TensorFlow Validation #3
Conversation
| # device was specified, then reload the model on the correct device. | ||
| # This is a bit awkward but I wanted to avoid all the tensorflow initialization | ||
| # stuff before loading a module, which could potentially be a PyTorch module. | ||
| # So we only get into the tensorflow device stuff if we found a tf module. |
There was a problem hiding this comment.
This whole device argument piece makes things a bit awkward when working with both tensorflow and pytorch models. One option would be to add another positional arg for the framework: "torch" or "tf". And then only import the framework that is required. That could simplify things. That being said, this is working fine for now.
There was a problem hiding this comment.
Where is the current self.model_type thing documented / called by the user?
There was a problem hiding this comment.
Maybe you can explain to me better about device differences in pytorch and tf
There was a problem hiding this comment.
The model_type isn't called by the user -- it is inferred based on the type of object returned by load_model. This will probably need to be updated to accept other objects like a kera.Model, etc... Alternative would be to change this to a command line arg and only import the required framework.
Adding support for TensorFlow model validation.