class AtomsTrainer:
def __init__(self, config):
self.config = config
self.pretrained = False
def load(self):
self.load_config()
self.load_rng_seed()
self.load_dataset()
self.load_model()
self.load_criterion()
self.load_optimizer()
self.load_logger()
self.load_extras()
self.load_skorch()
def load_config(self):
self.timestamp = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S")
self.identifier = self.config["cmd"].get("identifier", False)
if self.identifier:
self.identifier = self.timestamp + "-{}".format(self.identifier)
else:
self.identifier = self.timestamp
self.device = torch.device(self.config["optim"].get("device", "cpu"))
self.debug = self.config["cmd"].get("debug", False)
run_dir = self.config["cmd"].get("run_dir", "./")
os.chdir(run_dir)
if not self.debug:
self.cp_dir = os.path.join(run_dir, "checkpoints", self.identifier)
print(f"Results saved to {self.cp_dir}")
os.makedirs(self.cp_dir, exist_ok=True)
def load_rng_seed(self):
seed = self.config["cmd"].get("seed", 0)
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
def get_unique_elements(self, training_images):
elements = np.array(
[atom.symbol for atoms in training_images for atom in atoms])
elements = np.unique(elements)
return elements
def load_dataset(self):
training_images = self.config["dataset"]["raw_data"]
# TODO: Scalability when dataset to large to fit into memory
if isinstance(training_images, str):
training_images = ase.io.read(training_images, ":")
self.elements = self.config["dataset"].get(
"elements", self.get_unique_elements(training_images))
self.forcetraining = self.config["model"].get("get_forces", True)
self.fp_scheme = self.config["dataset"].get("fp_scheme",
"gaussian").lower()
self.fp_params = self.config["dataset"]["fp_params"]
self.cutoff_params = self.config["dataset"].get(
"cutoff_params", {"cutoff_func": "Cosine"})
self.train_dataset = AtomsDataset(
images=training_images,
descriptor_setup=(
self.fp_scheme,
self.fp_params,
self.cutoff_params,
self.elements,
),
forcetraining=self.forcetraining,
save_fps=self.config["dataset"].get("save_fps", True),
)
self.target_scaler = self.train_dataset.target_scaler
if not self.debug:
normalizers = {"target": self.target_scaler}
torch.save(normalizers, os.path.join(self.cp_dir,
"normalizers.pt"))
self.input_dim = self.train_dataset.input_dim
self.val_split = self.config["dataset"].get("val_split", 0)
print("Loading dataset: {} images".format(len(self.train_dataset)))
def load_model(self):
elements = list_symbols_to_indices(self.elements)
self.model = BPNN(elements=elements,
input_dim=self.input_dim,
**self.config["model"])
print("Loading model: {} parameters".format(self.model.num_params))
def load_extras(self):
callbacks = []
load_best_loss = train_end_load_best_loss(self.identifier)
self.split = CVSplit(cv=self.val_split) if self.val_split != 0 else 0
metrics = evaluator(
self.val_split,
self.config["optim"].get("metric", "mae"),
self.identifier,
self.forcetraining,
)
callbacks.extend(metrics)
if not self.debug:
callbacks.append(load_best_loss)
scheduler = self.config["optim"].get("scheduler", None)
if scheduler:
scheduler = LRScheduler(scheduler,
**self.config["optim"]["scheduler_params"])
callbacks.append(scheduler)
if self.config["cmd"].get("logger", False):
from skorch.callbacks import WandbLogger
callbacks.append(
WandbLogger(
self.wandb_run,
save_model=False,
keys_ignored="dur",
))
self.callbacks = callbacks
def load_criterion(self):
self.criterion = self.config["optim"].get("loss_fn", CustomLoss)
def load_optimizer(self):
self.optimizer = self.config["optim"].get("optimizer",
torch.optim.Adam)
def load_logger(self):
if self.config["cmd"].get("logger", False):
import wandb
self.wandb_run = wandb.init(
name=self.identifier,
config=self.config,
id=self.timestamp,
)
def load_skorch(self):
skorch.net.to_tensor = to_tensor
collate_fn = DataCollater(train=True, forcetraining=self.forcetraining)
self.net = NeuralNetRegressor(
module=self.model,
criterion=self.criterion,
criterion__force_coefficient=self.config["optim"].get(
"force_coefficient", 0),
criterion__loss=self.config["optim"].get("loss", "mse"),
optimizer=self.optimizer,
lr=self.config["optim"].get("lr", 1e-1),
batch_size=self.config["optim"].get("batch_size", 32),
max_epochs=self.config["optim"].get("epochs", 100),
iterator_train__collate_fn=collate_fn,
iterator_train__shuffle=True,
iterator_valid__collate_fn=collate_fn,
iterator_valid__shuffle=False,
device=self.device,
train_split=self.split,
callbacks=self.callbacks,
