Max_pool with dynamic ksize

[FilipeAleixo]

I have the following code for a convolutional layer. This layer is part a larger computational graph.

# Define the shape of the filter
filter_shape = [1,
                config.char_filter_size,
                config.dim_char,
                config.dim_char]

# Define the convolutional layer weights and biases
W_conv = tf.Variable(tf.truncated_normal(filter_shape, stddev=0.1),
                     name="W_conv")
b_conv = tf.Variable(tf.constant(0.1, shape=[config.dim_char]),
                     name="b_conv")
# Do 2d convolution
conv = tf.nn.conv2d(char_embeddings,
                    W_conv,
                    strides=[1, 1, 1, 1],
                    padding="VALID",
                    name="conv")
# Apply nonlinearity
# h_conv has the same shape as conv
h_conv = tf.nn.relu(tf.nn.bias_add(conv, b_conv),
                    name="conv_relu")
# Maxpooling h_conv over dim 2 (char dim)

# ERROR HERE
conv_pooled = tf.nn.max_pool(h_conv,
                             ksize=[1, 1, tf.shape(h_conv)[-2], 1],
                             strides=[1, 1, 1, 1],
                             padding='VALID',
                             name="conv_max_pool")

When trying to run, I get the error:

TypeError: Expected int for argument 'ksize' not tf.Tensor shape=() dtype=int32.

is tf.nn.max_pool unable to handle dynamic ksize?

[girving]

This looks like it should work via Derek's auto-stack machinery. I'll briefly try to reproduce...

[girving]

Ah, correct: tf.nn.max_pool does not support dynamic size. ksize is an attr, therefore must be a constant. This would be reasonable to fix, but would require a MaxPoolV2 op kernel similar to TopKV2. I'll leave it open as contributions welcome.

[girving]

Duplicate of #4746.

[nkszjx]

tf.reduce_max(activation, axis=1, keep_dims=True)

[nkszjx]

see https://stackoverflow.com/questions/46760927/using-the-shape-of-a-tensor-with-dynamic-shape-in-tensorflow-operations