Add one-click Colab training notebook with real Pokemon dataset
Browse files- colab_train_iris.py +455 -0
colab_train_iris.py
ADDED
|
@@ -0,0 +1,455 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""
|
| 3 |
+
IRIS Colab Training — One-Click, Real Dataset, Real Learning
|
| 4 |
+
=============================================================
|
| 5 |
+
|
| 6 |
+
Copy-paste into Google Colab (free tier T4) and run all cells.
|
| 7 |
+
Trains IRIS on Pokemon BLIP Captions (833 images + text).
|
| 8 |
+
|
| 9 |
+
Colab free tier specs (2025):
|
| 10 |
+
- GPU: NVIDIA T4 (16 GB VRAM)
|
| 11 |
+
- System RAM: ~12.7 GB
|
| 12 |
+
- Disk: ~78 GB
|
| 13 |
+
- PyTorch: 2.5+ preinstalled
|
| 14 |
+
- Runtime: ~12 hours max session
|
| 15 |
+
|
| 16 |
+
What this script does:
|
| 17 |
+
1. Installs dependencies (~30s)
|
| 18 |
+
2. Downloads IRIS source from HF Hub
|
| 19 |
+
3. Downloads DC-AE encoder (1.2 GB) + text encoder (87 MB)
|
| 20 |
+
4. Encodes all 833 Pokemon images to latents (~2 min on T4)
|
| 21 |
+
5. Encodes all captions to text embeddings (~5s)
|
| 22 |
+
6. Frees encoder VRAM
|
| 23 |
+
7. Trains IRIS-Small (40M params) for 3000 steps (~15 min on T4)
|
| 24 |
+
8. Generates sample images from trained model
|
| 25 |
+
9. Saves checkpoint
|
| 26 |
+
|
| 27 |
+
Total wall time: ~20 minutes for a trained model.
|
| 28 |
+
"""
|
| 29 |
+
|
| 30 |
+
# ============================================================
|
| 31 |
+
# CELL 1: Install dependencies
|
| 32 |
+
# ============================================================
|
| 33 |
+
print("Installing dependencies...")
|
| 34 |
+
import subprocess, sys
|
| 35 |
+
|
| 36 |
+
subprocess.check_call([sys.executable, "-m", "pip", "install", "-q",
|
| 37 |
+
"diffusers>=0.32.0",
|
| 38 |
+
"sentence-transformers",
|
| 39 |
+
"datasets",
|
| 40 |
+
"accelerate",
|
| 41 |
+
"huggingface_hub",
|
| 42 |
+
])
|
| 43 |
+
print("Done.")
|
| 44 |
+
|
| 45 |
+
# ============================================================
|
| 46 |
+
# CELL 2: Download IRIS source code
|
| 47 |
+
# ============================================================
|
| 48 |
+
print("Downloading IRIS architecture from HF Hub...")
|
| 49 |
+
from huggingface_hub import snapshot_download
|
| 50 |
+
import os, shutil
|
| 51 |
+
|
| 52 |
+
iris_path = snapshot_download(
|
| 53 |
+
repo_id="asdf98/iris-image-gen",
|
| 54 |
+
allow_patterns=["iris/*.py"],
|
| 55 |
+
local_dir="./iris_repo",
|
| 56 |
+
)
|
| 57 |
+
# Add to Python path
|
| 58 |
+
sys.path.insert(0, os.path.join(iris_path))
|
| 59 |
+
print(f"IRIS source at: {iris_path}")
|
| 60 |
+
|
| 61 |
+
# Verify import
|
| 62 |
+
from iris import IRIS, get_model_config, flow_matching_loss, euler_sample
|
| 63 |
+
from iris.flow_matching import DCAE_F32C32_SCALE
|
| 64 |
+
print("IRIS imported successfully.")
|
| 65 |
+
|
| 66 |
+
# ============================================================
|
| 67 |
+
# CELL 3: Detect hardware
|
| 68 |
+
# ============================================================
|
| 69 |
+
import torch
|
| 70 |
+
import gc
|
| 71 |
+
|
| 72 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 73 |
+
if device.type == "cuda":
|
| 74 |
+
gpu_name = torch.cuda.get_device_name(0)
|
| 75 |
+
gpu_mem = torch.cuda.get_device_properties(0).total_mem / 1e9
|
| 76 |
+
print(f"GPU: {gpu_name} ({gpu_mem:.1f} GB)")
|
| 77 |
+
else:
|
| 78 |
+
print("WARNING: No GPU detected. Training will be very slow.")
