feat: 冻结 MaskGIT 并联合训练

This commit is contained in:
unanmed 2026-04-22 22:15:05 +08:00
parent 4059f0e05a
commit 765cdcaeb0
3 changed files with 380 additions and 2 deletions

View File

@ -156,4 +156,77 @@ class GinkaHeatmapDataset(Dataset):
return {
"target_heatmap": heatmap,
"cond_heatmap": cond
}
class GinkaJointDataset(Dataset):
def __init__(self, data_path: str, min_mask=0, max_mask=0.8, blur_min=3, blur_max=6):
self.data = load_data(data_path)
self.blur_min = blur_min
self.blur_max = blur_max
self.min_mask = min_mask
self.max_mask = max_mask
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
item = self.data[idx]
target_map = np.array(item['map'])
heatmap = np.array(item['heatmap'], dtype=np.float32)
if np.random.rand() > 0.5:
k = np.random.randint(0, 4)
target_map = np.rot90(target_map, k)
for i in range(0, heatmap.shape[0]):
heatmap[i] = np.rot90(heatmap[i], k)
if np.random.rand() > 0.5:
target_map = np.fliplr(target_map)
for i in range(0, heatmap.shape[0]):
heatmap[i] = np.fliplr(heatmap[i])
if np.random.rand() > 0.5:
target_map = np.flipud(target_map)
for i in range(0, heatmap.shape[0]):
heatmap[i] = np.flipud(heatmap[i])
target_heatmap = heatmap.copy()
if random.random() < 0.5:
size = random.randint(self.blur_min, self.blur_max)
if size % 2 == 0:
size = size + 1 if random.random() < 0.5 else size - 1
target_heatmap = cv2.GaussianBlur(target_heatmap, (size, size), 0)
else:
sizeX = random.randint(self.blur_min, self.blur_max)
sizeY = random.randint(self.blur_min, self.blur_max)
if sizeX % 2 == 0:
sizeX = sizeX + 1 if random.random() < 0.5 else sizeX - 1
if sizeY % 2 == 0:
sizeY = sizeY + 1 if random.random() < 0.5 else sizeY - 1
target_heatmap = cv2.GaussianBlur(target_heatmap, (sizeX, sizeY), 0)
target_map = torch.LongTensor(target_map.copy())
target_heatmap = torch.FloatTensor(target_heatmap)
cond_heatmap = torch.FloatTensor(heatmap.copy())
channels, height, width = cond_heatmap.shape
for i in range(channels):
total = height * width
ratio = np.random.random() * (self.max_mask - self.min_mask) + self.min_mask
num = int(total * ratio)
masked_indices = np.random.choice(total, num, replace=False)
mask = np.zeros(total, dtype=bool)
mask[masked_indices] = True
mask = mask.reshape(height, width)
cond_heatmap[i, mask] = 0
return {
"target_map": target_map,
"target_heatmap": target_heatmap,
"cond_heatmap": cond_heatmap
}

View File

@ -18,6 +18,11 @@ class GinkaMaskGIT(nn.Module):
cond_channels = [d_model // 4, d_model // 2, d_model]
self.cond_encoder = GinkaMaskGITCond(input_channel=heatmap_channel, channels=cond_channels)
self.gate_encoder = nn.Sequential(
nn.Conv2d(cond_channels[2], cond_channels[2], 3, padding=1, padding_mode="replicate"),
nn.BatchNorm2d(cond_channels[2]),
nn.GELU()
)
self.cond_gate = nn.Sequential(
nn.Linear(cond_channels[2] * 2, cond_channels[2]),
nn.LayerNorm(cond_channels[2]),
@ -39,8 +44,9 @@ class GinkaMaskGIT(nn.Module):
# output: [B, H * W, num_classes]
heatmap = self.cond_encoder(heatmap) # [B, d_model, H, W]
B, C, H, W = heatmap.shape
heatmap_mean = F.avg_pool2d(heatmap, (H, W)) # [B, d_model, 1, 1]
heatmap_max = F.max_pool2d(heatmap, (H, W)) # [B, d_model, 1, 1]
heatmap_gate = self.gate_encoder(heatmap)
heatmap_mean = F.avg_pool2d(heatmap_gate, (H, W)) # [B, d_model, 1, 1]
heatmap_max = F.max_pool2d(heatmap_gate, (H, W)) # [B, d_model, 1, 1]
gate_input = torch.cat([heatmap_mean, heatmap_max], dim=1).squeeze(2).squeeze(2)
gate = self.cond_gate(gate_input) # [B, d_model]

