ASMR - Activation-Sharing Multi-Resolution Coordinate Networks for Efficient Inference
STS Luo
UofT CSC494 Winter 2024 Project
Report
This repo consists of all the code used to generate the figures in the report. Each figure is treated as an individual experiment and the corresponding code file contains both the training and plotting code.
Currently, there are four sets of experiments:
- Domain manipulation visualizer (
hash_visualizer.py
) - Domain scaling and performance (
exp_scale_up.py
) - Domain flipping, prediction segments, and signal bandwidth (
exp_bandwidth.py
) - Collision rates and performance (
exp_collision_error.py
) (not included in the report) - Visualizing how the MLP resolve hash collision via gradients (
exp_collision_gradient.py
) (not included in the report)
Each experiment code is structure as follow:
train(path, trial, n_seeds, **)
plot(model_path, figure_path, **)
if __name__ == "__main__"
block which defines and creates the directories, and iterates through the number of trials to calltrain()
andplot()
Simply run with a GPU the following code
python exp_<experiment_name>.py
Or download from dropbox
Attempted to develop a similarity score for signals that could determine how suitable a signal is for NGP to fit analytically. Related codes:
main.py
sim_score_plot.py
scorers/