[air/tf] Support TensorflowCheckpoint's saved model/h5 format #28474
Conversation
Signed-off-by: xwjiang2010 <[email protected]>
xwjiang2010
changed the title
…redictor_from_saved_model
Signed-off-by: xwjiang2010 <[email protected]>
xwjiang2010
force-pushed
the
tf_predictor_from_saved_model
branch
from
38a77e3
to
a3177f1
Compare
Signed-off-by: xwjiang2010 <[email protected]>
Signed-off-by: xwjiang2010 <[email protected]>
…redictor_from_saved_model
xwjiang2010
changed the title
There was a problem hiding this comment.
Thanks @xwjiang2010, left some comments.
Should we update our docs/examples to use saved model format?
…redictor_from_saved_model
Signed-off-by: xwjiang2010 <[email protected]>
Signed-off-by: xwjiang2010 <[email protected]>
Signed-off-by: xwjiang2010 <[email protected]>
xwjiang2010
force-pushed
the
tf_predictor_from_saved_model
branch
from
e8523f3
to
b8db76e
Compare
Signed-off-by: xwjiang2010 <[email protected]>
Checkpoints attributes are reset when passed to `TensorflowPredictor.from_checkpoint`. `TensorflowPredictor.from_checkpoint` calls `TensorflowCheckpoint.from_checkpoint`, which creates a new checkpoint with default object attributes -- even if the passed-in checkpoint is already a `TensorflowCheckpoint`. These changes are needed to merge #28474.
Signed-off-by: xwjiang2010 <[email protected]>
…redictor_from_saved_model
Signed-off-by: xwjiang2010 <[email protected]>
Do you mean to enhance the examples with saved model as well? Or change it to just saved model? |
|
running another suite of unit tests... |
Signed-off-by: xwjiang2010 <[email protected]>
| super().__init__(preprocessor) | ||
|
|
||
| def __repr__(self): | ||
| fn_name = getattr(self.model_definition, "__name__", self.model_definition) | ||
| fn_name = getattr(self._model, "__name__", self._model) | ||
| fn_name_str = "" |
There was a problem hiding this comment.
@amogkam @bveeramani let me know if this is OK. This is mostly to satisfy the length requirement.
It would be something like tf.keras.Sequential object with some cut-off.
| ) | ||
|
|
||
|
|
||
| def compare_weights(w1: List[ndarray], w2: List[ndarray]) -> bool: |
There was a problem hiding this comment.
This is needed when ndarray has multiple elements.
There was a problem hiding this comment.
can this just be return all([np.array_equal(i1, i2) for i1, i2 in zip(w1, w2)])?
@amogkam Actually looking at our current examples/docs, I also feel that it is not clear when to use framework specific checkpoint v.s. generic air checkpoint. Additionally, thinking about MLflow etc, it is not clear what should be done if I want to use MLflow to manage all my checkpoints when I use say TensorflowCheckpoint. Need to think more here... |
Signed-off-by: xwjiang2010 <[email protected]>
Signed-off-by: xwjiang2010 <[email protected]>
Signed-off-by: xwjiang2010 <[email protected]>
Signed-off-by: xwjiang2010 <[email protected]>
| ) | ||
|
|
||
|
|
||
| def compare_weights(w1: List[ndarray], w2: List[ndarray]) -> bool: |
There was a problem hiding this comment.
can this just be return all([np.array_equal(i1, i2) for i1, i2 in zip(w1, w2)])?
richardliaw
changed the title
) Checkpoints attributes are reset when passed to `TensorflowPredictor.from_checkpoint`. `TensorflowPredictor.from_checkpoint` calls `TensorflowCheckpoint.from_checkpoint`, which creates a new checkpoint with default object attributes -- even if the passed-in checkpoint is already a `TensorflowCheckpoint`. These changes are needed to merge ray-project#28474. Signed-off-by: Weichen Xu <[email protected]>
…oject#28474) Signed-off-by: Weichen Xu <[email protected]>
Signed-off-by: xwjiang2010 [email protected]
Why are these changes needed?
Support saved model and h5 formats.
This PR is dependent on landing #28935 first.
Related issue number
Closes #26933
Checks
git commit -s) in this PR.scripts/format.shto lint the changes in this PR.