File size: 18,898 Bytes
83a9309 | 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 | #!/usr/bin/env python3
import os
import sys
import yaml
import h5py
import torch
import argparse
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
import logging
from typing import Dict, List, Optional, Tuple, Any
# Need imports for classes/functions used
from torch_geometric.nn import ChebConv # For HNO definition
from torch_cluster import knn_graph # Needed if calculating z_ref
##############################################################
# Copied Class Definitions from Script M (Response #16)
##############################################################
class HNO(nn.Module):
def __init__(self, hidden_dim, K):
super().__init__()
self._debug_logged = False # For one-time debug logging
logger.debug(f"Initializing HNO with hidden_dim={hidden_dim}, K={K}")
sys.stdout.flush()
# Input dimension is 3 (x, y, z coordinates)
self.conv1 = ChebConv(3, hidden_dim, K=K)
self.conv2 = ChebConv(hidden_dim, hidden_dim, K=K)
self.conv3 = ChebConv(hidden_dim, hidden_dim, K=K)
self.conv4 = ChebConv(hidden_dim, hidden_dim, K=K)
# BatchNorm applied on the feature dimension (hidden_dim)
self.bano1 = nn.BatchNorm1d(hidden_dim)
self.bano2 = nn.BatchNorm1d(hidden_dim)
self.bano3 = nn.BatchNorm1d(hidden_dim)
# Final MLP maps back to 3D coordinates
self.mlpRep = nn.Linear(hidden_dim, 3)
def forward(self, x, edge_index, log_debug=False):
x = x.float()
x = self.conv1(x, edge_index)
x = self.bano1(F.leaky_relu(x))
x = self.conv2(x, edge_index)
x = self.bano2(F.leaky_relu(x))
x = self.conv3(x, edge_index)
x = self.bano3(F.relu(x)) # Note: ReLU here, LeakyReLU before
x = self.conv4(x, edge_index)
x = F.normalize(x, p=2.0, dim=1) # dim=1 is the feature dimension
x = self.mlpRep(x)
return x
def forward_representation(self, x, edge_index, log_debug=False):
x = x.float() # Ensure input is float
x = self.conv1(x, edge_index)
x = self.bano1(F.leaky_relu(x))
x = self.conv2(x, edge_index)
x = self.bano2(F.leaky_relu(x))
x = self.conv3(x, edge_index)
x = self.bano3(F.relu(x))
x = self.conv4(x, edge_index)
x = F.normalize(x, p=2.0, dim=1)
return x
# --- MLP Builder ---
def build_decoder_mlp(in_dim: int, out_dim: int, N_layers: int, h_dim: int = 128) -> nn.Sequential:
"""Builds MLP with BatchNorm."""
layers: List[nn.Module] = []; curr = in_dim
if N_layers<=0: raise ValueError("MLP layers must be >= 1.")
elif N_layers==1: layers.append(nn.Linear(curr, out_dim))
else:
layers.extend([nn.Linear(curr, h_dim), nn.BatchNorm1d(h_dim), nn.ReLU()]); curr = h_dim
for _ in range(N_layers - 2): layers.extend([nn.Linear(curr, h_dim), nn.BatchNorm1d(h_dim), nn.ReLU()])
layers.append(nn.Linear(curr, out_dim))
return nn.Sequential(*layers)
# --- Decoder2 Model Definition (Needed for loading state_dict) ---
class ProteinStateReconstructor2D(nn.Module):
_logged_fwd = False
def __init__(self, in_dim: int, N_nodes: int, cond_dim: int, pool_type: str = "blind", res_indices: Optional[List[List[int]]] = None, pool_size: Tuple[int, int] = (20, 4), mlp_h_dim: int = 128, mlp_layers: int = 2, pool2_size: Optional[Tuple[int, int]] = None, use_pool2: bool = False, use_attn: bool = False, attn_type: str = "global", logger: logging.Logger = logging.getLogger()): # Added logger default
super().__init__(); self.N_nodes=N_nodes; self.in_dim=in_dim; self.cond_dim=cond_dim; self.logger=logger; self.pool_type=pool_type
self.seg_indices: List[torch.LongTensor] = []
if pool_type=="blind": self.seg_indices.append(torch.arange(N_nodes,dtype=torch.long))
elif pool_type=="residue":
if not res_indices: self.logger.warning("Residue pooling selected but no indices provided during init."); self.seg_indices.append(torch.arange(N_nodes,dtype=torch.long)) # Fallback?
