y-kawagu / dcase2021_task2_baseline_ae Goto Github PK

Hi! Thanks for the amazing dataset, paper, and baseline. I just had a couple of dependency issues when installing it on a fresh conda environment (CUDA Linux machine with Slurm) and wanted to post what worked for me (simply pip install -r requirements.txt did not work because the GPU wasn't found by Keras-TF):

Create the environment and explicitly install GPU-enabled libraries:

conda create --name dcase21_t2 python=3.6
conda install numpy==1.18.1
conda install tensorflow-gpu==1.15.0
conda install -c anaconda keras-gpu==2.3.1

Then run pip install -r requirements_lite.txt with the following contents:

Keras-Applications==1.0.8
Keras-Preprocessing==1.0.5
matplotlib==3.0.3
scikit-learn==0.22.2.post1
numba==0.48
librosa==0.6.0
audioread==2.1.5
tqdm==4.43.0

Note the added numba dependency (needed by librosa, needs to be below 0.49). This was inspired by 2020 Task1 readme which has different instructions for pip and conda (similar issue as above).

Cheers!
Andres