Solution of the MSPD by FU-Berlin

by Nils Seitz, Luis Schulze, Alexander Schwarz

If you need help, send a mail to: [email protected]

Setup

1. Get testcases from original contest repo (and put inside contest/testcases)
2. Install necessary packages (conda create --name <env> --file requirements.txt, pip install -r requirements.txt)
3. Run inference (see Use)

Packages used

In case requirements.txt doesn't work, a quick list of necessary packages:

• spektral (GNN)
• numpy (arrays)
• pandas (dataframes)
• tensorflow (ML/NN)
• sklearn (specifically: clustering)

You should install them in this order, as e.g. spektral depends on TF and numpy.

Use

1. cd contest
2. python -m inference

Explanation of Code

... coming soon

Weight Issue

If results are of you might need to retrain weights shortly (local).
To do this, a train.py is provided in contest/utils, which is run with cd contest/utils && python -m train It generates weights to contest/utils/weights/best_weights.h5, which you can then copy to contest/weights (and rename to model_{obj_num}.h5) such that they get used in the inference.py.

At the top of train.py, there are adjustable settings:

DATA_PATH = './testcases'
SAVE_PATH = './weights'
params = {
    'classes' : 1,
    'col_name' : 'rating',
    'obj_num' : 3,
    # start: 0,
    'norm' : True,
    'TIMEIT': 1,
}

Here, you can also modify data and save paths. Note that you can adjust the 'obj_num', depending on which model_{obj_num}.h5 you want to (re-)train. Note also, that the column name is 'rating', which is a custom row we calculated from the given dataframes. This column norms all values (each source combination per net) for a given objective, e.g. 3W+S, between 0 and 1, such that 1 is the best source combination and 0 the worst.

In case you are interested in that data, you can send a mail - it's too large for GitHub.

You can basically adjust anything in the constructor of MSPDMixedNet and this function:

def calc_y(self, src_idx_enum):

    src_row = self.graph_df[self.graph_df['sourceIdx'] == src_idx_enum]
    s = (src_row[f'skew{self.objnum}'] - self.s_min) / self.s_max
    w = (src_row[f'wireLength{self.objnum}'] - self.w_min) / self.w_max
    #   return (1-w)

    x = np.array(src_row[[self.objective]], dtype=np.float32).reshape(1, )
    return x / 100
    # MIN_VAL = 50
    # x = (x-MIN_VAL)/ (100-MIN_VAL)
    #
    # return np.where(x > 0 , x , 0)

The logic should match with the col_name you chose for training in params. Additionally, if choosing something not normed, you would need to adjust the specific part in inference.py as well, since e.g. its written to use normed values, etc.