|
| 79 |
+
print("In Colab: Runtime -> Change runtime type -> T4 GPU")
|
| 80 |
+
|
| 81 |
+
use_amp = device.type == "cuda"
|
| 82 |
+
amp_dtype = torch.bfloat16 if (use_amp and torch.cuda.is_bf16_supported()) else torch.float16 if use_amp else torch.float32
|
| 83 |
+
print(f"AMP dtype: {amp_dtype}")
|
| 84 |
+
|
| 85 |
+
# ============================================================
|
| 86 |
+
# CELL 4: Load dataset
|
| 87 |
+
# ============================================================
|
| 88 |
+
print("\nLoading Pokemon BLIP Captions dataset...")
|
| 89 |
+
from datasets import load_dataset
|
| 90 |
+
|
| 91 |
+
ds = load_dataset("reach-vb/pokemon-blip-captions", split="train")
|
| 92 |
+
print(f"Loaded {len(ds)} images with captions.")
|
| 93 |
+
print(f"Example: '{ds[0]['text']}'")
|
| 94 |
+
|
| 95 |
+
# ============================================================
|
| 96 |
+
# CELL 5: Encode all images to DC-AE latents
|
| 97 |
+
# ============================================================
|
| 98 |
+
print("\nLoading DC-AE encoder (~1.2 GB)...")
|
| 99 |
+
from diffusers import AutoencoderDC
|
| 100 |
+
import torchvision.transforms as T
|
| 101 |
+
|
| 102 |
+
# Use float16 to save VRAM — stable for inference
|
| 103 |
+
ae = AutoencoderDC.from_pretrained(
|
| 104 |
+
"mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers",
|
| 105 |
+
torch_dtype=torch.float16,
|
| 106 |
+
).to(device).eval()
|
| 107 |
+
ae.requires_grad_(False)
|
| 108 |
+
|
| 109 |
+
SCALE = ae.config.scaling_factor # 0.41407
|
| 110 |
+
|
| 111 |
+
transform = T.Compose([
|
| 112 |
+
T.Resize(512, interpolation=T.InterpolationMode.BICUBIC, antialias=True),
|
| 113 |
+
T.CenterCrop(512),
|
| 114 |
+
T.ToTensor(),
|
| 115 |
+
T.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]),
|
| 116 |
+
])
|
| 117 |
+
|
| 118 |
+
print("Encoding images to latents...")
|
| 119 |
+
all_latents = []
|
| 120 |
+
import time
|
| 121 |
+
t0 = time.time()
|
| 122 |
+
|
| 123 |
+
batch_imgs = []
|
| 124 |
+
for i, example in enumerate(ds):
|
| 125 |
+
img = example["image"].convert("RGB")
|
| 126 |
+
tensor = transform(img)
|
| 127 |
+
batch_imgs.append(tensor)
|
| 128 |
+
|
| 129 |
+
# Process in batches of 8
|
| 130 |
+
if len(batch_imgs) == 8 or i == len(ds) - 1:
|
| 131 |
+
batch = torch.stack(batch_imgs).to(device, dtype=torch.float16)
|
| 132 |
+
with torch.no_grad():
|
| 133 |
+
latent = ae.encode(batch).latent.float() # encode in fp16, store in fp32
|
| 134 |
+
all_latents.append(latent.cpu())
|
| 135 |
+
batch_imgs = []
|
| 136 |
+
|
| 137 |
+
if (i + 1) % 100 == 0 or i == len(ds) - 1:
|
| 138 |
+
print(f" Encoded {i+1}/{len(ds)} images ({time.time()-t0:.1f}s)")
|
| 139 |
+
|
| 140 |
+
all_latents = torch.cat(all_latents, dim=0) # (N, 32, 16, 16)
|
| 141 |
+
print(f"All latents: {all_latents.shape}, took {time.time()-t0:.1f}s")
|
| 142 |
+
print(f"Latent stats: mean={all_latents.mean():.3f}, std={all_latents.std():.3f}")
|
| 143 |
+
|
| 144 |
+
# Free DC-AE VRAM
|
| 145 |
+
del ae
|
| 146 |
+
torch.cuda.empty_cache()
|
| 147 |
+
gc.collect()
|
| 148 |
+
print("DC-AE encoder freed from VRAM.")
|
| 149 |
+
|
| 150 |
+
# ============================================================
|
| 151 |
+
# CELL 6: Encode all captions to text embeddings
|
| 152 |
+
# ============================================================
|
| 153 |
+
print("\nLoading text encoder (~87 MB)...")
