Investigate all the Places that Involve Shape or Dimension Evaluation #11
Assignees
Labels
Milestone
Comments
DarrenZhang01
pushed a commit
that referenced
this issue
Jul 25, 2020
1. #11; 2. #10; 3. #6 Root problem: Although I have updated the returned shapes from `out_shape` to `np.zeros(out_shape)` in `ntk_init_fn`, this only works for non-activation layers in Neural Tangents like a standard Dense layer. For activation layers like Relu, it is still calling the `elementwise` function inside `tf_jax_stax`, where the output shape `input_shape` should also be revised to `np.zeros(input_shape)` for `eval_on_shapes` to be able to handle it.
|
Solved in 98be634, together with issues google/neural-tangents#52, #10 and #6 |
This was referenced Jul 25, 2020
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Since the current shape Info for the
Kernelobject is stored inside anndarrayin the form likeshape1 = ndarray<tf.Tensor([1024 2], shape=(2,), dtype=int64)>, there is currently error like #10, then there is possibility that there are chained problems caused by this shape evaluation that also leads to mis-evaluation on the dimensionality, etc. Shapes likendarray<tf.Tensor([2], shape=(1,), dtype=int64)>looks more or less like a mistake caused by the nested eval, something likeshape1.shape. Overall, need to investigate the places of shape evaluation and make revisions when necessary.The text was updated successfully, but these errors were encountered: