Florian Schmid

icon company Johannes Kepler University icon location Linz icon location University Assistant @CPJKU Linz Working on Audio Classification, Audio Tagging, Acoustic Scene Classification, Low-Complexity Models

Results 3 issues of Florian Schmid

Factor of 2 in count_flops.py

In the function [count_flop.py](https://github.com/qiuqiangkong/audioset_tagging_cnn/blob/master/pytorch/pytorch_utils.py), when calculating _kernel_ops_ and _weight_ops_ for convolution and linear layers: why do you need to multiply them by 2, if multiply_adds=True? Shouldn't that be the other...

Question wdgrl.py - gradient_penalty(critic, h_s, h_t)

1 comment

In wdgrl.py - gradient_penalty(critic, h_s, h_t): The interpolates created in line 29 are of size _(3 x batch_size x feature_size)_. Following that the gradients are also of size (3 x...

'synth_set' used twice

4 comment

Hi, I was looking through the code for the DCASE'24 Task 4 baseline system and noticed the following lines in the file [train_pretrained.py](https://github.com/DCASE-REPO/DESED_task/blob/1cb8c84f848240b77bc9ba6894c3d5f3b76283a7/recipes/dcase2024_task4_baseline/train_pretrained.py#L403): ``` strong_full_set = torch.utils.data.ConcatDataset([strong_set, synth_set]) tot_train_data =...