This Repository is dedicated to the CPJKU submission of DCASE 2022, Task 1, Low-Complexity Acoustic Scene Classification. It provides a lightweight version of the code we used to create our challenge submissions. The corresponding technical report can be found at DCASE2022.
The skeleton of the code is similar to previous CPJKU submissions and the PaSST repository.
An installation of conda is required on your system.
This repo uses forked versions of sacred for configuration and logging, pytorch-lightning as a convenient pytorch wrapper and ba3l as an integrating tool between mongodb, sacred and pytorch lightning.
To setup the environment Mamba is recommended and faster than conda:
conda install mamba -n base -c conda-forge
Now you can import the environment from environment.yml. This might take several minutes.
mamba env create -f environment.yml
Alternatively, you can also import the environment using conda:
conda env create -f environment.yml
An environment named dcase22_t1 has been created. Activate the environment:
conda activate dcase22_t1
Now install sacred, ba3l and pl-lightening:
# dependencies
pip install -e 'git+https://github.com/kkoutini/[email protected]#egg=ba3l'
pip install -e 'git+https://github.com/kkoutini/[email protected]#egg=pytorch-lightning'
pip install -e 'git+https://github.com/kkoutini/[email protected]#egg=sacred' Firstly, you need to create a reassembled version of the TAU Urban Acoustic Scenes 2022 Mobile development dataset. A short draft of how to reassemble the downloaded files is provided in files_reassemble.ipynb.
Secondly, you need to change all paths specfied in the dataset file to the correct locations on your system.
Thirdly, you need to download pre-trained PaSST models and 10-second teacher predictions from the github release provided in this repository and put them in the folder teacher_models.
After creating the environment and setting up external data resources the following is the simplest command to run:
python -m experiments.dcase22.t1.teachers_gang
And a short test run:
python -m experiments.dcase22.t1.teachers_gang with trainer.max_epochs=10 basedataset.subsample=100
A lot configuration possibilities can be set via the command line, here are examples of the configurations we used in our submissions:
Network configuration:
- cp_mini_resnet: named config to decrease the number of blocks in the network
- models.net.rho: changes receptive field of the network
- models.net.s2_group: grouping in second network stage
- models.net.cut_channels_s3: cutting the width on stage 3 of the network
Knowledge Distillation:
- temperature: temperature used to create soft targets
- soft_targets_weight: weighting of the distillation loss
- models.teacher_net.teachers_list: specifies list of teacher ids to be ensembled
- soft_targets_weight_long: weighting of the distillation loss reagrding a 10-second teacher
Mixup and Mixstyle:
- mixup_alpha: mixing coefficient for mixup
- mistyle_alpha: mixing coefficient for mixstyle
- mixstyle_p: probability of mixstyle being applied to a batch
Quantization:
- quantization: setting quantization=1 evaluates the quantized model on the test split of the development set after training
python -m experiments.dcase22.t1.teachers_gang with cp_mini_resnet models.net.rho=8 soft_targets_weight=50 soft_targets_weight_long=3.0 temperature=1 mixup_alpha=0.3 quantization=1 models.teacher_net.teachers_list='[253, 254, 255, 256]' models.net.s2_group=2 models.net.cut_channels_s3=36
python -m experiments.dcase22.t1.teachers_gang with cp_mini_resnet models.net.rho=8 soft_targets_weight=50 temperature=1 mixstyle_alpha=0.3 mixstyle_p=0.6 quantization=1 models.teacher_net.teachers_list='[253, 254, 255, 256]' models.net.s2_group=2 models.net.cut_channels_s3=36
Install pymongo:
pip install pymongo
If you have a running MongoDB-Server you can append the following to your command to log experiment details:
-p -m mongodb_server:[port]:[name] -c "Testing CPJKU Submission to DCASE22"