This is a ~50M parameter diffusion model trained on LibriSpeech train-clean-360.
It can roughly be reproduced by running the below commands: the first for 265K
iterations, the second for an additional 164K iterations.

    python3 -u train_diffusion.py \
        --predictor unet \
        --base-channels 64 \
        --grad-checkpoint \
        --batch 4 \
        --ema-rate 0.9999 \
        $DATA/LibriSpeech/train-clean-360
    python3 -u train_diffusion.py \
        --predictor unet \
        --base-channels 64 \
        --grad-checkpoint \
        --batch 4 \
        --lr 3e-5 \
        --ema-rate 0.9999 \
        $DATA/LibriSpeech/train-clean-360

The final quartile losses for the model were:

    q0=0.02067 q1=0.00118 q2=0.00008 q3=0.00001

For evaluations, I sampled with this command:

    python3 sample_diffusion.py \
        --checkpoint-path model_diffusion_ema.pt \
        --sample-steps 50 \
        --schedule 'lambda x: x**2' \
        --constrain \
        --num-samples 10000 \
        --batch-size 16 \
        --sample-path samples-50step-sqschedule

Class score: 69.0
Frechet score: 1834

I also evaluated the model before the second command. I expected it to be worse
but instead found that the Frechet score is much better. Perhaps the model was
overfitting by the second checkpoint.

Class score: 51.5
Frechet score: 855

I have put the early-stopped models in early_stopped/