Spaces:
Sleeping
Sleeping
Lazy-load model and heavy deps to avoid build-time installs
Browse files
app.py
CHANGED
|
@@ -13,7 +13,6 @@ import warnings
|
|
| 13 |
from pathlib import Path
|
| 14 |
|
| 15 |
import numpy as np
|
| 16 |
-
import torch
|
| 17 |
import yaml
|
| 18 |
import matplotlib
|
| 19 |
matplotlib.use('Agg')
|
|
@@ -34,32 +33,49 @@ except ImportError as _pyg_err:
|
|
| 34 |
PYGEOMETRIC_AVAILABLE = False
|
| 35 |
_pyg_err_msg = str(_pyg_err)
|
| 36 |
|
| 37 |
-
|
| 38 |
-
from src.data.mol_graph import smiles_to_graph
|
| 39 |
from src.data.protein_feat import residue_physicochemical_features, residue_domain_labels
|
| 40 |
|
| 41 |
-
#
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
DEVICE = torch.device('cpu')
|
| 46 |
_CONFIG_PATH = ROOT / 'configs' / 'default.yaml'
|
| 47 |
_CKPT_PATH = ROOT / 'checkpoints' / 'best.pt'
|
| 48 |
_REPORT_PATH = ROOT / 'examples' / 'interpretability_report.json'
|
| 49 |
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
_model
|
| 56 |
-
|
| 57 |
-
_model
|
| 58 |
-
_model.eval()
|
| 59 |
-
print(f"[BioInteract] Model ready — epoch {_ckpt.get('epoch','?')}, "
|
| 60 |
-
f"params = {sum(p.numel() for p in _model.parameters()):,}")
|
| 61 |
|
| 62 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 63 |
with open(_REPORT_PATH) as f:
|
| 64 |
_REPORT = json.load(f)
|
| 65 |
|
|
@@ -84,6 +100,7 @@ def _get_esm():
|
|
| 84 |
|
| 85 |
def compute_esm2_embedding(sequence: str, max_len: int = 512) -> torch.Tensor:
|
| 86 |
"""Run ESM-2 and return per-residue embeddings (L, 640)."""
|
|
|
|
| 87 |
seq = sequence[:max_len]
|
| 88 |
tokenizer, esm = _get_esm()
|
| 89 |
inputs = tokenizer(seq, return_tensors='pt', add_special_tokens=True)
|
|
@@ -299,6 +316,11 @@ def run_prediction(smiles: str, sequence: str, progress=gr.Progress()):
|
|
| 299 |
if not sequence:
|
| 300 |
return "Please enter a protein sequence (amino acids).", "", None, None
|
| 301 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 302 |
if not PYGEOMETRIC_AVAILABLE:
|
| 303 |
return (
|
| 304 |
f"torch-geometric is not available in this environment: {_pyg_err_msg}. "
|
|
@@ -308,11 +330,15 @@ def run_prediction(smiles: str, sequence: str, progress=gr.Progress()):
|
|
| 308 |
|
| 309 |
# --- Drug graph ---
|
| 310 |
progress(0.1, desc="Parsing SMILES …")
|
|
|
|
|
|
|
| 311 |
drug_graph = smiles_to_graph(smiles)
|
| 312 |
if drug_graph is None:
|
| 313 |
return "Invalid SMILES string — RDKit could not parse it.", "", None, None
|
| 314 |
|
| 315 |
-
|
|
|
|
|
|
|
| 316 |
|
| 317 |
# --- Protein features ---
|
| 318 |
seq = sequence[:MAX_SEQ_LEN]
|
|
@@ -326,10 +352,10 @@ def run_prediction(smiles: str, sequence: str, progress=gr.Progress()):
|
|
| 326 |
except Exception as e:
|
| 327 |
return f"ESM-2 error: {e}", "", None, None
|
| 328 |
|
| 329 |
-
esm2_emb = esm2_emb.unsqueeze(0).to(
|
| 330 |
-
physchem = residue_physicochemical_features(seq).unsqueeze(0).to(
|
| 331 |
-
domain = residue_domain_labels(L).unsqueeze(0).to(
|
| 332 |
-
prot_mask = torch.ones(1, L, dtype=torch.bool, device=
|
| 333 |
|
| 334 |
# --- Model forward ---
|
| 335 |
progress(0.85, desc="Running BioInteract model …")
|
|
|
|
| 13 |
from pathlib import Path
|
| 14 |
|
| 15 |
import numpy as np
|
|
|
|
| 16 |
import yaml
|
| 17 |
import matplotlib
|
| 18 |
matplotlib.use('Agg')
|
|
|
|
| 33 |
PYGEOMETRIC_AVAILABLE = False
|
| 34 |
_pyg_err_msg = str(_pyg_err)
|
| 35 |
|
| 36 |
+
# protein features are pure-python and safe to import eagerly
|
|
|
|
| 37 |
from src.data.protein_feat import residue_physicochemical_features, residue_domain_labels
|
| 38 |
|
| 39 |
+
# Model and heavy-dependency lazy loader
|
| 40 |
+
_model = None
|
| 41 |
+
_DEVICE = None
|
|
|
|
|
|
|
| 42 |
_CONFIG_PATH = ROOT / 'configs' / 'default.yaml'
|
| 43 |
_CKPT_PATH = ROOT / 'checkpoints' / 'best.pt'
|
| 44 |
_REPORT_PATH = ROOT / 'examples' / 'interpretability_report.json'
|
| 45 |
|
| 46 |
+
def _load_model():
|
| 47 |
+
"""Lazy-load the PyTorch model and return a ready-to-use instance.
