code/scripts/vis_kvasir_progression.py

"""
Kvasir-SEG: Training progression visualization
Compare generated images across different checkpoints
Layout: Mask | 10k | 30k | 50k | 70k | 100k | Real
"""
import sys
sys.path.insert(0, "/data/sichengli/Code/PixelGen")

import torch
import numpy as np
from PIL import Image, ImageDraw, ImageFont
import torchvision.transforms as transforms
import torchvision.transforms.functional as TF
import os, random, gc

from src.models.transformer.JiT_medical import JiTMedical

device = torch.device("cuda:0")

# Checkpoints to compare
ckpt_dir = "/data/sichengli/Code/PixelGen/medical_workdirs/exp_PixelGen_Medical_Kvasir"
checkpoints = [
    ("10k",  os.path.join(ckpt_dir, "epoch=1249-step=10000.ckpt")),
    ("30k",  os.path.join(ckpt_dir, "epoch=3749-step=30000.ckpt")),
    ("50k",  os.path.join(ckpt_dir, "epoch=6249-step=50000.ckpt")),
    ("70k",  os.path.join(ckpt_dir, "epoch=8749-step=70000.ckpt")),
    ("100k", os.path.join(ckpt_dir, "epoch=12499-step=100000.ckpt")),
]

model_kwargs = dict(
    input_size=256, patch_size=16, in_channels=3,
    hidden_size=768, depth=12, num_heads=12, mlp_ratio=4.0,
    attn_drop=0.0, proj_drop=0.1, num_classes=1,
    use_bottleneck=True, bottleneck_dim=128,
    in_context_len=32, in_context_start=4, mask_in_channels=1,
    mask_mode="spatial"
)


def load_ema_model(ckpt_path):
    ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)
    state_dict = ckpt["state_dict"]
    ema_state = {}
    for k, v in state_dict.items():
        if k.startswith("ema_denoiser."):
            new_k = k.replace("ema_denoiser.", "").replace("_orig_mod.", "")
            ema_state[new_k] = v
    model = JiTMedical(**model_kwargs)
    model.load_state_dict(ema_state, strict=False)
    model = model.to(device).eval().to(torch.float32)
    return model


def shift_respace_fn(t, shift=1.0):
    return t / (t + (1 - t) * shift)


@torch.no_grad()
def sample_no_cfg(model, noise, mask, num_steps=50, t_eps=0.05):
    batch_size = noise.shape[0]
    timesteps = torch.linspace(0.0, 1 - 1.0 / num_steps, num_steps)
    timesteps = torch.cat([timesteps, torch.tensor([1.0])], dim=0)
    timesteps = shift_respace_fn(timesteps, 1.0).to(noise.device)
    y = torch.zeros(batch_size, dtype=torch.long, device=noise.device)
    x = noise.clone()
    for i in range(len(timesteps) - 1):
        t_cur = timesteps[i]
        t_next = timesteps[i + 1]
        dt = t_next - t_cur
        t_batch = t_cur.repeat(batch_size)
        pred_img = model(x, t_batch, y, mask=mask)
        v = (pred_img - x) / (1.0 - t_batch.view(-1, 1, 1, 1)).clamp_min(t_eps)
        x = x + v * dt
    return x


# Load val samples
data_root = "/data2/sichengli/Data/test/Segmentation/Kvasir-SEG/Kvasir-SEG"
img_dir = os.path.join(data_root, "images")
mask_dir = os.path.join(data_root, "masks")
all_files = sorted([f for f in os.listdir(img_dir) if f.endswith((".jpg", ".png", ".jpeg"))])

random.seed(42)
indices = list(range(len(all_files)))
random.shuffle(indices)
val_indices = indices[int(len(indices) * 0.9):]
val_files = [all_files[i] for i in sorted(val_indices)]

random.seed(456)
selected = random.sample(val_files, min(6, len(val_files)))

images_list, masks_list = [], []
for fname in selected:
    img = Image.open(os.path.join(img_dir, fname)).convert("RGB")
    img = TF.resize(img, (256, 256))
    images_list.append(TF.to_tensor(img))
    mask = Image.open(os.path.join(mask_dir, fname)).convert("L")
    mask = TF.resize(mask, (256, 256), interpolation=transforms.InterpolationMode.NEAREST)
    masks_list.append(TF.to_tensor(mask))

real_images = torch.stack(images_list)
masks_tensor = torch.stack(masks_list).to(device)

# Use same noise for all checkpoints
torch.manual_seed(123)
shared_noise = torch.randn(len(selected), 3, 256, 256, device=device)

# Generate for each checkpoint
generated = {}
for name, ckpt_path in checkpoints:
    print(f"Generating with {name} checkpoint...")
    model = load_ema_model(ckpt_path)
    gen = sample_no_cfg(model, shared_noise, masks_tensor).clamp(-1, 1) * 0.5 + 0.5
    generated[name] = gen.cpu()
    del model
    gc.collect()
    torch.cuda.empty_cache()

# Build visualization grid
col_labels = ["Mask"] + [name for name, _ in checkpoints] + ["Real"]
n_rows = len(selected)
n_cols = len(col_labels)
h, w = 256, 256
pad = 4
label_h = 36

canvas_w = n_cols * w + (n_cols + 1) * pad
canvas_h = n_rows * h + (n_rows + 1) * pad + label_h
canvas = np.ones((canvas_h, canvas_w, 3), dtype=np.uint8) * 30

try:
    font = ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf", 20)
except Exception:
    font = ImageFont.load_default()

pil_canvas = Image.fromarray(canvas)
draw = ImageDraw.Draw(pil_canvas)
for col, label in enumerate(col_labels):
    x_pos = pad + col * (w + pad) + w // 2
    bbox = draw.textbbox((0, 0), label, font=font)
    text_w = bbox[2] - bbox[0]
    draw.text((x_pos - text_w // 2, 8), label, fill=(255, 255, 255), font=font)
canvas = np.array(pil_canvas)

for row in range(n_rows):
    y = label_h + pad + row * (h + pad)
    for col_idx, col_name in enumerate(col_labels):
        x = pad + col_idx * (w + pad)
        if col_name == "Mask":
            m = masks_tensor[row, 0].cpu().numpy()
            m_rgb = np.stack([m, m, m], axis=-1)
            canvas[y:y+h, x:x+w] = (m_rgb * 255).clip(0, 255).astype(np.uint8)
        elif col_name == "Real":
            r = real_images[row].permute(1, 2, 0).numpy()
            canvas[y:y+h, x:x+w] = (r * 255).clip(0, 255).astype(np.uint8)
        else:
            g = generated[col_name][row].permute(1, 2, 0).numpy()
            canvas[y:y+h, x:x+w] = (g * 255).clip(0, 255).astype(np.uint8)

out_dir = os.path.join(ckpt_dir, "val_samples")
os.makedirs(out_dir, exist_ok=True)
out_path = os.path.join(out_dir, "training_progression.png")
Image.fromarray(canvas).save(out_path)
print(f"\nSaved: {out_path} ({canvas_w}x{canvas_h})")