File size: 14,690 Bytes
db32e07 | 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 | #!/usr/bin/env python
"""
Comprehensive evaluation script for MolT5 RAG model.
Computes multiple metrics including Tanimoto similarity, validity, exact match, etc.
"""
from __future__ import annotations
import argparse
import json
import sys
from pathlib import Path
from typing import Dict, List, Tuple
import numpy as np
import torch
from tqdm import tqdm
ROOT = Path(__file__).resolve().parents[1]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from spec_rag.io import load_embeddings
from spec_rag.molt5_with_embeddings import MolT5WithEmbeddings
def load_jsonl(path: Path) -> List[dict]:
"""Load JSONL file."""
data = []
with open(path, "r") as f:
for line in f:
line = line.strip()
if line:
data.append(json.loads(line))
return data
def compute_molecular_metrics(
predictions: List[str],
ground_truths: List[str]
) -> Dict[str, float]:
"""
Compute comprehensive molecular generation metrics.
Returns:
Dictionary with metrics:
- exact_match: Raw string exact match
- canonical_exact_match: After canonicalization
- validity_rate: Percentage of valid SMILES
- tanimoto_similarity: Mean Tanimoto similarity
- tanimoto_similarity_valid: Mean Tanimoto (only valid predictions)
- top1_tanimoto: Mean max Tanimoto in top-1
- top10_tanimoto: Mean max Tanimoto in top-10 (if multiple predictions)
"""
try:
from rdkit import Chem
from rdkit.Chem import AllChem, DataStructs
from rdkit import rdBase
rdBase.DisableLog("rdApp.*")
except ImportError:
raise ImportError("RDKit is required for molecular metrics. Install with: conda install -c conda-forge rdkit")
if len(predictions) != len(ground_truths):
raise ValueError(f"Mismatch: {len(predictions)} predictions vs {len(ground_truths)} ground truths")
exact_matches = 0
canonical_matches = 0
valid_count = 0
tanimoto_sum = 0.0
tanimoto_valid_sum = 0.0
valid_predictions = []
valid_ground_truths = []
# For detailed analysis
tanimoto_scores = []
validity_flags = []
for pred, gt in zip(predictions, ground_truths):
pred = pred.strip()
gt = gt.strip()
# Exact match (raw string)
if pred == gt:
exact_matches += 1
# Try to parse as molecules
try:
mol_pred = Chem.MolFromSmiles(pred)
mol_gt = Chem.MolFromSmiles(gt)
if mol_pred is not None:
valid_count += 1
validity_flags.append(1)
pred_canon = Chem.MolToSmiles(mol_pred, canonical=True)
valid_predictions.append(mol_pred)
else:
validity_flags.append(0)
pred_canon = None
if mol_gt is not None:
gt_canon = Chem.MolToSmiles(mol_gt, canonical=True)
valid_ground_truths.append(mol_gt)
else:
gt_canon = None
# Canonical exact match
if pred_canon and gt_canon and pred_canon == gt_canon:
canonical_matches += 1
# Tanimoto similarity
if mol_pred is not None and mol_gt is not None:
fp_pred = AllChem.GetMorganFingerprintAsBitVect(mol_pred, radius=2, nBits=2048)
fp_gt = AllChem.GetMorganFingerprintAsBitVect(mol_gt, radius=2, nBits=2048)
tanimoto = DataStructs.TanimotoSimilarity(fp_pred, fp_gt)
tanimoto_sum += tanimoto
tanimoto_scores.append(tanimoto)
if mol_pred is not None:
tanimoto_valid_sum += tanimoto
else:
tanimoto_scores.append(0.0)
except Exception as e:
# Invalid SMILES or parsing error
validity_flags.append(0)
tanimoto_scores.append(0.0)
continue
n = len(predictions)
metrics = {
"exact_match": exact_matches / n if n > 0 else 0.0,
"canonical_exact_match": canonical_matches / n if n > 0 else 0.0,
"validity_rate": valid_count / n if n > 0 else 0.0,
"tanimoto_similarity": tanimoto_sum / n if n > 0 else 0.0,
"tanimoto_similarity_valid": tanimoto_valid_sum / valid_count if valid_count > 0 else 0.0,
"n_total": n,
"n_valid": valid_count,
}
# Additional statistics
if tanimoto_scores:
metrics["tanimoto_mean"] = np.mean(tanimoto_scores)
metrics["tanimoto_std"] = np.std(tanimoto_scores)
metrics["tanimoto_median"] = np.median(tanimoto_scores)
metrics["tanimoto_min"] = np.min(tanimoto_scores)
metrics["tanimoto_max"] = np.max(tanimoto_scores)
# Distribution analysis
metrics["tanimoto_ge_0.9"] = sum(1 for t in tanimoto_scores if t >= 0.9) / n
metrics["tanimoto_ge_0.8"] = sum(1 for t in tanimoto_scores if t >= 0.8) / n
metrics["tanimoto_ge_0.7"] = sum(1 for t in tanimoto_scores if t >= 0.7) / n
metrics["tanimoto_ge_0.5"] = sum(1 for t in tanimoto_scores if t >= 0.5) / n
metrics["tanimoto_ge_0.3"] = sum(1 for t in tanimoto_scores if t >= 0.3) / n
return metrics
def generate_predictions(
model,
tokenizer,
inputs: List[str],
spectrum_embeddings: torch.Tensor | None = None,
max_length: int = 512,
num_beams: int = 1,
batch_size: int = 8,
) -> List[str]:
"""Generate predictions in batches with spectrum embeddings."""
