This repo consists of the fixed-point analysis code for ECE498 DL in Hardware at UIUC
The code will generate bit-precisions and dynamic ranges of activations and weights for any trained DL networks through method presented here. The results are saved in .npy format.
The "compute_precision_offsets" function generates the resultant npy files. To test the accuracy of a quantized network, one should run "qfeedforward" function