File size: 9,724 Bytes
3cfa9a4 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 |
#!/usr/bin/env python3
"""
End-to-end test: data loading → model forward → backward.
Verifies that the full pipeline works before committing to long training.
Usage:
python test_pipeline.py
python test_pipeline.py --dataset active_matter --no-streaming --local_path /data/well
"""
import argparse
import sys
import time
import traceback
import torch
import torch.nn as nn
def fmt_mem():
if torch.cuda.is_available():
alloc = torch.cuda.memory_allocated() / 1e9
res = torch.cuda.memory_reserved() / 1e9
total = torch.cuda.get_device_properties(0).total_memory / 1e9
return f"alloc={alloc:.2f}GB, reserved={res:.2f}GB, total={total:.1f}GB"
return "CPU only"
def test_data_loading(args):
"""Test 1: Load data and print shapes."""
print("\n" + "=" * 60)
print("TEST 1: Data Loading")
print("=" * 60)
from data_pipeline import create_dataloader, prepare_batch, get_channel_info, get_data_info
t0 = time.time()
loader, dataset = create_dataloader(
dataset_name=args.dataset,
split="train",
batch_size=args.batch_size,
streaming=args.streaming,
local_path=args.local_path,
)
print(f" Dataset created in {time.time() - t0:.1f}s")
print(f" Dataset length: {len(dataset)}")
# Probe shapes
info = get_data_info(dataset)
print(f" Sample fields:")
for k, v in info.items():
print(f" {k}: {v}")
ch = get_channel_info(dataset)
print(f" Channel info: {ch}")
# Load one batch
t0 = time.time()
batch = next(iter(loader))
print(f" First batch loaded in {time.time() - t0:.1f}s")
print(f" Batch keys: {list(batch.keys())}")
for k, v in batch.items():
if isinstance(v, torch.Tensor):
print(f" {k}: {v.shape} ({v.dtype})")
# Prepare for model
device = "cuda" if torch.cuda.is_available() else "cpu"
x_in, x_out = prepare_batch(batch, device)
print(f" Model input: {x_in.shape} ({x_in.dtype})")
print(f" Model target: {x_out.shape} ({x_out.dtype})")
print(f" GPU memory: {fmt_mem()}")
return ch, x_in, x_out
def test_diffusion(ch, x_in, x_out):
"""Test 2: Diffusion model forward + backward."""
print("\n" + "=" * 60)
print("TEST 2: Diffusion Model")
print("=" * 60)
from unet import UNet
from diffusion import GaussianDiffusion
c_in = ch["input_channels"]
c_out = ch["output_channels"]
unet = UNet(
in_channels=c_out + c_in,
out_channels=c_out,
base_ch=64,
ch_mults=(1, 2, 4, 8),
n_res=2,
attn_levels=(3,),
)
model = GaussianDiffusion(unet, timesteps=1000)
device = x_in.device
model = model.to(device)
n_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f" Parameters: {n_params:,}")
print(f" GPU memory after model: {fmt_mem()}")
# Forward
t0 = time.time()
with torch.amp.autocast("cuda", dtype=torch.bfloat16):
loss = model.training_loss(x_out, x_in)
print(f" Forward pass: loss={loss.item():.4f} ({time.time() - t0:.3f}s)")
print(f" GPU memory after forward: {fmt_mem()}")
# Backward
t0 = time.time()
loss.backward()
print(f" Backward pass: ({time.time() - t0:.3f}s)")
print(f" GPU memory after backward: {fmt_mem()}")
# Quick sampling test (just 5 steps for speed)
model.eval()
model.T = 5 # temporarily reduce for testing
model.betas = model.betas[:5]
model.alphas = model.alphas[:5]
model.alpha_bar = model.alpha_bar[:5]
model.sqrt_alpha_bar = model.sqrt_alpha_bar[:5]
model.sqrt_one_minus_alpha_bar = model.sqrt_one_minus_alpha_bar[:5]
model.sqrt_recip_alpha = model.sqrt_recip_alpha[:5]
model.posterior_variance = model.posterior_variance[:5]
t0 = time.time()
with torch.no_grad():
sample = model.sample(x_in[:2], shape=(2, c_out, x_in.shape[2], x_in.shape[3]))
print(f" Sampling (5 steps, B=2): shape={sample.shape} ({time.time() - t0:.3f}s)")
del model
torch.cuda.empty_cache()
print(f" DIFFUSION OK")
def test_jepa(ch, x_in, x_out):
"""Test 3: JEPA forward + backward."""
