Mark Mazumder

icon indicating a website link https://markmaz.com

icon company Harvard University icon location Boston, MA icon location Machine Learning PhD candidate at Harvard

Results 17 issues of Mark Mazumder

Some percentage of wavs (~3%) are below 1s according to soxi, others can't be opened

1 comment

MSWC
high priority

re-run few shot experiments with same split of unknown/silence as the DSCNN tests

MSWC

verify 1-1 match between final audio files and splits

* this verification should only apply to words that have > the minimum number of clips for defining a split * currently the tarballs for audio has the full path...

MSWC

check for 16KHz in AudioDataset

both for mp3s and wav files

MSWC

Fine-tuning code for MP3D monodepth?

1 comment

Hi, I am using Habitat and read in your paper/code that this project has made some pretrained weights available [in `habitat_baselines/agents/slam_agents.py`](https://github.com/facebookresearch/habitat-api/blob/5adbf5c33c88eab379a730752d60698ddc340fa8/habitat_baselines/agents/slam_agents.py#L571-L577) for the ResNet50 variant of [this depth estimation network](https://github.com/JunjH/Revisiting_Single_Depth_Estimation)...

should we switch to non-stratified shufflesplit in selection.py?

i.e., should we have uniform class sizes? try both and compare

allow participants to extend config file

we should allow users to extend the configuration but we need a merging mechanism, e.g., https://github.com/rbgirshick/yacs#command-line-overrides so that we can externally enforce the number of allowed training samples during evaluation,...