else: self.seg_indices = [torch.tensor(idx, dtype=torch.long) for idx in res_indices if idx];
if not self.seg_indices: raise ValueError("Empty residue segments.")
else: raise ValueError(f"Unknown pool_type={pool_type}")
self.N_seg = len(self.seg_indices); #self.logger.info(f"Dec2 Init: Pool='{pool_type}', Segs={self.N_seg}")
self.seg_pools = nn.ModuleList([nn.AdaptiveAvgPool2d(pool_size) for _ in self.seg_indices])
self.prim_pool_dim = pool_size[0]*pool_size[1]; self.final_pool_dim = self.prim_pool_dim*self.N_seg # Calculate initial pooled dim based on primary pool concat
self.glob_pool2 = None
if use_pool2 and pool2_size and self.N_seg>0:
self.glob_pool2=nn.AdaptiveAvgPool2d(pool2_size);
self.final_pool_dim=pool2_size[0]*pool2_size[1]; # Recalculate if secondary pooling is used
#self.logger.info(f"Dec2 Init: Use Pool2. FinalPoolDim={self.final_pool_dim}")
#else: self.logger.info(f"Dec2 Init: Primary Pool only. ConcatPoolDim={self.final_pool_dim}")
self.attn = None;
if use_attn: self.logger.warning("Attn impl omitted.")
mlp_in_dim=self.cond_dim+self.final_pool_dim; # Input to MLP is conditioner + final_pooled_dim
self.decoder=build_decoder_mlp(mlp_in_dim, 3, mlp_layers, mlp_h_dim) # This is the part we need
self.logger.info(f"Dec2 Init MLP: In={mlp_in_dim}, Out=3, Layers={mlp_layers}, Hidden={mlp_h_dim}")
# get_pooled_latent and forward are NOT used directly in this script.
# --- End of Copied Class Definitions ---
##############################################################
# (D) Main Decoding Function
##############################################################
def main():
parser = argparse.ArgumentParser(description="Decode diffused GLOBAL pooled embeddings using trained Decoder2 MLP.")
parser.add_argument("--config", type=str, required=True, help="Path to YAML config used for TRAINING the main script (Script M).")
# Checkpoints from Script M training
parser.add_argument("--hno_ckpt", type=str, default="checkpoints/hno_checkpoint.pth", help="Path to trained HNO ckpt (needed only if conditioner_mode='z_ref' AND z_ref file is missing).")
parser.add_argument("--decoder2_ckpt", type=str, default="checkpoints/decoder2_checkpoint.pth", help="Path to trained Decoder2 ckpt (from Script M).")
# Inputs: Diffused embeddings and Conditioner references from Script M
parser.add_argument("--diff_emb_file", type=str, required=True, help="HDF5 file with diffused embeddings (output from diffusion script).")
parser.add_argument("--diff_emb_key", type=str, default="generated_embeddings", help="Dataset key for diffused embeddings in HDF5 file.")
parser.add_argument("--conditioner_x_ref_pt", type=str, default="structures/X_ref_coords.pt", help="Path to saved X_ref_coords.pt (generated by Script M).")
parser.add_argument("--conditioner_z_ref_pt", type=str, default="latent_reps/z_ref_embedding.pt", help="Path to saved z_ref_embedding.pt (generated by Script M, required if conditioner_mode='z_ref').")
# Output
parser.add_argument("--output_file", type=str, default="structures/full_coords_diff.h5", help="Output HDF5 file for decoded diffusion coordinates.")
parser.add_argument("--output_key", type=str, default="full_coords_diff", help="Dataset key for output coords in HDF5 file.")
