Channel order
Hi, thank you for sharing this pretrained model!
Could you clarify the channel order used during training? Is it the same as the order specified in this repository, https://github.com/jump-cellpainting/2024_Chandrasekaran_NatureMethods/tree/main?
ch01 - Alexa 647
ch02 - Alexa 568
ch03 - Alexa 488 long
ch04 - Alexa 488
ch05 - Hoechst 33342
ch06-8 - three brighfield z planes.
Thanks
Hi,
Here are the details. The raw JUMP filenames have a set of characters indicating channel order, i.e.:
cpg0016-jump/source_13/images/20220914_Run1/images/CP-CC9-R1-01/CP-CC9-R1-01_A02_T0001F001L01A01Z01C01.tif,
cpg0016-jump/source_13/images/20220914_Run1/images/CP-CC9-R1-01/CP-CC9-R1-01_A02_T0001F001L01A01Z01C02.tif,
cpg0016-jump/source_13/images/20220914_Run1/images/CP-CC9-R1-01/CP-CC9-R1-01_A02_T0001F001L01A02Z01C03.tif,
cpg0016-jump/source_13/images/20220914_Run1/images/CP-CC9-R1-01/CP-CC9-R1-01_A02_T0001F001L01A03Z01C04.tif,
cpg0016-jump/source_13/images/20220914_Run1/images/CP-CC9-R1-01/CP-CC9-R1-01_A02_T0001F001L01A04Z01C05.tif
correspond to the five channels in field F001 of well A02 in plate CP-CC9-R1-01. Leveraging this ordering, we re-index the channels during training with the following order:
rxrx3: [1, 2, 3, 4, 5, 6] (rxrx3 channels correspond to Nuclei, Endoplasmic Reticulum, F-Actin, Nucleoli & Cytoplasmic RNA, Mitochondria, and Golgi)
jump-orf: [5, 4, 2, 3, 1, 2] (cpg0016 - source 4)
jump-ko: [1, 3, 5, 3, 2, 1] (cpg0016 - source 13)
For inference the channel ordering doesn't matter as the model is channel agnostic. Passing the same set of channels with different permutations of the ordering will calculate the same embeddings. During training, the decoder order needs to be maintained, so we reordered the channels from different datasets to make them match visually.