naman9639 / aircraft_classification

Aircraft Classification using PyTorch

The dataset should be downloaded and kept in the following structure for the code to run properly. Use the rename_move.py file to rename all the images and move the images into the desired destination.

  -/FGVC-Aircraft/data/
                  └─── images
                          └─── 0034309.jpg
                          └─── 0034958.jpg
                          └─── ...
                  └─── variants.txt
                  └─── images_variant_trainval.txt
                  └─── images_variant_test.txt

The concept used behind the methodoly is as follows:

• We first train the model using images_variant_train.txt and variants.txt file. We use CNN to classify the images of the aircrafts into varients. Once the varient for a particular image is recieved we can directly map it to it's family.
• A dictionary is created that stores the family of the aircrafts and their varients in key:value format. We can use this dictionary to get the mapping between the varient and the family which is our final goal. The varient recieved from the above CNN can be used to get the family using this dict.

Note - The dictionary created above uses the images_variant_train.txt and images_family_train.txt files to map variants to families.

I have used transfer learning and Inception net for training the neural network. The validation accuracy achieved is 93% using WS-DAN method. WS-DAN also knows as Weakly Supervised Data Augmentation Network (WS-DAN) explores the potential of data augmentation. Specifically, for each training image, a attention map is generated to represent the object’s discriminative parts by weakly supervised learning. Next, the image is augmented guided by these attention maps, including attention cropping and attention dropping. WS-DAN improves the classification accuracy in two folds. In the first stage, images can be seen better since more discriminative parts’ features will be extracted. In the second stage, attention regions provide accurate location of object, which ensures our model to look at the object closer and further improve the performance.

We can improve the performance by increasing the number of epochs and the batch size. The batch size used was 12 on Titan Xp GPU but we can increase the batch size which will further improve the overall performance of the model. Not only that, I have just tried Inception net to train the model but a lot of other pretrained networks like resnet or vgg16 can be used and then we can compare the best architecture for solving this problem

Run

1. git clone this repo.

2. Prepare data and modify DATAPATH in datasets/aircraft_dataset.py.

3. Set configurations in config.py (Training Config, Model Config, Dataset/Path Config):

4. $ nohup python3 train.py > progress.bar & for training.

5. $ tail -f progress.bar to see training process (tqdm package is required. Other logs are written in <config.save_dir>/train.log).

6. Set configurations in config.py (Eval Config) and run $ python3 eval.py for evaluation and visualization.

aircraft_classification's Issues

run time error

what does it mean for the line 305 in train.py ?

result_arr.append((y_pred, d[y_pred]))

d[y_pred] is not corret. Should it be result_arr.append((y_pred, y)) instead?

folder should be updated and some files missing

Hi there, I run the code and find that the folders miss the 's' for both 'model' and 'dataset'. That causes complile failure. I guess you rename the folder but not updating the code? just like line 9 of wsdan.py.
What's more, there is no models.vgg and models.resnet defined. Are some files missing?
Thanks.