# Optional
parser.add_argument("--batch_size", type=int, default=64, help="Batch size for decoding.")
parser.add_argument("--cuda_device", type=int, default=0, help="GPU device index if CUDA available.")
args = parser.parse_args()
# Basic Logging Setup for this script
logging.basicConfig(level=logging.INFO, format='[%(asctime)s - %(levelname)s] %(message)s')
# --- Device Setup ---
if torch.cuda.is_available():
device = torch.device(f"cuda:{args.cuda_device}")
try:
torch.cuda.set_device(device) # Set default device
except Exception as e:
logging.warning(f"Could not set CUDA device {args.cuda_device}: {e}. Defaulting to cpu.")
device = torch.device("cpu")
else:
device = torch.device("cpu")
logging.info(f"Using Device: {device}")
# --- Load Training Config ---
try:
with open(args.config, "r") as f: config = yaml.safe_load(f)
logging.info(f"Loaded training config from: {args.config}")
except Exception as e: logging.error(f"Failed to load config: {e}"); sys.exit(1)
# --- Extract Necessary Parameters from Config ---
try:
hno_hdim = config["hno_encoder"]["hidden_dim"]
hno_k = config["hno_encoder"]["cheb_order"]
graph_cfg = config["graph"]; knn = graph_cfg["knn_value"] # Needed if calculating z_ref
d2s = config["decoder2_settings"]
d2_c_mode = d2s.get("conditioner_mode","z_ref"); d2_p_type = d2s.get("pooling_type","blind")
d2_ph, d2_pw = d2s.get("output_height",20), d2s.get("output_width",4); d2_mlp_h = d2s.get("mlp_hidden_dim",128)
d2_mlp_l = d2s.get("num_hidden_layers",2); d2_use_p2 = d2s.get("use_second_level_pooling",False);
d2_ph2, d2_pw2 = d2s.get("output_height2"), d2s.get("output_width2")
# --- FIX: Added missing lines ---
d2_use_a = d2s.get("use_cross_attention", False)
d2_a_type = d2s.get("cross_attention_type", "global")
# --- End Fix ---
except KeyError as e: logging.error(f"Missing key in config: {e}"); sys.exit(1)
# --- Load Conditioner Data ---
try:
X_ref_t = torch.load(args.conditioner_x_ref_pt, map_location='cpu').float() # Load X_ref [N, 3] on CPU
N_atoms = X_ref_t.shape[0]
logging.info(f"Loaded X_ref, N_atoms = {N_atoms}")
except Exception as e: logging.error(f"Failed to load X_ref from {args.conditioner_x_ref_pt}: {e}"); sys.exit(1)
conditioner_cpu = None; cond_dim = -1
if d2_c_mode == "X_ref":
conditioner_cpu = X_ref_t; cond_dim = 3; logging.info(f"Using X_ref conditioner (Dim={cond_dim})")
elif d2_c_mode == "z_ref":
if os.path.isfile(args.conditioner_z_ref_pt):
try:
conditioner_cpu = torch.load(args.conditioner_z_ref_pt, map_location='cpu').float()
cond_dim = conditioner_cpu.shape[1]
logging.info(f"Loaded pre-computed z_ref conditioner (Dim={cond_dim})")
except Exception as e: logging.error(f"Failed to load z_ref file {args.conditioner_z_ref_pt}: {e}"); sys.exit(1)
else: # Need to compute z_ref
logging.info(f"z_ref file not found ({args.conditioner_z_ref_pt}). Computing z_ref using HNO...")
try:
hno_model = HNO(hidden_dim=hno_hdim, K=hno_k).to(device)
if not os.path.isfile(args.hno_ckpt): raise FileNotFoundError(f"HNO checkpoint not found at {args.hno_ckpt}")
hno_ckpt_data = torch.load(args.hno_ckpt, map_location=device)
hno_model.load_state_dict(hno_ckpt_data["model_state_dict"]); hno_model.eval()
logging.info(f"Loaded HNO model from: {args.hno_ckpt}")
logging.info("Calculating edge index for X_ref...")
edge_index = knn_graph(X_ref_t.to(device), k=knn, loop=False)
logging.info("Calculating z_ref via HNO forward_representation...")
