This project implements several network models:
- A modified LeNet-5 [1] with parallel layers and achieves an accuracy of ~99% on the MNIST dataset
- A modified version of this net using an Inception-Layer [2] achieving an accuracy of ~99% on the MNIST dataset
Install all dependencies using the following command
$ pip install -r requirements.txt
[Optional] Start the visdom server for visualization
$ python -m visdom.server
Start the training procedure
$ python run.py MODEL DEVICE
with MODEL=parallenet/mininception and DEVICE=cpu/cuda
See epoch train loss live graph at http://localhost:8097.
The trained model will be exported as ONNX to lenet.onnx. The lenet.onnx file can be viewed with Netron.
[1] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. "Gradient-based learning applied to document recognition." Proceedings of the IEEE, 86(11):2278-2324, November 1998.
[2] Szegedy, Christian, et al. "Going deeper with convolutions." Proceedings of the IEEE conference on computer vision and pattern recognition. 2015.