Model2 - resnet34 commit

+import torch
+import torch.nn.functional as F
+from torch.autograd import Variable
+
+import pandas as pd
+import math
+import os
+import configure as c
+
+from DB_wav_reader import read_feats_structure
+from SR_Dataset import read_MFB, ToTensorTestInput
+from model.model2 import background_resnet
+
+def load_model(use_cuda, log_dir, cp_num, embedding_size, n_classes):
+    model = background_resnet(embedding_size=embedding_size, num_classes=n_classes)
+    if use_cuda:
+        model.cuda()
+    print('=> loading checkpoint')
+    # original saved file with DataParallel
+    checkpoint = torch.load(log_dir + '/checkpoint_' + str(cp_num) + '.pth')
+    # create new OrderedDict that does not contain `module.`
+    model.load_state_dict(checkpoint['state_dict'])
+    model.eval()
+    return model
+
+def split_enroll_and_test(dataroot_dir):
+    DB_all = read_feats_structure(dataroot_dir)
+    enroll_DB = pd.DataFrame()
+    test_DB = pd.DataFrame()
+    
+    enroll_DB = DB_all[DB_all['filename'].str.contains('enroll.p')]
+    test_DB = DB_all[DB_all['filename'].str.contains('test.p')]
+    
+    # Reset the index
+    enroll_DB = enroll_DB.reset_index(drop=True)
+    test_DB = test_DB.reset_index(drop=True)
+    return enroll_DB, test_DB
+
+def get_embeddings(use_cuda, filename, model, test_frames):
+    input, label = read_MFB(filename) # input size:(n_frames, n_dims)
+    
+    tot_segments = math.ceil(len(input)/test_frames) # total number of segments with 'test_frames' 
+    activation = 0
+    with torch.no_grad():
+        for i in range(tot_segments):
+            temp_input = input[i*test_frames:i*test_frames+test_frames]
+            
+            TT = ToTensorTestInput()
+            temp_input = TT(temp_input) # size:(1, 1, n_dims, n_frames)
+    
+            if use_cuda:
+                temp_input = temp_input.cuda()
+            temp_activation,_ = model(temp_input)
+            activation += torch.sum(temp_activation, dim=0, keepdim=True)
+    
+    activation = l2_norm(activation, 1)
+                
+    return activation
+
+def l2_norm(input, alpha):
+    input_size = input.size()  # size:(n_frames, dim)
+    buffer = torch.pow(input, 2)  # 2 denotes a squared operation. size:(n_frames, dim)
+    normp = torch.sum(buffer, 1).add_(1e-10)  # size:(n_frames)
+    norm = torch.sqrt(normp)  # size:(n_frames)
+    _output = torch.div(input, norm.view(-1, 1).expand_as(input))
+    output = _output.view(input_size)
+    # Multiply by alpha = 10 as suggested in https://arxiv.org/pdf/1703.09507.pdf
+    output = output * alpha
+    return output
+
+def enroll_per_spk(use_cuda, test_frames, model, DB, embedding_dir):
+    """
+    Output the averaged d-vector for each speaker (enrollment)
+    Return the dictionary (length of n_spk)
+    """
+    n_files = len(DB) # 10
+    enroll_speaker_list = sorted(set(DB['speaker_id']))
+    
+    embeddings = {}
+    
+    # Aggregates all the activations
+    print("Start to aggregate all the d-vectors per enroll speaker")
+    
+    for i in range(n_files):
+        filename = DB['filename'][i]
+        spk = DB['speaker_id'][i]
+        
+        activation = get_embeddings(use_cuda, filename, model, test_frames)
+        if spk in embeddings:
+            embeddings[spk] += activation
+        else:
+            embeddings[spk] = activation
+            
+        print("Aggregates the activation (spk : %s)" % (spk))
+        
+    if not os.path.exists(embedding_dir):
+        os.makedirs(embedding_dir)
+        
+    # Save the embeddings
+    for spk_index in enroll_speaker_list:
+        embedding_path = os.path.join(embedding_dir, spk_index+'.pth')
+        torch.save(embeddings[spk_index], embedding_path)
+        print("Save the embeddings for %s" % (spk_index))
+    return embeddings
+    
+def main():
+        
+    # Settings
+    use_cuda = True
+    log_dir = 'new_model2'
+    embedding_size = 128
+    cp_num = 24 # Which checkpoint to use?
