Support deploying Saved Tensorflow Models#328
Conversation
|
Can one of the admins verify this patch? |
|
jenkins test this please |
|
Test PASSed. |
|
@dcrankshaw supports serving Frozen tensor-flow models as well now. Can you pls review |
|
jenkins ok to test |
| if (len(is_frozen_graph) > 0): | ||
| os.makedirs(os.path.join(serialization_dir, "tfmodel")) | ||
| try: | ||
| shutil.copyfile( |
There was a problem hiding this comment.
@kakru's suggestion
I'd suggest using copytree instead of copyfile to keep the behavior consistent between model types (cf. line 207). Possible problem might arise if there is more than one file in the saved_model_path (check is for >0) and only the first .pb file is copied
|
Test PASSed. |
|
@dcrankshaw fixed the issue based on comment from @kakru |
|
Test FAILed. |
|
Test FAILed. |
|
@dcrankshaw i guess the tests failed due to some other sporadic error.Can you please re-trigger the build. |
|
Test PASSed. |
|
Test PASSed. |
3 participants
Current deployment of tensorflow models only supports deployment by passing the tensorflow session as a parameter.It should be possible to deploy any saved tensorflow model by passing the model directory path, like in the tfcifar example.This PR addresses the above issue.