|
| 154 |
+
from sentence_transformers import SentenceTransformer
|
| 155 |
+
|
| 156 |
+
text_encoder = SentenceTransformer(
|
| 157 |
+
"sentence-transformers/all-MiniLM-L6-v2",
|
| 158 |
+
device=str(device),
|
| 159 |
+
)
|
| 160 |
+
text_encoder.eval()
|
| 161 |
+
|
| 162 |
+
captions = [ex["text"] for ex in ds]
|
| 163 |
+
print(f"Encoding {len(captions)} captions...")
|
| 164 |
+
|
| 165 |
+
with torch.no_grad():
|
| 166 |
+
all_text_embs = text_encoder.encode(
|
| 167 |
+
captions,
|
| 168 |
+
convert_to_tensor=True,
|
| 169 |
+
normalize_embeddings=True,
|
| 170 |
+
batch_size=128,
|
| 171 |
+
show_progress_bar=True,
|
| 172 |
+
)
|
| 173 |
+
|
| 174 |
+
# Expand to sequence format: (N, 1, 384)
|
| 175 |
+
# The model projects 384 -> model_dim via registered context_proj
|
| 176 |
+
all_text_embs = all_text_embs.unsqueeze(1).cpu() # (N, 1, 384)
|
| 177 |
+
print(f"Text embeddings: {all_text_embs.shape}")
|
| 178 |
+
|
| 179 |
+
# Free text encoder VRAM
|
| 180 |
+
del text_encoder
|
| 181 |
+
torch.cuda.empty_cache()
|
| 182 |
+
gc.collect()
|
| 183 |
+
print("Text encoder freed from VRAM.")
|
| 184 |
+
|
| 185 |
+
# ============================================================
|
| 186 |
+
# CELL 7: Create dataset from precomputed features
|
| 187 |
+
# ============================================================
|
| 188 |
+
from torch.utils.data import Dataset, DataLoader
|
| 189 |
+
|
| 190 |
+
class PrecomputedLatentDataset(Dataset):
|
| 191 |
+
"""All latents and text embeddings precomputed — zero I/O during training."""
|
| 192 |
+
def __init__(self, latents, text_embs):
|
| 193 |
+
self.latents = latents
|
| 194 |
+
self.text_embs = text_embs
|
| 195 |
+
|
| 196 |
+
def __len__(self):
|
| 197 |
+
return len(self.latents)
|
| 198 |
+
|
| 199 |
+
def __getitem__(self, idx):
|
| 200 |
+
return {
|
| 201 |
+
"latent": self.latents[idx],
|
| 202 |
+
"text_embed": self.text_embs[idx],
|
| 203 |
+
}
|
| 204 |
+
|
| 205 |
+
train_ds = PrecomputedLatentDataset(all_latents, all_text_embs)
|
| 206 |
+
print(f"Training dataset: {len(train_ds)} samples")
|
| 207 |
+
print(f" Latent: {train_ds[0]['latent'].shape}")
|
| 208 |
+
print(f" Text: {train_ds[0]['text_embed'].shape}")
|
| 209 |
+
|
| 210 |
+
# ============================================================
|
| 211 |
+
# CELL 8: Create IRIS model
|
| 212 |
+
# ============================================================
|
| 213 |
+
print("\nCreating IRIS-Small model...")
|
| 214 |
+
|
| 215 |
+
model = IRIS(
|
| 216 |
+
**get_model_config("iris-small"),
|
| 217 |
+
gradient_checkpointing=True,
|
| 218 |
+
text_dim=384, # all-MiniLM-L6-v2 output dim — registered as proper nn.Module
|
| 219 |
+
).to(device)
|
| 220 |
+
|
| 221 |
+
counts = model.count_params()
|
| 222 |
+
print(f"Parameters: {counts['total']:,} ({counts['total']/1e6:.1f}M)")
|
| 223 |
+
print(f" Core: {counts['core']:,}")
|
| 224 |
+
print(f" Decoder: {counts['tiny_decoder']:,}")
|
| 225 |
+
|
| 226 |
+
if device.type == "cuda":
|
| 227 |
+
print(f"VRAM used: {torch.cuda.memory_allocated()/1e9:.2f} GB / {torch.cuda.get_device_properties(0).total_mem/1e9:.1f} GB")
