n_epoch = 10
batch_size = 1024
##############################
#### Training loop
loss_history = []
acc_history = []
for i in range(n_epoch):
# print(f"EPOCH: {i}")
for x_batch, y_batch in get_batches(train_x, train_y, batch_size):
net.zeroGradParameters()
# Forward
predictions = net.forward(x_batch)
loss = criterion.forward(predictions, y_batch)
# Backward
dp = criterion.backward(predictions, y_batch)
net.backward(x_batch, dp)
# Update weights
adam_optimizer(net.getParameters(), net.getGradParameters(),
optimizer_config, optimizer_state)
loss_history.append(loss)
acc = 100 * accuracy_score(predictions.argmax(axis=-1),
y_batch.argmax(axis=-1))
acc_history.append(acc)
n_epoch = 20
batch_size = 128
##############################
#### Training loop
loss_history = []
acc_history = []
for i in range(n_epoch):
print(f"EPOCH: {i}")
for x_batch, y_batch in get_batches(X, Y, batch_size):
net.zeroGradParameters()
# Forward
predictions = net.forward(x_batch)
loss = criterion.forward(predictions, y_batch)
# Backward
dp = criterion.backward(predictions, y_batch)
net.backward(x_batch, dp)
# Update weights
sgd_momentum(net.getParameters(), net.getGradParameters(),
optimizer_config, optimizer_state)
loss_history.append(loss)
acc_history.append(100 * accuracy_score(predictions.argmax(axis=-1),
y_batch.argmax(axis=-1)))
################################
# test_label = test_label.add(0.125).mul(0.8)
# ----- Implementation of the architecture -----
architecture = Sequential(Linear(2, 25, ReLU()), Linear(25, 25, ReLU()),
Linear(25, 25, ReLU()), Linear(25, 2, Sigma()))
# ----- Training -----
round = 1
prev_loss = math.inf
prev_prev_loss = math.inf
errors = []
for epoch in range(epochs):
for batch_start in range(0, nb_samples, batch_size):
features = train_features[batch_start:batch_start + batch_size, :]
labels = train_labels[batch_start:batch_start + batch_size]
tr_loss, tr_error = architecture.forward(train_features, train_labels)
architecture.backward()
architecture.update(eta)
loss, error = architecture.forward(test_features, test_labels)
print(' --- Epoch ', round, ' Test Loss: ', loss.item(), '---',
' Errors: ', error)
if prev_loss < tr_loss:
eta *= 0.999
if prev_prev_loss < tr_loss:
eta *= 0.998
errors.append(error)
prev_prev_loss = prev_loss
prev_loss = tr_loss
round += 1
plotCostAndData(nn_test, inp_train, target_train, cost_nn_seq)
# In[243]:
# Computing the test error
def compute_nb_errors(model, data_input, data_target):
mini_batch_size = 10
nb_data_errors = 0
for b in range(0, data_input.shape[0], mini_batch_size):
input_batch = data_input[b:b + mini_batch_size]
predicted_classes = model.forward(input_batch)
predicted_classes = predicted_classes > 0.5
for k in range(mini_batch_size):
if data_target[b + k, 0] != predicted_classes[k, 0]:
nb_data_errors += 1
return nb_data_errors
# In[278]:
nb_test_err = compute_nb_errors(nn_test, inp_test, target_test)
print('Number of errors on the test set : ', nb_test_err)
# In[291]:
prediction = nn_test.forward(inp_train)
plt.scatter(inp_train[:, 0], inp_test[:, 1], c=prediction[:, 0], s=1)
plt.title('Prediction on the test set')