How can i change the mini batch size when training my own dataset ?

[jihadbourassi]

Hello I am trying to train the Mask RCNN model for licence plate detection i followed the ballon detection examlpe and strated training without changing any of the hyperparameters (except multiprocessing=false since i am using cpu) by observing the training i noticed that the loss function started high in the first epochs and goes up and down a lot from one epoch step to another, and it only got down to 0.70 by the 23th epoch (out of 30) with an above 1 val_loss value still, so i figured i should increase the batch size (left batch size=1) since my image dataset is pretty diverse in order to reduce the variance of the loss function but i can't seem to do so since everytime i change the batch size on config.py and launch training it shows me BATCH SIZE=1 each time, i currentlly restarted training with the weights generated from the 23th epoch.

I am new to neural networks so bare with me please any help will be much appriciated, thank you all very much and have a good day.

[freezurbern]

I did not understand the BATCH_SIZE variable so instead, I used more STEPS_PER_EPOCH and VALIDATION_STEPS.

Try using say 300 steps and 50 validation steps. This allows more data per epoch before validating each time.

[cpoptic]

You can set the batch size in your config file, under class Config, in config.py. It's config.IMAGES_PER_GPU.

Actually your batch size, config.BATCH_SIZE is derived by multiplying config.IMAGES_PER_GPU * config.GPU_COUNT. So if you're working on a single GPU, and you've set your config.GPU_COUNT to 1, then BATCH_SIZE should be equal to IMAGES_PER_GPU.

[Prithvi0167]

When i change the value of IMAGES_PER_GPU and GPU_COUNT in config.py,still there is no change in the batch size.Please help ,what can i do for this

[ammarsyatbi]

@Prithvi0167 it is because the BATCH_SIZE is only assigned during initialization, you can see it in init function in config.py on line 216 self.BATCH_SIZE = self.IMAGES_PER_GPU * self.GPU_COUNT

To re-calculate batch size after you declare in your config

    config = YourConfig()
    config.BATCH_SIZE = config.IMAGES_PER_GPU * config.GPU_COUNT

or alternatively, you can just directly declare it in your config