skip k-means

.gitignore

@@ -4,6 +4,7 @@
 */audio
 */TextGrid
 *__pycache__
+*log
 humpback2/annots
 
 otter/pone.0112562.s003.xlsx

models.py

@@ -62,7 +62,7 @@ meta = {
   }
 }
 
-vgg16 = torchmodels.vgg16(weights=torchmodels.VGG16_Weights.DEFAULT)
+vgg16 = torchmodels.vgg16(pretrained=True) # weights=torchmodels.VGG16_Weights.DEFAULT)
 vgg16 = vgg16.features[:13]
 for nm, mod in vgg16.named_modules():
     if isinstance(mod, nn.MaxPool2d):
@@ -91,7 +91,7 @@ frontend = {
   'pcenMel': lambda sr, nfft, sampleDur, n_mel : nn.Sequential(
     STFT(nfft, int((sampleDur*sr - nfft)/128)),
     MelFilter(sr, nfft, n_mel, sr//nfft, sr//2),
-    PCENLayer(n_mel),
+    PCENLayer(n_mel, trainable=False),
     u.Croper2D(n_mel, 128)
   )
 }

train_AE.py

@@ -97,7 +97,10 @@ for epoch in range(100_000//len(loader)):
             mask = ~df.loc[idxs].label.isna()
             clusters, labels = clusters[mask], df.loc[idxs[mask]].label
             writer.add_scalar('NMI HDBSCAN', metrics.normalized_mutual_info_score(labels, clusters), step)
-            writer.add_scalar('ARI HDBSCAN', metrics.adjusted_rand_score(labels, clusters), step)
+            try:
+                writer.add_scalar('ARI HDBSCAN', metrics.adjusted_rand_score(labels, clusters), step)
+            except:
+                pass
             writer.add_scalar('Homogeneity HDBSCAN', metrics.homogeneity_score(labels, clusters), step)
             writer.add_scalar('Completeness HDBSCAN', metrics.completeness_score(labels, clusters), step)
             writer.add_scalar('V-Measure HDBSCAN', metrics.v_measure_score(labels, clusters), step)
@@ -112,7 +115,7 @@ for epoch in range(100_000//len(loader)):
             writer.add_histogram('HDBSCAN Precisions ', np.array(precs), step)
             writer.add_histogram('HDBSCAN Recalls ', np.array(recs), step)
             df.drop('cluster', axis=1, inplace=True)
-
+            continue
             print('\rRunning elbow method for K-Means...', end='')
             ks = (5*1.2**np.arange(20)).astype(int)
             distorsions = [cluster.KMeans(n_clusters=k).fit(encodings).inertia_ for k in ks]
@@ -126,7 +129,10 @@ for epoch in range(100_000//len(loader)):
             writer.add_scalar('Silhouette', metrics.silhouette_score(encodings, clusters), step)
             clusters, labels = clusters[mask], df.loc[idxs[mask]].label
             writer.add_scalar('NMI K-Means', metrics.normalized_mutual_info_score(labels, clusters), step)
-            writer.add_scalar('ARI K-Means', metrics.adjusted_rand_score(labels, clusters), step)
+            try:
+                writer.add_scalar('ARI K-Means', metrics.adjusted_rand_score(labels, clusters), step)
+            except:
+                pass
             writer.add_scalar('Homogeneity K-Means', metrics.homogeneity_score(labels, clusters), step)
             writer.add_scalar('Completeness K-Means', metrics.completeness_score(labels, clusters), step)
             writer.add_scalar('V-Measure K-Means', metrics.v_measure_score(labels, clusters), step)