lbortolotti

Results 9 comments of lbortolotti

BUG: not reproducible error `FloatingPointError: overflow encountered in multiply` in the following sequence: read_csv followed by to_datetime with pandas version 2.2.2

I'm seeing the exact same issue as the OP. No csv read/write in my case, just a pd.to_datetime call that randomly fails, but if re-run on the exact same data...

issue about XLA compile MirroredStrategy

Bumped into this exact same issue in TF 2.14! Is anyone using multi GPU at all?! I can't validate TF 2.15, as the official docker image has broken GPU support...

Inconsistent type handling between backends: ops.sqrt

That is the underlying issue - however at this point I'm a little unclear as to what the design goal of the keras_core.ops API is. If it's meant totally decouple...

[torch backend] ValueError: Expected an object of type `Trackable`, such as `tf.Module` or a subclass of the `Trackable` class, for export.

My understanding is that, if you want to "switch backends" like this, the only way is to save the model as .keras, and reload it after having enabled another. This...

JAX backend - exporting to TF savedmodel via JAX2TF errors (Cannot convert '(b, t, 1)' to a shape. Found invalid entry 't' AND ValueError: Invalid dtype: _DimExpr)

> 1. Save in keras format > > > ``` > import os > os.environ['KERAS_BACKEND'] = 'jax' > import keras > import numpy as np > > window_size = 100...

JAX backend - exporting to TF savedmodel via JAX2TF errors (Cannot convert '(b, t, 1)' to a shape. Found invalid entry 't' AND ValueError: Invalid dtype: _DimExpr)

> Similar issue with `torch` backend. #19017 I suspect that https://github.com/keras-team/keras/issues/19017 is impossible by design, it would require quite a lot of magic on keras' part! This issue, however, leverages...

JAX backend - exporting to TF savedmodel via JAX2TF errors (Cannot convert '(b, t, 1)' to a shape. Found invalid entry 't' AND ValueError: Invalid dtype: _DimExpr)

> Hi @lbortolotti, > > It seems that your workflow with jax2tf and TF SavedModel is exactly what the `model.export()` API is useful for. You can simply do `model.export("my_model")` or...

JAX backend - exporting to TF savedmodel via JAX2TF errors (Cannot convert '(b, t, 1)' to a shape. Found invalid entry 't' AND ValueError: Invalid dtype: _DimExpr)

Ah excellent! I see it here https://github.com/keras-team/keras/blob/c4dd4fab5bd9491b32b4ab1e360d290ad8e8a238/keras/export/export_lib.py#L267 It's roughly doing what I was trying to do "by hand", and manages to get the model exported. But the exported model is...

JAX backend - exporting to TF savedmodel via JAX2TF errors (Cannot convert '(b, t, 1)' to a shape. Found invalid entry 't' AND ValueError: Invalid dtype: _DimExpr)

These changes fix my specific example. I suppose the "proper" fix would be to: 1. Somehow expose `polymorphic_shape` to the user (or is there already a way in keras to...