def test_calculate_output_dims_with_invalid_padding_value(self):
# given
w = np.random.rand(5, 5, 3, 16)
b = np.random.rand(16)
padding = "lorem ipsum"
# when
layer = ConvLayer2D(w=w, b=b, padding=padding)
with pytest.raises(InvalidPaddingModeError):
_ = layer.calculate_output_dims((32, 11, 11, 3))
def test_calculate_pad_width_with_invalid_padding_value(self):
# given
w = np.random.rand(5, 5, 3, 16)
b = np.random.rand(16)
padding = 'lorem ipsum'
# when
layer = ConvLayer2D(w=w, b=b, padding=padding)
with pytest.raises(InvalidPaddingModeError):
_ = layer.calculate_pad_dims()
def test_forward_pass_with_invalid_padding_value(self):
# given
w = np.random.rand(5, 5, 3, 16)
b = np.random.rand(16)
activation = np.random.rand(16, 11, 11, 3)
padding = 'lorem ipsum'
# when
layer = ConvLayer2D(w=w, b=b, padding=padding)
with pytest.raises(InvalidPaddingModeError):
_ = layer.forward_pass(activation, training=True)
def test_calculate_output_dims_with_same_padding_symmetrical(self):
# given
w = np.random.rand(5, 5, 3, 16)
b = np.random.rand(16)
padding = 'same'
# when
layer = ConvLayer2D(w=w, b=b, padding=padding)
result = layer.calculate_output_dims((32, 11, 11, 3))
# then
assert result == (32, 11, 11, 16)
def test_calculate_pad_width_with_same_padding_asymmetrical(self):
# given
w = np.random.rand(5, 7, 3, 16)
b = np.random.rand(16)
padding = 'same'
# when
layer = ConvLayer2D(w=w, b=b, padding=padding)
result = layer.calculate_pad_dims()
# then
assert result == (2, 3)
def test_calculate_pad_width_with_valid_padding(self):
# given
w = np.random.rand(5, 5, 3, 16)
b = np.random.rand(16)
padding = 'valid'
# when
layer = ConvLayer2D(w=w, b=b, padding=padding)
result = layer.calculate_pad_dims()
# then
assert result == (0, 0)
def test_calculate_output_dims_with_valid_padding_asymmetrical(self):
# given
w = np.random.rand(3, 5, 3, 16)
b = np.random.rand(16)
padding = "valid"
# when
layer = ConvLayer2D(w=w, b=b, padding=padding)
result = layer.calculate_output_dims((32, 11, 11, 3))
# then
assert result == (32, 9, 7, 16)
def test_pad_symmetrical(self):
# given
array = np.random.rand(100, 28, 28, 3)
pad = 3, 3
# when
result = ConvLayer2D.pad(array=array, pad=pad)
# then
print(result.sum())
print(array.sum())
assert result.shape == (100, 34, 34, 3)
assert abs(result.sum() - array.sum()) < 1e-8
def test_backward_pass_only_size_valid_padding(self):
# given
activation = np.random.rand(64, 11, 11, 3)
w = np.random.rand(5, 5, 3, 16)
b = np.random.rand(16)
layer = ConvLayer2D(w=w, b=b, padding='valid')
# when
forward_result = layer.forward_pass(activation, training=True)
backward_result = layer.backward_pass(forward_result)
# then
assert backward_result.shape == activation.shape
def test_forward_pass_with_valid_padding(self):
# given
w = np.random.rand(5, 5, 3, 16)
b = np.random.rand(16)
activation = np.random.rand(16, 11, 11, 3)
padding = 'valid'
# when
layer = ConvLayer2D(w=w, b=b, padding=padding)
result = layer.forward_pass(activation, training=True)
assert result.shape == (16, 7, 7, 16)
expected_val = np.sum(
w[:, :, :, 0] * activation[0, 0:5, 0:5, :]) + b[0]
assert abs(expected_val - result[0, 0, 0, 0]) < 1e-8
import numpy as np
import pytest
from src.errors import InvalidPaddingModeError
from src.layers.convolutional import ConvLayer2D, FastConvLayer2D, SuperFastConvLayer2D
if __name__ == "__main__":
w = np.random.rand(5, 5, 3, 16)
b = np.random.rand(16)
conv_layer = ConvLayer2D(w, b)
print(conv_layer)