File size: 10,812 Bytes
97eb7cb |
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 |
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
from torch.optim.lr_scheduler import LinearLR, CosineAnnealingLR, SequentialLR
import json
import numpy as np
from tqdm import tqdm
# ---------------- Hyperparameters ----------------
ESM_DIM = 1280 # ESM-2 hidden dim (esm2_t33_650M_UR50D)
COMP_RATIO = 16 # compression factor
COMP_DIM = ESM_DIM // COMP_RATIO
MAX_SEQ_LEN = 50 # Actual sequence length from final_sequence_encoder.py
BATCH_SIZE = 32
EPOCHS = 30
BASE_LR = 1e-3 # initial learning rate
LR_MIN = 8e-5 # minimum learning rate for cosine schedule
WARMUP_STEPS = 10_000
DEPTH = 4 # total transformer layers (2 pre-pool, 2 post-pool)
HEADS = 8 # attention heads
DIM_FF = ESM_DIM * 4
POOLING = True # enforce ProtFlow hourglass pooling
# ---------------- Dataset for Pre-computed Embeddings ----------------
class PrecomputedEmbeddingDataset(Dataset):
def __init__(self, embeddings_path):
"""
Load pre-computed embeddings from the final_sequence_encoder.py output.
Args:
embeddings_path: Path to the directory containing individual .pt embedding files
"""
print(f"Loading pre-computed embeddings from {embeddings_path}...")
# Load all individual embedding files
import glob
import os
embedding_files = glob.glob(os.path.join(embeddings_path, "*.pt"))
embedding_files = [f for f in embedding_files if not f.endswith('metadata.json') and not f.endswith('sequence_ids.json')]
print(f"Found {len(embedding_files)} embedding files")
# Load and stack all embeddings
embeddings_list = []
for file_path in embedding_files:
try:
embedding = torch.load(file_path)
if embedding.dim() == 2: # (seq_len, hidden_dim)
embeddings_list.append(embedding)
else:
print(f"Warning: Skipping {file_path} - unexpected shape {embedding.shape}")
except Exception as e:
print(f"Warning: Could not load {file_path}: {e}")
if not embeddings_list:
raise ValueError("No valid embeddings found!")
self.embeddings = torch.stack(embeddings_list)
print(f"Loaded {len(self.embeddings)} embeddings with shape {self.embeddings.shape}")
# Ensure embeddings are the right shape
if len(self.embeddings.shape) != 3:
raise ValueError(f"Expected 3D tensor, got shape {self.embeddings.shape}")
if self.embeddings.shape[1] != MAX_SEQ_LEN:
print(f"Warning: Expected sequence length {MAX_SEQ_LEN}, got {self.embeddings.shape[1]}")
if self.embeddings.shape[2] != ESM_DIM:
print(f"Warning: Expected embedding dim {ESM_DIM}, got {self.embeddings.shape[2]}")
def __len__(self):
return len(self.embeddings)
def __getitem__(self, idx):
return self.embeddings[idx]
# ---------------- Compressor ----------------
class Compressor(nn.Module):
def __init__(self, in_dim=ESM_DIM, out_dim=COMP_DIM):
super().__init__()
self.norm = nn.LayerNorm(in_dim)
layer = lambda: nn.TransformerEncoderLayer(
d_model=in_dim, nhead=HEADS, dim_feedforward=DIM_FF,
batch_first=True)
# two layers before pool, two after
self.pre_tr = nn.TransformerEncoder(layer(), num_layers=DEPTH//2)
self.post_tr = nn.TransformerEncoder(layer(), num_layers=DEPTH//2)
self.proj = nn.Sequential(
nn.LayerNorm(in_dim),
nn.Linear(in_dim, out_dim),
nn.Tanh()
)
self.pooling = POOLING
def forward(self, x, stats=None):
if stats:
m, s, mn, mx = stats['mean'], stats['std'], stats['min'], stats['max']
# Move stats to the same device as x
m = m.to(x.device)
s = s.to(x.device)
mn = mn.to(x.device)
mx = mx.to(x.device)
x = torch.clamp((x - m) / s, -4, 4)
x = torch.clamp((x - mn) / (mx - mn + 1e-8), 0, 1)
x = self.norm(x)
x = self.pre_tr(x) # [B, L, D]
if self.pooling:
B, L, D = x.shape
if L % 2: x = x[:, :-1, :]
x = x.view(B, L//2, 2, D).mean(2) # halve sequence length
x = self.post_tr(x) # [B, L' , D]
return self.proj(x) # [B, L', COMP_DIM]
# ---------------- Decompressor ----------------
class Decompressor(nn.Module):
def __init__(self, in_dim=COMP_DIM, out_dim=ESM_DIM):
super().__init__()
self.proj = nn.Sequential(
nn.LayerNorm(in_dim),
nn.Linear(in_dim, out_dim)
)
layer = lambda: nn.TransformerEncoderLayer(
d_model=out_dim, nhead=HEADS, dim_feedforward=DIM_FF,
batch_first=True)
self.decoder = nn.TransformerEncoder(layer(), num_layers=DEPTH//2)
self.pooling = POOLING
def forward(self, z):
x = self.proj(z) # [B, L', D]
if self.pooling:
x = x.repeat_interleave(2, dim=1) # unpool to full length
return self.decoder(x) # [B, L, out_dim]
# ---------------- Training Loop ----------------
def train_with_precomputed_embeddings(embeddings_path, device='cuda'):
"""
Train compressor using pre-computed embeddings from final_sequence_encoder.py
"""
# Load dataset
ds = PrecomputedEmbeddingDataset(embeddings_path)
# Compute normalization statistics
print("Computing normalization statistics...")
