CODEWITHSUNDEEP
filter3dtensorflowreshape
Lingjing Wang
Lingjing Wang

Reputation: 9

How to reshape the pre-trained weights to input them to 3d convoluional neural network?

I have pre-trained weights for a 3d convolutional layer using Matlab. The weights is a 5d tensor with dimension (512,4,4,4,160). [out_channels, filter_depth, filter_height, filter_width, in_channels]

Now I want to input it as the initial weights for fine-tuning in tensorflow's tf.nn.conv3d. I see that the shape of weights are allowed for 3d convolutional neural networks should be: (4,4,4,160,512).[filter_depth, filter_height, filter_width, in_channels, out_channels]. Can I just use tf.Variable().reshape(4,4,4,160,512)? But I feel it is not the correct weights if I just use reshape.

Upvotes: 0

Views: 325

Answers (1)

dm0_
dm0_

Reputation: 2156

The tf.transpose operation can reorder axes: https://www.tensorflow.org/versions/r0.11/api_docs/python/array_ops.html#transpose

Provided that initial shape of tensor input is (512,4,4,4,160) the output tensor of tf.transpose(input, perm=[4,1,2,3,0]) will have shape (160,4,4,4,512).

Also you may need to reverse your weights along some axis or axes. In tensorflow convolutions are implemented as cross-correlations: https://www.tensorflow.org/versions/r0.11/api_docs/python/nn.html#convolution

Upvotes: 0

Related Questions