File size: 10,221 Bytes
ce9b7f3 | 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 | import torch
import h5py
import numpy as np
import matplotlib.pyplot as plt
import os
from torch.utils.data import Dataset
from pathlib import Path
from collections import OrderedDict
class DFTDatasetH5(Dataset):
"""
PyTorch Dataset for variable-length atomistic data stored in HDF5.
If `indices` is None, will try to read a split from the HDF5:
- if `split` is provided, load f["splits/<split>"]
- else try 'train' -> 'val'
- else fall back to all samples
"""
def __init__(self, h5_path, indices=None, split=None):
self.h5_path = h5_path
self.split_name = None
with h5py.File(h5_path, "r") as f:
self.offsets = f["offsets"][:] # int64, slice boundaries
self.num_atoms = f["num_atoms"][:] # int32, sequence lengths
self.energy = f["energy"][:] # float64
self.cell = f["cell"][:] # (N,3,3) float64
self._N_total = len(self.offsets) - 1
if indices is not None:
self.indices = np.asarray(indices, dtype=np.int64)
self.split_name = split # may be None if indices provided directly
else:
# Try to load indices from the file
loaded = False
if "splits" in f:
grp = f["splits"]
if split is not None:
ds = grp.get(split)
if ds is None:
raise KeyError(f"Requested split '{split}' not found in HDF5 ('splits' group exists but no '{split}').")
self.indices = ds[:].astype(np.int64, copy=False)
self.split_name = split
loaded = True
else:
for cand in ("train", "val"):
if cand in grp:
self.indices = grp[cand][:].astype(np.int64, copy=False)
self.split_name = cand
loaded = True
break
if not loaded:
# Fall back to all samples
self.indices = np.arange(self._N_total, dtype=np.int64)
self.split_name = "all"
self._h5 = None # lazy-open per worker
def _ensure_open(self):
if self._h5 is None:
self._h5 = h5py.File(self.h5_path, "r", libver="latest", swmr=True)
self._pos = self._h5["positions"] # float64
self._frc = self._h5["forces"] # float64
self._sym = self._h5["symbols"] # int64
def __len__(self):
return len(self.indices)
def __getitem__(self, i):
self._ensure_open()
j = int(self.indices[i]) # original sample index
s, e = self.offsets[j], self.offsets[j+1]
pos = torch.from_numpy(self._pos[s:e]).float() # (L,3)
frc = torch.from_numpy(self._frc[s:e]).float() # (L,3)
sym = torch.from_numpy(self._sym[s:e]) # (L,) int64
E = torch.tensor(self.energy[j]).float() # scalar
cell = torch.from_numpy(self.cell[j]).float() # (3,3)
L = int(self.num_atoms[j])
return {
"pos": pos,
"forces": frc,
"atomic_numbers": sym,
"energy": E.unsqueeze(0),
"cell": cell,
"natoms": L,
"pbc": torch.tensor([True, True, True]),
}
def __repr__(self):
return (f"{self.__class__.__name__}(path='{self.h5_path}', "
f"split='{self.split_name}', size={len(self)})")
def custom_collate_fn(batch):
"""
Collate function to batch graph data for EquiformerV2.
Args:
batch (list of dict): List of individual data samples.
Returns:
dict: Batched data.
"""
# Extract fields
pos = torch.cat([b["pos"] for b in batch], dim=0)
atomic_numbers = torch.cat([b["atomic_numbers"] for b in batch], dim=0)
cell = torch.stack([b["cell"] for b in batch if b["cell"] is not None]) if "cell" in batch[0] else None
pbc = torch.stack([b["pbc"] for b in batch if b["pbc"] is not None]) if "pbc" in batch[0] else None
natoms = torch.tensor([b["natoms"] for b in batch], dtype=torch.long)
forces = torch.cat([b["forces"] for b in batch], dim=0)
energy = torch.cat([b["energy"] for b in batch], dim=0)
# Build the batch index
batch_index = torch.cat([
torch.full((b["pos"].shape[0],), i, dtype=torch.long) for i, b in enumerate(batch)
], dim=0)
return {
"pos": pos,
"atomic_numbers": atomic_numbers,
"cell": cell,
"pbc": pbc,
"natoms": natoms,
"batch": batch_index,
"forces": forces,
"energy": energy
}
def count_parameters(model):
num_params = 0
for param in model.parameters():
if param.requires_grad:
num_params += param.numel()
print(f'num_params is: {num_params}')
def save_checkpoint(model, optimizer, epoch, best_val_metric, checkpoint_path, is_best=False):
"""
Save model checkpoint.