299
ginka/train_joint.py Normal file
View File

@ -0,0 +1,299 @@
import argparse
import math
import os
import sys
from datetime import datetime
import numpy as np
import torch
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
from tqdm import tqdm
from .dataset import GinkaJointDataset
from .heatmap.diffusion import Diffusion
from .heatmap.model import GinkaHeatmapModel
from .maskGIT.model import GinkaMaskGIT
BATCH_SIZE = 128
VAL_BATCH_DIVIDER = 64
NUM_CLASSES = 16
MASK_TOKEN = 15
GENERATE_STEP = 8
MAP_W = 13
MAP_H = 13
HEATMAP_CHANNEL = 9
LABEL_SMOOTHING = 0
NUM_LAYERS = 4
D_MODEL = 192
NUM_LAYERS_DIFFUSION = 4
D_MODEL_DIFFUSION = 128
T_DIFFUSION = 100
MIN_MASK = 0
MAX_MASK = 1
device = torch.device(
"cuda:1" if torch.cuda.is_available()
else "mps" if torch.mps.is_available()
else "cpu"
)
os.makedirs("result", exist_ok=True)
os.makedirs("result/joint", exist_ok=True)
disable_tqdm = not sys.stdout.isatty()
def parse_arguments():
parser = argparse.ArgumentParser(description="joint training codes")
parser.add_argument("--resume", type=bool, default=False)
parser.add_argument("--state_heatmap", type=str, default="result/heatmap/ginka-100.pth")
parser.add_argument("--train", type=str, default="ginka-dataset.json")
parser.add_argument("--validate", type=str, default="ginka-eval.json")
parser.add_argument("--epochs", type=int, default=50)
parser.add_argument("--checkpoint", type=int, default=5)
parser.add_argument("--load_optim", type=bool, default=True)
parser.add_argument("--maskgit_path", type=str, default="result/ginka_transformer.pth")
parser.add_argument("--ce_weight", type=float, default=1.0)
parser.add_argument("--cfg_drop_rate", type=float, default=0.2)
args = parser.parse_args()
return args
def load_heatmap_checkpoint(model, optimizer, args):
if not args.state_heatmap:
return
if not os.path.exists(args.state_heatmap):
raise FileNotFoundError(f"Heatmap checkpoint not found: {args.state_heatmap}")
checkpoint = torch.load(args.state_heatmap, map_location=device)
model.load_state_dict(checkpoint["model_state"], strict=False)
if args.resume and args.load_optim and checkpoint.get("optim_state") is not None:
optimizer.load_state_dict(checkpoint["optim_state"])
print("Loaded Diffusion model state.")
def freeze_module(module: torch.nn.Module):
module.eval()
for parameter in module.parameters():
parameter.requires_grad = False
def predict_x0(diffusion: Diffusion, x_t: torch.Tensor, pred_noise: torch.Tensor, t: torch.Tensor):
sqrt_ab = diffusion.sqrt_ab[t][:, None, None, None]
sqrt_one_minus_ab = diffusion.sqrt_one_minus_ab[t][:, None, None, None]
x0 = (x_t - sqrt_one_minus_ab * pred_noise) / sqrt_ab
return torch.clamp(x0, 0.0, 1.0)
def maskgit_joint_loss(maskgit, generated_heatmap: torch.Tensor, target_map: torch.Tensor):
batch_size, height, width = target_map.shape
target_tokens = target_map.view(batch_size, height * width)
canvas = torch.full_like(target_tokens, MASK_TOKEN)
losses = []
for step in range(GENERATE_STEP):
current_mask = canvas == MASK_TOKEN
if current_mask.sum().item() == 0:
break
logits = maskgit(canvas, generated_heatmap)
ce = F.cross_entropy(
logits.permute(0, 2, 1),
target_tokens,
reduction='none',
label_smoothing=LABEL_SMOOTHING
)
ce = (ce * current_mask).sum() / (current_mask.sum() + 1e-6)
losses.append(ce)
with torch.no_grad():
probs = F.softmax(logits, dim=-1)
sampled_tiles = torch.argmax(probs, dim=-1)
confidences = torch.gather(probs, -1, sampled_tiles.unsqueeze(-1)).squeeze(-1)
ratio = math.cos(((step + 1) / GENERATE_STEP) * math.pi / 2)
num_to_mask = math.floor(ratio * target_tokens.shape[1])
if num_to_mask > 0:
_, mask_indices = torch.topk(confidences, k=num_to_mask, largest=False)
sampled_tiles = sampled_tiles.scatter(1, mask_indices, MASK_TOKEN)
canvas = sampled_tiles
if not losses:
return torch.zeros((), device=generated_heatmap.device)
return torch.stack(losses).mean()
def validate(model, maskgit, diffusion, dataloader, ce_weight):
model.eval()
total_loss = 0.0
total_diffusion_loss = 0.0
total_maskgit_loss = 0.0
with torch.no_grad():
for batch in tqdm(dataloader, desc="Validating", leave=False, disable=disable_tqdm):
cond_heatmap = batch["cond_heatmap"].to(device)
target_heatmap = batch["target_heatmap"].to(device)
target_map = batch["target_map"].to(device)
batch_size = target_heatmap.shape[0]
t = torch.randint(1, T_DIFFUSION, [batch_size], device=device)
