D2NN-with-Pytorch

Environment:

torch==1.12.1 torchvision==0.13.1 numpy==1.23.5 matplotlib==3.7.1 tqdm==4.65.0

---

Reading Sequence

D2NN $\Rightarrow$ Beam-Diffraction $\Rightarrow$ D2NN-plus

D2NN-plus = D2NN + Beam-Diffraction

D2NN

1. Only 5 Phase Layers are concerned.
2. The Beam Propagation Evolution are not concerned.
3. The rate of training is Faster.
4. The Accuracy is limited to 70-80% for 5 diffractive layers.

D2NN-plus

1. The free space is Totally Meshed.
2. The Beam Propagation Evolution are concerned.
3. Only 5 meshed layers are attached with Phase Learning.
4. Similar Accuracy with D2NN.

The accuracy of D2NN model is troubling me for a long time. It seems difficult to be optimized to 90% accuracy.

---

Congratulations!

By using 4F system D2NN, I've got 97% Accuracy for MNIST validation dataset.

Now, I'm trying to apply D2NN on CIFAR-10 dataset with multichannel explorations. God Blessing Me !