def test_cnn(test_image_name):
test_img = mpimg.imread(test_image_name)
pkl_file = open('params.pkl', 'rb')
type = test_image_name.rsplit(".")[1]
params = pickle.load(pkl_file)
params1, params2, params3, params4, params5, params6, params7, params8, params9 = params
h, w, ip_channels = test_img.shape
out_channels = 64
filter_size = 3
padding = 1
layer_1 = Convolution(ip_channels, out_channels, filter_size, padding)
layer_1.weights = params1[0]
layer_1.bias=params1[1]
relu_1 = ReLU()
layer_2 = Convolution(out_channels, out_channels, filter_size, padding)
layer_2.weights = params2[0]
layer_2.bias = params2[1]
relu_2 = ReLU()
layer_3 = Convolution(out_channels, out_channels, filter_size, padding)
layer_3.weights = params3[0]
layer_3.bias = params3[1]
relu_3 = ReLU()
layer_4 = Convolution(out_channels, out_channels, filter_size, padding)
layer_4.weights = params4[0]
layer_4.bias = params4[1]
relu_4 = ReLU()
layer_5 = Convolution(out_channels, out_channels, filter_size, padding)
layer_5.weights = params5[0]
layer_5.bias = params5[1]
relu_5 = ReLU()
layer_6 = Convolution(out_channels, out_channels, filter_size, padding)
layer_6.weights = params6[0]
layer_6.bias = params6[1]
relu_6 = ReLU()
layer_7 = Convolution(out_channels, out_channels, filter_size, padding)
layer_7.weights = params7[0]
layer_7.bias = params7[1]
relu_7 = ReLU()
layer_8 = Convolution(out_channels, out_channels, filter_size, padding)
layer_8.weights = params8[0]
layer_8.bias = params8[1]
relu_8 = ReLU()
layer_9 = Convolution(out_channels, ip_channels, filter_size, padding)
layer_9.weights = params9[0]
layer_9.bias = params9[1]
layers = [(layer_1, relu_1), (layer_2, relu_2), (layer_3, relu_3), (layer_4, relu_4), (layer_5, relu_5),
(layer_6, relu_6),
(layer_7, relu_7), (layer_8, relu_8), (layer_9, None)]
cnn_obj = CNN(layers)
out_img = cnn_obj.forward(test_img)
output_file_name = test_image_name + "_sr" + str(h) + "." + type
cv2.imwrite(output_file_name, out_img)
def train_cnn(self, image_name_list, epochs=1):
self.train_data = []
self.train_data.extend(self.get_train_data(image_name_list))
shuffled_train_data = shuffle(self.train_data)
_, _, ip_channels = shuffled_train_data[0][0].shape
out_channels = 64
filter_size = 3
padding = 1
lr = 0.0001
layer_1 = Convolution(ip_channels, out_channels, filter_size, padding)
relu_1 = ReLU()
layer_2 = Convolution(out_channels, out_channels, filter_size, padding)
relu_2 = ReLU()
layer_3 = Convolution(out_channels, out_channels, filter_size, padding)
relu_3 = ReLU()
layer_4 = Convolution(out_channels, out_channels, filter_size, padding)
relu_4 = ReLU()
layer_5 = Convolution(out_channels, out_channels, filter_size, padding)
relu_5 = ReLU()
layer_6 = Convolution(out_channels, out_channels, filter_size, padding)
relu_6 = ReLU()
layer_7 = Convolution(out_channels, out_channels, filter_size, padding)
relu_7 = ReLU()
layer_8 = Convolution(out_channels, out_channels, filter_size, padding)
relu_8 = ReLU()
layer_9 = Convolution(out_channels, ip_channels, filter_size, padding)
layers = [(layer_1, relu_1), (layer_2, relu_2), (layer_3, relu_3), (layer_4, relu_4), (layer_5, relu_5),
(layer_6, relu_6),
(layer_7, relu_7), (layer_8, relu_8), (layer_9, None)]
cnn_obj = CNN(layers)
for epoch in range(epochs):
for x, y in self.train_data:
cnn_obj = CNN(layers)
loss, grads = cnn_obj.train_step(x, y)
print("loss:", loss)
params1, params2, params3, params4, params5, params6, params7, params8, params9 = grads
dw1, db1 = params1
dw2, db2 = params2
dw3, db3 = params3
dw4, db4 = params4
dw5, db5 = params5
dw6, db6 = params6
dw7, db7 = params7
dw8, db8 = params8
dw9, db9 = params9
layer_1.weights = layer_1.weights - (lr * dw1)
layer_1.bias = layer_1.bias - (lr * db1)
layer_2.weights = layer_2.weights - (lr * dw2)
layer_2.bias = layer_2.bias - (lr * db2)
layer_3.weights = layer_3.weights - (lr * dw3)
layer_3.bias = layer_3.bias - (lr * db3)
layer_4.weights = layer_4.weights - (lr * dw4)
layer_4.bias = layer_4.bias - (lr * db4)
layer_5.weights = layer_5.weights - (lr * dw5)
layer_5.bias = layer_5.bias - (lr * db5)
layer_6.weights = layer_6.weights - (lr * dw6)
layer_6.bias = layer_6.bias - (lr * db6)
layer_7.weights = layer_7.weights - (lr * dw7)
layer_7.bias = layer_7.bias - (lr * db7)
layer_8.weights = layer_8.weights - (lr * dw8)
layer_8.bias = layer_8.bias - (lr * db8)
layer_9.weights = layer_9.weights - (lr * dw9)
layer_9.bias = layer_9.bias - (lr * db9)
layers = [(layer_1, relu_1), (layer_2, relu_2), (layer_3, relu_3), (layer_4, relu_4), (layer_5, relu_5),
(layer_6, relu_6),
(layer_7, relu_7), (layer_8, relu_8), (layer_9, None)]
return cnn_obj.params, layer_9.A_curr