second commit, minor changes

src/evaluate_features.py

-import numpy as np
-import pickle as pkl
-from sklearn.metrics.pairwise import cosine_similarity
-from sklearn.metrics import roc_curve
-
-import os
-import sys
-import argparse
-import glob
-
-def load_features(input_folder):
-	# A lot of hardcoded stuff for librispeech format, ID probably should have been saved in extract features
-	feat=[]
-	ids=[]
-	for filename in sorted(glob.glob(args.input_folder+'/*')):
-		#print(filename)
-		ids.append(filename.split('/')[-1].split('-')[0])
-		ifile=open(filename,'rb')
-		feat.append(pkl.load(ifile))
-	return ids, feat
-if __name__ == "__main__":
-	parser = argparse.ArgumentParser(description="Compute similarity and EER on a set of embeddings")
-	parser.add_argument("-i", "--input_folder", type=str, help="Embeddings folder")
-	args = parser.parse_args()
-	ids, feats = load_features(args.input_folder)
-	
-	d = cosine_similarity(np.stack(feats))
-
-	dup=np.triu(d+1,1) ##I take only the upper part
-	y_score=d[dup>0]
-	#del d
-	#del dup
-	
-	labels=np.array([i==j for i in ids for j in ids], dtype=int ).reshape(-1,len(ids))
-	dup=np.triu(labels+1,1) ##I take only the upper part
-
-	y_true=labels[dup>0]
-
-
-	fpr, tpr, threshold = roc_curve(y_true, y_score)
-	fnr = 1 - tpr
-	
-	
-	EER = fpr[np.nanargmin(np.abs((fnr - fpr)))]
-	print(EER)
-	EER = fnr[np.nanargmin(np.absolute((fnr - fpr)))]
-	print(EER)

src/extract_features.py

@@ -47,13 +47,12 @@ for foldername in sorted(glob.glob(args.input_dir+'/*/')):
 		print(audio_file)
 		signal = utils.get_fft_spectrum(audio_file, buckets,SAMPLE_RATE,NUM_FFT,FRAME_LEN,FRAME_STEP,PREEMPHASIS_ALPHA)
 		feats.append(np.squeeze(model.predict(signal.reshape(1,*signal.shape,1))))
+		if not os.path.isdir(args.output_dir+'/'+sessionname):
+			os.mkdir(args.output_dir+'/'+sessionname)
+		fdump=open(audio_file.replace(args.input_dir,args.output_dir).replace('wav','pkl'),'wb')
+		pkl.dump(feats[-1], fdump, protocol=pkl.HIGHEST_PROTOCOL)
+		fdump.close()
 		
-		#fdump=open(args.output_dir+'/'+audio_file.split('/')[-1].split('.')[0]+'.pkl','wb')
-		#pkl.dump(embs, fdump, protocol=pkl.HIGHEST_PROTOCOL)
-		#fdump.close()
-
-
-
 d = cosine_similarity(np.stack(feats))
 print(d.shape)
 print(d)