Autogluon: Missing example to save and reload ObjectDetection model

[rebotnix]

I installed the requirements and the demo for object detection works fine.

Is there a way how I can save and reload a trained model after the task.fit call?

detector = task.fit(dataset_train,
num_trials=2,
epochs=epochs,
lr=ag.Categorical(5e-4, 1e-4),
ngpus_per_trial=1,
time_limits=time_limits)

I also tried to set a resume to False to initialize the task to predict directly, but training is always started from the beginng.

?

[zhanghang1989]

That's good question. We should improve the documentation.

When calling task.fit(...), the training will start. If resume=True, it will start will the existing search history.

If you would like to save/load the trained model, please use detector.save('detector.ag') and detector= Detector .load('detector.ag')

[rebotnix]

Thanks @zhanghang1989 for your response. It seems that save method in detector is not implemented yet?

I tried to call the save method after training, but receive an error.

# start training
detector = task.fit(dataset_train,
                    num_trials=2,
                    epochs=epochs,
                    lr=ag.Categorical(5e-4, 1e-4),
                    ngpus_per_trial=1,
                    time_limits=time_limits)

# save model
detector.save("motorbike.ag")

Error message:
/usr/lib/python3/dist-packages/apport/report.py:13: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses
import fnmatch, glob, traceback, errno, sys, atexit, locale, imp
Traceback (most recent call last):
File "demo_eval.py", line 24, in
detector.save("motorbike.ag")
File "/home/user/.local/lib/python3.7/site-packages/autogluon/task/object_detection/detector.py", line 154, in save
raise NotImplemented
TypeError: exceptions must derive from BaseException

• 
  I take a look at the original detector source of the object detector class and load and save are not implemented yet.

    def load(cls, checkpoint):
        raise NotImplemented

    def save(self, checkpoint):
        raise NotImplemented

[zhanghang1989]

Thanks for the feedback @rebotnix . We will prioritize the implementation of save/load method.