I have a custom dataset of 100 image classes with ~1250 images per class in S3. I'd like you to train an inception_v3 model from scratch with these images. I can give you access to my P3.8Xlarge instance on AWS for training purposes. It has 4 V100 GPUs and I think it can complete this training in a couple of days with proper configurations and code.
Dataset might need to be cleaned for a better performance but that's up to you.
1. Final trained models - .data, .meta, .index files of best model.
2. Frozen .pb of the best model mentioned in point 1.
3. Code to do inference on single images and batch images.
1. Preferably top 1 accuracy of 75% and I can't accept if it is less than 70%. The original model has ~75% accuracy on 1000 classes. So, I guess this shouldn't be a problem.
2. Previous knowledge in training inception v3 or similar models in Tensorflow using SLIM. Keras as frontend is also acceptable but I'd like it in SLIM.
3. Send me any of the links to your GitHub or StackOverflow.
Slim code to follow: [url removed, login to view]