File size: 16,086 Bytes
7a564e3 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 | #!/usr/bin/env python3
"""
Quick validation experiment: Single visual token vs Multi-token visual prefix.
Trains two TINY models (2-layer decoder, shared ViT) on 1K real images for 10 epochs.
Compares generation quality to determine if the architecture bottleneck is real.
Result interpretation:
- If BOTH collapse → data/training issue, not architecture
- If multi-token works but single doesn't → architecture bottleneck confirmed
- If single-token works → data scale was the issue all along
Usage:
python3 scripts/experiment_visual_prefix.py --data_dir data/downloads/stage1 --device cuda
python3 scripts/experiment_visual_prefix.py --data_dir data/downloads/stage2_fullscale --device cuda
"""
import argparse
import json
import os
import sys
import time
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader, Dataset
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
# ============================================================================
# Tiny Model A: Single visual token (current architecture)
# ============================================================================
class TinyEncoderA(nn.Module):
"""Simplified ViT → single embedding."""
def __init__(self, img_size=224, patch_size=16, hidden_dim=256, embed_dim=512):
super().__init__()
n_patches = (img_size // patch_size) ** 2
self.patch_embed = nn.Conv2d(3, hidden_dim, patch_size, patch_size)
self.pos_embed = nn.Parameter(torch.randn(1, n_patches, hidden_dim) * 0.02)
self.blocks = nn.ModuleList([
nn.TransformerEncoderLayer(hidden_dim, nhead=4, dim_feedforward=hidden_dim*4, batch_first=True)
for _ in range(2)
])
self.norm = nn.LayerNorm(hidden_dim)
self.proj = nn.Linear(hidden_dim, embed_dim)
def forward(self, images):
x = self.patch_embed(images).flatten(2).transpose(1, 2) # [B, N, D]
x = x + self.pos_embed[:, :x.shape[1], :]
for block in self.blocks:
x = block(x)
x = self.norm(x)
x = x.mean(dim=1) # average pool → single vector
x = self.proj(x)
x = F.normalize(x, p=2, dim=-1) # L2-normalize like current arch
return x # [B, embed_dim]
class TinyDecoderA(nn.Module):
"""Single-token decoder (current architecture)."""
def __init__(self, embed_dim=512, hidden_dim=256, vocab_size=1000, max_seq=64):
super().__init__()
self.embed_proj = nn.Linear(embed_dim, hidden_dim)
self.token_embed = nn.Embedding(vocab_size, hidden_dim, padding_idx=0)
self.pos_embed = nn.Parameter(torch.randn(1, max_seq + 1, hidden_dim) * 0.02)
self.blocks = nn.ModuleList([
nn.TransformerDecoderLayer(hidden_dim, nhead=4, dim_feedforward=hidden_dim*4, batch_first=True)
for _ in range(2)
])
self.norm = nn.LayerNorm(hidden_dim)
self.lm_head = nn.Linear(hidden_dim, vocab_size, bias=False)
self.hidden_dim = hidden_dim
self.vocab_size = vocab_size
def forward(self, pred_embedding, target_ids):
embed_token = self.embed_proj(pred_embedding).unsqueeze(1) # [B, 1, D]
token_embeds = self.token_embed(target_ids) # [B, T, D]
x = torch.cat([embed_token, token_embeds], dim=1) # [B, 1+T, D]
T = x.shape[1]
x = x + self.pos_embed[:, :T, :]
# Causal mask
mask = nn.Transformer.generate_square_subsequent_mask(T, device=x.device)
for block in self.blocks:
x = block(x, x, tgt_mask=mask, memory_mask=mask)
x = self.norm(x)
logits = self.lm_head(x)
shift_logits = logits[:, :-1, :].contiguous()
shift_targets = target_ids.contiguous()
loss = F.cross_entropy(shift_logits.view(-1, self.vocab_size), shift_targets.view(-1), ignore_index=0)
return loss
@torch.no_grad()
def generate(self, pred_embedding, max_tokens=32, eos_id=2):
embed_token = self.embed_proj(pred_embedding).unsqueeze(1)
x = embed_token
generated = []
for _ in range(max_tokens):
T = x.shape[1]
pos_x = x + self.pos_embed[:, :T, :]
mask = nn.Transformer.generate_square_subsequent_mask(T, device=x.device)
h = pos_x
for block in self.blocks:
h = block(h, h, tgt_mask=mask, memory_mask=mask)
h = self.norm(h)
logits = self.lm_head(h[:, -1, :])
next_token = logits.argmax(dim=-1)
generated.append(next_token.item())
if next_token.item() == eos_id:
break
next_embed = self.token_embed(next_token.unsqueeze(0))
x = torch.cat([x, next_embed], dim=1)
return generated
# ============================================================================
# Tiny Model B: Multi-token visual prefix (proposed fix)
# ============================================================================
class TinyEncoderB(nn.Module):
"""ViT → multi-token output (N visual tokens)."""
