File size: 15,783 Bytes
1d6f391 | 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 | #!/usr/bin/env python3
"""
Embed Benchmark Task Datasets with V5/V6 [CLS] Embeddings
Extracts frozen [CLS] embeddings for GlycanML benchmark task datasets
and produces t-SNE/UMAP visualizations colored by ground-truth labels.
Comparable to GlycanGT Figure 3.
Tasks:
1. Taxonomy (domain, kingdom)
2. Glycosylation type (N/O/free)
3. Immunogenicity (0/1)
Usage:
python embed_benchmark_tasks.py --model v5 [--splits val test] [--embed_all]
python embed_benchmark_tasks.py --model v6 [--splits val test] [--embed_all]
"""
import argparse
import json
import os
import sys
import warnings
from pathlib import Path
import numpy as np
import torch
import torch.nn.functional as F
import pandas as pd
warnings.filterwarnings('ignore')
PROJECT_ROOT = Path(__file__).resolve().parents[2]
sys.path.insert(0, str(PROJECT_ROOT))
sys.path.insert(0, str(PROJECT_ROOT / 'bert_training_v4'))
from model.multimodal_glycan_bert_v3 import MultimodalGlycanBERT, MultimodalGlycanBERTConfig
from downstream_tasks.utils.tokenizer import WURCSTokenizer
DATA_DIR = PROJECT_ROOT / 'bench' / 'GlycanML' / 'data'
VOCAB_PATH = PROJECT_ROOT / 'bert_training_v4' / 'data' / 'vocabulary.json'
CHECKPOINTS = {
'v5': PROJECT_ROOT / 'checkpoints_v5b_excluded' / 'best_v5b_excluded_model.pt',
'v6': PROJECT_ROOT / 'bert_v6_contrastive' / 'checkpoints' / 'phase_3_hard_checkpoint.pt',
}
# Try alternate V6 locations
_v6_alts = [
PROJECT_ROOT / 'bert_v6_contrastive' / 'checkpoints' / 'best_model.pt',
PROJECT_ROOT / 'bert_v6_contrastive' / 'checkpoints' / 'checkpoint_latest.pt',
PROJECT_ROOT / 'bert_v6_contrastive' / 'phase_3_hard_checkpoint.pt',
]
for _alt in _v6_alts:
if _alt.exists():
CHECKPOINTS['v6'] = _alt
break
TASKS = {
'domain': {
'csv': 'glycan_classification_wurcs_subset.csv',
'label_col': 'domain',
'wurcs_col': 'wurcs',
'split_cols': {'train': 'train', 'val': 'validation', 'test': 'test'},
'description': 'Taxonomy domain (Eukarya/Bacteria/Virus/Archaea)',
},
'kingdom': {
'csv': 'glycan_classification_wurcs_subset.csv',
'label_col': 'kingdom',
'wurcs_col': 'wurcs',
'split_cols': {'train': 'train', 'val': 'validation', 'test': 'test'},
'description': 'Taxonomy kingdom (11 classes)',
},
'link': {
'csv': 'glycan_link_wurcs_subset.csv',
'label_col': 'link',
'wurcs_col': 'wurcs',
'split_cols': {'train': 'train', 'val': 'valid', 'test': 'test'},
'description': 'Glycosylation type (N-linked/O-linked/free)',
},
'immunogenicity': {
'csv': 'glycan_immunogenicity_wurcs_subset.csv',
'label_col': 'immunogenicity',
'wurcs_col': 'wurcs',
'split_cols': {'train': 'train', 'val': 'valid', 'test': 'test'},
'description': 'Immunogenicity (0=non-immunogenic, 1=immunogenic)',
},
}
DOMAIN_COLORS = {
'Eukarya': '#2196F3', 'Bacteria': '#FF5722', 'Virus': '#9C27B0', 'Archaea': '#4CAF50'
}
KINGDOM_COLORS = {
'Plantae': '#4CAF50', 'Animalia': '#F44336', 'Fungi': '#FF9800',
'Protista': '#9C27B0', 'Viridiplantae': '#8BC34A', 'Metazoa': '#E91E63',
}
LINK_COLORS = {'N': '#2196F3', 'O': '#FF5722', 'free': '#4CAF50'}
IMMUNO_COLORS = {0.0: '#607D8B', 1.0: '#F44336', '0.0': '#607D8B', '1.0': '#F44336'}
def load_model(checkpoint_path, device='cuda'):
print(f"Loading model from {checkpoint_path}...")