+    n_classes = 241
+    test_frames = 200
+    
+    # Load model from checkpoint
+    model = load_model(use_cuda, log_dir, cp_num, embedding_size, n_classes)
+    
+    # Get the dataframe for enroll DB
+    enroll_DB, test_DB = split_enroll_and_test(c.TEST_FEAT_DIR)
+    
+    # Where to save embeddings
+    embedding_dir = 'enroll_embeddings2'
+    
+    # Perform the enrollment and save the results
+    enroll_per_spk(use_cuda, test_frames, model, enroll_DB, embedding_dir)
+    
+    """ Test speaker list
+    '103F3021', '207F2088', '213F5100', '217F3038', '225M4062', 
+    '229M2031', '230M4087', '233F4013', '236M3043', '240M3063'
+    """ 
+
+if __name__ == '__main__':
+    main()
\ No newline at end of file

+import torch
+import torch.nn.functional as F
+from torch.autograd import Variable
+
+import pandas as pd
+import math
+import os
+import configure as c
+
+from DB_wav_reader import read_feats_structure
+from SR_Dataset import read_MFB, ToTensorTestInput
+from model.model2 import background_resnet
+
+def load_model(use_cuda, log_dir, cp_num, embedding_size, n_classes):
+    model = background_resnet(embedding_size=embedding_size, num_classes=n_classes)
+    if use_cuda:
+        model.cuda()
+    print('=> loading checkpoint')
+    # original saved file with DataParallel
+    checkpoint = torch.load(log_dir + '/checkpoint_' + str(cp_num) + '.pth')
+    # create new OrderedDict that does not contain `module.`
+    model.load_state_dict(checkpoint['state_dict'])
+    model.eval()
+    return model
+
+def split_enroll_and_test(dataroot_dir):
+    DB_all = read_feats_structure(dataroot_dir)
+    enroll_DB = pd.DataFrame()
+    test_DB = pd.DataFrame()
+    
+    enroll_DB = DB_all[DB_all['filename'].str.contains('enroll.p')]
+    test_DB = DB_all[DB_all['filename'].str.contains('test.p')]
+    
+    # Reset the index
+    enroll_DB = enroll_DB.reset_index(drop=True)
+    test_DB = test_DB.reset_index(drop=True)
+    return enroll_DB, test_DB
+
+def load_enroll_embeddings(embedding_dir):
+    embeddings = {}
+    for f in os.listdir(embedding_dir):
+        spk = f.replace('.pth','')
+        # Select the speakers who are in the 'enroll_spk_list'
+        embedding_path = os.path.join(embedding_dir, f)
+        tmp_embeddings = torch.load(embedding_path)
+        embeddings[spk] = tmp_embeddings
+        
+    return embeddings
+
+def get_embeddings(use_cuda, filename, model, test_frames):
+    input, label = read_MFB(filename) # input size:(n_frames, n_dims)
+    
+    tot_segments = math.ceil(len(input)/test_frames) # total number of segments with 'test_frames' 
+    activation = 0
+    with torch.no_grad():
+        for i in range(tot_segments):
+            temp_input = input[i*test_frames:i*test_frames+test_frames]
+            
+            TT = ToTensorTestInput()
+            temp_input = TT(temp_input) # size:(1, 1, n_dims, n_frames)
+    
+            if use_cuda:
+                temp_input = temp_input.cuda()
+            temp_activation,_ = model(temp_input)
+            activation += torch.sum(temp_activation, dim=0, keepdim=True)
+    
+    activation = l2_norm(activation, 1)
+                
+    return activation
+
+def l2_norm(input, alpha):
+    input_size = input.size()  # size:(n_frames, dim)
+    buffer = torch.pow(input, 2)  # 2 denotes a squared operation. size:(n_frames, dim)
+    normp = torch.sum(buffer, 1).add_(1e-10)  # size:(n_frames)
+    norm = torch.sqrt(normp)  # size:(n_frames)
+    _output = torch.div(input, norm.view(-1, 1).expand_as(input))
+    output = _output.view(input_size)
+    # Multiply by alpha = 10 as suggested in https://arxiv.org/pdf/1703.09507.pdf
+    output = output * alpha
+    return output
+
+def perform_identification(use_cuda, model, embeddings, test_filename, test_frames, spk_list):
+    test_embedding = get_embeddings(use_cuda, test_filename, model, test_frames)
+    max_score = -10**8
+    best_spk = None
+    for spk in spk_list:
+        score = F.cosine_similarity(test_embedding, embeddings[spk])
+        score = score.data.cpu().numpy() 
+        if score > max_score:
+            max_score = score
+            best_spk = spk
+    #print("Speaker identification result : %s" %best_spk)
+    true_spk = test_filename.split('/')[-2].split('_')[0]
+    print("\n=== Speaker identification ===")
+    print("True speaker : %s\nPredicted speaker : %s\nResult : %s\n" %(true_spk, best_spk, true_spk==best_spk))
+    return best_spk
+
+def main():
+    
+    log_dir = 'new_model2' # Where the checkpoints are saved
+    embedding_dir = 'enroll_embeddings2' # Where embeddings are saved
+    test_dir = 'feat_logfbank_nfilt40/test/' # Where test features are saved
+    
+    # Settings
+    use_cuda = True # Use cuda or not
+    embedding_size = 128 # Dimension of speaker embeddings
+    cp_num = 30 # Which checkpoint to use?
