Resnet 18 - Custom PyTorch implementation

IndTechSpprt / Split_inference_microcontrollers_fork

This is the repository for my research internship with the Embedded Systems Group at TU Delft, based on Junyu Lu's master thesis - https://github.com/Eagleggs/Split_inference_microcontrollers

0 stars 0 forks source link

Resnet 18 - Custom PyTorch implementation #33

Closed IndTechSpprt closed 2 weeks ago

IndTechSpprt commented 3 weeks ago

ResNet18 has the following layers we currently don't support

ReLu
MaxPooling
Adaptive average pooling

Keeping this mind, make a custom implementation based on the following:

ReLu6 instead of ReLu
Average pooling using previous input and output instead of Adaptive Average Pooling

Compare this with the PyTorch implementation and report the observations

IndTechSpprt commented 2 weeks ago

As seen in the (generated by running resnet18-custom.py), the outputs post switching the layers out are within reasonable bounds, so ReLU can be replaced with ReLU6 and AdaptivveAvgPool2D can be replaced by AvgPool2D with a 7x7 kernel and a stride of 512.