File size: 9,780 Bytes
de3c862 | 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 | """
Train H4 geometric ranker with contrastive learning on SQuAD.
For each batch:
- Each question is paired with its correct passage (positive)
- All other passages in the batch are negatives (in-batch negatives)
- Loss: InfoNCE — correct passage should score highest
Metric: Recall@1 (does the top-ranked passage contain the answer?)
This is a much simpler task than extractive QA:
- Ranking maps two texts to a scalar (not text to text)
- 370K ternary params can learn this
- 5,928 SQuAD pairs provide enough signal
"""
import os
import math
import time
import random
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import sys
sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
from rag.ranking_model import H4Ranker
from rag.prepare_qa import download_squad_dev
from rag.tokenizer import BPETokenizer
# ---------------------------------------------------------------------------
# Hyperparameters
# ---------------------------------------------------------------------------
TIME_BUDGET = int(os.environ.get('RANKER_TIME', 600)) # default 10 min, override with env
D_MODEL = int(os.environ.get('RANKER_DMODEL', 128))
N_HEADS = 8
N_LAYERS = int(os.environ.get('RANKER_LAYERS', 2))
D_VALUE = D_MODEL // N_HEADS
D_FFN = D_MODEL * 4
USE_BITLINEAR = True
LR = float(os.environ.get('RANKER_LR', 3e-3))
WEIGHT_DECAY = 0.01
GRAD_CLIP = 1.0
BATCH_SIZE = int(os.environ.get('RANKER_BATCH', 32))
MAX_Q_LEN = 64
MAX_P_LEN = 192
TEMPERATURE = 0.15
EVAL_INTERVAL = 200
def pad_tokens(ids, max_len):
"""Pad or truncate token list to fixed length."""
ids = ids[:max_len]
return ids + [0] * (max_len - len(ids))
def contrastive_loss(q_h4, p_h4, temperature):
"""InfoNCE loss with in-batch negatives."""
B = q_h4.shape[0]
# Similarity matrix: (B, B)
sim = torch.mm(q_h4, p_h4.t()) / temperature
# Labels: positive is on diagonal
labels = torch.arange(B, device=sim.device)
loss = F.cross_entropy(sim, labels)
# Metrics
with torch.no_grad():
preds = sim.argmax(dim=1)
recall_at_1 = (preds == labels).float().mean().item()
top5 = sim.topk(min(5, B), dim=1).indices
recall_at_5 = sum(
labels[i].item() in top5[i].tolist() for i in range(B)
) / B
ranks = (sim.argsort(dim=1, descending=True) == labels.unsqueeze(1)).nonzero()[:, 1].float() + 1
mrr = (1.0 / ranks).mean().item()
return loss, recall_at_1, recall_at_5, mrr
def main():
t_start = time.time()
torch.manual_seed(42)
random.seed(42)
np.random.seed(42)
# Load SQuAD
squad = download_squad_dev()
if len(squad) < 100:
print("SQuAD not available. Run: python rag/prepare_qa.py")
return
print(f"SQuAD: {len(squad)} QA pairs")
# Build BPE tokenizer
tokenizer = BPETokenizer(max_vocab=4096)
all_texts = [qa['context'] + ' ' + qa['question'] for qa in squad[:2000]]
tokenizer.build_vocab(all_texts)
vocab_size = tokenizer.vocab_size
# Split train/val
indices = list(range(len(squad)))
random.shuffle(indices)
n_val = 200
val_indices = set(indices[:n_val])
train_data = [squad[i] for i in indices if i not in val_indices]
val_data = [squad[i] for i in indices if i in val_indices]
print(f"Train: {len(train_data)}, Val: {len(val_data)}")
# Create model
model = H4Ranker(
vocab_size=vocab_size,
d_model=D_MODEL,
n_heads=N_HEADS,
n_layers=N_LAYERS,
d_value=D_VALUE,
d_ffn=D_FFN,
use_bitlinear=USE_BITLINEAR,
max_seq_len=max(MAX_Q_LEN, MAX_P_LEN),
)
n_params = model.count_params()
print(f"Model: {n_params:,} params ({'ternary' if USE_BITLINEAR else 'float'})")
optimizer = torch.optim.AdamW(model.parameters(), lr=LR, weight_decay=WEIGHT_DECAY, betas=(0.9, 0.95))
def lr_schedule(step):
if step < 50:
return step / 50
progress = (step - 50) / max(1, 5000 - 50)
return 0.1 + 0.9 * 0.5 * (1 + math.cos(math.pi * min(progress, 1.0)))
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_schedule)
# Training loop
model.train()
step = 0
total_training_time = 0.0
best_r1 = 0.0
