added bpnet model (#191)

BPNet-OSKN/model.py

@@ -0,0 +1,75 @@
+from bpnet.seqmodel import SeqModel
+from keras.models import load_model
+import numpy as np
+import bpnet
+import tensorflow as tf
+from bpnet.functions import softmax
+import keras.backend as K
+import keras.layers as kl
+from kipoi.model import BaseModel
+
+def profile_contrib(p):
+    return kl.Lambda(lambda p:
+                     K.mean(K.sum(K.stop_gradient(tf.nn.softmax(p, dim=-2)) * p, axis=-2), axis=-1)
+                     )(p)
+
+
+class BPNetOldSeqModel(BaseModel, SeqModel):
+
+    preact_tensor_names = ['reshape_2/Reshape:0',
+                           'dense_1/BiasAdd:0',
+                           'reshape_4/Reshape:0',
+                           'dense_3/BiasAdd:0',
+                           'reshape_6/Reshape:0',
+                           'dense_5/BiasAdd:0',
+                           'reshape_8/Reshape:0',
+                           'dense_7/BiasAdd:0'
+                           ]
+
+    bottleneck_name = 'add_9/add:0'
+
+    target_names = ['Oct4/profile',
+                    'Oct4/counts',
+                    'Sox2/profile',
+                    'Sox2/counts',
+                    'Nanog/profile',
+                    'Nanog/counts',
+                    'Klf4/profile',
+                    'Klf4/counts']
+
+    seqlen = 1000
+
+    tasks = ['Oct4', 'Sox2', 'Nanog', 'Klf4']
+
+    postproc_fns = [softmax, None] * 4
+
+    def __init__(self, model_file):
+        self.model_file = model_file
+        K.clear_session()  # restart session
+        self.model = load_model(model_file, compile=False)
+        self.contrib_fns = {}
+
+    def predict_on_batch(self, seq):
+        preds = self.model.predict_on_batch({"seq": seq, **self.neutral_bias_inputs(len(seq), seqlen=seq.shape[1])})
+        pred_dict = {target: preds[i] for i, target in enumerate(self.target_names)}
+        return {task: softmax(pred_dict[f'{task}/profile']) * np.exp(pred_dict[f'{task}/counts'][:, np.newaxis])
+                for task in self.tasks}
+
+    def neutral_bias_inputs(self, length, seqlen):
+        """Compile a set of neutral bias inputs
+        """
+        return dict([('bias/' + target, np.zeros((length, seqlen, 4))
+                      if target.endswith("/profile")
+                      else np.zeros((length, 2)))
+                     for target in self.target_names])
+
+    def get_intp_tensors(self, preact_only=True, graph=None):
+        if graph is None:
+            graph = tf.get_default_graph()
+        intp_targets = []
+        for head_name, tensor_name in zip(self.target_names, self.preact_tensor_names):
+            tensor = graph.get_tensor_by_name(tensor_name)
+            if head_name.endswith("/profile"):
+                tensor = profile_contrib(tensor)
+            intp_targets.append((head_name, tensor))
+        return intp_targets

BPNet-OSKN/model.yaml

@@ -0,0 +1,67 @@
+defined_as: model.BPNetOldSeqModel
+args:
+    model_file:
+      # TODO - put to Zenodo
+      url: 'http://mitra.stanford.edu/kundaje/avsec/chipnexus/paper/modisco-comparison/v2-output/nexus,peaks,OSNK,0,10,1,FALSE,same,0.5,64,25,0.004,9,FALSE,[1,50],TRUE/model.calibrated.h5'
+      md5: bbe883baef261877bfad07d05feb627d
+
+default_dataloader:
+    defined_as: kipoiseq.dataloaders.SeqIntervalDl
+    default_args:
+        auto_resize_len: 1000
+        ignore_targets: True
+
+info:
+    authors: 
+        - name: Ziga Avsec
+          github: avsecz
+    doc: BPNet model predicting the ChIP-nexus profiles of Oct4, Sox2, Nanog and Klf4
+    cite_as: TODO
+    trained_on: ChIP-nexus data in mm10. test chromosomes 1, 8, 9, validation chromosomes 2, 3, 4
+    license: MIT
+
+dependencies:
+    channels:
+      - bioconda
+      - pytorch
+      - conda-forge
+      - defaults
+    conda:
+      - python=3.6
+      - bioconda::pybedtools>=0.7.10
+      - bioconda::bedtools>=2.27.1
+      - bioconda::pybigwig>=0.3.10
+      - bioconda::pysam>=0.14.0
+      - bioconda::genomelake>=0.1.4
+
+      - pytorch::pytorch  # optional for data-loading
+      - cython
+      - h5py>=2.7.0
+      - numpy
+
+      - pandas>=0.23.0
+      - fastparquet
+      - python-snappy
+
+      - nb_conda
+    pip:
+      - tensorflow>=1.0
+      - git+https://github.com/kundajelab/DeepExplain.git
+      - bpnet[extras]
+schema:
+    inputs:
+      shape: (1000, 4)
+      doc: "One-hot encoded DNA sequence."
+    targets:
+      Oct4:
+        shape: (1000,2)
+        doc: "Strand-specific ChIP-nexus data for Oct4."
+      Sox2:
+        shape: (1000,2)
+        doc: "Strand-specific ChIP-nexus data for Sox2."
+      Nanog:
+        shape: (1000,2)
+        doc: "Strand-specific ChIP-nexus data for Nanog."
+      Klf4:
+        shape: (1000,2)
+        doc: "Strand-specific ChIP-nexus data for Klf4."