verbose=self.config["cmd"].get("verbose", True),
)
print("Loading skorch trainer")
def train(self, raw_data=None):
if raw_data is not None:
self.config["dataset"]["raw_data"] = raw_data
if not self.pretrained:
self.load()
self.net.fit(self.train_dataset, None)
def predict(self, images, batch_size=32):
if len(images) < 1:
warnings.warn("No images found!", stacklevel=2)
return images
a2d = AtomsToData(
descriptor=self.train_dataset.descriptor,
r_energy=False,
r_forces=False,
save_fps=True,
fprimes=self.forcetraining,
cores=1,
)
data_list = a2d.convert_all(images, disable_tqdm=True)
self.net.module.eval()
collate_fn = DataCollater(train=False,
forcetraining=self.forcetraining)
predictions = {"energy": [], "forces": []}
for data in data_list:
collated = collate_fn([data])
energy, forces = self.net.module(collated)
energy = self.target_scaler.denorm(
energy, pred="energy").detach().tolist()
forces = self.target_scaler.denorm(forces,
pred="forces").detach().numpy()
predictions["energy"].extend(energy)
predictions["forces"].append(forces)
return predictions
def load_pretrained(self, checkpoint_path=None):
print(f"Loading checkpoint from {checkpoint_path}")
self.load()
self.net.initialize()
self.pretrained = True
try:
self.net.load_params(
f_params=os.path.join(checkpoint_path, "params.pt"),
f_optimizer=os.path.join(checkpoint_path, "optimizer.pt"),
f_criterion=os.path.join(checkpoint_path, "criterion.pt"),
f_history=os.path.join(checkpoint_path, "history.json"),
)
# TODO(mshuaibi): remove dataset load, use saved normalizers
except NotImplementedError:
print("Unable to load checkpoint!")
class AtomsTrainer:
def __init__(self, config={}):
self.config = config
self.pretrained = False
def load(self, load_dataset=True):
self.load_config()
self.load_rng_seed()
if load_dataset:
self.load_dataset()
self.load_model()
self.load_criterion()
self.load_optimizer()
self.load_logger()
self.load_extras()
self.load_skorch()
def load_config(self):
dtype = self.config["cmd"].get("dtype", torch.FloatTensor)
torch.set_default_tensor_type(dtype)
self.timestamp = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S")
self.identifier = self.config["cmd"].get("identifier", False)
if self.identifier:
self.identifier = self.timestamp + "-{}".format(self.identifier)
else:
self.identifier = self.timestamp
self.gpus = self.config["optim"].get("gpus", 0)
if self.gpus > 0:
self.output_device = 0
self.device = f"cuda:{self.output_device}"
else:
self.device = "cpu"
self.output_device = -1
self.debug = self.config["cmd"].get("debug", False)
run_dir = self.config["cmd"].get("run_dir", "./")
os.chdir(run_dir)
if not self.debug:
self.cp_dir = os.path.join(run_dir, "checkpoints", self.identifier)
print(f"Results saved to {self.cp_dir}")
os.makedirs(self.cp_dir, exist_ok=True)
def load_rng_seed(self):
seed = self.config["cmd"].get("seed", 0)
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
def get_unique_elements(self, training_images):
elements = np.array(
[atom.symbol for atoms in training_images for atom in atoms]
)
elements = np.unique(elements)
return elements
def load_dataset(self):
training_images = self.config["dataset"]["raw_data"]
# TODO: Scalability when dataset to large to fit into memory
if isinstance(training_images, str):
training_images = ase.io.read(training_images, ":")
del self.config["dataset"]["raw_data"]
self.elements = self.config["dataset"].get(
"elements", self.get_unique_elements(training_images)
)
self.forcetraining = self.config["model"].get("get_forces", True)
self.fp_scheme = self.config["dataset"].get("fp_scheme", "gaussian").lower()
self.fp_params = self.config["dataset"]["fp_params"]
self.save_fps = self.config["dataset"].get("save_fps", True)
self.cutoff_params = self.config["dataset"].get(
"cutoff_params", {"cutoff_func": "Cosine"}
)
descriptor_setup = (
self.fp_scheme,
self.fp_params,
self.cutoff_params,
self.elements,
)
self.train_dataset = AtomsDataset(
images=training_images,
descriptor_setup=descriptor_setup,
forcetraining=self.forcetraining,
save_fps=self.config["dataset"].get("save_fps", True),
scaling=self.config["dataset"].get(
"scaling", {"type": "normalize", "range": (0, 1)}
),
)
self.feature_scaler = self.train_dataset.feature_scaler
self.target_scaler = self.train_dataset.target_scaler
self.input_dim = self.train_dataset.input_dim
self.val_split = self.config["dataset"].get("val_split", 0)
self.config["dataset"]["descriptor"] = descriptor_setup
if not self.debug:
normalizers = {
