Current Implementation of 3DReconNet (MultiScale2DConcat) is validated for 64^3 and 128^3 output volume. This needs to be parameterized/flexible for future experiments.
Existing
# 128^3 output volume.128^2 input x-rays
model_config = {
"permute": True,
"dropout": 0.1,
"encoder": {
"initial_channel": 16,
"in_channels": [], # this will be filled in by autoconfig
"out_channels": [2, 4, 8, 16, 32, 64],
"encoder_count": 4,
"kernel_size": 3,
"act": "RELU",
"norm": "BATCH",
},
"decoder_2D": {
"in_channels": [], # this will be filled in by autoconfig
"out_channels": [4, 8, 16, 32, 64, 128],
# "out_channels": [2, 4, 8, 16, 32, 64],
"kernel_size": 3,
"act": "RELU",
"norm": "BATCH",
},
"fusion_3D": {
"in_channels": [], # this will be filled in by autoconfig
"out_channels": [32, 32, 32, 32, 32, 32],
"kernel_size": 3,
"act": "RELU",
"norm": "BATCH",
},
}
# 64^3 output volume, 64^2 input x-ray
model_config = {
"permute": True,
"dropout": 0.1,
"encoder": {
"initial_channel": 16,
"in_channels": [], # this will be filled in by autoconfig
"out_channels": [4, 8, 16, 32, 64],
"encoder_count": 4,
"kernel_size": 3,
"act": "RELU",
"norm": "BATCH",
},
"decoder_2D": {
"in_channels": [], # this will be filled in by autoconfig
# "out_channels": [8, 16, 32, 64, 128],
"out_channels": [4, 8, 16, 32, 64],
"kernel_size": 3,
"act": "RELU",
"norm": "BATCH",
},
"fusion_3D": {
"in_channels": [], # this will be filled in by autoconfig
"out_channels": [32, 32, 32, 32, 32],
"kernel_size": 3,
"act": "RELU",
"norm": "BATCH",
},
}
As shown above, current implementation take configuration dict as parameter whose parameters have to be carefully selected to make sure the network pipeline is valid.
Is it even possible to have input/output size other than power-of-2, say 96^2 or 196^2?
Description
Current Implementation of 3DReconNet (MultiScale2DConcat) is validated for 64^3 and 128^3 output volume. This needs to be parameterized/flexible for future experiments.
Existing
As shown above, current implementation take configuration
dictas parameter whose parameters have to be carefully selected to make sure the network pipeline is valid.Is it even possible to have input/output size other than power-of-2, say 96^2 or 196^2?
Files
XrayTo3DShape/architectures/twoDPermuteConcatMultiScale.py tests/test_multiscale_permuteconcat.py
Tasks