def test_model_cpu(self):
model = M3EP(config)
for step, batch in enumerate(data_loader):
batch = batch_to_device(batch, torch.device('cpu'))
pred = model(batch)
output_size = config['late_fusion']['output_size']
self.assertEqual(pred.shape[1], output_size)
self.assertEqual(pred.device, torch.device('cpu'))
def val_acc(model, val_loader, device):
val_batch = next(iter(val_loader))
val_batch = batch_to_device(val_batch, device)
val_pred = model(val_batch)
y_pred = val_pred.cpu().detach().numpy()
y_pred = [np.argmax(p) for p in y_pred]
score = accuracy_score(val_batch[1], y_pred)
return score
def test_loss(self):
loss_fn = torch.nn.CrossEntropyLoss()
for step, batch in enumerate(data_loader):
model = M3EP(config)
pred = model(batch_to_device(batch, torch.device('cpu')))
targets = batch[1]
loss = loss_fn(pred, targets)
self.assertEqual(type(loss).__name__, 'Tensor')
def test_correct_device(self):
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
model = M3EP(config)
model = model.to(device)
for step, batch in enumerate(data_loader):
batch = batch_to_device(batch, device)
pred = model(batch)
self.assertEqual(pred.device.type, device.type)
def test_correct_device_cuda_call(self):
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
model = M3EP(config)
if torch.cuda.is_available():
model.cuda()
for step, batch in enumerate(data_loader):
batch = batch_to_device(batch, device)
pred = model(batch)
# output_size = config['hp']['output_size']
self.assertEqual(pred.device.type, device.type)
def test_parallel_data_loader_batch(self):
data_loader = DataLoader(dataset,
batch_size=50,
collate_fn=dict_pad_collate)
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
for step, batch in enumerate(data_loader):
model = M3EP(config)
if torch.cuda.is_available():
model.cuda()
batch = batch_to_device(batch, device)
pred = model(batch)
self.assertEqual(pred.device.type, device.type)
num_workers=config['data_loader']['num_workers'],
collate_fn=dict_pad_collate,
shuffle=True)
loss_fn = nn.CrossEntropyLoss()
optimizer = optim.Adam(params=model.parameters(),
lr=float(config['learning_rate']))
for epoch in range(config['epochs']):
print('Epoch {} of {}:'.format(str(epoch + 1), config['epochs']))
loss_msg = {}
loss_values = []
with tqdm(total=len(train_loader)) as progress_bar:
for step, batch in enumerate(train_loader):
batch = batch_to_device(batch, device)
labels = batch[1]
prediction = model(batch)
loss = loss_fn(prediction, labels)
# noinspection PyUnresolvedReferences
loss_msg['loss'] = loss.item()
progress_bar.postfix = loss_msg
progress_bar.update()
optimizer.zero_grad()
loss.backward()
optimizer.step()
if step % config['log_frequency'] == 0 and len(
loss_values) > 0:
writer.add_scalar('train_loss',