futurelab-ai-2018-image

The code for FutureLab.AI 2018 scene image classification competition. Final score on TestB: 0.99186.

0. Requirements

• Ubuntu 16.04 or Ubuntu 14.04
• CUDA 8.0
• cuDNN 6.0
• Python 3.6 (Miniconda 3)

Use the following command to install Python 3 dependices:

pip install -r requirements.txt

1. Training-validation splitting

sh scripts/split_train_validation.sh

The script will split the training-list-0511.csv into 11 folds and then combined them into 11 different train-val splitting, like 11-fold-x-train.csv and 11-fold-x-val.csv in train_data_lists folder. For each splitting, the ratio of training set and validation set is 10:1. If time is enough, a full cross-validation should be employed. However, we only used 11-fold-1-train.csv and 11-fold-1-val.csv actually.

2. Train

Based on the pretrained ImageNet models in pretrained-models.pytorch, we trained inceptionv4 and se_resnext101_32x4d with PyTorch 0.3.1. Thanks for @Cadene's such a great repo, all pretrained models can be downloaded automatically when training.

sh scripts/train_inceptionv4.sh

sh scripts/train_se_resnext101_32x4d.sh

The data_root arguments in the above two training scripts should be designated to the training data root, which contains all the training images.

After training, the models with best top-3 accuracy performance will be saved to corresponding directory under checkpoints.

3. Test

(1) Download the models

Download the models we trained from Baidu Netdisk: https://pan.baidu.com/s/1vnUBGrOAaBoiZmCfLxLeSg. Extracting code: 8m5h.

After downloading, extract the zip file to checkpoints folder. The structure of checkpoints you should see:

checkpoints/
|-- inceptionv4_11-fold-1
|   `-- model_best.pth.tar
`-- se_resnext101_32x4d_11-fold-1
    `-- model_best.pth.tar

(2) Test the models

When testing, at first we resize the test image into a designated size, then we employ 10-crop to get a robust test output.

For the two models we trained, we averaged the test logits they output and got the final top-3 predictions.

sh scripts/test_inceptionv4_logits.sh

sh scripts/test_se_resnext101_32x4d_logits.sh

Similarly to training, the test_data_root argument should be designated with the root of images in TestB dataset.

The logits will be saved to corresponding directories under checkpoints, like the follwing structure:

checkpoints/
|-- inceptionv4_11-fold-1
|   |-- model_best.pth.tar
|   `-- test_logits.csv
`-- se_resnext101_32x4d_11-fold-1
    |-- model_best.pth.tar
    `-- test_logits.csv

(3) Average the logits

sh scripts/ensemble_logits.sh

Running this script will get two csv files under ensemble_results folder, which correspond to the averaged logits and the predicted top-3 labels respectively:

ensemble_results/
|-- logits_ensemble.csv
`-- test_pred_ensemble.csv

The test_pred_ensemble.csv is the result what we submit to the competition. This version of models got a 0.99186 score on TestB dataset.