François Chollet comments

Results 342 comments of


                                            François Chollet

How to save only the weights of trainable parameters?

> I want to save the trainable lora-related parameters instead of the whole model. How to do this? Assuming this is with a KerasNLP backbone, you can do: ```python backbone.save_lora_weights(filepath)...

Sharing weights across layers in keras 3 [feature request]

A simpler solution to your problem would be: 1. Instantiate the new Dense layer, e.g. `dense = Dense.from_config(...)`. (It doesn't have weights at that time) 2. Set `dense.kernel = old_layer.kernel`,...

Sharing weights across layers in keras 3 [feature request]

We'll add a setter for the kernel.

Sharing weights across layers in keras 3 [feature request]

The setter thing turned out to be problematic. What I would recommend is just direct setting but use `._kernel` instead of `.kernel`. Ref: https://github.com/keras-team/keras/pull/19469

Possible issue of gradients calculation in actor_critic_cartpole example

@apoorvnandan please take a look.

Add sparse input discretization test

Please take a look at the test failures.

Add sparse input discretization test

Sorry for the delay in merging, this was pending due to a failing test. Re-running now.

Add sparse input discretization test

Closing inactive stale PR -- feel free to reopen a new one.

Made masked losses compatible with masked nans

The reason has to do with parallelization. Conditional branches are much harder to handle than a single serial op even if it has more flops. @haifeng-jin can you advise here...

Made masked losses compatible with masked nans

Thanks, Haifeng! @jackd can you use the same code to benchmark the impact of this change?