test: add test for class weights (py_dataset adapter)

[edge7]

No description provided.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 82.54%. Comparing base (4c05e0c) to head (5a12688).

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #20638      +/-   ##
==========================================
+ Coverage   82.53%   82.54%   +0.01%     
==========================================
  Files         530      530              
  Lines       49058    49060       +2     
  Branches     7627     7627              
==========================================
+ Hits        40489    40497       +8     
+ Misses       6731     6726       -5     
+ Partials     1838     1837       -1     

Flag	Coverage Δ
keras	82.38% <100.00%> (+0.01%)	⬆️
keras-jax	65.74% <100.00%> (+0.01%)	⬆️
keras-numpy	60.62% <100.00%> (+0.01%)	⬆️
keras-tensorflow	66.52% <92.85%> (+0.01%)	⬆️
keras-torch	65.56% <100.00%> (+0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[edge7]

Simple PR to add tests for:

TODO: test class_weight

in py_dataset_adapter.PyDataset

Needed to add 2 calls to _standardize_batch:

• _finite_generator
• _infinite_generator

as I've noticed that _standardize_batch handles class_weights, but it was just called from get_tf_dataset. Hope there are no side effect, but tests are fine.

[fchollet]

Thanks for the PR!

[edge7]

Added a new commit where _standardize_batch is also called from:

• _finite_enqueuer_generator
• _infinite_enqueuer_generator

as those are probably even more used! Tested with real training, no crashes or anything like that

[edge7]

The PR turned out to be bigger than expected, as I initially thought it only required adding a test. However, I gradually noticed that class_weight was not actually being used in the case of PyDataset.
One note: astype is not supported with PyTorch tensors, so I had to adjust that case.

[fchollet]

y = np.round(y) should have converted the tensor to a numpy tensor, no?

[edge7]

I am not at the laptop right now to paste something to reproduce it, but it was failing for me. np.round was fine (but still returning a torch tensor)

[fchollet]

I really don't get this -- if you apply any np op to an input that isn't numpy (like a torch tensor) it will call __array__ on the object to turn it into a NumPy ndarray. You should never have a torch tensor past that point

[edge7]

I agree that isn't very clear.

Given this example:

import torch
import numpy as np
a = torch.from_numpy(np.array([1, 2.5, 3]))
b = np.round(a)

placing a breakpoint at the last line:

I get here:

which is here
it tries to apply the bound but:

so we enter _wrapit
which is here

Please note that following the debugging flow into:

conv = _array_converter(obj)
bring me here, i.e.: we are in the torch domain now.

then when back to _wrapit
the result then is indeed a numpy array:

the last call brings me to the following:

so I get back a Torch tensor again.

[edge7]

pip list | grep torch
torch                      2.5.1+cpu
torchvision                0.20.1+cpu

[notice] A new release of pip is available: 24.0 -> 24.3.1
[notice] To update, run: pip install --upgrade pip
❯ pip list | grep numpy
numpy                      2.0.2

[edge7]

Hi, am trying to get a better understanding and it looks like that array_wrap is indeed a way to get back the original object after the numpy conversion:

https://numpy.org/doc/2.1/user/basics.interoperability.html#returning-foreign-objects

Which seems to be exactly what torch has implemented.
If you scroll a bit the above link you will see that this torch behaviour is clearly documented with a couple of examples

[fchollet]

Ok, this is so messed up. Then we should use ops.convert_to_numpy at the start of the series of calls.

[edge7]

Do you mean at the start of this function or even before?

[edge7]

committed 'convert_to_numpy'

[fchollet]

LGTM, thank you!

[edge7]

Thanks for merging @fchollet
Do you remember what you meant here

keras/keras/src/trainers/data_adapters/py_dataset_adapter_test.py

Line 221 in c0a1f24

# TODO: test inference mode (single output)

?

Just passing batch equal to 1 or something else?