Newer
Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
import numpy as np
from .layer import Layer
from .weights import Weights
def assignParameter(parameter: int | tuple) -> tuple:
"""
this checks wether a parameter is a tuple or int and returns a tuple
"""
if isinstance(parameter, (int, float)):
if float(parameter).is_integer():
return (int(parameter), int(parameter))
else:
raise ValueError('the parameter should be a whole number')
return parameter
def getWindows(input: np.ndarray, kernelSize: tuple[int, int], outputSize: tuple[int, int], padding: tuple[int, int] = (0,0), stride: tuple[int, int] = (1,1), dilate: tuple[int, int] = (0,0)) -> np.ndarray:
"""
creates windows of input for convolution and pooling layer
this function is needed to avoid loops
"""
# getting shape parameters
batchSize, channels, height, width = input.shape
# dilate the input if necessary
if dilate[0] != 0:
input = np.insert(input, range(1, height), 0, axis=2)
if dilate[1] != 0:
input = np.insert(input, range(1, width), 0, axis=3)
# pad the input if necessary
if padding[0] != 0 or padding[1] != 0:
input = np.pad(input, pad_width=((0,), (0,), (padding[0],), (padding[1],)), mode='constant', constant_values=(0.,))
# getting the strides of input
batchStrides, channelStrides, kernelHeightStrides, kernelWidthStrides = input.strides
striding = (batchStrides, channelStrides, stride[0] * kernelHeightStrides, stride[1] * kernelWidthStrides, kernelHeightStrides, kernelWidthStrides)
# returning the windows
return np.lib.stride_tricks.as_strided(input, (*outputSize, kernelSize[0], kernelSize[1]), striding)
def checkDims(input: np.ndarray) -> None:
"""
Checks that the input tensor has the correct shape.
"""
# Check that the input tensor has 4 dimensions.
assert input.ndim == 4, f"Input tensor should have 4 dimensions, got {input.ndim}"
# Get the size of each dimension.
batchsize, channels, height, width = input.shape
# Check that all dimensions have a size greater than 0.
assert batchsize > 0 and channels > 0 and height > 0 and width > 0, "All dimensions should be greater than 0"
class Convolution2D(Layer):
"""
An implementation of the convolutional layer. We convolve the input with out_channels different filters
and each filter spans all channels in the input.
"""
__slots__ = ['inChannels', 'outChannels', 'kernelSize', 'stride', 'padding', 'windows', 'input', 'weights', 'bias']
def __init__(self, inChannels: int, outChannels: int, kernelSize: tuple = (3,3), padding: tuple = (0,0), stride: tuple = (1,1), weights: Weights = None, bias: Weights = None) -> None:
super().__init__()
self.inChannels = inChannels
self.outChannels = outChannels
self.kernelSize = assignParameter(kernelSize)
self.stride = assignParameter(stride)
self.padding = assignParameter(padding)
# learnable parameters
self.weights = Weights((self.outChannels, self.inChannels, self.kernelSize[0], self.kernelSize[1]), values=weights)
self.bias = Weights(self.outChannels, init='zeros', values=bias)
def params(self) -> tuple[Weights, Weights]:
"""
returns weights and bias in a python list, called by optimizers
"""
return [self.weights, self.bias]
def forward(self, input: np.ndarray) -> np.ndarray:
"""
The forward pass of convolution
"""
self.input = input
checkDims(input)
batchSize, channels, height, width = input.shape
outHeight = (height - self.kernelSize[0] + 2 * self.padding[0]) / self.stride[0] + 1
outWidth = (width - self.kernelSize[1] + 2 * self.padding[1]) / self.stride[1] + 1
outputSize = (batchSize, channels, int(outHeight), int(outWidth))
self.windows = getWindows(input, self.kernelSize, outputSize, self.padding, self.stride)
output = np.einsum('bihwkl,oikl->bohw', self.windows, self.weights.values) + self.bias.values[None, :, None, None]
return output
def backward(self, gradient: np.ndarray) -> np.ndarray:
"""
The backward pass of convolution
"""
padding = self.kernelSize - 1 if self.padding == 0 else self.padding
gradientWindows = getWindows(gradient, self.kernelSize, self.input.shape, padding=padding, stride=(1,1), dilate=(self.stride[0] - 1, self.stride[1] - 1))
rotatedKernel = np.rot90(self.weights.values, 2, axes=(2, 3))
self.weights.deltas = np.einsum('bihwkl,bohw->oikl', self.windows, gradient)
self.bias.deltas = np.sum(gradient, axis=(0, 2, 3))
return np.einsum('bohwkl,oikl->bihw', gradientWindows, rotatedKernel)
def __str__(self) -> str:
"""
used for print the layer in a human readable manner
"""
printString = self.name
printString += ' input channels: ' + str(self.inChannels)
printString += ' output channels: ' + str(self.outChannels)
printString += ' kernel size: ' + str(self.kernelSize)
printString += ' padding: ' + str(self.padding)
printString += ' stride: ' + str(self.stride)
return printString
class Unsqueeze(Layer):
"""
this layer type exists because I was too lazy adding/removing .reshape
to inputs, depending if there is a convolution or not as the first layer
it reshapes the input according to user specification
if no channel information is given, the class assumes 1 channel
"""
__slots__ = ['inputShape', 'orginialShape']
def __init__(self, inputShape: tuple[int, int, int]) -> None:
super().__init__()
# testing if inputShape provides (channels, height, width)
if type(inputShape) is not tuple:
raise TypeError('input shape should be a tuple')
if len(inputShape) == 2:
inputShape = (1, *inputShape)
elif len(inputShape) < 2 or len(inputShape) > 3:
raise ValueError('input shape not corrisponding to (channels, height, width)')
# class attributes
self.inputShape = inputShape
self.orginialShape = None
def forward(self, input: np.ndarray) -> np.ndarray:
"""
Reshapes input into an acceptable shape for convolutions
"""
if self.orginialShape is None:
self.orginialShape = input.shape[1:]
return input.reshape(-1, *self.inputShape)
def backward(self, gradient: np.ndarray) -> np.ndarray:
"""
Reshapes the upstream gradient into original shape
"""
return gradient.reshape(-1, *self.orginialShape)