search

DeepWok / mase

Machine-Learning Accelerator System Exploration Tools
Other
105 stars 52 forks source link

Added input channel pre-processing to enable vision models functioning on MNIST #64

Closed mau-mar closed 3 months ago

mau-mar commented 3 months ago

I'm Mauro Marino from Group 0 (Mauro Marino, William Powell) working on Project 1 (TensorRT integration in MASE).

An issue was opened in which errors in processing MNIST dataset with vision models was brought to attention. This is traceable to MNIST dataset being grayscale, hence providing only one channel, whilst MASE convolutional models expect 3 channels.

The proposed fix applies a series of checks to ascertain that the model-dataset combination requires intervention (since some feedforward neural network models are able to run on MNIST without further action), then, if needed, overrides the model architecture when using MNIST by including a single Conv2d, mapping the single input channel to 3 output channels, before the first convolutional layer.

jianyicheng commented 3 months ago

@mau-mar

Hi, we found there is a bug the yml files that stop CI running from forked repos. It has now been fixed and merged into the upstream. Could you merge the main branch of the upstream into this PR which should trigger the CI properly?

Thanks,

mau-mar commented 3 months ago

@jianyicheng

Hi, I just merged upstream changes into the PR branch. Does it work now?