with torch.no_grad(): z_ref_dev = hno_model.forward_representation(X_ref_t.to(device), edge_index)
conditioner_cpu = z_ref_dev.cpu(); cond_dim = conditioner_cpu.shape[1]
logging.info(f"Computed z_ref as conditioner (Dim={cond_dim})")
try: torch.save(conditioner_cpu, args.conditioner_z_ref_pt); logging.info(f"Saved computed z_ref to {args.conditioner_z_ref_pt}")
except Exception as save_e: logging.warning(f"Could not save computed z_ref: {save_e}")
del hno_model, edge_index # Clean up
if torch.cuda.is_available(): torch.cuda.empty_cache()
except Exception as e: logging.error(f"Failed to compute z_ref: {e}"); sys.exit(1)
else: logging.error(f"Invalid conditioner_mode: {d2_c_mode}"); sys.exit(1)
if conditioner_cpu is None or conditioner_cpu.shape[0] != N_atoms: logging.error(f"Conditioner failed validation (None or wrong N_atoms {conditioner_cpu.shape[0]} vs {N_atoms})."); sys.exit(1)
# --- Load Diffused Embeddings ---
try:
logging.info(f"Loading diffused embeddings: {args.diff_emb_file} [Key: {args.diff_emb_key}]")
with h5py.File(args.diff_emb_file, "r") as f: diff_emb_np = f[args.diff_emb_key][:]
diff_emb_t = torch.tensor(diff_emb_np, dtype=torch.float) # Keep on CPU initially for shape checks
N_gen = diff_emb_t.shape[0]
logging.info(f"Loaded diffusion embeddings shape: {diff_emb_t.shape}")
except Exception as e: logging.error(f"Failed to load diffused embeddings: {e}"); sys.exit(1)
# --- Instantiate Decoder2 to get parameters and MLP ---
res_indices_dummy = [list(range(N_atoms))] if d2_p_type == "blind" else None
if d2_p_type == "residue": logging.warning("Residue pooling selected, ensure diffusion embs match expected structure.")
temp_logger = logging.getLogger("dummy"); temp_logger.setLevel(logging.CRITICAL)
try:
decoder2_instance_for_state = ProteinStateReconstructor2D(
in_dim=hno_hdim, N_nodes=N_atoms, cond_dim=cond_dim, pool_type=d2_p_type,
res_indices=res_indices_dummy, pool_size=(d2_ph, d2_pw), mlp_h_dim=d2_mlp_h,
mlp_layers=d2_mlp_l, pool2_size=((d2_ph2, d2_pw2) if d2_use_p2 and d2_ph2 and d2_pw2 else None),
use_pool2=d2_use_p2, use_attn=d2_use_a, attn_type=d2_a_type, logger=temp_logger
)
final_pool_dim_expected = decoder2_instance_for_state.final_pool_dim
logging.info(f"Decoder expects final pooled dimension: {final_pool_dim_expected}")
except Exception as e: logging.error(f"Failed to instantiate dummy Decoder2 model structure: {e}"); sys.exit(1)
# --- Validate and Reshape loaded diffusion embeddings ---
diff_emb_global_t = None # Will hold final [N_gen, final_pool_dim_expected] tensor
if diff_emb_t.ndim == 3: # Shape [N_gen, R, PD_seg] - Assume from Script D
logging.info("Loaded embeddings seem per-segment/residue. Flattening to global representation.")
loaded_R, loaded_PD_seg = diff_emb_t.shape[1], diff_emb_t.shape[2]
# Flatten to [N_gen, R * PD_seg]
diff_emb_global_t = diff_emb_t.reshape(N_gen, -1)
loaded_pooled_dim = diff_emb_global_t.shape[1]
logging.info(f"Flattened per-segment embeddings to global shape: {diff_emb_global_t.shape}")
elif diff_emb_t.ndim == 2: # Shape [N_gen, SomeDim] - Assume this is already global
logging.info("Loaded embeddings seem global.")