checkpoint = torch.load(checkpoint_path, map_location='cpu', weights_only=False)
if 'model_state_dict' in checkpoint:
state_dict = checkpoint['model_state_dict']
else:
state_dict = checkpoint
backbone_sd = {k: v for k, v in state_dict.items() if not k.startswith('proj_head.')}
n_stripped = len(state_dict) - len(backbone_sd)
if n_stripped > 0:
print(f" Stripped {n_stripped} projection head keys")
vocab_size = backbone_sd['seq_embeddings.token_embeddings.weight'].shape[0]
ms_total_vocab = None
if 'ms_embeddings.token_embeddings.weight' in backbone_sd:
ms_total_vocab = backbone_sd['ms_embeddings.token_embeddings.weight'].shape[0]
config_kwargs = dict(
seq_vocab_size=vocab_size, seq_hidden_size=768, seq_num_layers=12,
seq_num_heads=12, seq_max_length=256, use_cnn_frontend=True, cnn_kernel_size=3,
)
if ms_total_vocab is not None:
config_kwargs['ms_vocab_size'] = ms_total_vocab - vocab_size
config = MultimodalGlycanBERTConfig(**config_kwargs)
model = MultimodalGlycanBERT(config)
model.load_state_dict(backbone_sd, strict=False)
model.to(device)
model.eval()
print(f" Model loaded: {sum(p.numel() for p in model.parameters()):,} params")
return model
def extract_cls_embeddings(model, tokenized_samples, device='cuda', batch_size=64, max_len=256):
all_embeddings = []
n_failed = 0
for i in range(0, len(tokenized_samples), batch_size):
batch = tokenized_samples[i:i + batch_size]
batch_tids, batch_bdeps, batch_ltypes = [], [], []
for sample in batch:
try:
tids = sample['token_ids']
bdeps = sample.get('branch_depths', [0] * len(tids))
ltypes = sample.get('linkage_types', [0] * len(tids))
tids_t = torch.tensor(tids[:max_len], dtype=torch.long)
bdeps_t = torch.tensor(bdeps[:max_len], dtype=torch.long)
ltypes_t = torch.tensor(ltypes[:max_len], dtype=torch.long)
min_len = min(len(tids_t), len(bdeps_t), len(ltypes_t))
tids_t, bdeps_t, ltypes_t = tids_t[:min_len], bdeps_t[:min_len], ltypes_t[:min_len]
if len(tids_t) < max_len:
pad_len = max_len - len(tids_t)
tids_t = F.pad(tids_t, (0, pad_len), value=0)
bdeps_t = F.pad(bdeps_t, (0, pad_len), value=0)
ltypes_t = F.pad(ltypes_t, (0, pad_len), value=0)
batch_tids.append(tids_t)
batch_bdeps.append(bdeps_t)
batch_ltypes.append(ltypes_t)
except Exception:
n_failed += 1
if not batch_tids:
continue
with torch.no_grad():
seq_out = model.seq_embeddings(
torch.stack(batch_tids).to(device),
branch_depths=torch.stack(batch_bdeps).to(device),
linkage_types=torch.stack(batch_ltypes).to(device)
)
all_embeddings.append(seq_out[:, 0, :].cpu().numpy())
if n_failed > 0:
print(f" Warning: {n_failed} samples failed")
return np.concatenate(all_embeddings, axis=0) if all_embeddings else np.array([])
def load_task_data(task_name, tokenizer, splits=None, embed_all=False):
task_cfg = TASKS[task_name]
csv_path = DATA_DIR / task_cfg['csv']
label_col = task_cfg['label_col']
wurcs_col = task_cfg['wurcs_col']
split_cols = task_cfg['split_cols']
print(f"\n{'='*60}")
print(f"Loading task: {task_name} ({task_cfg['description']})")
print(f" CSV: {csv_path}")
df = pd.read_csv(csv_path)
print(f" Total rows: {len(df)}")
target_splits = list(split_cols.keys()) if embed_all or splits is None else splits
results = []
n_tokenized = n_failed = n_ambiguous = 0
for _, row in df.iterrows():
split = 'unknown'
for split_name, col_name in split_cols.items():
if col_name in df.columns:
val = row.get(col_name)
if val == 1 or val == True or str(val).lower() in ('true', '1', '1.0'):
split = split_name
break
if split not in target_splits and not embed_all:
continue
label = row.get(label_col, '')
if pd.isna(label) or label == '' or label == 'nan':
label = 'Unknown'
wurcs = row.get(wurcs_col, '')
if pd.isna(wurcs) or wurcs == '' or not str(wurcs).startswith('WURCS'):
n_ambiguous += 1
continue
try:
tok = tokenizer.tokenize(str(wurcs), max_length=256)
results.append({
'token_ids': tok['token_ids'],
