How can I implement an early stopping criteria - Tensorflow Object Detection API

Question

Fairly new to the Object Detection API here, using tf-gpu==1.15 for training and 2.2.0 for evaluation as well as python 3.7.

I am able to utilize data augmentation as well as adjust the decay of the learning rate in the ssd_mobilenet_v1.config file, but I am not sure how to go about implementing a way for the model to stop training if I am confident the loss will not get below a certain value no matter how many more steps it trains.

How or where do I configure / implement early stopping criteria?

score 0 · Answer 1 · answered Jul 16 '20 at 15:28

0

You can make use of EarlyStopping callback from tensorflow. Refer tensorflow documentation here.

Create a callback using -

from tensorflow.keras.callbacks import EarlyStopping

es = EarlyStopping(
    monitor='val_loss', min_delta=0, patience=3, verbose=0, mode='auto',
    baseline=None, restore_best_weights=False
)

monitor - Change the monitor param to loss if u don't have validation data & want to stop training if training loss stops decreasing for 3 consecutive epochs. Also, u can change it to accuracy if u wish to stop training based on accuracy instead of loss.
patience - Number of epochs to wait before stopping the training if loss doesn't decrease or accuracy doesn't improve

Then pass it to fit function while training -

history = model.fit(X, y, epochs=10, callbacks=[es])

There are many examples online, u can check them!

answered Jul 16 '20 at 15:28

Aditya Mishra

1,687
2
15
24

Ok thank you, but where in the object detection API would it make sense for me to implement this? Not too familiar with tf callbacks. – anonymous Jul 16 '20 at 17:02
You can define callbacks anywhere in the script but before calling `fit` method. Usually people define the model, compile it, create callbacks if required & then call the fit method on it. – Aditya Mishra Jul 16 '20 at 18:34
The script used to train, like a train.py? – anonymous Jul 16 '20 at 18:58
Yeah, like train.py – Aditya Mishra Jul 16 '20 at 19:28
3

but there is no model.fit() method in tensorflow object detection. Any suggestion where to call the callback in train.py of tensorflow object detection ? – vishal yadav Nov 16 '20 at 15:45

How can I implement an early stopping criteria - Tensorflow Object Detection API

1 Answers1