def J(theta):
weights = pack_struct(theta, self.layer_units)
loss, grad = neural_net_loss(weights, X, y, reg)
grad = flatten_struct(grad)
return loss, grad
def progress(x):
nonlocal iter_feval, best_weights, best_val_acc
iter_feval += 1
# Loss history
weights = pack_struct(x, self.layer_units)
loss, grad = neural_net_loss(weights, X, y, reg)
loss_history.append(loss)
# Training accurary
y_pred_train = neural_net_predict(weights, X)
train_acc = np.mean(y_pred_train == y)
train_acc_history.append(train_acc)
# Validation accuracy
y_pred_val= neural_net_predict(weights, X_val)
val_acc = np.mean(y_pred_val == y_val)
val_acc_history.append(val_acc)
# Keep track of the best weights based on validation accuracy
if val_acc > best_val_acc:
best_val_acc = val_acc
n_weights = len(weights)
best_weights = [{} for i in range(n_weights)]
for i in range(n_weights):
for p in weights[i]:
best_weights[i][p] = weights[i][p].copy()
n_iters_verbose = max_iters / 20
if iter_feval % n_iters_verbose == 0:
print("iter: {:4d}, loss: {:8f}, train_acc: {:4f}, val_acc: {:4f}".format(iter_feval, loss, train_acc, val_acc))
def progress(x):
nonlocal iter_feval, best_weights, best_val_acc
iter_feval += 1
# Loss history
weights = pack_struct(x, self.layer_units)
loss, grad = mlp_loss(weights, X, y, reg)
loss_history.append(loss)
# Training accurary
y_pred_train = mlp_predict(weights, X)
train_acc = np.mean(y_pred_train == y)
train_acc_history.append(train_acc)
# Validation accuracy
y_pred_val = mlp_predict(weights, X_val)
val_acc = np.mean(y_pred_val == y_val)
val_acc_history.append(val_acc)
# Keep track of the best weights based on validation accuracy
if val_acc > best_val_acc:
best_val_acc = val_acc
n_weights = len(weights)
best_weights = [{} for i in range(n_weights)]
for i in range(n_weights):
for p in weights[i]:
best_weights[i][p] = weights[i][p].copy()
n_iters_verbose = max_iters / 20
if iter_feval % n_iters_verbose == 0:
print(
"iter: {:4d}, loss: {:8f}, train_acc: {:4f}, val_acc: {:4f}"
.format(iter_feval, loss, train_acc, val_acc))
def J(theta):
weights = pack_struct(theta, self.layer_units)
loss, grad = mlp_loss(weights, X, y, reg)
grad = flatten_struct(grad)
return loss, grad
def J(theta):
weights = pack_struct(theta, self.layer_units)
loss, grad = sparse_autoencoder_loss(weights,
X,
reg,
beta=beta,
sparsity_param=sparsity_param)
grad = flatten_struct(grad)
return loss, grad
def test_pack_struct():
layer_units = (4, 8, 4)
n_layers = len(layer_units)
num = 0
for i in range(n_layers - 1):
num += layer_units[i + 1] * layer_units[i] + layer_units[i + 1]
weights = np.random.randn(num)
actual = flatten_struct(pack_struct(weights, layer_units))
desired = weights
assert_allclose(actual, desired, atol=1e-8)
def test_pack_struct():
layer_units = (4, 8, 4)
n_layers = len(layer_units)
num = 0
for i in range(n_layers - 1):
num += layer_units[i+1]*layer_units[i] + layer_units[i+1]
weights = np.random.randn(num)
actual = flatten_struct(pack_struct(weights, layer_units))
desired = weights
assert_allclose(actual, desired, atol=1e-8)
def pack_struct(self, data):
return pack_struct(data, self.layer_units)
def pack_struct(self, data):
return pack_struct(data, self.layer_units)
