CNN training accelerator

The repo describes an HLS-based CNN training accelerator in floating-point format for a reference design, using the back-propagation algorithm with the SGD optimizer.

Structure

pytorch- Verification of back-propagation derivatives, including Conv, transposed Conv, dilated Conv, BN, ReLU, average pooling, and FC.

BP_function.py
fc_test.py

resnet20- HLS design of the accelerator with both input and output channel tiling.

design source files

bnn.h
conv_weights.h
dimension_def.h
layer.h
resnet20.cc or vgg.cc
typedefs.h

testbench files

conv_weights_tb.h
tb.cc
weights_tb.h

data_batch_1.bin
train.bin
(image data from http://www.cs.toronto.edu/~kriz/cifar.html)

Citation

@inproceedings{guo2023boost,
  title={BOOST: block minifloat-based on-device CNN training accelerator with transfer learning},
  author={Guo, Chuliang and Lou, Binglei and Liu, Xueyuan and Boland, David and Leong, Philip HW and Zhuo, Cheng},
  booktitle={2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD)},
  pages={1--9},
  year={2023},
  organization={IEEE}
}