Releases: jedlimlx/Muon-Tomography-AI
µ-Net Weights (Keras 3)
The model weights from the paper µ-Net: ConvNext-Based U-Nets for Cosmic Muon Tomography. The model is implemented using TensorFlow and trained using 2 T4 GPUs. These weights are updated to Keras 3, the previous release was for Keras 2.
Note: these weights have been trained for 20 epochs, not 15 epochs like what is mentioned by the paper. For some reason, the Keras 3 weights perform very slightly worse than the Keras 2 weights.
For the tiny model size,
model = Agg3D(
**{
'point_size': 1,
'downward_convs': [1, 2, 3, 4, 5],
'downward_filters': [8, 16, 32, 64, 128],
'upward_convs': [4, 3, 2, 1],
'upward_filters': [64, 32, 16, 8],
'resolution': 64,
'threshold': 1e-8
}
)
For the base model size,
model = Agg3D(
**{
'point_size': 1,
'downward_convs': [1, 2, 4, 4, 6],
'downward_filters': [16, 32, 64, 128, 256],
'upward_convs': [4, 4, 2, 1],
'upward_filters': [128, 64, 32, 16],
'resolution': 64,
'threshold': 1e-8
}
)
For the large model size,
model = Agg3D(
**{
'point_size': 1,
'downward_convs': [1, 2, 4, 6, 8],
'downward_filters': [24, 48, 96, 192, 384],
'upward_convs': [6, 4, 2, 1],
'upward_filters': [192, 96, 48, 24],
'resolution': 64,
'threshold': 1e-8
}
)
µ-Net Weights
The model weights from the paper µ-Net: ConvNext-Based U-Nets for Cosmic Muon Tomography. The model is implemented using TensorFlow and trained using 2 T4 GPUs.
For the tiny model size,
model = Agg3D(
**{
'point_size': 1,
'downward_convs': [1, 2, 3, 4, 5],
'downward_filters': [8, 16, 32, 64, 128],
'upward_convs': [4, 3, 2, 1],
'upward_filters': [64, 32, 16, 8],
'resolution': 64,
'threshold': 1e-8
}
)
For the base model size,
model = Agg3D(
**{
'point_size': 1,
'downward_convs': [1, 2, 4, 4, 6],
'downward_filters': [16, 32, 64, 128, 256],
'upward_convs': [4, 4, 2, 1],
'upward_filters': [128, 64, 32, 16],
'resolution': 64,
'threshold': 1e-8
}
)
For the large model size,
model = Agg3D(
**{
'point_size': 1,
'downward_convs': [1, 2, 4, 6, 8],
'downward_filters': [24, 48, 96, 192, 384],
'upward_convs': [6, 4, 2, 1],
'upward_filters': [192, 96, 48, 24],
'resolution': 64,
'threshold': 1e-8
}
)
$\mu$-Net weights
For a dosage of 1024, 2048 and 4096, use the following configuration with a point size of 3.
For a dosage of 8192 and 16384, use the following configuration with a point size of 1.
{
'point_size': <insert-point-size-here>,
'downward_convs': [1, 1, 2, 3, 5],
'downward_filters': [8, 16, 64, 128, 256],
'upward_convs': [4, 3, 2, 1],
'upward_filters': [128, 64, 16, 8],
'resolution': 64,
'noise_level': 0,
'threshold': 1e-3
}