diff_emb_global_t = diff_emb_t
loaded_pooled_dim = diff_emb_global_t.shape[1]
else: logging.error(f"Loaded diffusion embeddings unexpected ndim={diff_emb_t.ndim}."); sys.exit(1)
# Final dimension check
if loaded_pooled_dim != final_pool_dim_expected:
logging.error(f"Dimension mismatch! Loaded/Flattened diffusion emb dim ({loaded_pooled_dim}) != Expected decoder pooled dim ({final_pool_dim_expected}).")
sys.exit(1)
diff_emb_global_t = diff_emb_global_t.to(device) # Move final embeddings to device
# --- Load Decoder2 Checkpoint and Extract MLP ---
try:
logging.info(f"Loading Decoder2 checkpoint from: {args.decoder2_ckpt}")
decoder2_model = ProteinStateReconstructor2D( # Instantiate with correct parameters
in_dim=hno_hdim, N_nodes=N_atoms, cond_dim=cond_dim, pool_type=d2_p_type,
res_indices=res_indices_dummy, pool_size=(d2_ph, d2_pw), mlp_h_dim=d2_mlp_h,
mlp_layers=d2_mlp_l, pool2_size=((d2_ph2, d2_pw2) if d2_use_p2 and d2_ph2 and d2_pw2 else None),
use_pool2=d2_use_p2, use_attn=d2_use_a, attn_type=d2_a_type, logger=temp_logger
).to(device)
dec2_ckpt_data = torch.load(args.decoder2_ckpt, map_location=device)
decoder2_model.load_state_dict(dec2_ckpt_data["model_state_dict"])
decoder_mlp = decoder2_model.decoder # Extract the MLP part
decoder_mlp.eval()
logging.info("Successfully loaded Decoder2 MLP weights.")
except Exception as e: logging.error(f"Failed to load Decoder2 checkpoint/MLP: {e}"); sys.exit(1)
# --- Prepare Conditioner for Batches ---
conditioner_dev = conditioner_cpu.to(device) # Move final conditioner to device
# --- Decoding Loop ---
logging.info(f"Starting decoding for {N_gen} samples (Batch Size: {args.batch_size})...")
all_coords_list = []
with torch.no_grad():
for i in range(0, N_gen, args.batch_size):
b_start, b_end = i, min(i + args.batch_size, N_gen)
bs = b_end - b_start
pooled_batch = diff_emb_global_t[b_start:b_end] # [bs, final_pooled_dim]
cond_expanded = conditioner_dev.unsqueeze(0).expand(bs, -1, -1) # [bs, N, Cdim]
# Replicate the logic: Expand global pooled, concat with expanded conditioner
pooled_expanded = pooled_batch.unsqueeze(1).expand(-1, N_atoms, -1) # [bs, N, final_pooled_dim]
mlp_input_batch = torch.cat([cond_expanded, pooled_expanded], dim=-1) # [bs, N, Cdim + final_pool_dim]
mlp_input_flat = mlp_input_batch.view(bs * N_atoms, -1)
# Use the extracted MLP
coords_flat = decoder_mlp(mlp_input_flat) # [bs*N, 3]
coords_batch = coords_flat.view(bs, N_atoms, 3) # [bs, N, 3]
all_coords_list.append(coords_batch.cpu().numpy())
if (i // args.batch_size + 1) % 20 == 0: # Log progress less frequently
logging.info(f" Decoded up to sample {b_end}/{N_gen}")
# --- Concatenate and Save Results ---
try:
final_coords_np = np.concatenate(all_coords_list, axis=0)
logging.info(f"Concatenated decoded coords shape: {final_coords_np.shape}")
if final_coords_np.shape != (N_gen, N_atoms, 3):
logging.warning("Final coords shape mismatch!") # Sanity check
logging.info(f"Saving decoded structures to: {args.output_file} (Key: {args.output_key})")
os.makedirs(os.path.dirname(args.output_file), exist_ok=True)
with h5py.File(args.output_file, "w") as f:
f.create_dataset(args.output_key, data=final_coords_np, chunks=(1, N_atoms, 3), compression="gzip") # Add chunking/compression
logging.info(f"[SUCCESS] Wrote decoded coordinates.")
except Exception as e: logging.error(f"Failed to concatenate or save results: {e}"); sys.exit(1)
if __name__ == "__main__":
main()
|