print("\n" + "=" * 60)
print("TEST 3: JEPA Model")
print("=" * 60)
from jepa import JEPA
c_in = ch["input_channels"]
device = x_in.device
model = JEPA(
in_channels=c_in,
latent_channels=128,
base_ch=32,
pred_hidden=256,
).to(device)
n_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
total_params = sum(p.numel() for p in model.parameters())
print(f" Trainable parameters: {n_params:,}")
print(f" Total parameters (incl EMA target): {total_params:,}")
print(f" GPU memory after model: {fmt_mem()}")
# Forward
t0 = time.time()
with torch.amp.autocast("cuda", dtype=torch.bfloat16):
loss, metrics = model.compute_loss(x_in, x_out)
print(f" Forward: loss={loss.item():.4f}, metrics={metrics} ({time.time() - t0:.3f}s)")
print(f" GPU memory after forward: {fmt_mem()}")
# Backward
t0 = time.time()
loss.backward()
print(f" Backward: ({time.time() - t0:.3f}s)")
print(f" GPU memory after backward: {fmt_mem()}")
# EMA update
model.update_target()
print(f" EMA update: OK")
# Check latent shapes
model.eval()
with torch.no_grad():
z_pred, z_target = model(x_in[:2], x_out[:2])
print(f" Latent shapes: pred={z_pred.shape}, target={z_target.shape}")
del model
torch.cuda.empty_cache()
print(f" JEPA OK")
def test_training_step(ch, loader):
"""Test 4: Full training step with optimizer and grad scaling."""
print("\n" + "=" * 60)
print("TEST 4: Full Training Step")
print("=" * 60)
from data_pipeline import prepare_batch
from unet import UNet
from diffusion import GaussianDiffusion
c_in = ch["input_channels"]
c_out = ch["output_channels"]
device = "cuda" if torch.cuda.is_available() else "cpu"
unet = UNet(in_channels=c_out + c_in, out_channels=c_out, base_ch=64)
model = GaussianDiffusion(unet, timesteps=1000).to(device)
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4)
scaler = torch.amp.GradScaler("cuda")
model.train()
losses = []
for i, batch in enumerate(loader):
if i >= 3:
break
x_in, x_out = prepare_batch(batch, device)
optimizer.zero_grad(set_to_none=True)
with torch.amp.autocast("cuda", dtype=torch.bfloat16):
loss = model.training_loss(x_out, x_in)
scaler.scale(loss).backward()
scaler.unscale_(optimizer)
nn.utils.clip_grad_norm_(model.parameters(), 1.0)
scaler.step(optimizer)
scaler.update()
losses.append(loss.item())
print(f" Step {i}: loss={loss.item():.4f}, mem={fmt_mem()}")
print(f" 3 training steps completed. Losses: {[f'{l:.4f}' for l in losses]}")
del model, optimizer, scaler
torch.cuda.empty_cache()
print(f" TRAINING STEP OK")
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--dataset", default="turbulent_radiative_layer_2D")
parser.add_argument("--streaming", action="store_true", default=True)
parser.add_argument("--no-streaming", dest="streaming", action="store_false")
parser.add_argument("--local_path", default=None)
parser.add_argument("--batch_size", type=int, default=4)
args = parser.parse_args()
print("=" * 60)
print("THE WELL - Pipeline End-to-End Test")
print("=" * 60)
print(f"Dataset: {args.dataset}")
print(f"Streaming: {args.streaming}")
print(f"Batch: {args.batch_size}")
print(f"Device: {'cuda' if torch.cuda.is_available() else 'cpu'}")
if torch.cuda.is_available():
print(f"GPU: {torch.cuda.get_device_name(0)}")
print(f"Memory: {fmt_mem()}")
results = {}
# Test 1: Data
try:
ch, x_in, x_out = test_data_loading(args)
results["data"] = "PASS"
except Exception as e:
print(f" FAIL: {e}")
traceback.print_exc()
results["data"] = f"FAIL: {e}"
sys.exit(1)
# Test 2: Diffusion
try:
test_diffusion(ch, x_in, x_out)
results["diffusion"] = "PASS"
except Exception as e:
print(f" FAIL: {e}")
traceback.print_exc()
results["diffusion"] = f"FAIL: {e}"
# Test 3: JEPA
try:
test_jepa(ch, x_in, x_out)
results["jepa"] = "PASS"
except Exception as e:
print(f" FAIL: {e}")
traceback.print_exc()
results["jepa"] = f"FAIL: {e}"
# Test 4: Training step
try:
loader, _ = __import__("data_pipeline").create_dataloader(
dataset_name=args.dataset,
split="train",
batch_size=args.batch_size,
streaming=args.streaming,
local_path=args.local_path,
)
test_training_step(ch, loader)
results["training_step"] = "PASS"
except Exception as e:
print(f" FAIL: {e}")
traceback.print_exc()
results["training_step"] = f"FAIL: {e}"
# Summary
print("\n" + "=" * 60)
print("SUMMARY")
print("=" * 60)
all_pass = True
for name, status in results.items():
icon = "PASS" if status == "PASS" else "FAIL"
print(f" [{icon}] {name}")
if status != "PASS":
all_pass = False
if all_pass:
print("\nAll tests passed! Pipeline is ready for training.")
else:
print("\nSome tests failed. Check output above.")
sys.exit(1)
if __name__ == "__main__":
main()
|