setWeights fails with Layer weight shape not compatible error

TensorFlow.js version

1.2.2

Node version

v11.13.0

Describe the problem or feature request

Attempting to copy weights from one network to another fails with the following error: (node:16764) UnhandledPromiseRejectionWarning: Error: Layer weight shape 128 not compatible with provided weight shape 3,3,128,256

Code to reproduce the bug / link to feature request

Encountered this error while attempting to run the example snake-dqn app. Error is thrown from the following method in dqn.js:

export function copyWeights(destNetwork, srcNetwork) { destNetwork.setWeights(srcNetwork.getWeights()); }

Traced the error to tfjs-layers/src/engine/topology.js in the setWeights method.

It appears that the list of layers returned by getWeights() is in a different order than those returned by this.weights:

weights contains: 0 -> 3,3,2,128 conv2d_Conv2D1/kernel 1 -> 128 conv2d_Conv2D1/bias 2 -> 128 batch_normalization_BatchNormalization1/gamma 3 -> 128 batch_normalization_BatchNormalization1/beta 4 -> 3,3,128,256 conv2d_Conv2D2/kernel 5 -> 256 conv2d_Conv2D2/bias 6 -> 256 batch_normalization_BatchNormalization2/gamma 7 -> 256 batch_normalization_BatchNormalization2/beta 8 -> 3,3,256,256 conv2d_Conv2D3/kernel 9 -> 256 conv2d_Conv2D3/bias 10 -> 2304,100 dense_Dense1/kernel 11 -> 100 dense_Dense1/bias 12 -> 100,3 dense_Dense2/kernel 13 -> 3 dense_Dense2/bias 14 -> 128 batch_normalization_BatchNormalization1/moving_mean 15 -> 128 batch_normalization_BatchNormalization1/moving_variance 16 -> 256 batch_normalization_BatchNormalization2/moving_mean 17 -> 256 batch_normalization_BatchNormalization2/moving_variance

paramValues contains: 0 -> 3,3,2,128 conv2d_Conv2D4/kernel 1 -> 128 conv2d_Conv2D4/bias 2 -> 128 batch_normalization_BatchNormalization3/gamma 3 -> 128 batch_normalization_BatchNormalization3/beta 4 -> 128 batch_normalization_BatchNormalization3/moving_mean 5 -> 128 batch_normalization_BatchNormalization3/moving_variance 6 -> 3,3,128,256 conv2d_Conv2D5/kernel 7 -> 256 conv2d_Conv2D5/bias 8 -> 256 batch_normalization_BatchNormalization4/gamma 9 -> 256 batch_normalization_BatchNormalization4/beta 10 -> 256 batch_normalization_BatchNormalization4/moving_mean 11 -> 256 batch_normalization_BatchNormalization4/moving_variance 12 -> 3,3,256,256 conv2d_Conv2D6/kernel 13 -> 256 conv2d_Conv2D6/bias 14 -> 2304,100 dense_Dense3/kernel 15 -> 100 dense_Dense3/bias 16 -> 100,3 dense_Dense4/kernel 17 -> 3 dense_Dense4/bias

[jkovacs-dev]

I agree, but I’m not the maintainer. After some debugging, I traced it down to the fact that models which are marked as untrainable return their weights in a different order than trainable models. You can fix the example code by removing the call which marks the model as untrainable.

I was thinking of submitting a PR for the tfjs-examples repository so at least the example code might work OOTB.

On Aug 11, 2019, at 8:45 pm, Steven Weaver [e-mail] wrote:

Hi! I’m not sure how this isn’t considered a bug. The example doesn’t run… and there is an entire chapter dedicated to it in the upcoming release of Deep Learning with JavaScript.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[rthadur]

This question is better asked on StackOverflow since it is not a bug or feature request. There is also a larger community that reads questions there.