README for traint/test set generation.

Author: dohlee

.gitignore

@@ -0,0 +1 @@
+spliceai_train_code/

README.md

@@ -2,7 +2,7 @@
 
 ![model](img/banner.png)
 
-Implementation of SpliceAI, Illumina's deep neural network to predict variant effects on splicing, in PyTorch.
+Implementation of SpliceAI, Illumina's deep neural network to predict variant effects on splicing, in PyTorch. You can find the Illumina's official implementation [here](https://github.com/Illumina/SpliceAI).
 
 ## Installation
 
@@ -13,4 +13,21 @@ pip install spliceai-pytorch
 ## Usage
 ```python
 from spliceai_pytorch import SpliceAI
+
+```
+
+## Generating train/test sets
+
+First, download 'SpliceAI train code' directory from [here](https://basespace.illumina.com/s/5u6ThOblecrh) and unzip it to `spliceai_train_code` directory.
+Also, download human reference genome (version hg19) to `spliceai_train_code/reference` directory.
+
+Then, run the following command to generate train/test sets after moving into `spliceai_train_code/Canonical`.
+
+```bash
+chmod 755 grab_sequence.sh
+./grab_sequence.sh
+
+# Requires Python 2.7, with numpy and h5py installed
+python create_datafile.py train all  # ~4 miniutes, creates datafile_train_all.h5 (27G)
+python create_datafile.py test 0     # ~1 minute, creates datafile_test_0.h5 (2.4G)
 ```

spliceai_pytorch/spliceai_pytorch.py

@@ -225,4 +225,8 @@ def from_preconfigured(model_name):
 
     x = torch.randn([16, 4, 10000])
     model = SpliceAI_10k()
+    print(model(x).shape)
+
+    x = torch.randn([16, 4, 10000 + 5000 + 1000])
+    model = SpliceAI_10k()
     print(model(x).shape)