Training the Regressor

Four steps, in order. The first is a long multi-GPU job, so read the constants at the top of collect_data.py before starting it. With a collected coverage dataset, start at train.

python -m intelx.cli.collect_data    # coverage data, over your maps
<CHUNKING_SCRIPT>                    # the collected shows, as training chunks
python -m intelx.cli.train           # the U-Net
python -m intelx.cli.evaluate        # what it predicts, side by side

collect_data runs the 18 training maps listed in data/train_metadata.csv, each with the number of shows Table 1 of the paper gives it, in proportion to its complexity score. It reads the maps from train/ under the maps root, samples spawns, runs the exploration shows that record how much new area each goal would gain, and writes the result under INTELX_DATASET_ROOT. intelx.data.driver does one map at a time and is what the controller drives. <CHUNKING_SCRIPT> packs the collected shows into train_chunks/. train reads those chunks, at batch size 32 for up to 300 epochs, keeps the checkpoint with the best validation R², and writes each run, with the settings it used, to weights/runs/<name>. evaluate loads a checkpoint and writes prediction overlays and inference timings to results/evaluation_large.

Everything each step reads and writes is a class attribute at the top of its file. None of the three takes command-line arguments.