BachelorCode

Quantization of Image classifiers and Object detectors

ResNet - Done

• https://nervanasystems.github.io/distiller/
• PyTorch impl.

ResNet_2.ipynb

Quantization of ResNet18

• 8 bit, 7 bit, 6 bit, 5 bit
• 4 bit
• Mixed

ResNet18 - LAPQ v1

Loss-Aware Quantization of ResNet18

• 4bit

ResNet18 - LAPQ v2

Loss-Aware Quantization of ResNet18

• 4bit + 8bit

ResNet50.ipynb

Quantization of ResNet50

• 8bit
• 4bit
  • (maybe deeper network won't lose acc due to quant.)
• Mixed

SSD - Done

• https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Detection/SSD

SSD Nvidia impl.

• latest
• base AP 13 %
• 8 bit, 7 bit, 6 bit, 5 bit
• 4 bit quant AP 4 %
• from: https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Detection/SSD

Not working implementations

SSD_3

• used model with PyTorch Checkopint https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Detection/SSD#model-architecture

•   modified for quantization purposes
  

• problems with softmax (?), src/evaluate.py and src/utils.py

            # mask = score > 0.05 original
            mask = score > 0.02

            bboxes, score = bboxes_in[mask, :], score[mask]
            if score.size(0) == 0: continue

            score_sorted, score_idx_sorted = score.sort(dim=0)

• problems with masking probabilities in tensors (maybe needs full prec. layer)
• if mask = score > 0.02, then has some predictions, if mask = score > 0.05, then all probabalities are False

SSD.ipynb

• doesnt work

SSD_2.ipynb

• based on https://pytorch.org/hub/nvidia_deeplearningexamples_ssd/ tutorial

Resources:

• not all used

https://github.com/NervanaSystems/distiller https://nervanasystems.github.io/distiller/ https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Detection/SSD https://github.com/amdegroot/ssd.pytorch https://github.com/qfgaohao/pytorch-ssd