Just a simple model, which applies the Laplacian-operation to images to showcase the usage of variables, and alternating the inference-graph later.
| 38 | |
| 39 | |
| 40 | class Model(ModelDesc): |
| 41 | """Just a simple model, which applies the Laplacian-operation to images to showcase |
| 42 | the usage of variables, and alternating the inference-graph later. |
| 43 | """ |
| 44 | |
| 45 | def inputs(self): |
| 46 | return [tf.TensorSpec((None, SHAPE, SHAPE, CHANNELS), tf.uint8, 'input_img'), |
| 47 | tf.TensorSpec((None, SHAPE, SHAPE, CHANNELS), tf.uint8, 'target_img')] |
| 48 | |
| 49 | def make_prediction(self, img): |
| 50 | |
| 51 | img = tf.cast(img, tf.float32) |
| 52 | img = tf.image.rgb_to_grayscale(img) |
| 53 | |
| 54 | k = tf.get_variable('filter', dtype=tf.float32, |
| 55 | initializer=[[[[0.]], [[1.]], [[0.]]], [ |
| 56 | [[1.]], [[-4.]], [[1.]]], [[[0.]], [[1.]], [[0.]]]]) |
| 57 | prediction_img = tf.nn.conv2d(img, k, strides=[1, 1, 1, 1], padding='SAME') |
| 58 | return prediction_img |
| 59 | |
| 60 | def build_graph(self, input_img, target_img): |
| 61 | |
| 62 | target_img = tf.cast(target_img, tf.float32) |
| 63 | target_img = tf.image.rgb_to_grayscale(target_img) |
| 64 | |
| 65 | self.prediction_img = tf.identity(self.make_prediction(input_img), name='prediction_img') |
| 66 | |
| 67 | cost = tf.losses.mean_squared_error(target_img, self.prediction_img, |
| 68 | reduction=tf.losses.Reduction.MEAN) |
| 69 | return tf.identity(cost, name='total_costs') |
| 70 | |
| 71 | def optimizer(self): |
| 72 | lr = tf.get_variable('learning_rate', initializer=0.0, trainable=False) |
| 73 | return tf.train.AdamOptimizer(lr) |
| 74 | |
| 75 | |
| 76 | def get_data(subset): |
no outgoing calls
no test coverage detected