predictions = []
for i in tqdm(range(0, len(inputs), batch_size), desc="Generating predictions"):
batch = inputs[i:i + batch_size]
batch_embeddings = None
if spectrum_embeddings is not None:
batch_embeddings = spectrum_embeddings[i:i + batch_size].to(model.device)
# Tokenize inputs
encoded = tokenizer(
batch,
return_tensors="pt",
padding=True,
truncation=True,
max_length=max_length,
).to(model.device)
# Generate
with torch.no_grad():
if batch_embeddings is not None:
# Use model's generate with embeddings
outputs = model.generate(
**encoded,
spectrum_embeddings=batch_embeddings,
max_length=max_length,
num_beams=num_beams,
do_sample=(num_beams == 1),
pad_token_id=tokenizer.pad_token_id,
)
else:
outputs = model.generate(
**encoded,
max_length=max_length,
num_beams=num_beams,
do_sample=(num_beams == 1),
pad_token_id=tokenizer.pad_token_id,
)
# Decode
batch_preds = tokenizer.batch_decode(outputs, skip_special_tokens=True)
predictions.extend(batch_preds)
return predictions
def main():
parser = argparse.ArgumentParser(description="Evaluate MolT5 RAG model")
parser.add_argument("--model-path", required=True, help="Path to trained model directory")
parser.add_argument("--test-jsonl", required=True, help="Test JSONL file")
parser.add_argument("--output-json", default=None, help="Output JSON file for results")
parser.add_argument("--max-length", type=int, default=512, help="Max generation length")
parser.add_argument("--num-beams", type=int, default=1, help="Beam search size (1=greedy)")
parser.add_argument("--batch-size", type=int, default=8, help="Batch size for generation")
parser.add_argument("--device", default="cuda", help="Device (cuda/cpu)")
parser.add_argument("--save-predictions", default=None, help="Save predictions to JSONL file")
parser.add_argument("--spec-embeddings", required=True, help="Path to spectrum embeddings .npy file (required for generation)")
parser.add_argument("--prompt-len", type=int, default=10, help="Number of soft prompt tokens (must match training)")
args = parser.parse_args()
print("=" * 60)
print("MolT5 RAG Model Evaluation")
print("=" * 60)
# Load spectrum embeddings (required)
print(f"\nLoading spectrum embeddings from {args.spec_embeddings}...")
spec_embeddings = load_embeddings(args.spec_embeddings)
print(f"Loaded {len(spec_embeddings)} spectrum embeddings (shape: {spec_embeddings.shape})")
embedding_dim = spec_embeddings.shape[1]
# Load model and tokenizer
print(f"\nLoading model from {args.model_path}...")
base_model = AutoModelForSeq2SeqLM.from_pretrained(args.model_path)
tokenizer = AutoTokenizer.from_pretrained(args.model_path)
# Wrap model with embedding injection
print(f"Wrapping model with spectrum embedding injection (dim={embedding_dim}, prompt_len={args.prompt_len})")
model = MolT5WithEmbeddings(
base_model=base_model,
embedding_dim=embedding_dim,
prompt_len=args.prompt_len,
freeze_base=True, # Freeze during evaluation
)
model.to(args.device)
model.eval()
print(f"Model loaded on {args.device}")
# Load test data
print(f"\nLoading test data from {args.test_jsonl}...")
test_data = load_jsonl(Path(args.test_jsonl))
print(f"Loaded {len(test_data)} test examples")
# Extract inputs, ground truths, and spectrum IDs
# Remove "Target: ..." from input_text if present (data leakage prevention)
inputs = []
ground_truths = []
spectrum_ids = []
for item in test_data[:1000]:
input_text = item["input_text"]
# Remove target if present
if "\nTarget:" in input_text:
input_text = input_text.split("\nTarget:")[0].rstrip()
inputs.append(input_text)
ground_truths.append(item["target_text"])
# Get spectrum_id
spectrum_id = item.get("spectrum_id", len(spectrum_ids)) # Default to index if missing
spectrum_ids.append(spectrum_id)
# Map spectrum_ids to embeddings
print(f"\nMapping spectrum IDs to embeddings...")
