This is my first attempt at dealing with 3D data. The method used was a simple 3D convolution that downsampled the voxelgrid to a lower dimension space with a consequent fully connected sub-network. It really is a simple architecture but the technology has multiple real-world use cases, one prominent example is 3D Medical Imaging.
3D MNIST. The description of the data and helper code can be found on Kaggle itself, where the data was originally hosted. I have currently only used the full_dataset_vectors.h5 file and not the point clouds. The vectors were used to create a voxelgrid of dimensions (16,16,16,3).
main.ipynb- Training scriptmodel.py- Model architecture for the 3D-CNN.utils.py- Helper functions
The following text was created using torchsummary.
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv3d-1 [2000, 8, 14, 14, 14] 656
Conv3d-2 [2000, 32, 12, 12, 12] 6,944
MaxPool3d-3 [2000, 32, 3, 3, 3] 0
Conv3d-4 [2000, 32, 1, 1, 1] 27,680
BatchNorm3d-5 [2000, 32, 1, 1, 1] 64
Dropout-6 [2000, 32] 0
Linear-7 [2000, 10] 330
================================================================
Total params: 35,674
Trainable params: 35,674
Non-trainable params: 0
----------------------------------------------------------------
The pre-trained model can be found here
- Training and Testing losses
- Training and Testing Accuracies
As observed in the plots above, the model has converged with an accuracy of 67% on the training data and 65% on the test data. The training was carried out for 160 epochs with a learning rate of 0.00005 using an Adam Optimizer.
The testing accuracy can be increased by about (1.5-2.0)% upto ~67% but only at the cost of the model overfitting the training data. I analysed the results for 300 epochs with the same hyperparameters and the training accuracy increased upto 73% but there was a clear divergence observed between the training and testing loss at the latter end of the training.
- PyTorch 1.4.0
- Python 3.7
- Try using a point cloud based approach.
I am not actively working on this, it was more like a side project so I don't really know when I'll get to do this.