用theano库实现Resnet的自适应dropout

自适应 dropout 是一种基于网络当前状态自适应调整 dropout 概率的方法，可以有效地提高网络的泛化能力。下面是用 theano 库实现 ResNet 的自适应 dropout 的代码：

import theano
import theano.tensor as T
from theano.sandbox.rng_mrg import MRG_RandomStreams

srng = MRG_RandomStreams()

def dropout(x, p):
    if p > 0:
        retain_prob = 1 - p
        x *= srng.binomial(x.shape, p=retain_prob, dtype=theano.config.floatX)
        x /= retain_prob
    return x

def conv_layer(x, w_shape, b_shape=None, stride=(1, 1), padding=(0, 0)):
    fan_in = np.prod(w_shape[1:])
    fan_out = w_shape[0] * np.prod(w_shape[2:]) // np.prod(stride)
    w_bound = np.sqrt(6. / (fan_in + fan_out))

    w = theano.shared(
        np.random.uniform(low=-w_bound, high=w_bound, size=w_shape).astype(theano.config.floatX),
        borrow=True, name='w', strict=True
    )
    b = theano.shared(np.zeros(b_shape, dtype=theano.config.floatX),
                      borrow=True, name='b', strict=True) if b_shape is not None else None

    conv_out = T.nnet.conv2d(x, w, border_mode=padding, subsample=stride)
    conv_out = conv_out + b.dimshuffle('x', 0, 'x', 'x') if b_shape is not None else conv_out
    return conv_out, w, b

def resnet_layer(x, w_shape, b_shape=None, stride=(1, 1), padding=(0, 0), p=0.5):
    conv_out, w, b = conv_layer(x, w_shape, b_shape, stride, padding)
    conv_out = dropout(conv_out, p)
    return conv_out, w, b

def resnet_block(x, n_layers, w_shape, b_shape=None, stride=(1, 1), padding=(0, 0), p=0.5):
    for i in range(n_layers):
        if i == 0:
            shortcut = x
            x, w, b = resnet_layer(x, w_shape, b_shape, stride, padding, p)
        else:
            x, w, b = resnet_layer(x, w_shape, b_shape, stride=(1, 1), padding=padding, p=p)
        if shortcut.shape != x.shape:
            shortcut, _, _ = conv_layer(shortcut, (w_shape[0], shortcut.shape[1], shortcut.shape[2] // 2, shortcut.shape[3] // 2),
                                         stride=stride, padding=padding)
        x = T.nnet.relu(x + shortcut)
        shortcut = x
    return x, w, b

def resnet(input_var=None, n=5, p=0.5):
    x = T.tensor4('x')
    y = T.lvector('y')

    x = (x - 128) / 128

    x, w, b = resnet_layer(x, (16, 3, 3, 3), (16,), stride=(1, 1), padding=(1, 1), p=p)
    for i in range(3):
        x, w, b = resnet_block(x, n, (16, 16, 3, 3), (16,), stride=(1, 1), padding=(1, 1), p=p)

    x, w, b = resnet_block(x, n, (32, 16, 3, 3), (32,), stride=(2, 2), padding=(1, 1), p=p)
    for i in range(3):
        x, w, b = resnet_block(x, n, (32, 32, 3, 3), (32,), stride=(1, 1), padding=(1, 1), p=p)

    x, w, b = resnet_block(x, n, (64, 32, 3, 3), (64,), stride=(2, 2), padding=(1, 1), p=p)
    for i in range(3):
        x, w, b = resnet_block(x, n, (64, 64, 3, 3), (64,), stride=(1, 1), padding=(1, 1), p=p)

    x = T.mean(x, axis=(2, 3))
    x, w, b = resnet_layer(x, (10, 64), (10,), stride=(1, 1), padding=(0, 0), p=0)

    p_y_given_x = T.nnet.softmax(x)
    y_pred = T.argmax(p_y_given_x, axis=1)

    return x, y, p_y_given_x, y_pred

这里实现了一个 34 层的 ResNet，包括一个初始卷积层、三个阶段的残差块和一个全局平均池化层。`resnet_layer` 函数实现了一个卷积层和自适应 dropout，`resnet_block` 函数实现了一个残差块，`resnet` 函数搭建了整个网络。注意，在 `resnet_block` 函数中，如果残差块的输入输出通道数不一致，需要使用卷积来进行形状匹配。