File size: 8,768 Bytes
d3fa071 | 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 | """
Phase 3: Continue Pretraining
Demonstrates how to continue pretraining GeneMamba on your own data using masked LM objective.
Usage:
python examples/3_continue_pretraining.py
"""
import torch
import numpy as np
from torch.utils.data import Dataset
from transformers import (
AutoModelForMaskedLM,
AutoTokenizer,
Trainer,
TrainingArguments,
DataCollatorForLanguageModeling,
)
class PretrainingDataset(Dataset):
"""
Dataset for pretraining/continued pretraining.
Loads sequences and their lengths.
"""
def __init__(self, input_ids_list, max_length=2048):
self.input_ids_list = input_ids_list
self.max_length = max_length
def __len__(self):
return len(self.input_ids_list)
def __getitem__(self, idx):
input_ids = self.input_ids_list[idx]
# Pad or truncate to max_length
if len(input_ids) >= self.max_length:
input_ids = input_ids[:self.max_length]
else:
input_ids = np.pad(
input_ids,
(0, self.max_length - len(input_ids)),
constant_values=1 # Pad token ID
)
return {
"input_ids": torch.tensor(input_ids, dtype=torch.long),
}
def create_mock_pretraining_data(n_sequences=5000, seq_len=2048):
"""Create mock single-cell sequences for pretraining."""
print("Creating mock pretraining dataset...")
# Create ranked gene sequences
# In practice, these would come from your scRNA-seq data
sequences = []
for _ in range(n_sequences):
# Random ranked sequence
seq = np.random.randint(2, 25426, seq_len)
sequences.append(seq)
print(f"✓ Created {n_sequences} sequences of length {seq_len}")
return sequences
def main():
print("=" * 80)
print("GeneMamba Phase 3: Continue Pretraining")
print("=" * 80)
# ============================================================
# Step 1: Load pretrained model for masked LM
# ============================================================
print("\n[Step 1] Loading model for masked LM...")
try:
model = AutoModelForMaskedLM.from_pretrained(
"GeneMamba-24l-512d",
trust_remote_code=True,
local_files_only=True,
)
tokenizer = AutoTokenizer.from_pretrained(
"GeneMamba-24l-512d",
trust_remote_code=True,
local_files_only=True,
)
except Exception as e:
print(f"Note: Could not load from hub ({e})")
print("Using local initialization...")
# Initialize locally
from configuration_genemamba import GeneMambaConfig
from modeling_genemamba import GeneMambaForMaskedLM
config = GeneMambaConfig(
vocab_size=25426,
hidden_size=512,
num_hidden_layers=24,
)
model = GeneMambaForMaskedLM(config)
tokenizer = None
print(f"✓ Model loaded")
print(f" - Architecture: {model.config.num_hidden_layers} layers, "
f"hidden_size={model.config.hidden_size}")
# ============================================================
# Step 2: Prepare pretraining data
# ============================================================
print("\n[Step 2] Preparing pretraining dataset...")
sequences = create_mock_pretraining_data(n_sequences=5000, seq_len=2048)
# Split train/eval
train_size = int(0.9 * len(sequences))
train_sequences = sequences[:train_size]
eval_sequences = sequences[train_size:]
train_dataset = PretrainingDataset(train_sequences)
eval_dataset = PretrainingDataset(eval_sequences)
print(f"✓ Datasets created:")
print(f" - Training: {len(train_dataset)} samples")
print(f" - Evaluation: {len(eval_dataset)} samples")
# ============================================================
# Step 3: Set up data collator for MLM
# ============================================================
print("\n[Step 3] Setting up data collator...")
if tokenizer is not None:
data_collator = DataCollatorForLanguageModeling(
tokenizer=tokenizer,
mlm=True,
mlm_probability=0.15, # Mask 15% of tokens
)
else:
# Custom collator if no tokenizer available
class CustomDataCollator:
def __call__(self, batch):
input_ids = torch.stack([item["input_ids"] for item in batch])
# Create masked labels (for MLM loss)
labels = input_ids.clone()
mask = torch.rand(input_ids.shape) < 0.15
# Set input to [MASK] token (id=0)
input_ids[mask] = 0
# Set labels to -100 where not masked (loss ignores these)
labels[~mask] = -100
return {"input_ids": input_ids, "labels": labels}
data_collator = CustomDataCollator()
print(f"✓ Data collator ready (MLM probability: 0.15)")
# ============================================================
# Step 4: Set up training arguments
# ============================================================
print("\n[Step 4] Setting up training...")
output_dir = "./pretrain_results"
training_args = TrainingArguments(
output_dir=output_dir,
num_train_epochs=2,
per_device_train_batch_size=16,
per_device_eval_batch_size=16,
learning_rate=2e-5,
weight_decay=0.01,
warmup_steps=500,
logging_steps=100,
eval_strategy="epoch",
save_strategy="epoch",
load_best_model_at_end=True,
metric_for_best_model="eval_loss",
report_to="none", # Disable W&B
seed=42,
)
print(f"✓ Training config:")
print(f" - Output dir: {output_dir}")
print(f" - Epochs: {training_args.num_train_epochs}")
print(f" - Batch size: {training_args.per_device_train_batch_size}")
print(f" - Learning rate: {training_args.learning_rate}")
print(f" - MLM masking: 15%")
# ============================================================
# Step 5: Train
# ============================================================
print("\n[Step 5] Starting continued pretraining...")
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
data_collator=data_collator,
)
train_result = trainer.train()
print(f"✓ Training complete!")
print(f" - Final training loss: {train_result.training_loss:.4f}")
# ============================================================
# Step 6: Evaluate
# ============================================================
print("\n[Step 6] Evaluating on held-out set...")
eval_results = trainer.evaluate()
print(f"✓ Evaluation Results:")
for metric, value in eval_results.items():
if isinstance(value, (int, float)):
print(f" - {metric}: {value:.4f}")
# ============================================================
# Step 7: Save model
# ============================================================
print("\n[Step 7] Saving continued pretrained model...")
save_dir = "./genemamba_continued_pretrain"
model.save_pretrained(save_dir)
if tokenizer is not None:
tokenizer.save_pretrained(save_dir)
print(f"✓ Model saved to '{save_dir}'")
# ============================================================
# Step 8: Test model inference
# ============================================================
print("\n[Step 8] Testing inference on masked input...")
model.eval()
# Create sample input with masked tokens
sample_input = torch.randint(2, 25426, (1, 2048))
sample_input[0, :10] = 0 # Mask first 10 tokens
with torch.no_grad():
outputs = model(sample_input)
logits = outputs.logits
predictions = torch.argmax(logits, dim=-1)
print(f"✓ Sample predictions generated")
print(f" - Input shape: {sample_input.shape}")
print(f" - Output logits shape: {logits.shape}")
print(f" - Top predicted genes (tokens): {predictions[0, :10].tolist()}")
print("\n" + "=" * 80)
print("Phase 3 Complete! Model ready for downstream tasks or further training.")
print("=" * 80)
return model, trainer
if __name__ == "__main__":
model, trainer = main()
|