"target": self.target_scaler,
"feature": self.feature_scaler,
}
torch.save(normalizers, os.path.join(self.cp_dir, "normalizers.pt"))
# clean/organize config
self.config["dataset"]["fp_length"] = self.input_dim
torch.save(self.config, os.path.join(self.cp_dir, "config.pt"))
print("Loading dataset: {} images".format(len(self.train_dataset)))
def load_model(self):
elements = list_symbols_to_indices(self.elements)
self.model = BPNN(
elements=elements, input_dim=self.input_dim, **self.config["model"]
)
print("Loading model: {} parameters".format(self.model.num_params))
self.forcetraining = self.config["model"].get("get_forces", True)
collate_fn = DataCollater(train=True, forcetraining=self.forcetraining)
self.parallel_collater = ParallelCollater(self.gpus, collate_fn)
if self.gpus > 0:
self.model = DataParallel(
self.model,
output_device=self.output_device,
num_gpus=self.gpus,
)
def load_extras(self):
callbacks = []
load_best_loss = train_end_load_best_loss(self.identifier)
self.val_split = self.config["dataset"].get("val_split", 0)
self.split = CVSplit(cv=self.val_split) if self.val_split != 0 else 0
metrics = evaluator(
self.val_split,
self.config["optim"].get("metric", "mae"),
self.identifier,
self.forcetraining,
)
callbacks.extend(metrics)
if not self.debug:
callbacks.append(load_best_loss)
scheduler = self.config["optim"].get("scheduler", None)
if scheduler:
scheduler = LRScheduler(scheduler["policy"], **scheduler["params"])
callbacks.append(scheduler)
if self.config["cmd"].get("logger", False):
from skorch.callbacks import WandbLogger
callbacks.append(
WandbLogger(
self.wandb_run,
save_model=False,
keys_ignored="dur",
)
)
self.callbacks = callbacks
def load_criterion(self):
self.criterion = self.config["optim"].get("loss_fn", CustomLoss)
def load_optimizer(self):
self.optimizer = {
"optimizer": self.config["optim"].get("optimizer", torch.optim.Adam)
}
optimizer_args = self.config["optim"].get("optimizer_args", False)
if optimizer_args:
self.optimizer.update(optimizer_args)
def load_logger(self):
if self.config["cmd"].get("logger", False):
import wandb
self.wandb_run = wandb.init(
name=self.identifier,
config=self.config,
)
def load_skorch(self):
skorch.net.to_tensor = to_tensor
self.net = NeuralNetRegressor(
module=self.model,
criterion=self.criterion,
criterion__force_coefficient=self.config["optim"].get(
"force_coefficient", 0
),
criterion__loss=self.config["optim"].get("loss", "mse"),
lr=self.config["optim"].get("lr", 1e-1),
batch_size=self.config["optim"].get("batch_size", 32),
max_epochs=self.config["optim"].get("epochs", 100),
iterator_train__collate_fn=self.parallel_collater,
iterator_train__shuffle=True,
iterator_train__pin_memory=True,
iterator_valid__collate_fn=self.parallel_collater,
iterator_valid__shuffle=False,
iterator_valid__pin_memory=True,
device=self.device,
train_split=self.split,
callbacks=self.callbacks,
verbose=self.config["cmd"].get("verbose", True),
**self.optimizer,
)
print("Loading skorch trainer")
def train(self, raw_data=None):
if raw_data is not None:
self.config["dataset"]["raw_data"] = raw_data
if not self.pretrained:
self.load()
stime = time.time()
self.net.fit(self.train_dataset, None)
elapsed_time = time.time() - stime
print(f"Training completed in {elapsed_time}s")
def predict(self, images, disable_tqdm=True):
if len(images) < 1:
warnings.warn("No images found!", stacklevel=2)
return images
self.descriptor = construct_descriptor(self.config["dataset"]["descriptor"])
a2d = AtomsToData(
descriptor=self.descriptor,
r_energy=False,
r_forces=False,
save_fps=self.config["dataset"].get("save_fps", True),
fprimes=self.forcetraining,
cores=1,
)
data_list = a2d.convert_all(images, disable_tqdm=disable_tqdm)
self.feature_scaler.norm(data_list, disable_tqdm=disable_tqdm)
self.net.module.eval()
collate_fn = DataCollater(train=False, forcetraining=self.forcetraining)
predictions = {"energy": [], "forces": []}
for data in data_list:
collated = collate_fn([data]).to(self.device)
energy, forces = self.net.module([collated])
energy = self.target_scaler.denorm(
energy.detach().cpu(), pred="energy"
).tolist()
forces = self.target_scaler.denorm(
forces.detach().cpu(), pred="forces"
).numpy()
predictions["energy"].extend(energy)
predictions["forces"].append(forces)
return predictions
def load_pretrained(self, checkpoint_path=None, gpu2cpu=False):
"""
Args:
checkpoint_path: str, Path to checkpoint directory
gpu2cpu: bool, True if checkpoint was trained with GPUs and you
wish to load on cpu instead.