try:
indices = [int(sid) if isinstance(sid, (int, str)) and str(sid).isdigit() else i
for i, sid in enumerate(spectrum_ids)]
indices = [min(i, len(spec_embeddings) - 1) for i in indices]
eval_embeddings = torch.tensor([spec_embeddings[i] for i in indices], dtype=torch.float32)
print(f" Mapped {len(eval_embeddings)} embeddings (shape: {eval_embeddings.shape})")
except Exception as e:
raise ValueError(f"Failed to map spectrum IDs to embeddings: {e}")
print(f"\nGenerating predictions with spectrum embeddings...")
print(f" Batch size: {args.batch_size}")
print(f" Max length: {args.max_length}")
print(f" Num beams: {args.num_beams}")
print(f" Prompt length: {args.prompt_len}")
print(f" Embedding dimension: {embedding_dim}")
predictions = generate_predictions(
model,
tokenizer,
inputs,
spectrum_embeddings=eval_embeddings,
max_length=args.max_length,
num_beams=args.num_beams,
batch_size=args.batch_size,
)
# Save predictions if requested
if args.save_predictions:
print(f"\nSaving predictions to {args.save_predictions}...")
with open(args.save_predictions, "w") as f:
for inp, pred, gt in zip(inputs, predictions, ground_truths):
f.write(json.dumps({
"input": inp,
"prediction": pred,
"ground_truth": gt,
}, ensure_ascii=False) + "\n")
# Compute metrics
print(f"\nComputing metrics...")
metrics = compute_molecular_metrics(predictions, ground_truths)
# Print results
print("\n" + "=" * 60)
print("EVALUATION RESULTS")
print("=" * 60)
print(f"\nDataset: {len(test_data)} examples")
print(f"\nBasic Metrics:")
print(f" Exact Match (raw): {metrics['exact_match']:.4f} ({metrics['exact_match']*100:.2f}%)")
print(f" Canonical Exact Match: {metrics['canonical_exact_match']:.4f} ({metrics['canonical_exact_match']*100:.2f}%)")
print(f" Validity Rate: {metrics['validity_rate']:.4f} ({metrics['validity_rate']*100:.2f}%)")
print(f" Valid Predictions: {metrics['n_valid']}/{metrics['n_total']}")
print(f"\nTanimoto Similarity:")
print(f" Mean (all): {metrics['tanimoto_similarity']:.4f}")
print(f" Mean (valid only): {metrics['tanimoto_similarity_valid']:.4f}")
if 'tanimoto_mean' in metrics:
print(f" Median: {metrics['tanimoto_median']:.4f}")
print(f" Std Dev: {metrics['tanimoto_std']:.4f}")
print(f" Min: {metrics['tanimoto_min']:.4f}")
print(f" Max: {metrics['tanimoto_max']:.4f}")
print(f"\nTanimoto Similarity Distribution:")
if 'tanimoto_ge_0.9' in metrics:
print(f" ≥ 0.9 (excellent): {metrics['tanimoto_ge_0.9']:.4f} ({metrics['tanimoto_ge_0.9']*100:.2f}%)")
print(f" ≥ 0.8 (very good): {metrics['tanimoto_ge_0.8']:.4f} ({metrics['tanimoto_ge_0.8']*100:.2f}%)")
print(f" ≥ 0.7 (good): {metrics['tanimoto_ge_0.7']:.4f} ({metrics['tanimoto_ge_0.7']*100:.2f}%)")
print(f" ≥ 0.5 (moderate): {metrics['tanimoto_ge_0.5']:.4f} ({metrics['tanimoto_ge_0.5']*100:.2f}%)")
print(f" ≥ 0.3 (low): {metrics['tanimoto_ge_0.3']:.4f} ({metrics['tanimoto_ge_0.3']*100:.2f}%)")
# Save results
if args.output_json:
print(f"\nSaving results to {args.output_json}...")
with open(args.output_json, "w") as f:
json.dump({
"model_path": args.model_path,
"test_file": args.test_jsonl,
"n_examples": len(test_data),
"metrics": metrics,
"config": {
"max_length": args.max_length,
"num_beams": args.num_beams,
"batch_size": args.batch_size,
}
}, f, indent=2)
print("Results saved!")
print("\n" + "=" * 60)
print("Evaluation complete!")
print("=" * 60)
if __name__ == "__main__":
main()
|