Wanted to know more information about the ranked sequence of input ids
Hey Thanks so much for your contributions! I have a question regarding the input sequences so the dataset I am looking at is for the https://huggingface.co/datasets/ctheodoris/Genecorpus-30M/tree/main/example_input_files/cell_classification/disease_classification/human_dcm_hcm_nf.dataset
total genes n = 17,903 and 27,406,216 filtered cells
Q) Is the input ids are the ranked genes from 1 cell to create a token or multiple cells.
Q) If there are more than 2048 genes present in that cell how would it be split? would it be split into another token.
how many genes and cells are present in the embedding?
Thank you for your questions.
Firstly: the dataset you mentioned has less than 27M cells - the 27M cell dataset is the full corpus. The number of genes detected in the full corpus is more than 17k.
Regarding your questions:
- Each example in the dataset is 1 cell.
- Each gene is 1 token. There is no splitting of genes. The rank value encoding is truncated to 2048 or 4096 genes, depending on which model you are using (30M or 95M, respectively).
- The model outputs embeddings of the genes. You can obtain an embedding of the cell either by aggregating the gene embeddings (for example, mean pooling) or by using the CLS token embedding at the front of the rank value encoding in the 95M model.