def __init__(self, img_size=224, patch_size=16, hidden_dim=256, n_visual_tokens=16):
super().__init__()
n_patches = (img_size // patch_size) ** 2
self.patch_embed = nn.Conv2d(3, hidden_dim, patch_size, patch_size)
self.pos_embed = nn.Parameter(torch.randn(1, n_patches, hidden_dim) * 0.02)
self.blocks = nn.ModuleList([
nn.TransformerEncoderLayer(hidden_dim, nhead=4, dim_feedforward=hidden_dim*4, batch_first=True)
for _ in range(2)
])
self.norm = nn.LayerNorm(hidden_dim)
# Learned queries for downsampling
self.queries = nn.Parameter(torch.randn(1, n_visual_tokens, hidden_dim) * 0.02)
self.cross_attn = nn.MultiheadAttention(hidden_dim, num_heads=4, batch_first=True)
self.query_norm = nn.LayerNorm(hidden_dim)
def forward(self, images):
x = self.patch_embed(images).flatten(2).transpose(1, 2)
x = x + self.pos_embed[:, :x.shape[1], :]
for block in self.blocks:
x = block(x)
x = self.norm(x) # [B, N_patches, D]
# Downsample to N visual tokens via cross-attention
B = x.shape[0]
queries = self.queries.expand(B, -1, -1)
out, _ = self.cross_attn(queries, x, x)
return self.query_norm(out) # [B, N_visual, D]
class TinyDecoderB(nn.Module):
"""Multi-token visual prefix decoder."""
def __init__(self, hidden_dim=256, vocab_size=1000, max_seq=64):
super().__init__()
self.token_embed = nn.Embedding(vocab_size, hidden_dim, padding_idx=0)
self.pos_embed = nn.Parameter(torch.randn(1, max_seq + 32, hidden_dim) * 0.02)
self.blocks = nn.ModuleList([
nn.TransformerDecoderLayer(hidden_dim, nhead=4, dim_feedforward=hidden_dim*4, batch_first=True)
for _ in range(2)
])
self.norm = nn.LayerNorm(hidden_dim)
self.lm_head = nn.Linear(hidden_dim, vocab_size, bias=False)
self.hidden_dim = hidden_dim
self.vocab_size = vocab_size
def forward(self, visual_prefix, target_ids):
# visual_prefix: [B, N, D] — N visual tokens
N = visual_prefix.shape[1]
token_embeds = self.token_embed(target_ids) # [B, T, D]
x = torch.cat([visual_prefix, token_embeds], dim=1) # [B, N+T, D]
T_total = x.shape[1]
x = x + self.pos_embed[:, :T_total, :]
mask = nn.Transformer.generate_square_subsequent_mask(T_total, device=x.device)
for block in self.blocks:
x = block(x, x, tgt_mask=mask, memory_mask=mask)
x = self.norm(x)
logits = self.lm_head(x)
# Loss: only on answer positions (after visual prefix)
answer_logits = logits[:, N-1:-1, :].contiguous() # [B, T, vocab]
shift_targets = target_ids.contiguous()
loss = F.cross_entropy(answer_logits.view(-1, self.vocab_size), shift_targets.view(-1), ignore_index=0)
return loss
@torch.no_grad()
def generate(self, visual_prefix, max_tokens=32, eos_id=2):
x = visual_prefix # [1, N, D]
generated = []
for _ in range(max_tokens):
T = x.shape[1]
pos_x = x + self.pos_embed[:, :T, :]
mask = nn.Transformer.generate_square_subsequent_mask(T, device=x.device)
h = pos_x
for block in self.blocks:
h = block(h, h, tgt_mask=mask, memory_mask=mask)
h = self.norm(h)
logits = self.lm_head(h[:, -1, :])
next_token = logits.argmax(dim=-1)
generated.append(next_token.item())
if next_token.item() == eos_id:
break
next_embed = self.token_embed(next_token.unsqueeze(0))
x = torch.cat([x, next_embed], dim=1)
return generated
# ============================================================================
# Simple Dataset
# ============================================================================
class SimpleVQADataset(Dataset):
"""Load N real image-text pairs from JSONL."""
