Native Keras support #74
Conversation
src/sparseml/keras/__init__.py
Outdated
| version = [int(v) for v in tensorflow.__version__.split(".")] | ||
| if version[0] != 2 or version[1] < 2: | ||
| raise Exception | ||
| raise RuntimeError("TensorFlow >= 2.2 is required, found {}".format(version)) |
There was a problem hiding this comment.
Why are we still restricted on 2.2 right now?
There was a problem hiding this comment.
Updated to support from TF 2.1: for TF2.0 there's a bug in TF model save.
natuan
force-pushed
the
tuan/native_keras_tf2x
branch
3 times, most recently
from
9e83e5c to
eece7bc
Compare
| from tensorflow.keras.optimizers.schedules import LearningRateSchedule | ||
|
|
||
|
|
||
| try: |
There was a problem hiding this comment.
need to change this over to our shared conditional keras import
| from typing import List, Union | ||
|
|
||
| import tensorflow | ||
| import tensorflow as tf |
There was a problem hiding this comment.
NIT: remove the abbreviation for tensorflow
| import tensorflow as tf | ||
|
|
||
|
|
||
| try: |
There was a problem hiding this comment.
need to change this over to our shared conditional keras import
| from typing import Dict, List, Union | ||
|
|
||
| import tensorflow | ||
| import tensorflow as tf |
There was a problem hiding this comment.
NIT: remove abbreviation for tensorflow
| tensor | ||
| ) | ||
| mask = tf.cast(tf.not_equal(tensor, 0.0), tensor.dtype) | ||
| sparsity = tf.math.reduce_sum(1.0 - mask).numpy() / tf.size(tensor) |
There was a problem hiding this comment.
this is going to return a numpy float and not a python float. Need to update the types or use .item.
Additionally, the sparsity is set to None and returned and the return type does not reflect that for this function
src/sparseml/keras/utils/exporter.py
Outdated
|
|
||
| from tensorflow.keras import Model | ||
|
|
||
| try: |
There was a problem hiding this comment.
need to change this over to our shared conditional keras import
tests/sparseml/keras/optim/mock.py
Outdated
| from tensorflow.keras.models import Model | ||
|
|
||
|
|
||
| try: |
There was a problem hiding this comment.
need to change this over to our shared conditional keras import
| import tensorflow as tf | ||
|
|
||
|
|
||
| try: |
There was a problem hiding this comment.
need to change this over to our shared conditional keras import
natuan
force-pushed
the
tuan/native_keras_tf2x
branch
from
eece7bc to
4c8561c
Compare
3 participants
Allow to use native Keras. Only Keras >= 2.4.3 is supported due to the following reasons:
clone_modelfunction does not have the paramclone_layerrequired by the modifiers;fitfunctions in 2.4.0/2.4.1/2.4.2 do not work with the current custom callbacks such asLRModifierCallbackw/ errorAttributeError: 'LRModifierCallback' object has no attribute '_implements_train_batch_hooks'For TF2.0, model save hit an assertion inside TF, and therefore we only enable >=TF2.1.