noise = torch.randn_like(target_heatmap)
x_t = diffusion.q_sample(target_heatmap, t, noise)
pred_noise = model(x_t, cond_heatmap, t)
diffusion_loss = F.mse_loss(pred_noise, noise)
generated_heatmap = predict_x0(diffusion, x_t, pred_noise, t)
maskgit_loss = maskgit_joint_loss(maskgit, generated_heatmap, target_map)
loss = diffusion_loss + ce_weight * maskgit_loss
total_loss += loss.item()
total_diffusion_loss += diffusion_loss.item()
total_maskgit_loss += maskgit_loss.item()
size = max(len(dataloader), 1)
return {
"loss": total_loss / size,
"diffusion_loss": total_diffusion_loss / size,
"maskgit_loss": total_maskgit_loss / size,
}
def train():
print(f"Using {device.type} to train model.")
args = parse_arguments()
maskgit = GinkaMaskGIT(
num_classes=NUM_CLASSES,
heatmap_channel=HEATMAP_CHANNEL,
num_layers=NUM_LAYERS,
d_model=D_MODEL,
).to(device)
if not os.path.exists(args.maskgit_path):
raise FileNotFoundError(f"MaskGIT checkpoint not found: {args.maskgit_path}")
maskgit_state = torch.load(args.maskgit_path, map_location=device)
maskgit.load_state_dict(maskgit_state["model_state"])
freeze_module(maskgit)
print("Loaded and froze MaskGIT model state.")
model = GinkaHeatmapModel(
T=T_DIFFUSION,
heatmap_dim=HEATMAP_CHANNEL,
d_model=D_MODEL_DIFFUSION,
num_layers=NUM_LAYERS_DIFFUSION,
).to(device)
diffusion = Diffusion(device, T=T_DIFFUSION)
dataset = GinkaJointDataset(args.train, min_mask=MIN_MASK, max_mask=MAX_MASK)
dataset_val = GinkaJointDataset(args.validate, min_mask=MIN_MASK, max_mask=MAX_MASK)
dataloader = DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=True)
dataloader_val = DataLoader(
dataset_val,
batch_size=max(1, BATCH_SIZE // VAL_BATCH_DIVIDER),
shuffle=True,
)
optimizer = optim.AdamW(model.parameters(), lr=1e-4, weight_decay=1e-2)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer,
T_max=args.epochs,
eta_min=1e-6,
)
load_heatmap_checkpoint(model, optimizer, args)
for epoch in tqdm(range(args.epochs), desc="Joint Training", disable=disable_tqdm):
model.train()
epoch_loss = 0.0
epoch_diffusion_loss = 0.0
epoch_maskgit_loss = 0.0
for batch in tqdm(dataloader, leave=False, desc="Epoch Progress", disable=disable_tqdm):
cond_heatmap = batch["cond_heatmap"].to(device)
target_heatmap = batch["target_heatmap"].to(device)
target_map = batch["target_map"].to(device)
batch_size = target_heatmap.shape[0]
optimizer.zero_grad()
t = torch.randint(1, T_DIFFUSION, [batch_size], device=device)
noise = torch.randn_like(target_heatmap)
x_t = diffusion.q_sample(target_heatmap, t, noise)
cond_for_diffusion = cond_heatmap
use_unconditional_branch = False
if np.random.rand() < args.cfg_drop_rate:
cond_for_diffusion = torch.zeros_like(cond_heatmap)
use_unconditional_branch = True
pred_noise = model(x_t, cond_for_diffusion, t)
diffusion_loss = F.mse_loss(pred_noise, noise)
pred_noise_for_joint = pred_noise
if use_unconditional_branch:
pred_noise_for_joint = model(x_t, cond_heatmap, t)
generated_heatmap = predict_x0(diffusion, x_t, pred_noise_for_joint, t)
maskgit_loss = maskgit_joint_loss(maskgit, generated_heatmap, target_map)
loss = diffusion_loss + args.ce_weight * maskgit_loss
loss.backward()
optimizer.step()
epoch_loss += loss.item()
epoch_diffusion_loss += diffusion_loss.item()
epoch_maskgit_loss += maskgit_loss.item()
scheduler.step()
train_size = max(len(dataloader), 1)
tqdm.write(
f"[Epoch {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}] "
f"E: {epoch + 1} | "
f"Loss: {epoch_loss / train_size:.6f} | "
f"Diffusion: {epoch_diffusion_loss / train_size:.6f} | "
f"MaskGIT: {epoch_maskgit_loss / train_size:.6f} | "
f"LR: {scheduler.get_last_lr()[0]:.6f}"
)
if (epoch + 1) % args.checkpoint == 0:
checkpoint_path = f"result/joint/ginka-joint-{epoch + 1}.pth"
torch.save(
{
"model_state": model.state_dict(),
"optim_state": optimizer.state_dict(),
},
checkpoint_path,
)
metrics = validate(model, maskgit, diffusion, dataloader_val, args.ce_weight)
tqdm.write(
f"[Validate {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}] "
f"E: {epoch + 1} | "
f"Loss: {metrics['loss']:.6f} | "
f"Diffusion: {metrics['diffusion_loss']:.6f} | "
f"MaskGIT: {metrics['maskgit_loss']:.6f}"
)
print("Train ended.")
torch.save(
{
"model_state": model.state_dict(),
"optim_state": optimizer.state_dict(),
},
"result/ginka_joint_heatmap.pth",
)
if __name__ == "__main__":
torch.set_num_threads(4)
train()