hello

[DinaOmar02]

Hello,
I’m fine-tuning the MoViNet A2 base model on my custom dataset and using your code and MoVinet tutorial as a reference. I wanted to ask whether you trained the model on a GPU or CPU.

I’m trying to train it on a GPU, but it still runs on the CPU, that takes much more time.
can you help me for this problem?

[engares]

Hey there! Glad you ask. Some of the models were trained on CPU while I was testing my CUDA support. I got tons of problems getting it to work.

The training code explicitly hides the GPU for testing reasons. tf.config.set_visible_devices([],'GPU'). You can re-run it without this line and see if it works:

print("GPUs:", tf.config.list_physical_devices('GPU'))

Best regards

[DinaOmar02]

Ok, now i train it on gpu
I ask for your help if you can
I'm fine-tuning the model on driving front-view videos to classify driver actions (turn left, turn right, go straight). My dataset consists of 1,260 training videos and 312 validation videos.(i sampled each video to sequence of 10 frames). but my custom dataset significantly different from the original dataset the model was trained on. After training the last five layers with the classifier layer for 30 epochs, the validation top-1 accuracy reaches only 50%. What can I do to improve accuracy?

[engares]

Okay, given that you're using the weights of a model trained on Kinetics (for human action recognition) you may find it easier to train the full model (or a light version) by yourself.

It will take longer, but since I haven't found a pretrained movinet for that purpose, it can always be an option.

You're not getting the expected results due to the change on the domain of expertise of the model. You can also try to fine-tune other architectures trained on driving video data.

Regards