def fit(model):
optimizer = SGD(model.parameters(), model.grads(), lr=0.1)
losses = []
print('Epoch | Loss')
for epoch in range(500):
epoch_loss = 0
for b in range(num_batches):
batch_input = train_input[b * batch_size:(b + 1) * batch_size]
batch_target = train_target[b * batch_size:(b + 1) * batch_size]
batch_output = model(batch_input)
batch_loss = criterion(batch_output, batch_target)
epoch_loss += batch_loss
output_grad = criterion.backward()
model.backward(output_grad)
optimizer.step()
optimizer.zero_grad()
losses.append(epoch_loss.item() / num_batches)
print(f'{epoch+1:>5} | {epoch_loss.item() / num_batches:.5f}')
train_output = model(train_input)
print(
f'\nTrain Error: {sum(train_output.argmax(1) != train_target.argmax(1)).item() / 1000}'
)
test_output = model(test_input)
print(
f'Test Error: {sum(test_output.argmax(1) != test_target.argmax(1)).item() / 1000}'
)
return losses
def our_fit(model, epochs=500, verbose=False):
criterion = MSE()
optimizer = SGD(model.parameters(), model.grads(), lr=0.1)
start = time()
losses = []
if verbose: print('Epoch | Loss')
for epoch in range(epochs):
epoch_loss = 0
for b in range(num_batches):
batch_input = train_input[b * batch_size:(b + 1) * batch_size]
batch_target = train_target[b * batch_size:(b + 1) * batch_size]
batch_output = model(batch_input)
batch_loss = criterion(batch_output, batch_target)
epoch_loss += batch_loss
output_grad = criterion.backward()
model.backward(output_grad)
optimizer.step()
optimizer.zero_grad()
losses.append(epoch_loss.item() / num_batches)
if verbose:
print(f'{epoch+1:>5} | {epoch_loss.item() / num_batches:.5f}')
end = time()
train_output = model(train_input)
print(
f'\nTrain Error: {sum(train_output.argmax(1) != train_target.argmax(1)).item() / len(train_output)}'
)
test_output = model(test_input)
print(
f'Test Error: {sum(test_output.argmax(1) != test_target.argmax(1)).item() / len(test_output)}'
)
return losses, end - start
avg_loss = avg_acc = 0
for e in range(epoch):
if e and e % 3 == 0:
optimizer.lr *= 0.1
train_loss = train_acc = 0
e_data, e_labels = shuffle(train_data, train_labels)
with tqdm(total=epoch_steps) as pbar:
for x, t in zip(np.array_split(e_data, epoch_steps),
np.array_split(e_labels, epoch_steps)):
x = Tensor(x[:, None])
t = Tensor(t)
optimizer.zero_grad()
logits = my_net(x)
loss, grad = cross_entropy_loss(logits, t)
acc = accuracy_score(t, logits.argmax(1))
logits.backward(grad)
optimizer.step()
loss_history.append(loss)
train_loss += loss
train_acc += acc
if not avg_loss:
avg_loss = loss
else:
avg_loss *= 0.98