def experiment(threshold,
iterations,
train_loss,
n_conv,
optimizer,
batch_size=1,
batch_norm=False,
learning_rate=1e-3,
summary_dir=None):
model = ConvNet(filters=4,
n_conv=n_conv,
train_loss=train_loss,
batch_norm=batch_norm,
optimizer=optimizer,
learning_rate=learning_rate,
summary_dir=summary_dir)
print('train_loss:', train_loss.value, 'optimizer:', optimizer.value,
'n_conv:', n_conv, 'batch_norm:', batch_norm, 'batch_size:',
batch_size, 'learning_rate:', learning_rate)
ret = dict()
val_input_batch, val_output_batch = get_data(threshold, 100, verbose=True)
best_accuracy = (0.0, 0)
for i in tqdm(range(iterations)):
input_batch, output_batch = get_data(threshold, batch_size)
out = model.train(input_batch, output_batch)
if i == 0:
for k in out.keys():
ret[k] = []
ret['accuracy'] = []
for k, v in out.items():
ret[k].append(v)
if i % 250 == 0:
accuracy = model.accuracy(val_input_batch, val_output_batch)
if accuracy > best_accuracy[0]:
best_accuracy = (accuracy, i)
ret['accuracy'].append((i, accuracy))
#print('[%d] accuracy: %.3g' % (i, accuracy))
print('Best accuracy %.3g at iteration %d.' % best_accuracy)
return ret
train_acc_list = []
test_acc_list = []
iters_num = 10000
batch_size = 100
train_size = x_train.shape[0]
iter_per_epoch = max(train_size / batch_size, 1)
net = ConvNet()
optim = SGD(net.params, lr=0.1, momentum=0.9)
# optim = AdaGrad(net.params)
# optim = Adam(net.params)
for i in range(iters_num):
batch_mask = np.random.choice(train_size, batch_size)
x_batch = x_train[batch_mask]
t_batch = t_train[batch_mask]
grad = net.gradient(x_batch, t_batch)
net.params = optim.update(net.params, grad)
loss = net.loss(x_batch, t_batch)
train_loss_list.append(loss)
if i % iter_per_epoch == 0:
train_acc = net.accuracy(x_train, t_train)
test_acc = net.accuracy(x_test, t_test)
train_acc_list.append(train_acc)
test_acc_list.append(test_acc)
print("train acc, test acc | " + str(train_acc) + ", " + str(test_acc))