataLoader(
        dataset,
        shuffle=True,
        batch_size=args.batch_size,
        num_workers=args.dataloader_num_workers
    )

    # model
    model = LightningModel(
        pretrained_weights=[args.pretrained_path],
        torch_dtype=torch.float32 if args.precision == "32" else torch.float16,
        learning_rate=args.learning_rate,
        lora_rank=args.lora_rank,
        lora_alpha=args.lora_alpha,
        use_gradient_checkpointing=args.use_gradient_checkpointing
    )

    # train
    trainer = pl.Trainer(
        max_epochs=args.max_epochs,
        accelerator="gpu",
        devices="auto",
        precision=args.precision,
        strategy=args.training_strategy,
        default_root_dir=args.output_path,
        accumulate_grad_batches=args.accumulate_grad_batches,
        callbacks=[pl.pytorch.callbacks.ModelCheckpoint(save_top_k=-1)]
    )
    trainer.fit(model=model, train_dataloaders=train_loader)