Parameters:
model (torch.nn.Module): The model to save.
optimizer (torch.optim.Optimizer): The optimizer state.
epoch (int): The current epoch number.
best_val_metric (float): The best validation metric achieved so far.
checkpoint_path (str): Directory where the checkpoint will be saved.
is_best (bool): Whether this checkpoint is the best model so far.
"""
state = {
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'best_val_metric': best_val_metric,
}
# If this is the best model so far, save it separately
if is_best:
best_model_file = os.path.join(checkpoint_path, 'best_model.pth')
torch.save(state, best_model_file)
else:
# Save the checkpoint
checkpoint_file = os.path.join(checkpoint_path, f'checkpoint_epoch_{epoch}.pth')
torch.save(state, checkpoint_file)
def load_checkpoint(checkpoint_path, model, optimizer=None, load_optimizer=True):
"""
Load model checkpoint.
Parameters:
checkpoint_path (str): Path to the checkpoint file.
model (torch.nn.Module): The model to load the state_dict into.
optimizer (torch.optim.Optimizer, optional): The optimizer to load the state_dict into (if required).
load_optimizer (bool): Whether to load the optimizer state_dict (default: True).
Returns:
model (torch.nn.Module): Model with loaded weights.
optimizer (torch.optim.Optimizer, optional): Optimizer with loaded state_dict (if provided).
epoch (int): The epoch at which the checkpoint was saved.
best_val_metric (float): The best validation metric at the time of saving the checkpoint.
"""
# Load the checkpoint from file
checkpoint = torch.load(checkpoint_path, map_location=torch.device('cpu'), weights_only=True)
# Load model state_dict
state_dict = checkpoint['model_state_dict']
new_state_dict = OrderedDict()
for k, v in state_dict.items():
if k.startswith('module.'):
name = k[7:] # remove 'module.' prefix
else:
name = k
new_state_dict[name] = v
model.load_state_dict(new_state_dict)
# Load optimizer state_dict if applicable
if load_optimizer and optimizer is not None:
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
# Extract additional information from the checkpoint
epoch = checkpoint.get('epoch', -1)
best_val_metric = checkpoint.get('best_val_metric', None)
return model, optimizer, epoch, best_val_metric
def load_PT_equiformer_weights(path, model):
ckpt = torch.load(path, map_location="cpu", weights_only=True)
state_dict = ckpt["state_dict"] # <-- confirmed key from inspection
# remove double "module.module." prefix
new_state = OrderedDict()
for k, v in state_dict.items():
if k.startswith("module.module."):
new_state[k[len("module.module."):]] = v
elif k.startswith("module."):
new_state[k[len("module."):]] = v
else:
new_state[k] = v
missing, unexpected = model.load_state_dict(new_state, strict=False)
print(f"[ckpt] load strict=False | missing={len(missing)} unexpected={len(unexpected)}")
if missing and len(missing) < 12:
print("Missing keys:", missing)
if unexpected and len(unexpected) < 12:
print("Unexpected keys:", unexpected)
return model
def validate_args(args):
# (model, select_test_dataset, load_OC20_pt)
allowed = {
("orig", "without_rep", False),
("orig", "without_rep", True), # PT only allowed here
("small", "without_rep", False),
("small", "with_rep", False),
}
key = (args.model, args.select_test_dataset, bool(args.load_OC20_pt))
if key not in allowed:
raise ValueError(
"Invalid argument combination:\n"
f" --model {args.model}, --select-test-dataset {args.select_test_dataset}, "
f"--load_OC20_pt {bool(args.load_OC20_pt)}\n\n"
"Allowed combinations are:\n"
" • model=orig, select-test-dataset=without_rep, load_OC20_pt=False\n"
" • model=orig, select-test-dataset=without_rep, load_OC20_pt=True\n"
" • model=small, select-test-dataset=without_rep, load_OC20_pt=False\n"
" • model=small, select-test-dataset=with_rep, load_OC20_pt=False\n"
"\nNote: small + PT (load_OC20_pt=True) is not allowed."
)
if args.load_OC20_pt:
pt_path = os.path.join(args.data_root, "eq2_31M_ec4_allmd.pt")
if not os.path.isfile(pt_path):
raise FileNotFoundError(
f"Pretrained checkpoint not found:\n"
f" {pt_path}\n\n"
"Make sure eq2_31M_ec4_allmd.pt is inside the folder given by --data_root.\n"
)
|