Nucleus

devlog#4, also again long time no see, been really busy with bunch of school works, but still also a bit of development almost everyday so here i modify the refseqs requests only 25 organisms and 6 genomes for each organism, also modified bunch of files to support the new refseqs.json style, and also trained the classifier model many mistakes and 100minutes * 10 epoch = 1000 minutes totaling, take so long to finish (around 16.67 hours *no more 67 atp) the model itself has 4 conv1d blocks and 2 layers of transformer, and the dataset itself has ~523M tokens (as of i remember) and the output is astonishing, as you can see in the image below, 0.062 loss, 98.11% accuracy, and validation loss of 0.7708, and validation accuracy of 88.84% (although it has 9% gap between acc and vacc but still nevermind this aint production grade and should not be use in production or other associated) also i test it with test data (from the dataloader split) and other dataset with the same organism but different genomes, i got avg loss at ~1.25 and avg test accuracy at ~82% pretty gooddd, that’s all for now, next is like genomic analysis, like promoter region prediction and open reading frame prediction (ORF) (for finding a sequence that potentially encodes a protein) and done genomes-level prediction and out off 306 genomes, got 292 correct which is 95.42% holy really welllllllllll!!!

devlog#3, long time no see, been really busy with school works and practices for the incoming sport events and competitions. as of now i have created tokenizes.py first time i scale for 6-mers but then ;ater on i design to abuse my entire laptop and increase it to 8 and now my tokens.bin or basically binary of those token IDs has taken up 16.45GB of my disk space and it’s really huge, this gonna take so long to just train the CNN but nevermind, follow along!

devlog#2 last time i forgot to add and commit the changes, so here it is and also create utils/split.py for splitting train, test and val with 80/10/10 ratio for the next development, i will do splitting each genomes into chunks at around 64 kb per chunk and might scale bigger if my vram wont hit oom error, and if i pass i will need to convert 6-mers into int, like mapping tokenization (ATGCGC -> 12 and more)

devlog#1, initialized the project and done some brief project configurations, this took just minutes and then continue progressing on downloading the dataset from official NCBI rest api (documented here) after that is merging the .fna files (total of .fna might mot match total fasta files as of esch FASTA record might contains many .fna) and for the next development is k-mers tokenization which I will do it using overlapping (stride=1) 6-mers tokenization and use CNN + Transformer (and attention) for the best result and learning opportunity, follow along I guesd this might be my masterpiece