From 3461d3b07f77a1605d234dbb79a99e0d8a58aad3 Mon Sep 17 00:00:00 2001
From: barthes <nicolas.barthes@cefe.cnrs.fr>
Date: Tue, 16 Apr 2024 15:44:09 +0200
Subject: [PATCH] HDBSCAN working with metadata & supervised UMAP + fixed
widget width
---
Class_Mod/HDBSCAN_Clustering.py | 21 ++++++++++++++-------
pages/1-samples_selection.py | 33 +++++++++++++++++++++------------
2 files changed, 35 insertions(+), 19 deletions(-)
diff --git a/Class_Mod/HDBSCAN_Clustering.py b/Class_Mod/HDBSCAN_Clustering.py
index 1a9df2d..63b32a4 100644
--- a/Class_Mod/HDBSCAN_Clustering.py
+++ b/Class_Mod/HDBSCAN_Clustering.py
@@ -1,9 +1,9 @@
from Packages import *
class Hdbscan:
"""
- Runs an automatic optimized sklearn.HDBSCAN clustering on Dimensionality reducted space.
+ Runs an automatic optimized sklearn.HDBSCAN clustering on Dimensionality reduced space.
Vars:
- data: the Dimensionality reducted space, raw result of the UMAP.fit()
+ data: the Dimensionality reduced space, raw result of the UMAP.fit()
param_dist: the HDBSCAN optimization parameters to test
Density-Based Clustering Validation - DBCV (https://github.com/christopherjenness/DBCV/tree/master ;
Moulavi, Davoud, et al. "Density-based clustering validation." Proceedings of the 2014 SIAM
@@ -17,14 +17,21 @@ class Hdbscan:
The HDBSCAN_scores_ @property return the cluster number of each sample (_labels) and the DBCV best score.
"""
def __init__(self, data):
+ # Really fast
# self._param_dist = {'min_samples': [1],
- # 'min_cluster_size':[5,10],
+ # 'min_cluster_size':[5],
# 'metric' : ['euclidean','manhattan'],
# }
- self._param_dist = {'min_samples': [1,5,10,],
- 'min_cluster_size':[5,25,50,],
- 'metric' : ['euclidean','manhattan'],
- }
+ # Medium
+ self._param_dist = {'min_samples': [1,10],
+ 'min_cluster_size':[5,50],
+ 'metric' : ['euclidean','manhattan'],
+ }
+ # Complete
+ # self._param_dist = {'min_samples': [1,5,10,],
+ # 'min_cluster_size':[5,25,50,],
+ # 'metric' : ['euclidean','manhattan'],
+ # }
self._clusterable_embedding = data
diff --git a/pages/1-samples_selection.py b/pages/1-samples_selection.py
index 8f11ea6..f97708c 100644
--- a/pages/1-samples_selection.py
+++ b/pages/1-samples_selection.py
@@ -111,7 +111,7 @@ if not t.empty:
ncluster = scores.number_input(min_value=2, max_value=30, value=3, label = 'Select the desired number of clusters')
cl_model = Sk_Kmeans(tcr, max_clusters = 30)
fig2 = px.scatter(cl_model.inertia_.T, y = 'inertia')
- scores.plotly_chart(fig2)
+ scores.plotly_chart(fig2, use_container_width=True)
data, labels = cl_model.fit_optimal(nclusters = ncluster)
elif clus_method == cluster_methods[2]:
@@ -131,7 +131,7 @@ if not t.empty:
# scores plot with metadata
elif len(list(labels)) == 0 and not meta_data.empty:
filter = meta_data.columns[1:]
- col = st.selectbox('Group by:', options= filter)
+ col = st.selectbox('Color by:', options= filter)
if col == 0:
fig = px.scatter_3d(tcr, x=axis1, y=axis2, z = axis3)
else:
@@ -145,7 +145,7 @@ if not t.empty:
else:
filter = meta_data.columns[1:].insert(0,'None')
- col = st.selectbox('Group by:', options= filter)
+ col = st.selectbox('Color by:', options= filter)
if col == "None":
fig = px.scatter_3d(tcr, x=axis1, y=axis2, z = axis3)
elif col == clus_method:
@@ -156,7 +156,7 @@ if not t.empty:
else:
fig = px.scatter_3d(tcr, x=axis1, y=axis2, z = axis3)
fig.update_traces(marker=dict(size=4))
- st.plotly_chart(fig)
+ st.plotly_chart(fig, use_container_width=True)
@@ -178,13 +178,22 @@ if not spectra.empty:
leverage = dr_model.leverage_
residuals = dr_model.residuals_
fig = px.scatter(x=leverage[ax1], y=residuals[ax1], color = leverage[ax1]*residuals[ax1]).update_layout(xaxis_title="Leverage",yaxis_title="Residuals")
- st.plotly_chart(fig)
+ st.plotly_chart(fig, use_container_width=True)
with hotelling:
- st.write('T²-Hotelling vs Q residuals plot')
- hotelling = dr_model.hotelling_
- ax2 = st.selectbox("Component", options=dr_model.scores_.columns, index=4)
-
- hotelling = dr_model.hotelling_
- fig = px.scatter(t, x=hotelling[ax2], y=residuals[ax2]).update_layout(xaxis_title="T²",yaxis_title="Residuals")
- st.plotly_chart(fig)
\ No newline at end of file
+ st.write('T²-Hotelling vs Q residuals plot')
+ hotelling = dr_model.hotelling_
+ ax2 = st.selectbox("Component", options=dr_model.scores_.columns, index=4)
+
+ hotelling = dr_model.hotelling_
+ fig = px.scatter(t, x=hotelling[ax2], y=residuals[ax2]).update_layout(xaxis_title="T²",yaxis_title="Residuals")
+ st.plotly_chart(fig, use_container_width=True)
+
+ if dim_red_method == dim_red_methods[2] and clus_method == cluster_methods[2]: # UMAP clustered by HDBSCAN
+ with loadings: # Display some clustering metrics
+ st.write('Clustering metrics:')
+ clusters_number = set(labels)
+ clusters_number.remove(-1)
+ st.write('Optimal number of clusters = ' + str(len(clusters_number)))
+ st.write('DBCV score (-1 to 1 - higher is better) = ' + str(round(hdbscan_score,3)))
+ st.write('Unclassified samples: ' + str(len(t[labels==-1])) + ' on ' + str(len(t)) + ' samples (' + str(round(len(t[labels==-1])/len(t)*100, 1)) + '%).')
--
GitLab