kaggle website: https://www.kaggle.com/c/cs-ioc5008-hw1/overview
Brief introduction:
The task of the homework is the scene recognition. I train a convolution neural network using transfer learning to complete this competition. Moreover, I do some data augmentation to improve accuracy. And I get the final score of 0.97211 in Kaggle competition.
Methodology:
I choose “ResNext-101-32x8d” as my CNN model because it has the lowest classification error rate in ImageNet dataset. And I think that the tasks of these two datasets are slightly different, so when I do transfer learning, I don’t fix any layer. I just use pre-train model and retrain it. Besides, my input size is 256x256 because it is the size of most training set.
At first, I get the accuracy of 0.96346. Then I make use of data augmentation by random horizontal flip with 0.5 probability and obtain 96.826% accuracy. Finally, I also use random rotation with 7 angle degree to get the score of 0.97211.
The loss function is cross-entropy loss, and the optimizer is SGD. The following table is about the value of some hyperparameters:
| Hyperparameter | Batch size | Learning rate | Momentum |
|---|---|---|---|
| value | 20 | 0.001 | 0.9 |