pyFlowSOM 0.1.4

A Python implementation of the SOM training functionality of FlowSOM

pyFlowSOM

Python runner for the FlowSOM library.

Basic usage:

import pandas as pd
from pyFlowSOM import map_data_to_nodes, som

df = pd.read_csv('examples/example_som_input.csv')
example_som_input_arr = df.to_numpy()
node_output = som(example_som_input_arr, xdim=10, ydim=10, rlen=10)
clusters, dists = map_data_to_nodes(node_output, example_som_input_arr)

To put the data back into dataframes:

eno = pd.DataFrame(data=node_output, columns=df.columns)
eco = pd.DataFrame(data=clusters, columns=["cluster"])

To export to csv:

eno.to_csv('examples/example_node_output.csv', index=False)
eco.to_csv('examples/example_clusters_output.csv', index=False)

To plot the output as a heatmap:

import seaborn as sns

# Append results to the input data
example_som_input_df['cluster'] = clusters

# Find mean of each cluster
df_mean = example_som_input_df.groupby(['cluster']).mean()

# Make heatmap
sns_plot = sns.clustermap(df_mean, z_score=1, cmap="vlag", center=0, yticklabels=True)
sns_plot.figure.savefig(f"example_cluster_heatmap.png")

Develop

Continually build and test while developing. This will automatically create your virtual env

./build.sh && ./test.sh

The C code (pyFlowSOM/flosom.c) is wrapped using Cython (pyFlowSOM/cyFlowSOM.c).

Tests do an approximate exact comparison to cluster id groundtruth and an approximate comparison to node values only because of floating point differences. All randomness has stubbed out in in the y2kbugger/FlowSOM fork and works in tandem to the deterministic flag to the som function.

To regenerate test data, which may be required if you changed any sources of randomness:

python -m pyFlowSOM.generate_test_outputs

To generate heatmaps for manual comparison:

python -m pyFlowSOM.generate_test_heatmaps