def resolve_uri_locally(self, artifact_uri: Text, path: Text = None):
"""
Takes a URI that points within the artifact store, downloads the
URI locally, then returns local URI.
Args:
artifact_uri: uri to artifact.
path: optional path to download to. If None, is inferred.
"""
if not path_utils.is_remote(artifact_uri):
# Its already local
return artifact_uri
if path is None:
path = os.path.join(
GlobalConfig.get_config_dir(),
self.unique_id,
ArtifactStore.get_component_name_from_uri(artifact_uri),
path_utils.get_parent(artifact_uri) # unique ID from MLMD
)
# Create if not exists and download
path_utils.create_dir_recursive_if_not_exists(path)
path_utils.copy_dir(artifact_uri, path, overwrite=True)
return path
def run_fn(self):
train_dataset = self.input_fn(self.train_files,
self.tf_transform_output)
eval_dataset = self.input_fn(self.eval_files, self.tf_transform_output)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = self.model_fn(train_dataset, eval_dataset)
model.to(device)
criterion = nn.BCEWithLogitsLoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)
model.train()
for e in range(1, self.epoch + 1):
epoch_loss = 0
epoch_acc = 0
step_count = 0
for x, y in train_dataset:
step_count += 1
X_batch, y_batch = x.to(device), y.to(device)
optimizer.zero_grad()
y_pred = model(X_batch)
loss = criterion(y_pred, y_batch)
acc = binary_acc(y_pred, y_batch)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
epoch_acc += acc.item()
print(f'Epoch {e + 0:03}: | Loss: '
f'{epoch_loss / step_count:.5f} | Acc: '
f'{epoch_acc / step_count:.3f}')
path_utils.create_dir_if_not_exists(self.serving_model_dir)
if path_utils.is_remote(self.serving_model_dir):
temp_model_dir = '__temp_model_dir__'
temp_path = os.path.join(os.getcwd(), temp_model_dir)
if path_utils.is_dir(temp_path):
raise PermissionError('{} is used as a temp path but it '
'already exists. Please remove it to '
'continue.')
torch.save(model, temp_path)
path_utils.copy_dir(temp_path, self.serving_model_dir)
path_utils.rm_dir(temp_path)
else:
torch.save(model, os.path.join(self.serving_model_dir, 'model.pt'))
def run_fn(self):
train_dataset = self.input_fn(self.train_files,
self.tf_transform_output)
eval_dataset = self.input_fn(self.eval_files, self.tf_transform_output)
class LitModel(pl.LightningModule):
def __init__(self):
super().__init__()
self.l1 = torch.nn.Linear(8, 64)
self.layer_out = torch.nn.Linear(64, 1)
def forward(self, x):
x = torch.relu(self.l1(x))
x = self.layer_out(x)
return x
def training_step(self, batch, batch_idx):
x, y = batch
y_hat = self(x)
loss = F.binary_cross_entropy_with_logits(y_hat, y)
tensorboard_logs = {'train_loss': loss}
return {'loss': loss, 'log': tensorboard_logs}
def configure_optimizers(self):
return torch.optim.Adam(self.parameters(), lr=0.001)
def train_dataloader(self):
return train_dataset
def validation_step(self, batch, batch_idx):
x, y = batch
y_hat = self(x)
return {
'val_loss': F.binary_cross_entropy_with_logits(y_hat, y)
}
def validation_epoch_end(self, outputs):
avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
tensorboard_logs = {'val_loss': avg_loss}
return {'avg_val_loss': avg_loss, 'log': tensorboard_logs}
def val_dataloader(self):
return eval_dataset
model = LitModel()
# most basic trainer, uses good defaults
trainer = Trainer(
default_root_dir=self.log_dir,
max_epochs=self.epoch,
)
trainer.fit(model)
path_utils.create_dir_if_not_exists(self.serving_model_dir)
if path_utils.is_remote(self.serving_model_dir):
temp_model_dir = '__temp_model_dir__'
temp_path = os.path.join(os.getcwd(), temp_model_dir)
if path_utils.is_dir(temp_path):
raise PermissionError('{} is used as a temp path but it '
'already exists. Please remove it to '
'continue.')
trainer.save_checkpoint(os.path.join(temp_path, 'model.cpkt'))
path_utils.copy_dir(temp_path, self.serving_model_dir)
path_utils.rm_dir(temp_path)
else:
trainer.save_checkpoint(
os.path.join(self.serving_model_dir, 'model.ckpt'))
def run_fn(self):
train_dataset = self.input_fn(self.train_files,
self.tf_transform_output)
eval_dataset = self.input_fn(self.eval_files,
self.tf_transform_output)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = self.model_fn(train_dataset, eval_dataset)
model.to(device)
criterion = nn.BCEWithLogitsLoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)
writer = SummaryWriter(self.log_dir)
model.train()
total_count = 0
for e in range(1, self.epochs + 1):
epoch_loss = 0
epoch_acc = 0
step_count = 0
for x, y, _ in train_dataset:
step_count += 1
total_count += 1
x_batch = torch.cat([v.to(device) for v in x.values()], dim=-1)
y_batch = torch.cat([v.to(device) for v in y.values()], dim=-1)
optimizer.zero_grad()
y_pred = model(x_batch)
loss = criterion(y_pred, y_batch)
acc = binary_acc(y_pred, y_batch)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
epoch_acc += acc.item()
if e == 1 and step_count == 1:
writer.add_graph(model, x_batch)
writer.add_scalar('training_loss', loss, total_count)
writer.add_scalar('training_accuracy', acc, total_count)
print(f'Epoch {e + 0:03}: | Loss: '
f'{epoch_loss / step_count:.5f} | Acc: '
f'{epoch_acc / step_count:.3f}')
# test
test_results = self.test_fn(model, eval_dataset)
utils.save_test_results(test_results, self.test_results)
path_utils.create_dir_if_not_exists(self.serving_model_dir)
if path_utils.is_remote(self.serving_model_dir):
temp_model_dir = '__temp_model_dir__'
temp_path = os.path.join(os.getcwd(), temp_model_dir)
if path_utils.is_dir(temp_path):
raise PermissionError('{} is used as a temp path but it '
'already exists. Please remove it to '
'continue.')
torch.save(model, temp_path)
path_utils.copy_dir(temp_path, self.serving_model_dir)
path_utils.rm_dir(temp_path)
else:
torch.save(model, os.path.join(self.serving_model_dir, 'model.pt'))