Train on custom dataset

[robosina]

Today I tried to use this implementation on custom dataset, as you said in readme file for training on custom dataset we have to add path to training dataset by -r flag , but in the train.py there is no such flag? So could you please help me about how to train efficientnet-b0 on custom dataset?

[narumiruna]

It seems that I forgot to update README.md file. Please change the path to the training dataset in the config file.

[robosina]

Thanks , did you mean to add new CustomDataLoader or we can just set path of training data?

[narumiruna]

If the directory structure of your dataset is the same as the ImageNet, then you can just set path of training data. Otherwise, you should add a new custom data loader.

[robosina]

@narumiruna ,Thanks I add my dataset to the data folder and training is started, but did you get good performance on custom dataset? I have a dataset of two class with images with size 44, but my accuracy don't increase from 78%.

[narumiruna]

Maybe you can disable the augmentation by replacing

efficientnet-pytorch/efficientnet/datasets/imagenet.py

Lines 37 to 50 in 42e95f5

	if train:
	transform = transforms.Compose([
	transforms.RandomResizedCrop(image_size, interpolation=Image.BICUBIC),
	transforms.RandomHorizontalFlip(),
	transforms.ToTensor(),
	normalize,
	])
	else:
	transform = transforms.Compose([
	transforms.Resize(image_size + 32, interpolation=Image.BICUBIC),
	transforms.CenterCrop(image_size),
	transforms.ToTensor(),
	normalize,
	])

with

transform = transforms.Compose([
            transforms.Resize(image_size + 32, interpolation=Image.BICUBIC),
            transforms.CenterCrop(image_size),
            transforms.ToTensor(),
            normalize,
        ])

or use less augmentation.

[robosina]

Thanks, I will test it