|
| 48 |
+
This delays importing heavy packages (torch, torch_geometric, rdkit)
|
| 49 |
+
until a user requests a custom prediction.
|
| 50 |
+
"""
|
| 51 |
+
global _model, _DEVICE
|
| 52 |
+
if _model is not None:
|
| 53 |
+
return _model
|
|
|
|
|
|
|
|
|
|
| 54 |
|
| 55 |
+
try:
|
| 56 |
+
import torch
|
| 57 |
+
except Exception as e:
|
| 58 |
+
raise RuntimeError(f"PyTorch not available: {e}")
|
| 59 |
+
|
| 60 |
+
# Import model code only when torch is available
|
| 61 |
+
from src.models.biointeract import BioInteract
|
| 62 |
+
|
| 63 |
+
print("[BioInteract] Loading model config …")
|
| 64 |
+
with open(_CONFIG_PATH) as f:
|
| 65 |
+
cfg = yaml.safe_load(f)
|
| 66 |
+
|
| 67 |
+
print("[BioInteract] Loading pretrained weights …")
|
| 68 |
+
device = torch.device('cpu')
|
| 69 |
+
model = BioInteract(cfg['model']).to(device)
|
| 70 |
+
ckpt = torch.load(_CKPT_PATH, map_location='cpu', weights_only=False)
|
| 71 |
+
model.load_state_dict(ckpt['model_state_dict'])
|
| 72 |
+
model.eval()
|
| 73 |
+
_model = model
|
| 74 |
+
_DEVICE = device
|
| 75 |
+
print(f"[BioInteract] Model ready — epoch {ckpt.get('epoch','?')}, params = {sum(p.numel() for p in _model.parameters()):,}")
|
| 76 |
+
return _model
|
| 77 |
+
|
| 78 |
+
# Load pre-computed case study report (safe, JSON-only)
|
| 79 |
with open(_REPORT_PATH) as f:
|
| 80 |
_REPORT = json.load(f)
|
| 81 |
|
|
|
|
| 100 |
|
| 101 |
def compute_esm2_embedding(sequence: str, max_len: int = 512) -> torch.Tensor:
|
| 102 |
"""Run ESM-2 and return per-residue embeddings (L, 640)."""
|
| 103 |
+
import torch
|
| 104 |
seq = sequence[:max_len]
|
| 105 |
tokenizer, esm = _get_esm()
|
| 106 |
inputs = tokenizer(seq, return_tensors='pt', add_special_tokens=True)
|
|
|
|
| 316 |
if not sequence:
|
| 317 |
return "Please enter a protein sequence (amino acids).", "", None, None
|
| 318 |
|
| 319 |
+
try:
|
| 320 |
+
_load_model()
|
| 321 |
+
except Exception as e:
|
| 322 |
+
return (f"Model not available: {e}", "", None, None)
|
| 323 |
+
|
| 324 |
if not PYGEOMETRIC_AVAILABLE:
|
| 325 |
return (
|
| 326 |
f"torch-geometric is not available in this environment: {_pyg_err_msg}. "
|
|
|
|
| 330 |
|
| 331 |
# --- Drug graph ---
|
| 332 |
progress(0.1, desc="Parsing SMILES …")
|
| 333 |
+
# import heavy SMILES->graph conversion lazily
|
| 334 |
+
from src.data.mol_graph import smiles_to_graph
|
| 335 |
drug_graph = smiles_to_graph(smiles)
|
| 336 |
if drug_graph is None:
|
| 337 |
return "Invalid SMILES string — RDKit could not parse it.", "", None, None
|
| 338 |
|
| 339 |
+
import torch
|
| 340 |
+
device = _DEVICE
|
| 341 |
+
drug_batch = PyGBatch.from_data_list([drug_graph]).to(device)
|
| 342 |
|
| 343 |
# --- Protein features ---
|
| 344 |
seq = sequence[:MAX_SEQ_LEN]
|
|
|
|
| 352 |
except Exception as e:
|
| 353 |
return f"ESM-2 error: {e}", "", None, None
|
| 354 |
|
| 355 |
+
esm2_emb = esm2_emb.unsqueeze(0).to(device) # (1, L, 640)
|
| 356 |
+
physchem = residue_physicochemical_features(seq).unsqueeze(0).to(device) # (1, L, 4)
|
| 357 |
+
domain = residue_domain_labels(L).unsqueeze(0).to(device) # (1, L)
|
| 358 |
+
prot_mask = torch.ones(1, L, dtype=torch.bool, device=device)
|
| 359 |
|
| 360 |
# --- Model forward ---
|
| 361 |
progress(0.85, desc="Running BioInteract model …")
|