'branch_depths': tok.get('branch_depths', [0] * len(tok['token_ids'])),
'linkage_types': tok.get('linkage_types', [0] * len(tok['token_ids'])),
'label': str(label), 'split': split, 'wurcs': str(wurcs),
})
n_tokenized += 1
except Exception:
n_failed += 1
print(f" Tokenized: {n_tokenized}, Failed: {n_failed}, Ambiguous: {n_ambiguous}")
for s in target_splits:
s_data = [r for r in results if r['split'] == s]
labels = {}
for r in s_data:
labels[r['label']] = labels.get(r['label'], 0) + 1
print(f" Split '{s}': {len(s_data)} samples, labels: {labels}")
return results
def plot_embeddings(embeddings, labels, task_name, model_name, output_dir, method='tsne',
color_map=None, split_labels=None):
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from sklearn.metrics import silhouette_score, calinski_harabasz_score
print(f" Plotting {method.upper()} for {task_name} ({model_name})...")
if method == 'tsne':
from sklearn.manifold import TSNE
perplexity = min(30, len(embeddings) - 1)
coords = TSNE(n_components=2, perplexity=perplexity, max_iter=1000,
init='pca', random_state=42, learning_rate='auto').fit_transform(embeddings)
else:
import umap
coords = umap.UMAP(n_neighbors=15, min_dist=0.1, random_state=42).fit_transform(embeddings)
unique_labels = sorted(set(labels))
label_to_int = {l: i for i, l in enumerate(unique_labels)}
int_labels = np.array([label_to_int[l] for l in labels])
metrics = {}
if 2 <= len(unique_labels) < len(embeddings):
try: metrics['silhouette'] = float(silhouette_score(embeddings, int_labels))
except: metrics['silhouette'] = None
try: metrics['calinski_harabasz'] = float(calinski_harabasz_score(embeddings, int_labels))
except: metrics['calinski_harabasz'] = None
metrics['n_samples'] = len(embeddings)
metrics['n_classes'] = len(unique_labels)
metrics['classes'] = unique_labels
fig, ax = plt.subplots(1, 1, figsize=(10, 8))
for label in unique_labels:
mask = np.array(labels) == label
color = color_map.get(label, None) if color_map else None
ax.scatter(coords[mask, 0], coords[mask, 1], c=color,
label=f'{label} (n={mask.sum()})', s=15, alpha=0.7, edgecolors='none')
if split_labels is not None:
for split in sorted(set(split_labels)):
mask = np.array(split_labels) == split
if split == 'test':
ax.scatter(coords[mask, 0], coords[mask, 1], facecolors='none',
edgecolors='black', s=40, linewidths=0.5, alpha=0.3,
label=f'test split (n={mask.sum()})')
sil_str = f"Sil={metrics.get('silhouette', 'N/A'):.3f}" if metrics.get('silhouette') is not None else "Sil=N/A"
ch_str = f"CH={metrics.get('calinski_harabasz', 'N/A'):.1f}" if metrics.get('calinski_harabasz') is not None else "CH=N/A"
ax.set_title(f"{task_name} - {model_name.upper()} [CLS] ({method.upper()})\n{sil_str} | {ch_str} | n={len(embeddings)}", fontsize=13)
ax.set_xlabel(f'{method.upper()}-1')
ax.set_ylabel(f'{method.upper()}-2')
ax.legend(loc='best', fontsize=8, framealpha=0.8)
ax.set_aspect('equal', adjustable='box')
plt.tight_layout()
fname = f'{task_name}_{model_name}_{method}.png'
plt.savefig(os.path.join(output_dir, fname), dpi=200, bbox_inches='tight')
plt.close()
print(f" Saved: {fname}")
return metrics
def main():
parser = argparse.ArgumentParser(description='Embed benchmark tasks with V5/V6')
parser.add_argument('--model', choices=['v5', 'v6'], required=True)
parser.add_argument('--splits', nargs='+', default=['val', 'test'])
parser.add_argument('--embed_all', action='store_true')
parser.add_argument('--tasks', nargs='+', default=list(TASKS.keys()))
parser.add_argument('--method', choices=['tsne', 'umap', 'both'], default='tsne')
parser.add_argument('--output_dir', default=None)
parser.add_argument('--device', default='cuda')
args = parser.parse_args()
if args.output_dir is None:
args.output_dir = str(PROJECT_ROOT / 'bert_v6_contrastive' / 'analysis' / 'benchmark_embeddings')
os.makedirs(args.output_dir, exist_ok=True)
print(f"Loading tokenizer from {VOCAB_PATH}...")
