In a new open dataset,the performance of the tsm is poor.Which lead to this?

[ZJU-lishuang]

I train the tsm on the fight dataset.
And the best top1 acc is 0.53.
When I use other small open dataset,the top1 acc is nearly to 1.0.
Is the fight dataset problem lead to the problem?

[ZJU-lishuang]

config
tsm_r50_1x1x8_50e_customdataset_fight_rgb320.py
log
20210220_081407.log

Can you help me?

[innerlee]

try i3d?

[innerlee]

And use pre-trained model on k400

[ZJU-lishuang]

Oh,I forget to use pre-trained model on k400.
I will have a try.

[ZJU-lishuang]

@innerlee When I do as you said ,the top1 acc of last model is higher than past.
But the top1 acc value is smaller and smaller as the training continues.
i3d log with k400 pre-train
20210223_092126.log

tsm log with k400 pre-train
20210223_081901.log

Is the dataset problem leading to low top1 acc?

[innerlee]

The dataset is very tiny.

• Use small lr. Divide the original lr by 10 or 100
• Use strong augmentation. See imgaug examples [Feature] Support Imgaug for augmentations in the data pipeline. #492

[ZJU-lishuang]

because of the small dataset.
These two strategies are useless.

[innerlee]

• Fix the first two stages of the backbone

[ZJU-lishuang]

@innerlee How to do it?
There isn't a tutorial to Freeze Backbone.
Is The relative parameter frozen_stages? I don't know how to use it.
Set frozen_stages=2?

[irvingzhang0512]

It's overfitting. For small datasets, fewer epochs (i.e. 20) are enough.
You can also set interval=1 for both evaluation and checkpoint_config in your configs.

[ZJU-lishuang]

early stop is a good idea.

[ZJU-lishuang]

using pre-trained model on k400 and early stop get a good result.