getting started notebook

[gagneur]

• Package description: ”Standard set of data-loaders for training and making predictions for DNA sequence-based models.” -> “Set of data-loaders for Kipoi sequence-based models”. We don’t inforce a standard, do we? Not just DNA.

• kipoiseq -> Shall we have “data loader” in the package name -> kipoidlseq / kipoiseqdl / kipoi_dlseq? Not sure.

Notebook

• have a quick intro "This notebook gives a short introduction to kipoiseq, a package providing data-loaders for Kipoi sequence-based models. Examples for accessing sequences on disk, batch training of sequence-based models, and for applying a trained sequence-based model on data by batch are provided”.

• the genomic intervals are 0-based which is maybe expected from a python point of view but not necessarily obvious from out there (GFF format uses 1-based coordinates). Write ”You can see that the first interval matches the one provided in the intervals file (kipoiseq assumes the 0-based convention).”. You may want to stress this elsewhere too in the doc. A typical pitfall.

• have “We will now apply a model on these sequences that assume all sequences to be of length 10. Since the intervals are of variable length, we have to resize them. Calling SeqDataset() with the auto_resize_len parameter returns intervals centered on the center of input intervals and with the desired length.”

• How does auto_resize_len deal with odd / even interval length. Why not truncating from the left? What happens when you go outside the chromosome range?

• Write “You can load the whole dataset as dictionary into memory using load_all()”

• after the call to load_all(), I found more useful to show the full data object.

• Have “seqname” instead of “chr”. We used seq_name in GenomeIntervals and GenomicRanges used seqnames for a good reason: Sequences do not have to be chromosomes, eg. vector, construct, MPRA, contigs, transcripts, protein, …

• have “We will now train a convolutional neural network on these data. To this end, we define a Keras model consisting of a first convolutional layer with 2 filters of size 3, ReLU activation, max pooling, and a final dense layer.”

• have “This model will be trained using the Adam optimizer and binary crossentropy, as loss function, and accuracy as performance metric.”

• have “Training is done by batches of size 2. ”

• [ ] have “We will now create an iterator from the data loader that returns data by batches. Number of CPU workers is set to 4.” Before deinfing the iterator
  iterator = dl.batch_train_iter(batch_size=batch_size, num_workers=4)”

• [ ] explain “What does the iterator return?” after initiating it. Afterward, do the call for model fitting that uses the iterator.