tokenizer = WURCSTokenizer(str(VOCAB_PATH))
print(f" Vocab size: {tokenizer.vocab_size}")
ckpt_path = CHECKPOINTS[args.model]
if not ckpt_path.exists():
print(f"ERROR: Checkpoint not found: {ckpt_path}")
sys.exit(1)
model = load_model(str(ckpt_path), device=args.device)
color_maps = {'domain': DOMAIN_COLORS, 'kingdom': KINGDOM_COLORS,
'link': LINK_COLORS, 'immunogenicity': IMMUNO_COLORS}
all_metrics = {}
for task_name in args.tasks:
if task_name not in TASKS:
print(f"WARNING: Unknown task '{task_name}', skipping")
continue
data = load_task_data(task_name, tokenizer,
splits=args.splits if not args.embed_all else None,
embed_all=args.embed_all)
if len(data) < 10:
print(f" Skipping {task_name}: too few samples ({len(data)})")
continue
print(f" Extracting [CLS] embeddings for {len(data)} samples...")
embeddings = extract_cls_embeddings(model, data, device=args.device)
labels = [d['label'] for d in data]
split_labels = [d['split'] for d in data]
valid_mask = [l != 'Unknown' for l in labels]
embeddings = embeddings[valid_mask]
labels = [l for l, v in zip(labels, valid_mask) if v]
split_labels = [s for s, v in zip(split_labels, valid_mask) if v]
if len(embeddings) < 10:
print(f" Skipping {task_name}: too few labeled samples")
continue
print(f" Embeddings shape: {embeddings.shape}")
npz_path = os.path.join(args.output_dir, f'{task_name}_{args.model}_embeddings.npz')
np.savez_compressed(npz_path, embeddings=embeddings,
labels=np.array(labels), splits=np.array(split_labels))
print(f" Saved: {npz_path}")
methods = ['tsne', 'umap'] if args.method == 'both' else [args.method]
task_metrics = {}
for method in methods:
m = plot_embeddings(embeddings, labels, task_name, args.model,
args.output_dir, method=method,
color_map=color_maps.get(task_name, None),
split_labels=split_labels)
task_metrics[method] = m
all_metrics[task_name] = task_metrics
metrics_path = os.path.join(args.output_dir, f'benchmark_metrics_{args.model}.json')
with open(metrics_path, 'w') as f:
json.dump(all_metrics, f, indent=2, default=str)
print(f"\nAll metrics saved to: {metrics_path}")
print(f"\n{'='*60}")
print(f"SUMMARY - {args.model.upper()}")
print(f"{'='*60}")
for task, tmetrics in all_metrics.items():
for method, m in tmetrics.items():
sil = m.get('silhouette', 'N/A')
sil_str = f"{sil:.4f}" if isinstance(sil, float) else str(sil)
print(f" {task:20s} ({method:5s}): Silhouette={sil_str}, n={m.get('n_samples',0)}, classes={m.get('n_classes',0)}")
if __name__ == '__main__':
main()
|