print(f"\nTraining for {TIME_BUDGET}s, metric=Recall@1")
print(f"{'step':>6} {'loss':>8} {'R@1':>8} {'R@5':>8} {'MRR':>8} {'lr':>10}")
print("-" * 56)
while True:
t0 = time.time()
# Sample batch
batch = random.sample(train_data, min(BATCH_SIZE, len(train_data)))
q_ids = torch.tensor(
[pad_tokens(tokenizer.encode(qa['question']), MAX_Q_LEN) for qa in batch],
dtype=torch.long,
)
p_ids = torch.tensor(
[pad_tokens(tokenizer.encode(qa['context']), MAX_P_LEN) for qa in batch],
dtype=torch.long,
)
# Encode
q_h4 = model.encode(q_ids)
p_h4 = model.encode(p_ids)
# Loss
loss, r1, r5, mrr = contrastive_loss(q_h4, p_h4, TEMPERATURE)
optimizer.zero_grad()
loss.backward()
if GRAD_CLIP > 0:
torch.nn.utils.clip_grad_norm_(model.parameters(), GRAD_CLIP)
optimizer.step()
scheduler.step()
dt = time.time() - t0
if step > 2:
total_training_time += dt
if step % EVAL_INTERVAL == 0:
# Quick val eval
model.eval()
val_r1s = []
with torch.no_grad():
for vi in range(0, min(len(val_data), 100), BATCH_SIZE):
vbatch = val_data[vi:vi + BATCH_SIZE]
if len(vbatch) < 2:
continue
vq = torch.tensor(
[pad_tokens(tokenizer.encode(qa['question']), MAX_Q_LEN) for qa in vbatch],
dtype=torch.long,
)
vp = torch.tensor(
[pad_tokens(tokenizer.encode(qa['context']), MAX_P_LEN) for qa in vbatch],
dtype=torch.long,
)
vq_h4 = model.encode(vq)
vp_h4 = model.encode(vp)
_, vr1, _, _ = contrastive_loss(vq_h4, vp_h4, TEMPERATURE)
val_r1s.append(vr1)
val_r1 = sum(val_r1s) / len(val_r1s) if val_r1s else 0
if val_r1 > best_r1:
best_r1 = val_r1
current_lr = scheduler.get_last_lr()[0]
print(f"{step:6d} {loss.item():8.4f} {val_r1:8.3f} {r5:8.3f} {mrr:8.3f} {current_lr:10.6f}")
model.train()
step += 1
if step > 2 and total_training_time >= TIME_BUDGET:
break
# Final evaluation
model.eval()
print("\n" + "=" * 60)
print("FINAL RANKING EVALUATION:")
all_r1 = []
all_r5 = []
all_mrr = []
with torch.no_grad():
for vi in range(0, min(len(val_data), 200), BATCH_SIZE):
vbatch = val_data[vi:vi + BATCH_SIZE]
if len(vbatch) < 2:
continue
vq = torch.tensor(
[pad_tokens(tokenizer.encode(qa['question']), MAX_Q_LEN) for qa in vbatch],
dtype=torch.long,
)
vp = torch.tensor(
[pad_tokens(tokenizer.encode(qa['context']), MAX_P_LEN) for qa in vbatch],
dtype=torch.long,
)
vq_h4 = model.encode(vq)
vp_h4 = model.encode(vp)
_, vr1, vr5, vmrr = contrastive_loss(vq_h4, vp_h4, TEMPERATURE)
all_r1.append(vr1)
all_r5.append(vr5)
all_mrr.append(vmrr)
final_r1 = sum(all_r1) / len(all_r1) if all_r1 else 0
final_r5 = sum(all_r5) / len(all_r5) if all_r5 else 0
final_mrr = sum(all_mrr) / len(all_mrr) if all_mrr else 0
# Show some examples
print(f"\nSample rankings (batch of {min(BATCH_SIZE, 8)}):")
sample_batch = val_data[:min(BATCH_SIZE, 8)]
sq = torch.tensor([pad_tokens(tokenizer.encode(qa['question']), MAX_Q_LEN) for qa in sample_batch], dtype=torch.long)
sp = torch.tensor([pad_tokens(tokenizer.encode(qa['context']), MAX_P_LEN) for qa in sample_batch], dtype=torch.long)
with torch.no_grad():
sq_h4 = model.encode(sq)
sp_h4 = model.encode(sp)
sim = torch.mm(sq_h4, sp_h4.t())
for i in range(min(3, len(sample_batch))):
scores = sim[i].tolist()
ranked = sorted(range(len(scores)), key=lambda j: -scores[j])
correct = i
rank_of_correct = ranked.index(correct) + 1
print(f" Q: {sample_batch[i]['question'][:60]}")
print(f" Correct passage rank: {rank_of_correct}/{len(scores)}")
print(f" Top passage: {sample_batch[ranked[0]]['context'][:60]}...")
print()
print("=" * 60)
print("\n---")
print(f"val_recall_at_1: {final_r1:.4f}")
print(f"val_recall_at_5: {final_r5:.4f}")
print(f"val_mrr: {final_mrr:.4f}")
print(f"best_recall_at_1: {best_r1:.4f}")
print(f"training_seconds: {total_training_time:.1f}")
print(f"total_seconds: {time.time() - t_start:.1f}")
print(f"num_steps: {step}")
print(f"num_params: {n_params}")
print(f"ternary: {'yes' if USE_BITLINEAR else 'no'}")
print(f"batch_size: {BATCH_SIZE}")
print(f"temperature: {TEMPERATURE}")
if __name__ == '__main__':
main()
|