How to increase accuracy of model from tensorflow model zoo?

Situation:

My dataset is 70k images of people wearing clothes. Images are labelled: bbox position and class. There are 10 classes. I did 80:20 split. Categories are balanced with exception of one category, but I can accept poor performance on one category.

The goal is cloth recognition in images. When I feed an image of a person wearing pants and a t-shirt, I want to see two bboxes of these clothes.

My problems:

I already trained a few models from tf model zoo. I did over 100k steps on ssd mobilenet v1 and faster rcnn resnet 101.

The problem with ssd is that it won't converge. Loss is not getting below stable 2 and accuracy is bad. The problem with faster rcnn is that loss is below 1 but it's varying a lot and sometimes it jumps over 1.

What I've done:

I tried different batch sizes for ssd with no luck. FRCNN is locked with batch size 1. I improved the dataset multiple times. I went from 50 unbalanced classes to 10 balanced classes. I didn't tweak hyperparameters from models configs besides batch size.

My access to strong GPU is limited for me so I can't just randomly try different hyperparameters combinations with the hope that it will work. Could you suggest me few things that I can do in order to improve my models? I would be very thankful.

Topic object-detection convolutional-neural-network tensorflow deep-learning machine-learning

Category Data Science

1
answered at

Are you training the models from scratch? If yes, then can you try using pre-trained models and fine-tune for your specific dataset.

You'll have the experiment with hyperparameters. Learning rate and optimizer (e.g., sgd, adam, rmsprop, adadelta) will have largest effect on model training and performance.

About

Geeks Mental is a community that publishes articles and tutorials about Web, Android, Data Science, new techniques and Linux security.