x = tfg.Placeholder(name="x")
target = tfg.Placeholder(name="target")
n = tfn.Multi_Layer_Network(
input_size=input_size,
hidden_size_list=[hidden_layer1_size, hidden_layer2_size],
output_size=output_size,
input_node=x,
target_node=target,
init_mean=0.0,
init_sd=0.01,
activator=tfe.Activator.ReLU.value,
optimizer=tfe.Optimizer.Adam.value,
learning_rate=0.01,
model_params_dir=model_params_dir
)
#n.draw_and_show()
data = mnist.MNIST_Data()
forward_final_output = n.feed_forward(input_data=data.test_input, is_numba=False)
print(forward_final_output.shape)
print(tff.accuracy(forward_final_output, data.test_target))
batch_size = 1000
n.learning(max_epoch=100, data=data, batch_size=batch_size, print_period=1, is_numba=False, verbose=False)
forward_final_output = n.feed_forward(input_data=data.test_input, is_numba=False)
print(tff.accuracy(forward_final_output, data.test_target))
def set_learning_process_specification(self, data, batch_size, epoch,
print_period, is_numba, fold_idx,
max_epoch, verbose):
batch_mask = np.random.choice(data.num_train_data, batch_size)
i_batch = data.train_input[batch_mask]
t_batch = data.train_target[batch_mask]
train_error = self.session.run(self.error, {
self.input_node: i_batch,
self.target_node: t_batch
},
is_numba=is_numba,
verbose=False)
self.train_error_list.append(train_error)
validation_error = self.session.run(
self.error, {
self.input_node: data.validation_input,
self.target_node: data.validation_target
},
is_numba=is_numba,
verbose=False)
self.validation_error_list.append(validation_error)
min_flag = False
if validation_error < self.min_validation_error:
self.min_validation_error = validation_error
self.min_validation_error_epoch = epoch + fold_idx * max_epoch
self.min_fold_idx = fold_idx
self.save_params()
min_flag = True
forward_final_output = self.feed_forward(input_data=data.test_input,
is_numba=is_numba)
test_accuracy = tff.accuracy(forward_final_output, data.test_target)
self.test_accuracy_list.append(test_accuracy)
for idx in range(self.hidden_layer_num + 1):
d = self.get_param_describe(layer_num=idx, kind="W")
self.param_mean_list['W'][idx].append(d.mean)
self.param_variance_list['W'][idx].append(d.variance)
self.param_skewness_list['W'][idx].append(d.skewness)
self.param_kurtosis_list['W'][idx].append(d.kurtosis)
d = self.get_param_describe(layer_num=idx, kind="b")
self.param_mean_list['b'][idx].append(d.mean)
self.param_variance_list['b'][idx].append(d.variance)
self.param_skewness_list['b'][idx].append(d.skewness)
self.param_kurtosis_list['b'][idx].append(d.kurtosis)
d = self.get_activation_describe(layer_num=idx, kind="affine")
self.output_mean_list['affine'][idx].append(d.mean)
self.output_variance_list['affine'][idx].append(d.variance)
self.output_skewness_list['affine'][idx].append(d.skewness)
self.output_kurtosis_list['affine'][idx].append(d.kurtosis)
if idx != self.hidden_layer_num:
d = self.get_activation_describe(layer_num=idx,
kind="activation")
self.output_mean_list['activation'][idx].append(d.mean)
self.output_variance_list['activation'][idx].append(d.variance)
self.output_skewness_list['activation'][idx].append(d.skewness)
self.output_kurtosis_list['activation'][idx].append(d.kurtosis)
if epoch % print_period == 0:
print(
"Epoch {:3d} Completed - Train Error:{:6.5f} - Validation Error:{:6.5f} - Test Accuracy:{:6.5f}"
.format(epoch, float(train_error), float(validation_error),
float(test_accuracy)),
end="")
if min_flag:
print(" <== Minimal Val. Error")
else:
print()
if verbose:
self.draw_params_histogram()
for idx in range(self.hidden_layer_num + 1):
desc_obj = self.get_param_describe(layer_num=idx, kind="W")
num = "{:10d}".format(desc_obj.nobs)
min = "{:5.4f}".format(desc_obj.minmax[0])
max = "{:5.4f}".format(desc_obj.minmax[1])
mean = "{:5.4f}".format(desc_obj.mean)