flat = ds.embeddings.view(-1, ESM_DIM)
stats = {
'mean': flat.mean(0),
'std': flat.std(0) + 1e-8,
'min': torch.clamp((flat - flat.mean(0)) / (flat.std(0) + 1e-8), -4,4).min(0)[0],
'max': torch.clamp((flat - flat.mean(0)) / (flat.std(0) + 1e-8), -4,4).max(0)[0]
}
# Save statistics for later use
torch.save(stats, 'normalization_stats.pt')
print("Saved normalization statistics to normalization_stats.pt")
# Create data loader
dl = DataLoader(ds, batch_size=BATCH_SIZE, shuffle=True)
# Initialize models
comp = Compressor().to(device)
decomp = Decompressor().to(device)
# Initialize optimizer
opt = optim.AdamW(list(comp.parameters()) + list(decomp.parameters()), lr=BASE_LR)
# LR scheduling: warmup -> cosine
warmup_sched = LinearLR(opt, start_factor=1e-8, end_factor=1.0, total_iters=WARMUP_STEPS)
cosine_sched = CosineAnnealingLR(opt, T_max=EPOCHS*len(dl), eta_min=LR_MIN)
sched = SequentialLR(opt, [warmup_sched, cosine_sched], milestones=[WARMUP_STEPS])
print(f"Starting training for {EPOCHS} epochs...")
print(f"Device: {device}")
print(f"Batch size: {BATCH_SIZE}")
print(f"Total batches per epoch: {len(dl)}")
# Training loop
for epoch in range(1, EPOCHS+1):
total_loss = 0
comp.train()
decomp.train()
for batch_idx, x in enumerate(tqdm(dl, desc=f"Epoch {epoch}/{EPOCHS}")):
x = x.to(device)
z = comp(x, stats)
xr = decomp(z)
loss = (x - xr).pow(2).mean()
opt.zero_grad()
loss.backward()
opt.step()
sched.step()
total_loss += loss.item()
# Print progress every 100 batches
if batch_idx % 100 == 0:
print(f" Batch {batch_idx}/{len(dl)} - Loss: {loss.item():.6f}")
avg_loss = total_loss / len(dl)
print(f"Epoch {epoch}/{EPOCHS} — Average MSE: {avg_loss:.6f}")
# Save checkpoint every 5 epochs
if epoch % 5 == 0:
torch.save({
'epoch': epoch,
'compressor_state_dict': comp.state_dict(),
'decompressor_state_dict': decomp.state_dict(),
'optimizer_state_dict': opt.state_dict(),
'loss': avg_loss,
}, f'checkpoint_epoch_{epoch}.pth')
# Save final models
torch.save(comp.state_dict(), 'compressor_final.pth')
torch.save(decomp.state_dict(), 'decompressor_final.pth')
print("Training completed! Models saved as compressor_final.pth and decompressor_final.pth")
# ---------------- Utility Functions ----------------
def load_and_test_models(compressor_path, decompressor_path, embeddings_path, device='cuda'):
"""
Load trained models and test reconstruction quality
"""
print("Loading trained models...")
comp = Compressor().to(device)
decomp = Decompressor().to(device)
comp.load_state_dict(torch.load(compressor_path))
decomp.load_state_dict(torch.load(decompressor_path))
comp.eval()
decomp.eval()
# Load test data
ds = PrecomputedEmbeddingDataset(embeddings_path)
test_loader = DataLoader(ds, batch_size=16, shuffle=False)
# Load normalization stats
stats = torch.load('normalization_stats.pt')
print("Testing reconstruction quality...")
total_mse = 0
total_samples = 0
with torch.no_grad():
for batch in tqdm(test_loader, desc="Testing"):
x = batch.to(device)
z = comp(x, stats)
xr = decomp(z)
mse = (x - xr).pow(2).mean()
total_mse += mse.item() * len(x)
total_samples += len(x)
avg_mse = total_mse / total_samples
print(f"Average reconstruction MSE: {avg_mse:.6f}")
return avg_mse
# ---------------- Entrypoint ----------------
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser(description='Train protein compressor with pre-computed embeddings')
parser.add_argument('--embeddings', type=str, default='/data2/edwardsun/flow_project/compressor_dataset/peptide_embeddings.pt',
help='Path to pre-computed embeddings from final_sequence_encoder.py')
parser.add_argument('--device', type=str, default='cuda', help='Device to use (cuda/cpu)')
parser.add_argument('--test', action='store_true', help='Test existing models instead of training')
args = parser.parse_args()
device = torch.device(args.device if torch.cuda.is_available() else 'cpu')
print(f"Using device: {device}")
if args.test:
# Test existing models
load_and_test_models('compressor_final.pth', 'decompressor_final.pth', args.embeddings, device)
else:
# Train new models
train_with_precomputed_embeddings(args.embeddings, device) |