|
| 228 |
+
|
| 229 |
+
# ============================================================
|
| 230 |
+
# CELL 9: Train!
|
| 231 |
+
# ============================================================
|
| 232 |
+
import math
|
| 233 |
+
from iris.train import CosineWarmupScheduler
|
| 234 |
+
from iris.flow_matching import flow_matching_loss
|
| 235 |
+
|
| 236 |
+
# Training config — tuned for Colab T4 with 833 Pokemon images
|
| 237 |
+
NUM_STEPS = 3000 # ~15 min on T4
|
| 238 |
+
BATCH_SIZE = 16 # fits T4 with IRIS-Small + grad checkpoint
|
| 239 |
+
LR = 3e-4 # slightly higher LR for small dataset
|
| 240 |
+
WARMUP_STEPS = 200
|
| 241 |
+
GRAD_CLIP = 1.0
|
| 242 |
+
NUM_ITERS = 3 # refinement iterations (3 is good for speed/quality)
|
| 243 |
+
LOG_EVERY = 50
|
| 244 |
+
SAVE_EVERY = 1000
|
| 245 |
+
|
| 246 |
+
loader = DataLoader(
|
| 247 |
+
train_ds,
|
| 248 |
+
batch_size=BATCH_SIZE,
|
| 249 |
+
shuffle=True,
|
| 250 |
+
num_workers=2,
|
| 251 |
+
pin_memory=True,
|
| 252 |
+
drop_last=True,
|
| 253 |
+
persistent_workers=True,
|
| 254 |
+
)
|
| 255 |
+
|
| 256 |
+
optimizer = torch.optim.AdamW(
|
| 257 |
+
model.parameters(),
|
| 258 |
+
lr=LR,
|
| 259 |
+
weight_decay=0.01,
|
| 260 |
+
betas=(0.9, 0.999),
|
| 261 |
+
)
|
| 262 |
+
scheduler = CosineWarmupScheduler(optimizer, WARMUP_STEPS, NUM_STEPS, min_lr_ratio=0.05)
|
| 263 |
+
scaler = torch.amp.GradScaler(enabled=(use_amp and amp_dtype == torch.float16))
|
| 264 |
+
|
| 265 |
+
model.train()
|
| 266 |
+
step = 0
|
| 267 |
+
epoch = 0
|
| 268 |
+
running_loss = 0.0
|
| 269 |
+
loss_history = []
|
| 270 |
+
best_loss = float("inf")
|
| 271 |
+
t_start = time.time()
|
| 272 |
+
|
| 273 |
+
print(f"\n{'='*60}")
|
| 274 |
+
print(f"Training IRIS-Small on Pokemon BLIP Captions")
|
| 275 |
+
print(f" {len(train_ds)} images, {NUM_STEPS} steps, BS={BATCH_SIZE}, R={NUM_ITERS}")
|
| 276 |
+
print(f" LR={LR}, warmup={WARMUP_STEPS}, AMP={amp_dtype}")
|
| 277 |
+
print(f"{'='*60}\n")
|
| 278 |
+
|
| 279 |
+
while step < NUM_STEPS:
|
| 280 |
+
epoch += 1
|
| 281 |
+
for batch in loader:
|
| 282 |
+
if step >= NUM_STEPS:
|
| 283 |
+
break
|
| 284 |
+
|
| 285 |
+
latent = batch["latent"].to(device, non_blocking=True)
|
| 286 |
+
text_embed = batch["text_embed"].to(device, non_blocking=True)
|
| 287 |
+
|
| 288 |
+
with torch.amp.autocast(device_type=device.type, dtype=amp_dtype, enabled=use_amp):
|
| 289 |
+
losses = flow_matching_loss(
|
| 290 |
+
model, latent, text_embed,
|
| 291 |
+
num_iterations=NUM_ITERS,
|
| 292 |
+
timestep_sampling="logit_normal",
|
| 293 |
+
scale_factor=SCALE,
|
| 294 |
+
)
|
| 295 |
+
loss = losses["loss"]
|
| 296 |
+
|
| 297 |
+
optimizer.zero_grad(set_to_none=True)
|
| 298 |
+
if scaler.is_enabled():
|
| 299 |
+
scaler.scale(loss).backward()
|
| 300 |
+
scaler.unscale_(optimizer)
|
| 301 |
+
gn = torch.nn.utils.clip_grad_norm_(model.parameters(), GRAD_CLIP)
|
| 302 |
+
scaler.step(optimizer)
|
| 303 |
+
scaler.update()
|
| 304 |
+
else:
|
| 305 |
+
loss.backward()
|
| 306 |
+
gn = torch.nn.utils.clip_grad_norm_(model.parameters(), GRAD_CLIP)
|
| 307 |
+
optimizer.step()
|
| 308 |
+
|
| 309 |
+
scheduler.step()
|
| 310 |
+
step += 1
|
| 311 |
+
lv = loss.item()
|
| 312 |
+
running_loss += lv
|
| 313 |
+
loss_history.append(lv)
|
| 314 |
+
|
| 315 |
+
if step % LOG_EVERY == 0:
|
| 316 |
+
avg = running_loss / LOG_EVERY
|
| 317 |
+
elapsed = time.time() - t_start
|
| 318 |
+
sps = step / elapsed
|
| 319 |
+
eta = (NUM_STEPS - step) / sps
|
| 320 |
+
lr = scheduler.get_lr()[0]
|
| 321 |
+
gn_val = gn.item() if isinstance(gn, torch.Tensor) else gn
|
| 322 |
+
tag = "OK" if not (math.isnan(avg) or math.isinf(avg)) else "!!"