variance = "{:5.4f}".format(desc_obj.variance)
skewness = "{:5.4f}".format(desc_obj.skewness)
kurtosis = "{:5.4f}".format(desc_obj.kurtosis)
print(
'W' + str(idx) + '-',
"num:{:10s}, min:{:5s}, max:{:5s}, mean:{:5s}, variance:{:5s}, skewness:{:5s}, kurtosis:{:5s}"
.format(num, min, max, mean, variance, skewness,
kurtosis))
for idx in range(self.hidden_layer_num + 1):
desc_obj = self.get_param_describe(layer_num=idx, kind="b")
num = "{:10d}".format(desc_obj.nobs)
min = "{:5.4f}".format(desc_obj.minmax[0])
max = "{:5.4f}".format(desc_obj.minmax[1])
mean = "{:5.4f}".format(desc_obj.mean)
variance = "{:5.4f}".format(desc_obj.variance)
skewness = "{:5.4f}".format(desc_obj.skewness)
kurtosis = "{:5.4f}".format(desc_obj.kurtosis)
print(
'b' + str(idx) + '-',
"num:{:10s}, min:{:5s}, max:{:5s}, mean:{:5s}, variance:{:5s}, skewness:{:5s}, kurtosis:{:5s}"
.format(num, min, max, mean, variance, skewness,
kurtosis))
print()
def set_learning_process_parameters(self, data, batch_size, epoch, print_period, verbose):
batch_mask = np.random.choice(data.num_train_data, batch_size)
i_batch = data.train_input[batch_mask]
t_batch = data.train_target[batch_mask]
train_error = self.session.run(self.error,
{
self.input_node: i_batch,
self.target_node: t_batch
}, False)
self.train_error_list.append(train_error)
validation_error = self.session.run(self.error,
{
self.input_node: data.validation_input,
self.target_node: data.validation_target
}, False)
self.validation_error_list.append(validation_error)
forward_final_output = self.feed_forward(input_data=data.test_input)
test_accuracy = tff.accuracy(forward_final_output, data.test_target)
self.test_accuracy_list.append(test_accuracy)
for idx in range(self.hidden_layer_num + 1):
d = self.get_param_describe(layer_num=idx, kind="W")
self.param_mean_list['W' + str(idx)].append(d.mean)
self.param_variance_list['W' + str(idx)].append(d.variance)
self.param_skewness_list['W' + str(idx)].append(d.skewness)
self.param_kurtosis_list['W' + str(idx)].append(d.kurtosis)
d = self.get_param_describe(layer_num=idx, kind="b")
self.param_mean_list['b' + str(idx)].append(d.mean)
self.param_variance_list['b' + str(idx)].append(d.variance)
self.param_skewness_list['b' + str(idx)].append(d.skewness)
self.param_kurtosis_list['b' + str(idx)].append(d.kurtosis)
if epoch % print_period == 0:
print(
"Epoch {:3d} Completed - Train Error: {:7.6f} - Validation Error: {:7.6f} - Test Accuracy: {:7.6f}".format(
epoch,
float(train_error),
float(validation_error),
float(test_accuracy)
))
if verbose:
self.draw_params_histogram()
for idx in range(self.hidden_layer_num + 1):
desc_obj = self.get_param_describe(layer_num=idx, kind="W")
num = "{:10d}".format(desc_obj.nobs)
min = "{:5.4f}".format(desc_obj.minmax[0])
max = "{:5.4f}".format(desc_obj.minmax[1])
mean = "{:5.4f}".format(desc_obj.mean)
variance = "{:5.4f}".format(desc_obj.variance)
skewness = "{:5.4f}".format(desc_obj.skewness)
kurtosis = "{:5.4f}".format(desc_obj.kurtosis)
print('W' + str(idx) + '-',
"num:{:10s}, min:{:5s}, max:{:5s}, mean:{:5s}, variance:{:5s}, skewness:{:5s}, kurtosis:{:5s}".format(
num, min, max, mean, variance, skewness, kurtosis
)
)
for idx in range(self.hidden_layer_num + 1):
desc_obj = self.get_param_describe(layer_num=idx, kind="b")
num = "{:10d}".format(desc_obj.nobs)
min = "{:5.4f}".format(desc_obj.minmax[0])
max = "{:5.4f}".format(desc_obj.minmax[1])
mean = "{:5.4f}".format(desc_obj.mean)
variance = "{:5.4f}".format(desc_obj.variance)
skewness = "{:5.4f}".format(desc_obj.skewness)
kurtosis = "{:5.4f}".format(desc_obj.kurtosis)
print('b' + str(idx) + '-',
"num:{:10s}, min:{:5s}, max:{:5s}, mean:{:5s}, variance:{:5s}, skewness:{:5s}, kurtosis:{:5s}".format(
num, min, max, mean, variance, skewness, kurtosis
)
)
print()
