03_Trained

Finally, I train my own models to be able to classify single-cell images into their respective perturbations. This so-called Weakly Supervised Representation Learning leads the pre-trained model to become more sensitive to cell inputs and partially displays better metric scores.

Content

Results

Speeds of different infrastructure

Check EC2 instance types for their technical details Technical details of the Nvidia A100-SXM4-40GB (CHTC server)

Inference / Profiling

Inference times are compared by their inference speed of a site in seconds. Profiling the subsection of the LINCS dataset is around 90,000 sites or around 26 plates (or 1/5 of the entire LINCS dataset). A quick way to predict how long the inference step will take is to multiply the seconds taken to infer a site and multiply with 90,000.

Changing the batch size in the config file will slightly affect the time estimates below.

CPU With 4 cores, the profiling runs at ~100 seconds per site. This would result in a very, very long inference time

P2 GPU The cheapest GPU on AWS has an Nvidia Tesla K80 which infers a site in ~10s. Still very slow.

P3 GPU The next EC2 generation (NVIDIA Tesla V100) can infer a site in ~2.5 seconds. This allows inferring the LINCS subsection within a handful of days.

CHTC On the CTHC server (NVIDA Ampere 100) the speed goes down to ~0.2 seconds per site. The overall time drops to 5 hours.

Training