Attach random seed to gensim word2vec (#78)

* Attach random_state to gensim word2vec seed

* Add script for testing reproducibility

Author: RemyLau

demo/reproducibility.sh

@@ -0,0 +1,14 @@
+#!/bin/bash --login
+# reproducibility.sh
+# Test the reproducibility of PecanPy between runs.
+
+source ~/.bashrc
+
+rs=100
+export PYTHONHASHSEED=$rs
+
+conda activate pecanpy-dev
+pecanpy --input karate.edg --output karate1.emd --mode FirstOrderUnweighted --workers 1 --random_state $rs
+pecanpy --input karate.edg --output karate2.emd --mode FirstOrderUnweighted --workers 1 --random_state $rs
+cmp karate1.emd karate2.emd
+rm -f karate1.emd karate2.emd

src/pecanpy/cli.py

@@ -298,6 +298,7 @@ def learn_embeddings(args, walks):
         sg=1,
         workers=args.workers,
         epochs=args.epochs,
+        seed=args.random_state,
     )
 
     output_path = args.output

src/pecanpy/pecanpy.py

@@ -77,7 +77,10 @@ def __init__(
                 weights added to the average edge weights as the noisy edge
                 threashold, only used by node2vec+ (default: 0)
             random_state (int, optional): Random seed for generating random
-                walks (default: :obj:`None`).
+                walks. Note that to fully ensure reproducibility, use single
+                thread (i.e., workers=1), and potentially need to set the
+                Python environment variable ``PYTHONHASHSEED`` to match the
+                random_state (default: :obj:`None`).
 
         """
         super().__init__()
@@ -254,6 +257,7 @@ def embed(
             min_count=0,
             workers=self.workers,
             epochs=epochs,
+            seed=self.random_state,
         )
 
         # index mapping back to node IDs