Max Pooling now prettier!

cvnn/_version.py

@@ -1 +1 @@
-__version__ = '1.0.0'
+__version__ = '1.0.1'

cvnn/layers.py

@@ -945,17 +945,12 @@ def pool_function(self, inputs, ksize, strides, padding, data_format):
         # The max is calculated with the absolute value. This will still work on real values.
         abs_in = tf.math.abs(inputs)
         output, argmax = tf.nn.max_pool_with_argmax(input=abs_in, ksize=ksize, strides=strides,
-                                                    padding=padding, data_format=data_format)
+                                                    padding=padding, data_format=data_format,
+                                                    include_batch_in_index=True)
         if inputs.shape[0] is None:
             return output
-        # Flatten input and argmax to make them equivalent
-        flat_in = tf.reshape(inputs, [tf.shape(inputs)[0], tf.math.reduce_prod(tf.shape(inputs)[1:])])
-        flat_argmax = tf.reshape(argmax, [tf.shape(argmax)[0], tf.math.reduce_prod(tf.shape(argmax)[1:])])
-        # TODO: Using numpy in the next line seems quite inneficient, is there a better way?
-        # Get the max values using the indeces of argmax
-        res = [tf.convert_to_tensor(flat_in.numpy()[i][arg_ind]) for i, arg_ind in enumerate(flat_argmax.numpy())]
-        # for i, arg_ind in enumerate(flat_argmax.numpy()):
-        tf_res = tf.reshape(res, output.shape)
+        shape = output.shape
+        tf_res = tf.reshape(tf.gather(tf.reshape(inputs, [-1]), argmax), shape)
         # assert np.all(tf_res == output)             # For debugging when the input is real only!
         assert tf_res.dtype == inputs.dtype
         return tf_res
@@ -989,7 +984,7 @@ def get_real_equivalent(self):
 __copyright__ = 'Copyright 2020, {project_name}'
 __credits__ = ['{credit_list}']
 __license__ = '{license}'
-__version__ = '1.0.0'
+__version__ = '1.0.1'
 __maintainer__ = 'J. Agustin BARRACHINA'
 __email__ = 'joseagustin.barra@gmail.com; jose-agustin.barrachina@centralesupelec.fr'
 __status__ = '{dev_status}'

docs/index.rst

@@ -3,7 +3,7 @@ Complex-Valued Neural Network (CVNN)
 ====================================
 
 :Author:  J. Agustin Barrachina
-:Version: 1.0.0 of 01/05/2021
+:Version: 1.0.1 of 01/20/2021
 
 
 Content

tests/test_custom_layers.py

@@ -136,7 +136,6 @@ def dropout():
                              [7.5 + 7.5j, 8.75 + 8.75j],
                              [10. + 10.j, 0. + 0.j]]
                             )
-    set_trace()
     assert np.all(data == layer(data, training=False))
     assert np.all(outputs == expected_out)
     ds_train, ds_test = get_dataset()
@@ -161,7 +160,7 @@ def get_img():
         [0, 2, 2],
         [0, 5, 7]
     ], [
-        [0, 4, 5],
+        [0, 7, 5],
         [3, 7, 9],
         [4, 5, 3]
     ]]).astype(np.float32)
@@ -181,7 +180,7 @@ def get_img():
 
 def complex_max_pool_2d():
     img = get_img()
-    max_pool = ComplexMaxPooling2D(strides=1)
+    max_pool = ComplexMaxPooling2D(strides=1, data_format="channels_last")
     res = max_pool(img.astype(np.complex64))
     expected_res = np.array([
         [[
@@ -190,13 +189,13 @@ def complex_max_pool_2d():
             [[2. + 7.j],
              [2. + 9.j]]],
         [[
-            [7. + 2.j],
+            [7. + 4.j],
             [9. + 2.j]],
             [
-                [5. + 8.j],
-                [3. + 9.j]]
-        ]])
-    assert (res == expected_res.astype(np.complex64)).numpy().all()
+            [5. + 8.j],
+            [3. + 9.j]]]
+        ])
+    assert (res.numpy() == expected_res.astype(np.complex64)).all()
     x = tf.constant([[1., 2., 3.],
                      [4., 5., 6.],
                      [7., 8., 9.]])
@@ -211,15 +210,15 @@ def complex_avg_pool():
     avg_pool = ComplexAvgPooling2D(strides=1)
     res = avg_pool(img.astype(np.complex64))
     expected_res = np.array([[[[0.75 + 3.5j], [1.75 + 6.25j]], [[1.75 + 4.75j], [4. + 6.j]]],
-                             [[[3.5 + 2.25j], [6.25 + 3.25j]], [[4.75 + 4.25j], [6. + 5.25j]]]])
-    assert (res == expected_res.astype(np.complex64)).numpy().all()
+                             [[[4.25+2.25j], [7 + 3.25j]], [[4.75 + 4.25j], [6. + 5.25j]]]])
+    assert (res.numpy() == expected_res.astype(np.complex64)).all()
 
 
 def test_layers():
+    complex_max_pool_2d()
     dropout()
     complex_avg_pool()
     shape_ad_dtype_of_conv2d()
-    complex_max_pool_2d()
     dense_example()
 
