CNNNet_mnist_mnistClassifier_net.py 3.97 KB
Newer Older
Nicola Gatto's avatar
Nicola Gatto committed
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
import mxnet as mx
import numpy as np
from mxnet import gluon

class Softmax(gluon.HybridBlock):
    def __init__(self, **kwargs):
        super(Softmax, self).__init__(**kwargs)

    def hybrid_forward(self, F, x):
        return F.softmax(x)


class Split(gluon.HybridBlock):
    def __init__(self, num_outputs, axis=1, **kwargs):
        super(Split, self).__init__(**kwargs)
        with self.name_scope():
            self.axis = axis
            self.num_outputs = num_outputs

    def hybrid_forward(self, F, x):
        return F.split(data=x, axis=self.axis, num_outputs=self.num_outputs)


class Concatenate(gluon.HybridBlock):
    def __init__(self, dim=1, **kwargs):
        super(Concatenate, self).__init__(**kwargs)
        with self.name_scope():
            self.dim = dim

    def hybrid_forward(self, F, *x):
        return F.concat(*x, dim=self.dim)


class ZScoreNormalization(gluon.HybridBlock):
    def __init__(self, data_mean, data_std, **kwargs):
        super(ZScoreNormalization, self).__init__(**kwargs)
        with self.name_scope():
            self.data_mean = self.params.get('data_mean', shape=data_mean.shape,
                init=mx.init.Constant(data_mean.asnumpy().tolist()), differentiable=False)
            self.data_std = self.params.get('data_std', shape=data_mean.shape,
                init=mx.init.Constant(data_std.asnumpy().tolist()), differentiable=False)

    def hybrid_forward(self, F, x, data_mean, data_std):
        x = F.broadcast_sub(x, data_mean)
        x = F.broadcast_div(x, data_std)
        return x


class Padding(gluon.HybridBlock):
    def __init__(self, padding, **kwargs):
        super(Padding, self).__init__(**kwargs)
        with self.name_scope():
            self.pad_width = padding

    def hybrid_forward(self, F, x):
        x = F.pad(data=x,
            mode='constant',
            pad_width=self.pad_width,
            constant_value=0)
        return x


class NoNormalization(gluon.HybridBlock):
    def __init__(self, **kwargs):
        super(NoNormalization, self).__init__(**kwargs)

    def hybrid_forward(self, F, x):
        return x


class Net(gluon.HybridBlock):
    def __init__(self, data_mean=None, data_std=None, **kwargs):
        super(Net, self).__init__(**kwargs)
        with self.name_scope():
            if not data_mean is None:
                assert(not data_std is None)
                self.input_normalization = ZScoreNormalization(data_mean=data_mean, data_std=data_std)
            else:
                self.input_normalization = NoNormalization()

            self.conv1_ = gluon.nn.Conv2D(channels=20,
                kernel_size=(5,5),
                strides=(1,1),
                use_bias=True)
            # conv1_, output shape: {[20,24,24]}

            self.pool1_ = gluon.nn.MaxPool2D(
                pool_size=(2,2),
                strides=(2,2))
            # pool1_, output shape: {[20,12,12]}

            self.conv2_ = gluon.nn.Conv2D(channels=50,
                kernel_size=(5,5),
                strides=(1,1),
                use_bias=True)
            # conv2_, output shape: {[50,8,8]}

            self.pool2_ = gluon.nn.MaxPool2D(
                pool_size=(2,2),
                strides=(2,2))
            # pool2_, output shape: {[50,4,4]}

            self.fc2_flatten = gluon.nn.Flatten()
            self.fc2_ = gluon.nn.Dense(units=500, use_bias=True)
            # fc2_, output shape: {[500,1,1]}

            self.relu2_ = gluon.nn.Activation(activation='relu')
            self.fc3_ = gluon.nn.Dense(units=10, use_bias=True)
            # fc3_, output shape: {[10,1,1]}


        self.last_layer = 'softmax'


    def hybrid_forward(self, F, x):
        image = self.input_normalization(x)
        conv1_ = self.conv1_(image)
        pool1_ = self.pool1_(conv1_)
        conv2_ = self.conv2_(pool1_)
        pool2_ = self.pool2_(conv2_)
        fc2_flatten_ = self.fc2_flatten(pool2_)
        fc2_ = self.fc2_(fc2_flatten_)
        relu2_ = self.relu2_(fc2_)
        fc3_ = self.fc3_(relu2_)
        return fc3_