|
| 323 |
+
|
| 324 |
+
print(
|
| 325 |
+
f"[{tag}] step {step:>5d}/{NUM_STEPS} | "
|
| 326 |
+
f"loss={avg:.4f} | "
|
| 327 |
+
f"grad={gn_val:.3f} | "
|
| 328 |
+
f"lr={lr:.1e} | "
|
| 329 |
+
f"{sps:.1f} steps/s | "
|
| 330 |
+
f"ETA {eta/60:.0f}min"
|
| 331 |
+
)
|
| 332 |
+
|
| 333 |
+
if avg < best_loss:
|
| 334 |
+
best_loss = avg
|
| 335 |
+
running_loss = 0.0
|
| 336 |
+
|
| 337 |
+
if step % SAVE_EVERY == 0:
|
| 338 |
+
os.makedirs("./iris_checkpoints", exist_ok=True)
|
| 339 |
+
p = f"./iris_checkpoints/iris_pokemon_step{step}.pt"
|
| 340 |
+
torch.save({
|
| 341 |
+
"step": step,
|
| 342 |
+
"model_state_dict": model.state_dict(),
|
| 343 |
+
"loss_history": loss_history,
|
| 344 |
+
"config": get_model_config("iris-small"),
|
| 345 |
+
}, p)
|
| 346 |
+
print(f" Saved: {p}")
|
| 347 |
+
|
| 348 |
+
# Final save
|
| 349 |
+
os.makedirs("./iris_checkpoints", exist_ok=True)
|
| 350 |
+
final_path = "./iris_checkpoints/iris_pokemon_final.pt"
|
| 351 |
+
torch.save({
|
| 352 |
+
"step": step,
|
| 353 |
+
"model_state_dict": model.state_dict(),
|
| 354 |
+
"loss_history": loss_history,
|
| 355 |
+
"config": get_model_config("iris-small"),
|
| 356 |
+
}, final_path)
|
| 357 |
+
|
| 358 |
+
total_time = time.time() - t_start
|
| 359 |
+
f50 = sum(loss_history[:50]) / min(50, len(loss_history))
|
| 360 |
+
l50 = sum(loss_history[-50:]) / min(50, len(loss_history))
|
| 361 |
+
print(f"\n{'='*60}")
|
| 362 |
+
print(f"Training complete!")
