11/01/2023 - predictor function added

configs/parameters.yaml

 # PATHS
 wav_path: '../data/wavs/SPEECOMCO/'
 data_path: './data/'
-embedding_path: './data/embeddings/ecapa/'
+embedding_path: './data/embeddings/xvec/'
 model_path: './models/'
 
-train_set_file: './data/TRAIN_LEC_SEGAUG_C2SI_98_ECAPA.csv'            #'./data/TRAINING.txt'
-validation_set_file: './data/VALIDATION_LEC_SEGAUG_C2SI_10_ECAPA.csv'
-test_set_file: './data/TEST_DAP_UNSEGMENTED_ECAPA.csv'
+train_set_file: './data/TRAIN_LEC_SEGAUG_C2SI_98_XVEC.csv'            #'./data/TRAINING.txt'
+validation_set_file: './data/VALIDATION_LEC_SEGAUG_C2SI_10_XVEC.csv'
+test_set_file: './data/TEST_DAP_UNSEGMENTED_XVEC.csv'
 
 
 
@@ -21,7 +21,7 @@ dropout: 0.2
 
 # MODEL PARAMETERS
 
-first_layer: 192
+first_layer: 512
 second_layer: 128
 third_layer: 64
 
@@ -30,4 +30,4 @@ third_layer: 64
 # Types of embeddings supported: 'ecapa_tdnn' or 'x-vector'
 # ecapa_tdnn: dim = 192 (change first_layer dim)
 # x-vector: dim = 512   (change first_layer dim)
-embedding_type: ecapa_tdnn
+embedding_type: x-vector

data/.~lock.FULL_TRAIN_LEC_SEGAUG_C2SI_108_ECAPA.csv#

-,sebastiao,Valinor,10.01.2023 13:49,file:///homelocal/sebastiao/.config/libreoffice/4;
\ No newline at end of file

data/.~lock.VALIDATION_LEC_SEGAUG_C2SI_10_ECAPA.csv#

-,sebastiao,Valinor,10.01.2023 13:54,file:///homelocal/sebastiao/.config/libreoffice/4;
\ No newline at end of file

predictor/model.py

+import yaml
+import torch
+import torch.nn as nn
+import torch.nn.functional as F 
+
+with open("./parameters.yaml", "r") as ymlfile:
+    cfg = yaml.safe_load(ymlfile)
+
+
+class model_embedding_snn(nn.Module):
+    def __init__(self):
+        super(model_embedding_snn,self).__init__()
+
+        self.relu = nn.ReLU()
+        #self.dropout = nn.Dropout2d(cfg['dropout'])
+        self.dropout = nn.Dropout(cfg['dropout'])
+
+        self.batch_norm1 = nn.BatchNorm1d(cfg['first_layer'])
+        self.batch_norm2 = nn.BatchNorm1d(cfg['second_layer'])
+        self.batch_norm3 = nn.BatchNorm1d(cfg['third_layer'])
+
+        self.fc1 = nn.Linear(cfg['first_layer'],cfg['second_layer'])
+        self.fc2 = nn.Linear(cfg['second_layer'],cfg['third_layer'])
+        
+        self.fc_voix = nn.Linear(cfg['third_layer'],1)
+        self.fc_res = nn.Linear(cfg['third_layer'],1)
+        self.fc_pros = nn.Linear(cfg['third_layer'],1)
+        self.fc_pd = nn.Linear(cfg['third_layer'],1)
+
+        self.fc_sev = nn.Linear(cfg['third_layer'],1)
+        self.fc_int = nn.Linear(cfg['third_layer'],1)        
+
+    def forward(self, input_embs):
+
+        x = self.batch_norm1(input_embs)
+        x = self.fc1(x)
+        x = self.dropout(x)
+        x = self.relu(x)
+        x = self.batch_norm2(x)
+        x = self.fc2(x)
+        x = self.dropout(x)
+        x = self.relu(x)
+        x = self.batch_norm3(x)
+
+        v = self.fc_voix(x)
+        v = self.relu(v)
+        r = self.fc_res(x)
+        r = self.relu(r)
+        p = self.fc_pros(x)
+        p = self.relu(p)
+        pd = self.fc_pd(x)
+        pd = self.relu(pd)
+
+        INT = self.fc_int(x)
+        INT = self.relu(INT)
+
+        SEV = self.fc_sev(x)
+        SEV = self.relu(SEV)
+
+        return SEV, INT, v, r, p, pd