"""
self.pretrained = True
print(f"Loading checkpoint from {checkpoint_path}")
assert os.path.isdir(
checkpoint_path
), f"Checkpoint: {checkpoint_path} not found!"
if not self.config:
# prediction only
self.config = torch.load(os.path.join(checkpoint_path, "config.pt"))
self.config["cmd"]["debug"] = True
self.elements = self.config["dataset"]["descriptor"][-1]
self.input_dim = self.config["dataset"]["fp_length"]
if gpu2cpu:
self.config["optim"]["gpus"] = 0
self.load(load_dataset=False)
else:
# prediction+retraining
self.load(load_dataset=True)
self.net.initialize()
if gpu2cpu:
params_path = os.path.join(checkpoint_path, "params_cpu.pt")
if not os.path.exists(params_path):
params = torch.load(
os.path.join(checkpoint_path, "params.pt"),
map_location=torch.device("cpu"),
)
new_dict = OrderedDict()
for k, v in params.items():
name = k[7:]
new_dict[name] = v
torch.save(new_dict, params_path)
else:
params_path = os.path.join(checkpoint_path, "params.pt")
try:
self.net.load_params(
f_params=params_path,
f_optimizer=os.path.join(checkpoint_path, "optimizer.pt"),
f_criterion=os.path.join(checkpoint_path, "criterion.pt"),
f_history=os.path.join(checkpoint_path, "history.json"),
)
normalizers = torch.load(os.path.join(checkpoint_path, "normalizers.pt"))
self.feature_scaler = normalizers["feature"]
self.target_scaler = normalizers["target"]
except NotImplementedError:
print("Unable to load checkpoint!")
def get_calc(self):
return AMPtorch(self)
cp = Checkpoint(dirname='segnet_mse_no_sigmoid_sgd_150ep_b8_lr_0.01_30enc/checkpoints')
train_end_cp = TrainEndCheckpoint(dirname='segnet_mse_no_sigmoid_sgd_150ep_b8_lr_0.01_30enc/checkpoints')
net = NeuralNetRegressor(
SegNet,
module__encoding_size=30,
device=torch.device("cuda:0" if torch.cuda.is_available() else "cpu"),
max_epochs=150,
batch_size=8,
criterion=MSELoss,
lr=0.01,
iterator_train__shuffle=True,
optimizer=torch.optim.SGD,
optimizer__momentum=.9,
callbacks=[cp, train_end_cp]
)
net.initialize()
net.load_params(checkpoint=cp)
mean = np.array([0.5020, 0.4690, 0.4199])
std = np.array([0.2052, 0.2005, 0.1966])
inverse_transform = transforms.Normalize(
mean=(-mean) / std,
std=1 / std
)
transform = transforms.Compose([transforms.Resize((224, 224),
interpolation=3),
transforms.ToTensor(),
transforms.Normalize(mean, std),
])
for i in range(10):
img = transform(Image.open(f"../cropped_data/000{i}.png").convert('RGB'))
input_image = (img * torch.tensor(std).view(3, 1, 1) + torch.tensor(mean).view(3, 1, 1)).numpy().transpose(1, 2,
#out = self.dp(out)
out = self.l1(out)
out = self.fc2(out)
return out
net_regr = NeuralNetRegressor(
Net(hidden_size=500),
max_epochs=200,
lr=0.003,
device='cuda',
optimizer=torch.optim.Adam,
train_split=None,
verbose=1,
)
net_regr.initialize()
net_regr.load_params('step1result')
res = net_regr.fit(t_d_inp, t_d_oup)
# save
net_regr.save_params(f_params='step2result')
pred = net_regr.predict(test_inp)
mse = ((test_oup - pred)**2).mean()
print('test error(NN) = ' + str(mse))
mse2 = ((test_oup - predicted_gp3)**2).mean()
print('test error(GP) = ' + str(mse2))
# plot 1 loss
loss = net_regr.history[:, 'train_loss']
plt.figure()
plt.plot(loss)
plt.ylabel('loss')