How to import the saved/trained model to test on a testing dataset?

[Smallyolk2024]

Hi, @cweill I have trained a model on my datasets with NASNet-A as the subnetwork, and have saved some checkpoints. If I want to import the saved model (model.ckpt**) again to test on a testing dataset, what should I do?

[cweill]

@SmallyolkLiu: Did you find the answer to your original question?

[Smallyolk2024]

@cweill Not yet, I have an more interesting issue instead, that is , if the program suddenly breaks down, how to rerun it based on the saved results previously? For the case of NASNet as a subnetwork, many many training steps (hundreds of thousands) are necessary, it will a problem in the absence of an effective solution to multi-GPU training.

[cweill]

If you want to use the latest checkpoint, just recreate the same adanet.Estimator with the same model_dir. When you call Estimator#predict it will load the variables from the latest checkpoint.

If you want to load a previous checkpoint, there is a checkpoint file created automatically in your model_dir.

It should look something like:

model_checkpoint_path: "model.ckpt-400000"
all_model_checkpoint_paths: "model.ckpt-270000"
all_model_checkpoint_paths: "model.ckpt-280000"
all_model_checkpoint_paths: "model.ckpt-290000"
all_model_checkpoint_paths: "increment.ckpt-3"
all_model_checkpoint_paths: "model.ckpt-300000"
all_model_checkpoint_paths: "model.ckpt-310000"
all_model_checkpoint_paths: "model.ckpt-320000"
all_model_checkpoint_paths: "model.ckpt-330000"
all_model_checkpoint_paths: "model.ckpt-340000"
all_model_checkpoint_paths: "model.ckpt-350000"
all_model_checkpoint_paths: "model.ckpt-360000"
all_model_checkpoint_paths: "model.ckpt-370000"
all_model_checkpoint_paths: "model.ckpt-380000"
all_model_checkpoint_paths: "model.ckpt-390000"
all_model_checkpoint_paths: "model.ckpt-400000"

Just update model_checkpoint_path to point to the checkpoint you want the Estimator to read and continue from.

[Smallyolk2024]

@cweill Thanks.

[AAbusnina]

My problem is not mainly loading the estimator, actually it is creating it properly and using it I tried to do what is suggested here https://github.com/tensorflow/adanet/issues/22#issuecomment-437567652, but my generator is empty. Can you please explain this the correct approach to make prediction on test data in the tutorial, that will help alot? Big thanks

[AAbusnina]

May be it is simpler if you can tell how to load a saved model?