cebnn

A personal project I use to train neural networks. See cmds.zsh for the kinds of things it can do.

Setup

1. Install python 3.10
2. Create a venv:

python3.10 -m venv .venv --system-site-packages

3. Activate the venv:

source .venv/bin/activate

4. Install the requirements:

pip install -r requirements.txt

torchvision may need to be installed outside of pip if it fails to build.

Basic usage example

1. Activate the environment (assumes zsh):

source cmds.zsh

2. Gather images to train on:

time ./gather_images.py data/trial1 ::/path/to/negative_class :labela:/path/to/a_class :labelb:/path/to/b_class

3. Setup data:

./setup_data.sh data/trial1 /mnt/big_hdd/scaled

4. Edit cmds.zsh to point to the dataset (dataset=trial1)
5. Train models:

train 1.0 10 20e-4 --sublayers=.91 --resample='+:8:all' --classifier-dropout=.65 --criterion=digamma
train 2.0 10 20e-4 --sublayers=.91 --resample='+:8:all' --classifier-dropout=.65 --criterion=digamma --seed=44
train 3.0 15 8e-4 --cv-fold=0 --sublayers=.4 --resample=none --inner-dropout=.05 --classifier-dropout=.65 --criterion=digamma --optimizer=aadamw01 --gamma=.1 --sch-period=5

6. Evaluate the models:

mcc nets/trial1/*.torch

7. Generate the binary accuracy information:

get_correct fbeta nets/trial1/*.torch

8. Run majority voting trials:

time ./best_majvote.py 2 0 1 nets/trial1/correctfbeta/*

9. Generate test binary accuracy information:

get_correct_test fbeta nets/trial1/*.torch

10. Evaluate the selected majority vote for each label:

./best_majvote_eval.py 0 0 nets/trial1/correctfbeta_test/rexnet_200_{1,2,3}.0.torch_correct.pkl
./best_majvote_eval.py 0 1 nets/trial1/correctfbeta_test/rexnet_200_{1,2,3}.0.torch_correct.pkl

11. Run inference on a given dataset for each model:

./infer.py nets/trial1/rexnet_200_1.0.torch infer/trial1/file_list.txt infer/trial1/file_list_results_trial1_rexnet_200_1.0_1.pkl; done
./infer.py nets/trial1/rexnet_200_2.0.torch infer/trial1/file_list.txt infer/trial1/file_list_results_trial1_rexnet_200_2.0_1.pkl; done
./infer.py nets/trial1/rexnet_200_3.0.torch infer/trial1/file_list.txt infer/trial1/file_list_results_trial1_rexnet_200_3.0_1.pkl; done

12. (A) Compute majority vote and print results:

./infer_print_majvote.py "$(<data/trial1/classes.txt)" 0 0 $(for n in rexnet_200_{1,2,3}.0; do echo infer/trial1/file_list_results_trial1_${n}_1.pkl nets/trial1/correctfbeta/${n}.torch_correct.pkl; done) >infer/trial1/mydata_results_trial1_majvote_1.csv

12. (B) Or, print the results for a single net/label with a given threshold:

./infer_print.py "$(<data/trial1/classes.txt)" 1 .4837 infer/trial1/file_list_results_trial1_rexnet_200_1.0_1.pkl >infer/trial1/file_list_results_trial1_rexnet_200_1.0_1.csv