def test_dataloader(self):
# OPTIONAL
test_dir = os.path.join(self.hparams.data, 'test')
split_idx = 0
raw_data = read_smiles_multiclass('%s/raw.csv' % self.hparams.data)
data_splits = read_splits('%s/split_%d.txt' %
(self.hparams.data, split_idx))
test_dataset = get_loader(raw_data,
data_splits['test'],
self.hparams,
shuffle=True) #TODO
if self.use_ddp:
test_sampler = DistributedSampler(test_dataset)
else:
test_sampler = None
test_loader = DataLoader(dataset=test_dataset,
batch_size=self.hparams.batch_size,
shuffle=(test_sampler is None),
num_workers=0,
sampler=test_sampler)
return test_loader
def train_dataloader(self):
# REQUIRED
train_dir = os.path.join(self.hparams.data, 'train')
split_idx = 0
raw_data = read_smiles_multiclass('%s/raw.csv' % self.hparams.data)
data_splits = read_splits('%s/split_%d.txt' %
(self.hparams.data, split_idx))
train_dataset = get_loader(raw_data,
data_splits['train'],
self.hparams,
shuffle=True)
if self.use_ddp:
train_sampler = DistributedSampler(train_dataset)
else:
train_sampler = None
"""
train_loader = DataLoader(dataset=train_dataset,
batch_size=self.hparams.batch_size,
shuffle=(train_sampler is None),
num_workers=0,
sampler=train_sampler)
"""
train_loader = get_loader(raw_data,
data_splits['train'],
self.hparams,
shuffle=True)
return train_loader
def test_dataloader(self):
# OPTIONAL
test_dir = os.path.join(self.hparams.data, 'test')
split_idx = 0
if self.hparams.multi:
raw_data = read_smiles_multiclass('%s/raw.csv' % self.hparams.data)
n_classes = len(raw_data[0][1])
else:
raw_data = read_smiles_from_file('%s/raw.csv' % self.hparams.data)
n_classes = 1
data_splits = read_splits('%s/split_%d.txt' %
(self.hparams.data, split_idx))
test_dataset = get_loader(raw_data,
data_splits['test'],
self.hparams,
num_workers=0,
shuffle=False) #TODO"""
if self.use_ddp:
test_sampler = DistributedSampler(test_dataset)
else:
test_sampler = None
test_loader = get_loader(raw_data,
data_splits['test'],
self.hparams,
num_workers=0,
shuffle=(test_sampler is None),
sampler=test_sampler)
return test_loader
def val_dataloader(self):
# OPTIONAL
val_dir = os.path.join(self.hparams.data, 'valid')
split_idx = 0
if self.hparams.multi:
raw_data = read_smiles_multiclass('%s/raw.csv' % self.hparams.data)
n_classes = len(raw_data[0][1])
else:
raw_data = read_smiles_from_file('%s/raw.csv' % self.hparams.data)
n_classes = 1
data_splits = read_splits('%s/split_%d.txt' %
(self.hparams.data, split_idx))
val_dataset = get_loader(raw_data,
data_splits['valid'],
self.hparams,
shuffle=True,
batch_size=self.hparams.batch_size,
num_workers=self.hparams.n_workers) #TODO
if self.use_ddp:
val_sampler = DistributedSampler(val_dataset)
else:
val_sampler = None
val_loader = get_loader(raw_data,
data_splits['valid'],
self.hparams,
shuffle=False,
sampler=val_sampler)
return val_loader
def main(hparams):
# init module
raw_data = read_smiles_multiclass('%s/raw.csv' % hparams.data)
n_classes = len(raw_data[0][1])
model = PropPredictor(hparams, n_classes=n_classes)
model.version = hparams.dataset
load_shortest_paths(hparams)
#model.half()
comet_logger = CometLogger(
api_key=os.environ["COMET_KEY"],
experiment_name=f'{hparams.dataset}-{str(time.time())}',
log_graph=True,
project_name="tcvaemolgen",
workspace=os.environ["COMET_WKSP"]
)
# most basic trainer, uses good defaults
trainer = Trainer(
check_val_every_n_epoch=1,
default_save_path=f'data/05_model_outputs/{hparams.dataset}',
distributed_backend=hparams.distributed_backend,
max_nb_epochs=hparams.max_nb_epochs,
early_stop_callback=None,
gpus=hparams.gpus,
gradient_clip_val=10,
nb_gpu_nodes=hparams.nodes,
logger=comet_logger,
log_save_interval=100,
row_log_interval=10,
show_progress_bar=True,
track_grad_norm=2
)
for round_idx in range(hparams.n_rounds):
model.split_idx = round_idx
log.info(f'Split {round_idx}')
trainer.fit(model)
trainer.test()
return
def train_dataloader(self):
split_idx = self.split_idx
# REQUIRED
train_dir = os.path.join(self.hparams.data, 'train')
if self.hparams.multi:
raw_data = read_smiles_multiclass('%s/raw.csv' % self.hparams.data)
n_classes = len(raw_data[0][1])
print(f'N_Classes: {n_classes}')
else:
raw_data = read_smiles_from_file('%s/raw.csv' % self.hparams.data)
n_classes = 1
data_splits = read_splits('%s/split_%d.txt' %
(self.hparams.data, split_idx))
train_dataset = get_loader(raw_data,
data_splits['train'],
self.hparams,
shuffle=True)
if self.use_ddp:
train_sampler = DistributedSampler(train_dataset)
else:
train_sampler = None
"""
train_loader = DataLoader(dataset=train_dataset,
batch_size=self.hparams.batch_size,
shuffle=(train_sampler is None),
num_workers=0,
sampler=train_sampler)
"""
train_loader = get_loader(raw_data,
data_splits['train'],
self.hparams,
shuffle=True,
sampler=train_sampler,
batch_size=self.hparams.batch_size,
num_workers=self.hparams.n_workers)
return train_loader