Added support for converting large models #1090
Conversation
TomWildenhain-Microsoft
force-pushed
the
tom/ConvertLargeModels
branch
2 times, most recently
from
40c1d93 to
4bf08f4
Compare
tf2onnx/tf_loader.py
Outdated
| @@ -96,7 +96,19 @@ def inputs_without_resource(sess, input_names): | |||
| return input_names | |||
|
|
|||
|
|
|||
| def from_function(func, input_names, output_names): | |||
| def from_function(func, input_names, output_names, large_model): | |||
There was a problem hiding this comment.
maybe use large_model=None since one might call this from some random python code
There was a problem hiding this comment.
other than that looks super lgtm.
TomWildenhain-Microsoft
force-pushed
the
tom/ConvertLargeModels
branch
from
4bf08f4 to
297352a
Compare
|
Are there any plans for this support to be added to TF 1.X aswell? |
|
I believe converting saved models for TF 1.X should already support large models though I haven't tested it. Do you have a specific model you would like to convert? If so, please open a GitHub issue. |
|
I've tried but i get the error: "ValueError: You passed in an iterable attribute but I cannot figure out its applicable type". It's odd because it seems to compile fine until i get the protobuf size error when i don't pass --saved_model. I do have a specific model in mind. I'm fairly new to ML, so i suppose it could be user error aswell. |
|
Can you make a new GitHub issue provide the full error output? Issue link: |
|
Just added it |
|
If this is fixed, why am I seeing it in TF 2.4.1? see tensorflow/recommenders#274 (comment) |
|
Hi @dgoldenberg-audiomack, are you trying to convert your model from tensorflow to onnx? This repo is the tf2onnx repo. Our fix is a workaround/hack that lets us convert large TF models. It isn't a fix to tensorflow itself. |
No description provided.