CODEWITHSUNDEEP
pythonnumpytensorflowtensorboard
BinRoot
BinRoot

Reputation: 694

Visualize conv2d filter for TensorBoard image_summary

I want to visualize the filter weights of my CNN. They are of size heightxwidthxinputxoutput.

However, TensorBoard requires the image_summary to be a Tensor of shape batchesxheightxwidthxchannels.

How can I convert my filter weights to the correct form?

Some context:

W1 = tf.Variable(tf.random_normal([5, 5, 1, 64]), name='W1')
conv = tf.nn.conv2d(x, W1, strides=[1, 1, 1, 1], padding='SAME')

Upvotes: 1

Views: 4406

Answers (1)

yuefengz
yuefengz

Reputation: 3358

A normal image batch has shape [batch, height, width, 3] so you can make Tensorboard show a batch of colored images for the first convolution layer by transposing the filters to [output, height, width, 3]. This answer has the code: How to visualize learned filters on tensorflow.

For weights in other layers, you can only show input * output grayscale images. You first need to split the tensor along input/output channel, transpose and concatenate the tensor to shape [input * output, height, width, 1]. You can find some example code here: https://github.com/tensorflow/tensorflow/issues/908

Upvotes: 1

Related Questions