 

tests/test_several_datasets.py

@@ -4,6 +4,7 @@
 from tensorflow.keras import datasets, models
 from cvnn.initializers import ComplexGlorotUniform
 from cvnn.layers import ComplexDense, ComplexFlatten, ComplexInput
+import cvnn.layers as complex_layers
 from cvnn import layers
 from cvnn.montecarlo import run_gaussian_dataset_montecarlo
 
@@ -84,12 +85,13 @@ def cifar10_test():
     test_images = test_images.astype(dtype_1)
     train_labels = train_labels.astype(dtype_1)
     test_labels = test_labels.astype(dtype_1)
+
     model = models.Sequential()
     model.add(layers.ComplexInput(input_shape=(32, 32, 3), dtype=dtype_1))     # Never forget this!!!
     model.add(layers.ComplexConv2D(32, (3, 3), activation='cart_relu'))
-    model.add(layers.ComplexAvgPooling2D((2, 2)))
+    model.add(layers.ComplexMaxPooling2D((2, 2)))   # TODO: This is changing the dtype!
     model.add(layers.ComplexConv2D(64, (3, 3), activation='cart_relu'))
-    model.add(layers.ComplexMaxPooling2D((2, 2)))       # TODO: This is changing the dtype!
+    model.add(layers.ComplexAvgPooling2D((2, 2)))
     model.add(layers.ComplexConv2D(64, (3, 3), activation='cart_relu'))
     model.add(layers.ComplexFlatten())
     model.add(layers.ComplexDense(64, activation='cart_relu'))
@@ -101,12 +103,46 @@ def cifar10_test():
     history = model.fit(train_images, train_labels, epochs=2, validation_data=(test_images, test_labels))
 
 
+def random_dataset():
+    x_train = np.complex64(tf.complex(tf.random.uniform([640, 65, 82, 1]), tf.random.uniform([640, 65, 82, 1])))
+    x_test = np.complex64(tf.complex(tf.random.uniform([200, 65, 82, 1]), tf.random.uniform([200, 65, 82, 1])))
+    y_train = np.uint8(np.random.randint(5, size=(640, 1)))
+    y_test = np.uint8(np.random.randint(5, size=(200, 1)))
+
+    model = tf.keras.models.Sequential()
+    model.add(complex_layers.ComplexInput(input_shape=(65, 82, 1)))  # Always use ComplexInput at the start
+    model.add(complex_layers.ComplexConv2D(8, (5, 5), activation='cart_relu'))
+    model.add(complex_layers.ComplexMaxPooling2D((2, 2)))
+    model.add(complex_layers.ComplexConv2D(16, (5, 5), activation='cart_relu'))
+    model.add(complex_layers.ComplexFlatten())
+    model.add(complex_layers.ComplexDense(256, activation='cart_relu'))
+    model.add(complex_layers.ComplexDropout(0.1))
+    model.add(complex_layers.ComplexDense(64, activation='cart_relu'))
+    model.add(complex_layers.ComplexDropout(0.1))
+    model.add(complex_layers.ComplexDense(5, activation='convert_to_real_with_abs'))
+    # An activation that casts to real must be used at the last layer.
+    # The loss function cannot minimize a complex number
+
+    # Compile it
+    model.compile(optimizer='adam',
+                  loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
+                  metrics=['accuracy'],
+                  # run_eagerly=True
+                  )
+    model.summary()
+
+    # Train and evaluate
+    history = model.fit(x_train, y_train, epochs=100, validation_data=(x_test, y_test))
+    test_loss, test_acc = model.evaluate(x_test, y_test, verbose=2)
+
+
 def test_datasets():
     cifar10_test()
-    fashion_mnist_example()
-    mnist_example()
-    run_gaussian_dataset_montecarlo(epochs=2, iterations=1)
+    random_dataset()
+    # fashion_mnist_example()
+    # mnist_example()
+    # run_gaussian_dataset_montecarlo(epochs=2, iterations=1)
 
 
 if __name__ == '__main__':
-    test_datasets()
+    test_datasets()