predictor/parameters.yaml

+# PATHS
+wav_path: '../data/wavs/SPEECOMCO/'
+data_path: './data/'
+embedding_path: './data/embeddings/xvec/'
+model_path: './models/'
+
+train_set_file: './data/TRAIN_LEC_SEGAUG_C2SI_98_XVEC.csv'            #'./data/TRAINING.txt'
+validation_set_file: './data/VALIDATION_LEC_SEGAUG_C2SI_10_XVEC.csv'
+test_set_file: './data/TEST_DAP_UNSEGMENTED_XVEC.csv'
+
+
+
+# TRAINING PARAMETERS
+sampling_rate: 16000
+batch_size: 8
+learning_rate: 0.001
+epochs: 20
+dropout: 0.2
+
+
+
+# MODEL PARAMETERS
+
+first_layer: 512
+second_layer: 128
+third_layer: 64
+
+
+
+# Types of embeddings supported: 'ecapa_tdnn' or 'x-vector'
+# ecapa_tdnn: dim = 192 (change first_layer dim)
+# x-vector: dim = 512   (change first_layer dim)
+embedding_type: x-vector

predictor/predictor.py

+import torch
+#import yaml
+import model
+import torchaudio
+
+from speechbrain.pretrained import EncoderClassifier
+
+
+#with open("./configs/parameters.yaml", "r") as ymlfile:
+#    cfg = yaml.safe_load(ymlfile)
+
+MODEL_PATH = './model_snn_x-vector'
+
+
+class PythonPredictor():
+
+    def __init__(self):
+        super(PythonPredictor,self).__init__()
+
+        self.model = model.model_embedding_snn().cuda()
+        #self.model.load_state_dict(torch.load(cfg['model_path']+'model_snn_'\
+        #    +cfg['embedding_type']))
+        self.model.load_state_dict(torch.load(MODEL_PATH))
+        self.model.eval()
+
+        self.emb_extractor = EncoderClassifier.from_hparams(source=\
+                "speechbrain/spkrec-xvect-voxceleb", savedir=\
+                "pretrained_models/spkrec-xvect-voxceleb")
+
+    def predict(self,wavfile):
+
+        signal, fs =torchaudio.load(wavfile)
+        emb = self.emb_extractor.encode_batch(signal)
+        emb = emb.squeeze().unsqueeze(0).cuda()
+        
+        sev_,int_,v_,r_,p_,pd_=self.model(emb)
+              
+        print('\nFilename: {}\n'.format(wavfile))
+
+        print('Speech Disorder Severity:   {}'.format(round(sev_.item(),3)))
+        print('Speech Intelligibility:     {}\n'.format(round(int_.item(),3)))
+        

predictor/pretrained_models/spkrec-xvect-voxceleb/classifier.ckpt

+/homelocal/sebastiao/.cache/huggingface/hub/models--speechbrain--spkrec-xvect-voxceleb/snapshots/e2cc27f853f99bd5d539432f0cba3f124c059f71/classifier.ckpt
\ No newline at end of file

predictor/pretrained_models/spkrec-xvect-voxceleb/embedding_model.ckpt

+/homelocal/sebastiao/.cache/huggingface/hub/models--speechbrain--spkrec-xvect-voxceleb/snapshots/e2cc27f853f99bd5d539432f0cba3f124c059f71/embedding_model.ckpt
\ No newline at end of file

predictor/pretrained_models/spkrec-xvect-voxceleb/hyperparams.yaml

+/homelocal/sebastiao/.cache/huggingface/hub/models--speechbrain--spkrec-xvect-voxceleb/snapshots/e2cc27f853f99bd5d539432f0cba3f124c059f71/hyperparams.yaml
\ No newline at end of file