ZeroDiffusion/run.sh at main · ysims/ZeroDiffusion · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
#!/bin/bash
# Run baseline ZeroDiffusion experiments on core datasets.
# Best method: MSE + Interpolated Noise

mkdir -p results
export PYTHONUNBUFFERED=1

echo "Running baseline MSE+Interp experiments..."

# # ESC-50: 4 validation folds and a held out test fold
# uv run python train.py ../ZeroShotESCData/pickles/esc50/ESC_50_synonyms/fold04.pickle --dataset ESC-50 --split fold0 --cls_dataset_size 40 > results/escfold0.txt
# uv run python train.py ../ZeroShotESCData/pickles/esc50/ESC_50_synonyms/fold14.pickle --dataset ESC-50 --split fold1 --cls_dataset_size 40 > results/escfold1.txt
# uv run python train.py ../ZeroShotESCData/pickles/esc50/ESC_50_synonyms/fold24.pickle --dataset ESC-50 --split fold2 --cls_dataset_size 40 > results/escfold2.txt
# uv run python train.py ../ZeroShotESCData/pickles/esc50/ESC_50_synonyms/fold34.pickle --dataset ESC-50 --split fold3 --cls_dataset_size 40 > results/escfold3.txt
# uv run python train.py ../ZeroShotESCData/pickles/esc50/ESC_50_synonyms/fold4.pickle --dataset ESC-50 --split test --cls_dataset_size 40 > results/esctest.txt

# # GTZAN - music classification dataset
# uv run python train.py ../ZeroShotESCData/pickles/gtzan/YAMNet_synonyms_test.pickle --dataset GTZAN --split test --cls_dataset_size 100 > results/gtzan.txt

# # UrbanSound8k - small number of classes but many samples per class
# uv run python train.py ../ZeroShotESCData/pickles/urbansound8k/YAMNet_synonyms_test.pickle --dataset UrbanSound8k --split test --cls_dataset_size 873 > results/urbansound8k.txt

# FSC22 - split into train, val, test, less classes than ESC-50 but more than GTZAN and UrbanSound8k
uv run python train.py ../ZeroShotESCData/pickles/fsc22/YAMNet_synonyms_val.pickle --dataset FSC22 --split val --cls_dataset_size 75 > results/fscval.txt
uv run python train.py ../ZeroShotESCData/pickles/fsc22/YAMNet_synonyms_test.pickle --dataset FSC22 --split test --cls_dataset_size 75 > results/fsctest.txt

# ARCA23K-FSD - larger dataset with multiple validation folds
uv run python train.py ../ZeroShotESCData/pickles/arca23kfsd/YAMNet_normal_fold0.pickle --dataset ARCA23K-FSD --split fold0 --cls_dataset_size 339 > results/arcafold0.txt
uv run python train.py ../ZeroShotESCData/pickles/arca23kfsd/YAMNet_normal_fold1.pickle --dataset ARCA23K-FSD --split fold1 --cls_dataset_size 339 > results/arcafold1.txt
uv run python train.py ../ZeroShotESCData/pickles/arca23kfsd/YAMNet_normal_fold2.pickle --dataset ARCA23K-FSD --split fold2 --cls_dataset_size 339 > results/arcafold2.txt
uv run python train.py ../ZeroShotESCData/pickles/arca23kfsd/YAMNet_normal_fold3.pickle --dataset ARCA23K-FSD --split fold3 --cls_dataset_size 339 > results/arcafold3.txt
uv run python train.py ../ZeroShotESCData/pickles/arca23kfsd/YAMNet_normal_fold4.pickle --dataset ARCA23K-FSD --split fold4 --cls_dataset_size 339 > results/arcafold4.txt
uv run python train.py ../ZeroShotESCData/pickles/arca23kfsd/YAMNet_normal_fold5.pickle --dataset ARCA23K-FSD --split fold5 --cls_dataset_size 339 > results/arcafold5.txt
uv run python train.py ../ZeroShotESCData/pickles/arca23kfsd/YAMNet_normal_fold6.pickle --dataset ARCA23K-FSD --split test --cls_dataset_size 339 > results/arcatest.txt

echo "Experiments complete! Results saved to results/ directory."