The kaggle_protein_classification from xuzhengxiao

kaggle_protein_classification's Introduction

This project is code of Human Protein Atlas Image Classification competition

Getting Started

dataset.py is self customed data loader solution independent of Deep learning frameworks.When applied in ensemble model,the last model's memory can't be released completely,thus causing increasing memeory usage.So I finally replace it with pytorch's data loader in pytorch_dataset.py.
resnet.py trains model.
transforms.py. image augmentation.
utils.py handles csv file,calculates class weight and splits train and validation data.
predict.py predicts test data

Some appcoaches taken

learning rate schedule is warm restarts,which is described in SGDR: Stochastic Gradient Descent with Warm Restarts.
bce loss combined with log-dampned weight works better.

Attention

If your hardware can't deal with large batch,you can try accumulating gradient tech.

for iteration,(images,labels) in enumerate(train_loader):
   images = images.cuda()
   labels = labels.cuda()
   inputs = Variable(images)
   targets = Variable(labels)
   outputs=model(inputs)
   loss=criterion(outputs,targets)
   # loss regularization
   loss = loss/accumulation_steps   
   loss.backward()
   if(iteration%accumulation_steps)==0:
       optimizer.step()        # update parameters of net
       optimizer.zero_grad()   # reset gradient

we trained our model only in train data,no external data nor leaked data. External data can boost about 0.1 in lb.However,someone in competition's discussion claimed that he could achieve 0.6 without external data,so we didn't take external data into consideration.
our final result ends with public lb 0.474,private lb 0.443,wihch is obvious not good.Because this is our first competition,we do not konw lots of techniques useful in such competition.We still have a long way to go.

Recommend Projects