This competition requires you to distinguish between real and fake videos, both in video and in sound. Final position is based upon your ability to filter the real vs fakes from a 4000 video dataset. You are provided with approximately 100,000 training/test set videos.
Submitted to qualify for free GCP + TPU credits, hence the quick whip-up, which will be needed when the über-TPU-version with some esoteric architecture makes its debut ). So, no hyperparameter tuning on this quick’n'(very)dirty couple of hours effort. Log loss was so bad my eyes could not comprehend such an error existed (I naively assumed log loss didn’t stretch to infinity!). It did however enable me to familiarise myself with the Kaggle submission process and shatter my illusions a crap CNN model trained on a few hundred items or so would even be useful.
Now pulling/extracting the 470 GB training zip to a GCP VM attached Disk. Initially intend to train on a reasonable subset before the entire data-set. I was using preemptible instances but found being kicked off mostly after minutes was just unusable, therefore have provisioned a 4-Core 16 GB 1.5 TB standard instance with a P100 GPU to start off.
Incidentally I initially used fuse + GCP buckets for storage, but the fuse driver is simply unreliable and buggy, resulting in issues that I decided I could not waste time and effort solving, hence I finally just attached a disk.