search

open-mmlab / mmaction

An open-source toolbox for action understanding based on PyTorch
https://open-mmlab.github.io/
Apache License 2.0
1.86k stars 352 forks source link

Feature extraction #72

Open sparsh-b opened 5 years ago

sparsh-b commented 5 years ago

Dear @zhaoyue-zephyrus ,

I want to extract features (for a custom dataset) generated by TSN. I reckon that the tensor containing features is present here (https://github.com/open-mmlab/mmaction/blob/eb7ace3458df6d551b99a1026abee2b209376a6b/mmaction/models/recognizers/TSN2D.py#L123). While running my dataset on TSN2D, the dimension of the tensor that i get after the above referenced line is (32768 x 1). And there are 22 such tensors getting generated for each epoch of training. I dont understand what these numbers (22 & 32768)mean. Kindly help me.

TESTING: I reckon that the features getting generated for videos in test split will be present here(https://github.com/open-mmlab/mmaction/blob/eb7ace3458df6d551b99a1026abee2b209376a6b/mmaction/models/recognizers/TSN2D.py#L155). By printing this, I'm getting tensors of dimension (62 x 1). And there are 384 such tensors getting generated.

And the values in test set tensors were mostly negative... Do they mean anything? Screenshot from 2019-08-11 20-17-01

Plz help me understand these tensors & their dimensions.

P.S: My dataset has 61 classes. In the training split, each class has number of videos in the range of (5-200). Number of vids in test split of approximately 1/3rd of those.

akashsharma29 commented 4 years ago

Hi, @sparsh-b did you find solutions to your question? I am also trying to figure out the same, please let me know.

sparsh-b commented 4 years ago

Hello @akashsharma29 ! I'm sorry. I wasn't able to figure it out either. I regret being unable to help you.