如何在TensorFlow中清除tf.Tensor的形状信息？

Question

Suppose during training a network, we resize all images to 512*512, so there might be a tf.Tensor named input:0 , which is of shape (batch_size, 512, 512, 3) .

However, when making predictions, it is possible to feed images of multiple sizes into the network. So the shape of tensor input:0 should be something like (batch_size, None, None, 3) , since the size of images are unknown.

So if I have a Tensor of shape (batch_size, 512, 512, 3) , how do I "reshape" it to (batch_size, None, None, 3) ? I tried

inputs=tf.reshape(inputs, (batch_size, tf.shape(inputs)[1], tf.shape(inputs)[2], 3)

but the output is still of shape (batch_size, 512, 512, 3) .

Answer 1

I don't believe you can resize/rescale the weight/bias terms in a neural network. But it would be pretty easy to resize your image input to 512*512. Have you considered that?