Fig. 6
From: Contrastive self-supervised clustering of scRNA-seq data
Execution time and scalability analysis. Average execution time for all benchmarked methods (a). The relation between the performance of contrastive + KMeans and the other methods has also been computed as a speed factor. All methods have been run on GPU. For comparison, the contrastive methods have also been benchmarked on CPU (contrastive + KM CPU and contrastive + LD CPU). b Depicts how our method scales with an increasing number of cells (from 1000 to 50,000); c illustrates how our method scales with an increasing number of input variables (from 500 to 250,000); the number of epochs needed for contrastive-sc to reach maximum performance measured as ARI (d) and Silhouette (e) scores. The annotated values in c and d represent the mean score. For most datasets, 30 epochs are enough for the model to learn meaningful representations, which brings a computational speed gain compared to other deep-learning competitor methods. This plot depicts three runs of the proposed method for each of the selected numbers of epochs