|
| 363 |
+
print(f" {step} steps in {total_time/60:.1f} min ({step/total_time:.1f} steps/s)")
|
| 364 |
+
print(f" Loss: {f50:.4f} -> {l50:.4f} ({(1-l50/f50)*100:.1f}% reduction)")
|
| 365 |
+
print(f" Best: {best_loss:.4f}")
|
| 366 |
+
print(f" Saved: {final_path}")
|
| 367 |
+
print(f"{'='*60}")
|
| 368 |
+
|
| 369 |
+
# ============================================================
|
| 370 |
+
# CELL 10: Plot training loss
|
| 371 |
+
# ============================================================
|
| 372 |
+
try:
|
| 373 |
+
import matplotlib.pyplot as plt
|
| 374 |
+
|
| 375 |
+
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(14, 5))
|
| 376 |
+
|
| 377 |
+
ax1.plot(loss_history, alpha=0.3, color="blue", linewidth=0.5)
|
| 378 |
+
window = 50
|
| 379 |
+
if len(loss_history) > window:
|
| 380 |
+
smoothed = [sum(loss_history[max(0,i-window):i+1])/min(i+1, window) for i in range(len(loss_history))]
|
| 381 |
+
ax1.plot(smoothed, color="red", linewidth=2, label=f"Smoothed (w={window})")
|
| 382 |
+
ax1.set_xlabel("Step")
|
| 383 |
+
ax1.set_ylabel("Flow Matching Loss")
|
| 384 |
+
ax1.set_title("Training Loss")
|
| 385 |
+
ax1.legend()
|
| 386 |
+
ax1.grid(True, alpha=0.3)
|
| 387 |
+
|
| 388 |
+
chunks = [loss_history[i:i+100] for i in range(0, len(loss_history), 100)]
|
| 389 |
+
if len(chunks) > 1:
|
| 390 |
+
ax2.boxplot([c for c in chunks], positions=list(range(len(chunks))))
|
| 391 |
+
ax2.set_xlabel("Step (x100)")
|
| 392 |
+
ax2.set_ylabel("Loss")
|
| 393 |
+
ax2.set_title("Loss Distribution Over Time")
|
| 394 |
+
ax2.grid(True, alpha=0.3)
|
| 395 |
+
|
| 396 |
+
plt.tight_layout()
|
| 397 |
+
plt.savefig("./iris_checkpoints/training_loss.png", dpi=100)
|
| 398 |
+
plt.show()
|
| 399 |
+
print("Loss plot saved.")
|
| 400 |
+
except ImportError:
|
| 401 |
+
print("matplotlib not available, skipping loss plot")
|
| 402 |
+
|
| 403 |
+
# ============================================================
|
| 404 |
+
# CELL 11: Generate sample images from trained model
|
| 405 |
+
# ============================================================
|
| 406 |
+
print("\nGenerating sample images from trained model...")
|
| 407 |
+
|
| 408 |
+
# Reload DC-AE decoder for visualization
|
| 409 |
+
ae_decoder = AutoencoderDC.from_pretrained(
|
| 410 |
+
"mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers",
|
| 411 |
+
torch_dtype=torch.float16,
|
| 412 |
+
).to(device).eval()
|
| 413 |
+
ae_decoder.requires_grad_(False)
|
| 414 |
+
|
| 415 |
+
# Reload text encoder for new prompts
|
| 416 |
+
text_enc = SentenceTransformer(
|
| 417 |
+
"sentence-transformers/all-MiniLM-L6-v2",
|
| 418 |
+
device=str(device),
|
| 419 |
+
)
|
| 420 |
+
|
| 421 |
+
model.eval()
|
| 422 |
+
|
| 423 |
+
sample_prompts = [
|
| 424 |
+
"a blue water pokemon with fins",
|
| 425 |
+
"a fire dragon pokemon with wings",
|
| 426 |
+
"a cute pink pokemon with big eyes",
|
| 427 |
+
"a green grass pokemon",
|
| 428 |
+
]
|
| 429 |
+
|
| 430 |
+
for i, prompt in enumerate(sample_prompts):
|
| 431 |
+
with torch.no_grad():
|
| 432 |
+
txt_emb = text_enc.encode(
|
| 433 |
+
[prompt], convert_to_tensor=True, normalize_embeddings=True
|
| 434 |
+
).unsqueeze(1).to(device) # (1, 1, 384)
|
| 435 |
+
|
| 436 |
+
noise = torch.randn(1, 32, 16, 16, device=device)
|
| 437 |
+
|
| 438 |
+
with torch.no_grad():
|
| 439 |
+
z_pred = euler_sample(
|
| 440 |
+
model, noise, txt_emb,
|
| 441 |
+
num_steps=20,
|
| 442 |
+
num_iterations=NUM_ITERS,
|
| 443 |
+
cfg_scale=1.0,
|
| 444 |
+
scale_factor=SCALE,
|
| 445 |
+
)
|
| 446 |
+
img = ae_decoder.decode(z_pred.half()).sample
|
| 447 |
+
img = (img.float().clamp(-1, 1) * 0.5 + 0.5)
|
| 448 |
+
|
| 449 |
+
from torchvision.utils import save_image
|
| 450 |
+
fname = f"./iris_checkpoints/sample_{i}_{prompt[:20].replace(' ','_')}.png"
|
| 451 |
+
save_image(img, fname)
|
| 452 |
+
print(f" Sample {i}: '{prompt}' -> {fname}")
|
| 453 |
+
|
| 454 |
+
print("\nAll samples saved to ./iris_checkpoints/")
|
| 455 |
+
print("NOTE: Trained on 833 images for 3000 steps — quality improves with more data + steps.")
|