kirvlasenkov / doggy_transfer_learning

job interview task (Sber or Alpha - idgaf)

doggy_transfer_learning

Solution for ImageWoof dataset classification problem using transfer Learning technique.

Requirements

pip3 install torch torchvision tqdm numpy pandas matplotlib seaborn

or

pip3 install -r requirements.txt

Usage

Just install jupyter notebook and have fun.

Description of a problem

Imagewoof is a subset of 10 classes from Imagenet that aren't so easy to classify, since they're all dog breeds. The breeds are: Australian terrier, Border terrier, Samoyed, Beagle, Shih-Tzu, English foxhound, Rhodesian ridgeback, Dingo, Golden retriever, Old English sheepdog.

There you can see some examples from this dataset (already with augmentation).

Description of a training process

As this dataset is not so easy to classify even for humans' eyes, I've decided to use pretrained ResNet50 model as a basic approach. The training data was also augmented and other basic methods were also made. In addition, but just for fun, I've decided to use SGD+Momentum as an optimization method, because Adaptive gradient algorithms you can see in any other solutions :) (Really, I've tested AdamW on this problem and I have not noticed any improvement in convergence rate as an instance.)

Whereupon, there are two general steps:

• Firstly, ResNet was used as a feature extractor, i. e. all the layers of the net were frozen, the FC was cut off and replaced with a linear layer with required number of out_features. This stage of training lasted for 25 epochs.

• Secondly, It's the time for model fine-tuning, but with discriminative learning rates technique. The network trained for 20 epochs.

Result

Below you can see a plot of an accuracy score for finetuned model on a validation data. In my opinion, it's a good score for such a small time of learning. But before fine-tuning, when the model was only a feature extractor, accuracy on a test a data was about 0.896156. I guess, learning only one layer (head) and freezing of a backbone have some limit value of accuracy and for further improvment it's necessary to slightly correct weights on all layers of NN.

Also, I'm sure that it's possible to find more stable and powerful model for this task, but generally it's just a question of a time. Maybe EfficientNet-b1 or InceptionV3 are our future friends?