def __init__(self, jsonl_dir, max_samples=1000, img_size=224, vocab_size=1000):
from PIL import Image
from torchvision import transforms
self.transform = transforms.Compose([
transforms.Resize((img_size, img_size)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
self.samples = []
self.vocab_size = vocab_size
# Build simple word→id vocab from data
all_words = set()
raw_samples = []
for fname in sorted(os.listdir(jsonl_dir)):
if fname.endswith('.jsonl'):
with open(os.path.join(jsonl_dir, fname)) as f:
for line in f:
try:
item = json.loads(line.strip())
img_path = item.get("image_path")
answer = item.get("answer", "")
if img_path and os.path.exists(img_path) and answer:
raw_samples.append((img_path, answer))
all_words.update(answer.lower().split()[:20])
if len(raw_samples) >= max_samples:
break
except json.JSONDecodeError:
continue
if len(raw_samples) >= max_samples:
break
# Simple vocab: PAD=0, BOS=1, EOS=2, then words
self.word2id = {"<pad>": 0, "<bos>": 1, "<eos>": 2}
for w in sorted(all_words):
if len(self.word2id) < vocab_size:
self.word2id[w] = len(self.word2id)
self.id2word = {v: k for k, v in self.word2id.items()}
# Process samples
for img_path, answer in raw_samples:
try:
img = self.transform(Image.open(img_path).convert("RGB"))
ids = [1] # BOS
for w in answer.lower().split()[:10]:
ids.append(self.word2id.get(w, 0))
ids.append(2) # EOS
# Pad to 16
ids = (ids + [0] * 16)[:16]
self.samples.append((img, torch.tensor(ids, dtype=torch.long)))
except Exception:
continue
print(f" Loaded {len(self.samples)} samples, vocab {len(self.word2id)} words")
def __len__(self):
return len(self.samples)
def __getitem__(self, idx):
return self.samples[idx]
def decode(self, ids):
return " ".join(self.id2word.get(i, "?") for i in ids if i > 2)
# ============================================================================
# Experiment
# ============================================================================
def run_experiment(data_dir, device, max_samples=1000, epochs=10):
print("=" * 70)
print("EXPERIMENT: Single Visual Token vs Multi-Token Visual Prefix")
print("=" * 70)
dataset = SimpleVQADataset(data_dir, max_samples=max_samples, img_size=224)
if len(dataset) < 10:
print("FATAL: Not enough data. Need at least 10 real image-text pairs.")
return
loader = DataLoader(dataset, batch_size=16, shuffle=True, drop_last=True)
vocab_size = len(dataset.word2id)
# ---- Model A: Single visual token ----
print("\n--- Model A: Single Visual Token (Current Architecture) ---")
enc_a = TinyEncoderA(img_size=224, patch_size=16, hidden_dim=256, embed_dim=512).to(device)
dec_a = TinyDecoderA(embed_dim=512, hidden_dim=256, vocab_size=vocab_size).to(device)
opt_a = torch.optim.AdamW(list(enc_a.parameters()) + list(dec_a.parameters()), lr=1e-3)
for epoch in range(epochs):
total_loss = 0
n = 0
for images, targets in loader:
images, targets = images.to(device), targets.to(device)
emb = enc_a(images) # [B, 512] single vector
loss = dec_a(emb, targets)
opt_a.zero_grad()
loss.backward()
opt_a.step()
total_loss += loss.item()
n += 1
print(f" Epoch {epoch+1}/{epochs}: loss={total_loss/n:.4f}")
# Test generation
enc_a.eval()
dec_a.eval()
print("\n Generation test (Model A):")
for i in range(5):
img, target = dataset[i]
emb = enc_a(img.unsqueeze(0).to(device))
gen_ids = dec_a.generate(emb, max_tokens=16)
gen_text = dataset.decode(gen_ids)
target_text = dataset.decode(target.tolist())
print(f" [{i}] Target: {target_text}")
print(f" Generated: {gen_text}")
print(f" Raw IDs: {gen_ids[:10]}")
# ---- Model B: Multi-token visual prefix ----
print("\n--- Model B: Multi-Token Visual Prefix (Proposed Fix) ---")
enc_b = TinyEncoderB(img_size=224, patch_size=16, hidden_dim=256, n_visual_tokens=16).to(device)
dec_b = TinyDecoderB(hidden_dim=256, vocab_size=vocab_size).to(device)
opt_b = torch.optim.AdamW(list(enc_b.parameters()) + list(dec_b.parameters()), lr=1e-3)
for epoch in range(epochs):
total_loss = 0
n = 0
for images, targets in loader:
images, targets = images.to(device), targets.to(device)
vis = enc_b(images) # [B, 16, 256] multi-token
loss = dec_b(vis, targets)
opt_b.zero_grad()
loss.backward()
opt_b.step()
total_loss += loss.item()
n += 1
print(f" Epoch {epoch+1}/{epochs}: loss={total_loss/n:.4f}")
# Test generation
enc_b.eval()
dec_b.eval()
print("\n Generation test (Model B):")
for i in range(5):
img, target = dataset[i]
vis = enc_b(img.unsqueeze(0).to(device))
gen_ids = dec_b.generate(vis, max_tokens=16)
gen_text = dataset.decode(gen_ids)
target_text = dataset.decode(target.tolist())
print(f" [{i}] Target: {target_text}")
print(f" Generated: {gen_text}")
print(f" Raw IDs: {gen_ids[:10]}")
print("\n" + "=" * 70)
print("EXPERIMENT COMPLETE")
print("=" * 70)
print()
print("Interpretation:")
print(" - If BOTH collapse → data or training issue")
print(" - If B works but A doesn't → architecture bottleneck confirmed")
print(" - If A works too → data scale was the issue, not architecture")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--data_dir", default="data/downloads/stage1")
parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
parser.add_argument("--max_samples", type=int, default=1000)
parser.add_argument("--epochs", type=int, default=10)
args = parser.parse_args()
run_experiment(args.data_dir, args.device, args.max_samples, args.epochs)
|