“ Encoder / conv6 / Conv2D”的2减去3导致的负尺寸大小

Question

I am trying to implement an AutoEncoder in Tensorflow. I am a beginner in Python as well as StackOverflow. These two are my encoder and decoder.My train_data.shape is (42000,28,28,1) (mnist dataset).

def Network(Input):
with tf.name_scope("Encoder"):
    #encoder starts here
    conv1 = tf.layers.conv2d(Input, filters = 64, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'conv1')
    conv2 = tf.layers.conv2d(conv1, filters = 64, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'conv2')
    pool1 = tf.layers.max_pooling2d(conv2, pool_size = 2, strides = 2, name = 'pool1')

    conv3 = tf.layers.conv2d(pool1, filters = 128, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'conv3')
    conv4 = tf.layers.conv2d(conv3, filters = 128, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'conv4')
    pool2 = tf.layers.max_pooling2d(conv4, pool_size = 2, strides = 2, name = 'pool2')


    conv5 = tf.layers.conv2d(pool2, filters = 256, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'conv5')
    conv6 = tf.layers.conv2d(conv5, filters = 256, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'conv6')
    conv7 = tf.layers.conv2d(conv6, filters = 256, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'conv7')
    pool3 = tf.layers.max_pooling2d(conv7, pool_size = 2, strides = 2, name = 'pool3')

    conv8 = tf.layers.conv2d(pool3, filters = 512, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'conv8')
    conv9 = tf.layers.conv2d(conv8, filters = 512, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'conv9')
    conv10 = tf.layers.conv2d(conv9, filters = 512, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'conv10')
    pool4 = tf.layers.max_pooling2d(conv10, pool_size = 2, strides = 2, name = 'pool4')

    conv11 = tf.layers.conv2d(pool4, filters = 512, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'conv11')
    conv12 = tf.layers.conv2d(conv11, filters = 512, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'conv12')
    conv13 = tf.layers.conv2d(conv12, filters = 512, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'conv13')

    pool5 = tf.layers.max_pooling2d(conv13, pool_size = 2, strides = 2, name = 'pool5')

return pool5

..

def Decoder(pool5):
with tf.name_scope("Decoder"):

    deconv1=tf.layers.conv2d_transpose(pool5, filters = 512, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'deconv1')
    deconv2=tf.layers.conv2d_transpose(deconv1, filters = 512, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'deconv2')
    deconv3=tf.layers.conv2d_transpose(deconv2, filters = 512, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'deconv3')

    pool6=tf.layers.max_pooling2d(deconv3, pool_size = 2, strides = 2, name = 'pool6')
    deconv4 = tf.layers.conv2d_transpose(pool6, filters = 512, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'deconv4')
    deconv5 = tf.layers.conv2d_transpose(deconv4, filters = 512, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'deconv5')
    deconv6 = tf.layers.conv2d_transpose(deconv5, filters = 512, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'deconv6')

    pool7 = tf.layers.max_pooling2d(deconv6, pool_size = 2, strides = 2, name = 'pool7')
    deconv7 = tf.layers.conv2d_transpose(pool7, filters = 256, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'deconv7')
    deconv8 = tf.layers.conv2d_transpose(deconv7, filters = 256, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'deconv8')
    deconv9 = tf.layers.conv2d_transpose(deconv8, filters = 256, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'deconv9')

    pool8 = tf.layers.max_pooling2d(deconv9, pool_size = 2, strides = 2, name = 'pool8')
    deconv10 = tf.layers.conv2d_transpose(pool8, filters = 128, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'deconv10')
    deconv11 = tf.layers.conv2d_transpose(deconv10, filters = 128, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'deconv11')

    pool9= tf.layers.max_pooling2d(deconv11, pool_size = 2, strides = 2, name = 'pool9')
    deconv12 = tf.layers.conv2d_transpose(pool9, filters = 64, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'deconv12')
    deconv13 = tf.layers.conv2d_transpose(deconv12, filters = 64, kernel_size = 3, strides = 1, activation = tf.nn.relu, name = 'deconv13')


    flat = tf.contrib.layers.flatten(deconv13)
    fc1 = tf.layers.dense(flat, units = 1024, activation = tf.nn.relu, name = 'fc1')
    fc2 = tf.layers.dense(fc1, units = 10, activation = None, name = 'fc2')

return fc2

I run into this error:

 ----------------------------------------------------------------------
 ValueError                                Traceback (most recent call    last)


ValueError: Negative dimension size caused by subtracting 3 from 2 for 'Encoder/conv6/Conv2D' (op: 'Conv2D') with input shapes: [?,2,2,256],                   [3,3,256,256].

I feel there is some error in my Input.Please suggest how to visualise the shape of tensor after each conv2d operation.

Input = tf.placeholder(dtype = tf.float32, shape = [None, 28, 28, 1])

Answer 1

To visualize the shape of the tensor t you use t.get_shape() . The error explicitly says that you're trying to use a too big kernel on a too small input. From Stanford's CS231n class notes you can get the equation to compute the exact shapes of your data given certain kernel size, stride and padding You should either decrease the kernel size, increase the stride, or get bigger input to be able to perform that many convolutions. Keep in mind that your max_pooling will also shrink the input as many times as large is your pool size in that layer, so if it's 2 then your data will shrink 2 times