Merger Challenge

William J. Pearson

Code for William J. Pearson's entry to the merger challenge.

Method

Architecture

The swin_tiny_patch4_window7_224_fe Swin Transformer of Sayak Paul that was pre-trained on the ImageNet-1k dataset was used. A fully-connected layer with 256 neurons is added along with an output later of 1 neuron (binary) or 4 neurons (quad).

Training

1. TNG100 training Images are cropped to 128x128 pixels with split_crop_train.py
2. Images are split ~80%-20% for training-validation, keeping images from the same merger history in each split
3. In the training scripts ( TNG100-swin-transfer-###-cut.py ) images are:
   1. Cropped again to 112x112 pixels
   2. Scaled between 0 and 1
   3. Stacked with itself to form a 3 channel image
   4. Scaled to 224 with nearest neighbour
4. During training, training images (but not validation images) are:
   • Randomly rotated by 90^o AND
   • Randomly flipped left/right AND
   • Randomly flipped up/down
5. The binary model with the lowest vaidation loss and the quad model with the lowest validation loss and highest validation accuracy are saved

Testing

1. TNG100 test, HSC and Horizon-AGN images are cropped to 128x128 with split_crop_test_TNG100.py or split_crop_train.py (for HSC and Horizon-AGN)
2. In the test scripts ( TNG100-swin-transfer-###-cut-predict.py ) images are:
   1. Cropped again to 112x112 pixels
   2. Scaled between 0 and 1
   3. Stacked with itself to form a 3 channel image
   4. Scaled to 224 with nearest neighbour