model-template.yaml

defined_as: kipoi.model.PyTorchModel
args:
  module_class: {{ module_class }}  # model_arch.WrappedDeeperDeepSEA 
  {% if module_kwargs %}
  module_kwargs: 
    {% for k, v in module_kwargs.items(): -%}
      {{ k }}: {{ v }}
    {% endfor %}
  {% endif %}
  weights: {{ model_weights }}  # remove this line after uploading to Zenodo
  #  To share the model with others, upload the `.pth.tar` to Zenodo
  #  or Figshare, fill in the correct URL below, comment out `selene_model.pth.tar` above,
  #  and uncomment these two lines (`url` and `md5` should still be nested under `weights`):
  #    url: https://zenodo.org/record/<zenodo id>/files/selene_model.pth?download=1
  #    md5: {{ selene_model_md5 }} 
info:
  authors:
    {% for a in authors -%}
      - name: {{ a['name'] }}
      github: {{ a['github'] }} 
    {% endfor %} 
  license: {{ license }}  
  doc: >
      model_name = {{ model_name }}
      ....
  cite_as: /     # update with the DOI url to the publication
  # e.g. trained_on: all chromosomes except validation chromosomes 6 & 7
  #                  and test chromosomes 8 & 9
  trained_on: {{ trained_on_description }} 
  training_procedure: Using Selene version {{ selene_version }}
  {% if tags -%}
  tags:
    {% for t in tags -%}
      - {{ t }}
    {% endfor %}
  {% endif %}
default_dataloader:
  defined_as: kipoiseq.dataloaders.SeqIntervalDl
  default_args:
    auto_resize_len: {{ seq_len }}
    alphabet_axis: 0
    dtype: np.float32

dependencies:
  conda:
    - h5py
    - pytorch::pytorch-cpu=={{ pytorch_version }}
  pip:
    - kipoiseq

schema:
  inputs:
    name: seq
    special_type: DNASeq
    shape: (4, {{ seq_len }})
    doc: DNA sequence
    associated_metadata: ranges
  targets:
    name: Output tasks
    shape: ({{ n_tasks }}, )   # (i.e. n_classes or n_genomic_features)
    doc: Probability for a specific output task
    column_labels:
      - {{ predictor_names }}   # list of classes predicted by the model

# Run `kipoi test <model_dir> -o expect.h5` to generate the file
# test:
# expect:
#   url: <URL>/expect.h5
#   md5: <md5 hash>
#   precision_decimal: 6