AutoML-Kaggle-Competition/description.txt at master · leihchen/AutoML-Kaggle-Competition · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
The main function is in version2.py and parsing helper functions are in parser_1.py.
We parse the structure of the neural network and then reconstruct the pytorch model.
Our constructed feature and original feature are included in 'data matrix.csv.'
Then we use xbgboosting model to fit the data by different weight.
We tune the paramters using GridSearchCV.

List of constructed feature:
tot_flops: Number of FLOPS involved in neural network provided.
ReLU...MaxPool2d: Counter of all operation in neural network provided.
init_b_mu_Conv2d...init_b_l2_BatchNorm2d: Sum of all initial params for different operations.
val_accs_diff0...val_accs_diff6: Differences between val_accs over all epochs averaged over a 7 * 7 grid.
val_losses_diff0...val_losses_diff6: Differences between val_losses over all epochs averaged over a 7 * 7 grid.
train_accs_diff0...train_accs_diff6: Differences between train_accs over all epochs averaged over a 7 * 7 grid.
train_losses_diff0...train_losses_diff6: Differences between train_losses over all epochs averaged over a 7 * 7 grid.