This is our solution for BAAI Astrodata 2019. We implemented a simple but powerful deep conv model and ranked 2rd in the test set.

• Codes are neat and clean using skorch.
• Using a 6-block-layer Resnet
• Using Squeeze layer for faster convergence
• Using CEloss for training, L1loss for finetuning
• Scores are stable when training

You can refer to the slides here or the Bilibili video here. The Bilibili video quickly goes through the task background and the data distribution. And it also provides an easy-to-start baseline using LightGBM and hand-crafted features.

1. Open the jupyter notebook jupyter lab or jupyter notebook
2. Cleaning the Dataset Run 01*.ipynb. In this notebook, we cleaned the dataset and convert the data type into float32 to save memory usage. Also, we used the binary npy files as the storage format.
3. Training Run 03*.ipynb, In our experiment, we trained 16 models using 4 gpus sharing the same model and hyperparameter. That is, we repeatly run the same code for 16 times to reduce the varience and boost the performance of our models.
4. Inference Run04*.ipynb, We simply infered the result using 1 GPU and average the output from 16 models as our final result. And this leads to our final result - Macro-F1 0.98627835085156.

Haha, it is a solo team: Chengxuan Ying, Dalian University of Technology (应承轩 大连理工大学)