AttributeError for VideoClassifier instantiation with backbone slow_r50_detection

[kskdermolab]

🐛 Bug

Hi!
I need to train the slow_50 model for spatio temporal detection. I didn't find a VideoClassificationData for loading AVA dataset, so, I decided to build a new DataModule using as reference the one on the flash_examples (custom_data_loading.py). I don't know if is going to work because I don't know how the trainer pass the input to the model (have two parameters: the Tensor with the clip, and the list of bounding boxes).
I tried to load the model using the VideoClassifier:
model = VideoClassifier( backbone="slow_r50_detection", pretrained=True )
and it fails with the error:

*** AttributeError: 'DetectionBBoxNetwork' object has no attribute 'blocks'

I didn't find any issue or documentation about how to train or load this models, are you planning to make a tutorial about that?

[ethanwharris]

Hey @MGReyes thanks for reporting this! There are a few things to unpack here that we should work on:

• documentation for using tasks with your own datamodules (expect this soon in our upcoming 0.8 release)
• resolving the issue with the slow_r50_detection backbone (we can target this for a 0.8.x patch release)
• supporting more formats for video classification data (some small additions coming in 0.8 but we can scope some more for 0.9 also)

I'll keep you updated as we make progress with these efforts. Hope that helps! 😃