François Chollet comments

Results 342 comments of


                                            François Chollet

Add GeM layer

IMO it seems fine to keep `power` as a static constructor argument.

Add `training` parameter to `compute_metrics`

I've re-read through the code, but I find the factoring in the PR quite confusing. > IMO I would decouple those logics with two public methods update_compiled_metrics and collect_metrics with...

Use outputs for model output names if outputs are dict

> Maybe make tf.keras.layers.Layer name it's outputs based on return dict just like tf.keras.Model does? For example we have layer called my_layer and it's call function returns {"a": a, "b":...

Applications: Updated f-string methods in all possible scripts

Thanks for the PR. > Lint / Check the code format (pull_request) Failing after 2m — Check the code format Please fix.

Applications: Updated f-string methods in all possible scripts

Please sign the CLA.

Applications: Updated f-string methods in all possible scripts

The CLA check is still failure, as well as the code formatting check. Please address.

Modify EarlyStopping on_train_end

@lbosk are you able to address the comment? Note that for formatting you can just use `black`.

Use outputs dictionary keys as output_names if provided

Thanks for the PR. Yes, so it would be inconsistent to use `Input.name` as the input name for every situation (list of Inputs, tuples of inputs...) except the dict case....

Use outputs dictionary keys as output_names if provided

If you keep two sets of names to refer to the same objects, that's on you. The framework cannot read your mind to determine which set you intended to use....

Thanks for the PR. First of all we should fix the current error message: > ValueError: Unexpected result of `train_function` (Empty logs). Please use `Model.compile(..., run_eagerly=True)`, or `tf.config.run_functions_eagerly(True)` for more...