YuanGongND
commented
2 years ago Hi there,
What I noticed is that mixup should be in the range of [0-1], so 8 is out of range. Specifically for ESC-50, we use the cross-entropy loss, so you should set mixup=0, or if you want to use mixup, change the loss in traintest.py to BCELoss, we recommend to set mixup=0. You also used a smaller batch-size, which might has some impact (in that case, you need to use a smaller learning rate). There might be some other things, the model is not trained at all.
Anyway, the easiest way to reproduce the result is using the hyper-parameters in the recipe, if you don't have enough GPU memory (we use 4 * 12GB GPUs), it is better to set the tstride and fstride = 16 (which saves a lot of GPU memory) and keep other things unchanged. You won't get the same result as in the paper for larger strides, but it should be similar.
-Yuan
nikhilbyte
There must be some mistake from my side. Can someone help me identify it?
This is how I'm training:
!python -W ignore /content/ast/src/run.py --model ast --dataset esc50 \ --data-train /content/data/datafiles/esc_train_data_1.json --data-val /content/data/datafiles/esc_eval_data_1.json --exp-dir /content/expdir/fold1 \ --label-csv /content/ast/egs/esc50/data/esc_class_labels_indices.csv --n_class 50 \ --lr 1e-5 --n-epochs 25 --batch-size 12 --save_model False \ --freqm 24 --timem 96 --mixup 8 --bal None \ --tstride 10 --fstride 10 --imagenet_pretrain True --audioset_pretrain True