Convolutional layer

class layers.convolutional_layer.Convolutional_layer(filters, size, stride=None, input_shape=None, weights=None, bias=None, pad=False, activation=<class 'NumPyNet.activations.Activations'>, **kwargs)

Bases: NumPyNet.layers.base.BaseLayer

Convolutional Layer

Parameters

• filters (int) – Number of filters to be slided over the input, and also the number of channels of the output (channels_out)

• size (tuple) – Size of the kernel of shape (kx, ky).

• stride (tuple (default=None)) – Step of the kernel, with shape (st1, st2). If None, stride is assigned size values.

• input_shape (tuple (default=None)) – Shape of the input in the format (batch, w, h, c), None is used when the layer is part of a Network model.

• weights (array-like (default=None)) – Filters of the convolutionanl layer, with shape (kx, ky, channels_in, filters). If None, random weights are initialized

• bias (array-like (default=None)) – Bias of the convolutional layer. If None, bias init is random with shape (filters, )

• pad (bool (default=False)) – If False the image is cutted along the last raws and columns, if True the input is padded following keras SAME padding

• activation (str or Activation object) – Activation function of the layer.

Example

>>>  import os
>>>  from PIL import Image
>>>  import pylab as plt
>>>  from NumPyNet import activations
>>>
>>>  img_2_float = lambda im : ((im - im.min()) * (1./(im.max() - im.min()) * 1.)).astype(float)
>>>  float_2_img = lambda im : ((im - im.min()) * (1./(im.max() - im.min()) * 255.)).astype(np.uint8)
>>>
>>>  filename = os.path.join(os.path.dirname(__file__), '..', '..', 'data', 'dog.jpg')
>>>  inpt = np.asarray(Image.open(filename), dtype=float)
>>>  inpt.setflags(write=1)
>>>  inpt = img_2_float(inpt)
>>>  # Relu activation constrain
>>>  inpt = inpt * 2 - 1
>>>
>>>  inpt = np.expand_dims(inpt, axis=0) # shape from (w, h, c) to (1, w, h, c)
>>>
>>>  channels_out = 10
>>>  size         = (3, 3)
>>>  stride       = (1, 1)
>>>  pad          = False
>>>
>>>  layer_activation = activations.Relu()
>>>
>>>  np.random.seed(123)
>>>
>>>  b, w, h, c = inpt.shape
>>>  filters    = np.random.uniform(-1., 1., size = (size[0], size[1], c, channels_out))
>>>  # bias       = np.random.uniform(-1., 1., size = (channels_out,))
>>>  bias = np.zeros(shape=(channels_out,))
>>>
>>>  layer = Convolutional_layer(input_shape=inpt.shape,
>>>                              filters=channels_out,
>>>                              weights=filters,
>>>                              bias=bias,
>>>                              activation=layer_activation,
>>>                              size=size,
>>>                              stride=stride,
>>>                              pad=pad)
>>>
>>>  # FORWARD
>>>
>>>  layer.forward(inpt)
>>>  forward_out = layer.output.copy()
>>>
>>>  # after the forward to load all the attribute
>>>  print(layer)
>>>
>>>  # BACKWARD
>>>
>>>  layer.delta = np.ones(layer.out_shape, dtype=float)
>>>  delta = np.zeros(shape=inpt.shape, dtype=float)
>>>  layer.backward(delta)
>>>
>>>  # layer.update()
>>>
>>>  # Visualization
>>>
>>>  fig, (ax1, ax2, ax3) = plt.subplots(nrows=1, ncols=3, figsize=(10, 5))
>>>  fig.subplots_adjust(left=0.1, right=0.95, top=0.95, bottom=0.15)
>>>
>>>  fig.suptitle('Convolutional Layer')
>>>
>>>  ax1.imshow(float_2_img(inpt[0]))
>>>  ax1.set_title('Original image')
>>>  ax1.axis('off')
>>>  # here every filter effect on the image can be shown
>>>  ax2.imshow(float_2_img(forward_out[0, :, :, 1]))
>>>  ax2.set_title('Forward')
>>>  ax2.axis('off')
>>>
>>>  ax3.imshow(float_2_img(delta[0]))
>>>  ax3.set_title('Backward')
>>>  ax3.axis('off')
>>>
>>>  fig.tight_layout()
>>>  plt.show()

References

• https://arxiv.org/abs/1603.07285

• https://cs231n.github.io/convolutional-networks/

• https://stackoverflow.com/questions/42463172/how-to-perform-max-mean-pooling-on-a-2d-array-using-numpy

• https://docs.scipy.org/doc/numpy/reference/generated/numpy.lib.stride_tricks.as_strided.html

• https://stackoverflow.com/questions/53819528/how-does-tf-keras-layers-conv2d-with-padding-same-and-strides-1-behave

backward(delta, copy=False)

Backward function of the Convolutional layer. Source: https://arxiv.org/abs/1603.07285

Parameters

• delta (array-like) – delta array of shape (batch, w, h, c). Global delta to be backpropagated.

• copy (bool (default=False)) – States if the activation function have to return a copy of the input or not.

Return type

self

forward(inpt, copy=False)

Forward function of the Convolutional Layer: it convolves an image with ‘channels_out’ filters with dimension (kx, ky, channels_in). In doing so, it creates a view of the image with shape (batch, out_w, out_h, in_c, kx, ky) in order to perform a single matrix multiplication with the reshaped filters array, which shape is (in_c * kx * ky, out_c).

Parameters

• inpt (array-like) – Input batch of images in format (batch, in_w, in_h, in _c)

• copy (bool (default=False)) – If False the activation function modifies its input, if True make a copy instead

Return type

self

load_weights(chunck_weights, pos=0)

Load weights from full array of model weights

Parameters

• chunck_weights (array-like) – model weights and bias

• pos (int (default=0)) – Current position of the array

Returns

pos – Updated stream position.

Return type

int

property out_shape

Get the output shape as (batch, out_w, out_h, out_channels)

save_weights()

Return the biases and weights in a single ravel fmt to save in binary file

update()

Update function for the convolution layer. Optimizer must be assigned externally as an optimizer object.

Return type

self