Official repository for the paper: "Characterizing Learning in Deep Neural Networks using Tractable Algorithmic Complexity Analysis".
Results in Figure 4 in the paper can be reproduced by running this script:
python train.py
Computes whole-model complexity for a list of timm models, comparing pretrained vs. random weights. Model names are read from model_names_100.txt. Results are saved to results/complexities_100models.json.
python complexity_per_model.py
Computes bitplane complexity per layer for 5 pretrained models (ResNet18, ResNet50, ViT-B/16, EfficientNet-B0, MobileNetV3), comparing pretrained vs. random weights. Results are saved to results/complexity_per_layer_5models.json.
python complexity_per_layer.py
Evaluates post-training quantization (PTQ) accuracy for 5 pretrained models (ResNet18, ResNet50, ViT-B/16, EfficientNet-B0, MobileNetV3) on the ImageNet-1K validation set. Compares FP32, FP16, and per-channel uniform PTQ at bit depths 1–8. Results are saved to ptq.json.
python ptq.py
Utility module providing quantizer classes (UniformQuantizer, UniformQuantizer_per_channel) and helper functions (attach_weight_quantizers, detach_weight_quantizers, toggle_quantization) used by ptq.py. Not intended to be run directly.