def forward(self, forward_final_output_value, target_value, is_numba):
self.target_value = target_value
self.y = tff.softmax(forward_final_output_value, is_numba=is_numba)
loss = tff.cross_entropy_error(self.y,
self.target_value,
is_numba=is_numba)
return loss
def draw_filtered_images(self, test_inputs, figsize=(20, 5)):
self.feed_forward(input_data=test_inputs, is_train=False, is_numba=False)
plt.figure(figsize=figsize)
for idx in range(len(test_inputs)):
plt.subplot(100 + len(test_inputs) * 10 + idx + 1)
img = test_inputs[idx]
img.shape = (28, 28)
print(img.shape)
plt.imshow(img, cmap='gray')
plt.show()
for layer_name, layer in self.layers.items():
if layer_name.startswith("conv"):
print("[Convolution Layer: {:s}]".format(layer_name))
plt.figure(figsize=(layer.output[0].shape[2], int(layer.output[0].shape[1] / len(layer.output))))
for idx in range(len(layer.output)):
plt.subplot(100 + len(layer.output) * 10 + idx + 1)
img = layer.output[idx]
if (img.shape[0] == 1):
img = np.reshape(img, (img.shape[1], img.shape[2]))
print(img.shape)
plt.imshow(img, cmap='gray')
elif (img.shape[0] == 3 or img.shape[0] == 4):
img = np.transpose(img, (1, 2, 0))
print(img.shape)
plt.imshow(img, cmap=cm.PRGn)
else:
print("Image Channel Size (Filter Num) should be 1, 3, 4")
plt.show()
activation_layer_name = "activation" + layer_name[4:]
activation_layer = self.layers[activation_layer_name]
print("[Activation Layer: {:s}]".format(activation_layer_name))
plt.figure(figsize=(activation_layer.output[0].shape[2], int(activation_layer.output[0].shape[1] / len(activation_layer.output))))
for idx in range(len(activation_layer.output)):
plt.subplot(100 + len(activation_layer.output) * 10 + idx + 1)
img = activation_layer.output[idx]
if (img.shape[0] == 1):
img = np.reshape(img, (img.shape[1], img.shape[2]))
print(img.shape)
plt.imshow(img, cmap='gray')
elif (img.shape[0] == 3 or img.shape[0] == 4):
img = np.transpose(img, (1, 2, 0))
print(img.shape)
plt.imshow(img, cmap=cm.PRGn)
else:
print("Image Channel Size (Filter Num) should be 1, 3, 4")
plt.show()
if layer_name.startswith("pooling"):
print("[Pooling Layer: {:s}]".format(layer_name))
plt.figure(figsize=(layer.output[0].shape[2], int(layer.output[0].shape[1] / len(layer.output))))
for idx in range(len(layer.output)):
plt.subplot(100 + len(layer.output) * 10 + idx + 1)
img = layer.output[idx]
if (img.shape[0] == 1):
img = np.reshape(img, (img.shape[1], img.shape[2]))
print(img.shape)
plt.imshow(img, cmap='gray')
elif (img.shape[0] == 3 or img.shape[0] == 4):
img = np.transpose(img, (1, 2, 0))
print(img.shape)
plt.imshow(img, cmap=cm.PRGn)
else:
print("Image Channel Size (Filter Num) should be 1, 3, 4")
plt.show()
if layer_name.startswith("reshape"):
print("[Pooling Layer: {:s}]".format(layer_name))
for idx in range(len(layer.output)):
plt.figure(figsize=(layer.output.shape[1], int(layer.output.shape[0])))
img = layer.output[idx]
img.shape = (1, len(img))
print(img.shape)
plt.imshow(img, cmap='gray')
plt.show()
if layer_name.startswith("affine"):
print("[Affine Layer: {:s}]".format(layer_name))
for idx in range(len(layer.output)):
plt.figure(figsize=(layer.output.shape[1], int(layer.output.shape[0])))
img = layer.output[idx]
img.shape = (1, len(img))
print(img.shape)
plt.imshow(img, cmap='gray')
plt.show()
print("[Softmax Layer]")
np.set_printoptions(precision=3)
for idx in range(len(layer.output)):
plt.figure(figsize=(layer.output.shape[1], int(layer.output.shape[0])))
img = layer.output[idx]
softmax_img = tff.softmax(img, is_numba=False)
softmax_img.shape = (1, len(softmax_img))
print(softmax_img.shape, ":", softmax_img)
plt.imshow(softmax_img, cmap='gray')
plt.show()