HunyuanPortraitLCM/train_lcm_lora.py at main · kkakkkka/HunyuanPortraitLCM · GitHub

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
import os
import argparse
import numpy as np
import torch
import torch.nn.functional as F
from tqdm import tqdm
from accelerate import Accelerator
from diffusers import DDPMScheduler
from peft import LoraConfig, get_peft_model
from src.models.condition.unet_3d_svd_condition_ip import UNetSpatioTemporalConditionModel
from torch.utils.data import DataLoader

class DDIMSolver:
    def __init__(self, alpha_cumprods, timesteps=1000, ddim_timesteps=50):
        step_ratio = timesteps // ddim_timesteps
        self.ddim_timesteps = (np.arange(1, ddim_timesteps + 1) * step_ratio).round().astype(np.int64) - 1
        self.ddim_alpha_cumprods = alpha_cumprods[self.ddim_timesteps]
        self.ddim_alpha_cumprods_prev = np.asarray(
            [alpha_cumprods[0]] + alpha_cumprods[self.ddim_timesteps[:-1]].tolist()
        )
        self.ddim_sigmas = np.zeros_like(self.ddim_timesteps)

    def ddim_step(self, pred_x0, pred_noise, timestep_index):
        alpha_cumprod_prev = self.ddim_alpha_cumprods_prev[timestep_index]
        dir_xt = (1.0 - alpha_cumprod_prev - self.ddim_sigmas[timestep_index] ** 2).sqrt() * pred_noise
        x_prev = alpha_cumprod_prev.sqrt() * pred_x0 + dir_xt
        return x_prev

def main(args):
    accelerator = Accelerator(mixed_precision="fp16")

    teacher_unet = UNetSpatioTemporalConditionModel.from_pretrained(
        args.pretrained_model_path, subfolder="unet"
    )
    teacher_unet.requires_grad_(False)
    teacher_unet.to(accelerator.device, dtype=torch.float16)

    student_unet = UNetSpatioTemporalConditionModel.from_pretrained(
        args.pretrained_model_path, subfolder="unet"
    )
    student_unet.requires_grad_(False)

    lora_config = LoraConfig(
        r=args.lora_rank,
        lora_alpha=args.lora_rank,
        init_lora_weights="gaussian",
        target_modules=["to_k", "to_q", "to_v", "to_out.0"],
    )
    student_unet = get_peft_model(student_unet, lora_config)
    student_unet.train()

    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_path, subfolder="scheduler")
    alpha_cumprods = noise_scheduler.alphas_cumprod.numpy()
    solver = DDIMSolver(alpha_cumprods, timesteps=noise_scheduler.config.num_train_timesteps)

    optimizer = torch.optim.AdamW(
        student_unet.parameters(),
        lr=args.learning_rate,
        weight_decay=1e-4
    )

    train_dataset = []
    train_dataloader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True)

    student_unet, optimizer, train_dataloader = accelerator.prepare(
        student_unet, optimizer, train_dataloader
    )

    for epoch in range(args.num_epochs):
        for batch in tqdm(train_dataloader, disable=not accelerator.is_local_main_process):
            latents = batch["latents"].to(accelerator.device, dtype=torch.float16)
            encoder_hidden_states = batch["encoder_hidden_states"].to(accelerator.device, dtype=torch.float16)
            added_time_ids = batch["added_time_ids"].to(accelerator.device, dtype=torch.float16)

            bsz = latents.shape[0]

            topk = noise_scheduler.config.num_train_timesteps // args.num_ddim_timesteps
            index = torch.randint(0, args.num_ddim_timesteps, (bsz,), device=accelerator.device).long()
            start_timesteps = solver.ddim_timesteps[index]

            c_skip_routing = args.num_ddim_timesteps // args.lcm_steps
            next_index = torch.clamp(index - c_skip_routing, min=0)
            next_timesteps = solver.ddim_timesteps[next_index]

            noise = torch.randn_like(latents)
            noisy_latents = noise_scheduler.add_noise(latents, noise, start_timesteps)

            student_pred = student_unet(
                noisy_latents,
                start_timesteps,
                encoder_hidden_states=encoder_hidden_states,
                added_time_ids=added_time_ids
            ).sample

            with torch.no_grad():
                teacher_pred = teacher_unet(
                    noisy_latents,
                    start_timesteps,
                    encoder_hidden_states=encoder_hidden_states,
                    added_time_ids=added_time_ids
                ).sample

                pred_x0 = (noisy_latents - solver.ddim_sigmas[index].view(-1,1,1,1,1) * teacher_pred) / solver.ddim_alpha_cumprods_prev[index].sqrt().view(-1,1,1,1,1)
                target_latents = solver.ddim_step(pred_x0, teacher_pred, next_index)

            loss = F.huber_loss(student_pred.float(), target_latents.float(), delta=args.huber_c)

            accelerator.backward(loss)
            optimizer.step()
            optimizer.zero_grad()

        if accelerator.is_local_main_process and (epoch + 1) % args.save_every == 0:
            unwrapped_model = accelerator.unwrap_model(student_unet)
            unwrapped_model.save_pretrained(os.path.join(args.output_dir, f"lcm-lora-epoch-{epoch+1}"))

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--pretrained_model_path", type=str, required=True)
    parser.add_argument("--output_dir", type=str, default="./lcm_lora_output")
    parser.add_argument("--batch_size", type=int, default=1)
    parser.add_argument("--lora_rank", type=int, default=64)
    parser.add_argument("--learning_rate", type=float, default=1e-4)
    parser.add_argument("--num_epochs", type=int, default=10)
    parser.add_argument("--save_every", type=int, default=1)
    parser.add_argument("--num_ddim_timesteps", type=int, default=50)
    parser.add_argument("--lcm_steps", type=int, default=4)
    parser.add_argument("--huber_c", type=float, default=0.001)
    args = parser.parse_args()
    main(args)