+    n_classes = 241 # How many speakers in training data?
+    test_frames = 100 # Split the test utterance 
+
+    # Load model from checkpoint
+    model = load_model(use_cuda, log_dir, cp_num, embedding_size, n_classes)
+
+    # Get the dataframe for test DB
+    enroll_DB, test_DB = split_enroll_and_test(c.TEST_FEAT_DIR)
+    
+    # Load enroll embeddings
+    embeddings = load_enroll_embeddings(embedding_dir)
+    
+    """ Test speaker list
+    '103F3021', '207F2088', '213F5100', '217F3038', '225M4062', 
+    '229M2031', '230M4087', '233F4013', '236M3043', '240M3063'
+    """ 
+    
+    spk_list = ['103F3021', '207F2088', '213F5100', '217F3038', '225M4062',\
+    '229M2031', '230M4087', '233F4013', '236M3043', '240M3063','777M7777','778M8777']
+    
+    # Set the test speaker
+    test_speaker = '236M3043' 
+    
+    test_path = os.path.join(test_dir, test_speaker, 'test.p')
+    
+    # Perform the test 
+    best_spk = perform_identification(use_cuda, model, embeddings, test_path, test_frames, spk_list)
+
+if __name__ == '__main__':
+    main()

+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.autograd import Function
+import model.resnet as resnet
+
+
+class background_resnet(nn.Module):
+    def __init__(self, embedding_size, num_classes, backbone='resnet34'):
+        super(background_resnet, self).__init__()
+        self.backbone = backbone
+        # copying modules from pretrained models
+        if backbone == 'resnet50':
+            self.pretrained = resnet.resnet50(pretrained=False)
+        elif backbone == 'resnet101':
+            self.pretrained = resnet.resnet101(pretrained=False)
+        elif backbone == 'resnet152':
+            self.pretrained = resnet.resnet152(pretrained=False)
+        elif backbone == 'resnet18':
+            self.pretrained = resnet.resnet18(pretrained=False)
+        elif backbone == 'resnet34':
+            self.pretrained = resnet.resnet34(pretrained=False)
+        else:
+            raise RuntimeError('unknown backbone: {}'.format(backbone))
+            
+        self.fc0 = nn.Linear(128, embedding_size)
+        self.bn0 = nn.BatchNorm1d(embedding_size)
+        self.relu = nn.ReLU()
+        self.last = nn.Linear(embedding_size, num_classes)
+
+    def forward(self, x):
+        # input x: minibatch x 1 x 40 x 40
+        x = self.pretrained.conv1(x)
+        x = self.pretrained.bn1(x)
+        x = self.pretrained.relu(x)
+        
+        x = self.pretrained.layer1(x)
+        x = self.pretrained.layer2(x)
+        x = self.pretrained.layer3(x)
+        x = self.pretrained.layer4(x)
+        
+        out = F.adaptive_avg_pool2d(x,1) # [batch, 128, 1, 1]
+        out = torch.squeeze(out) # [batch, n_embed]
+        # flatten the out so that the fully connected layer can be connected from here
+        out = out.view(x.size(0), -1) # (n_batch, n_embed)
+        spk_embedding = self.fc0(out)
+        out = F.relu(self.bn0(spk_embedding)) # [batch, n_embed]
+        out = self.last(out)
+        
+        return spk_embedding, out
\ No newline at end of file

+import torch
+import torch.nn.functional as F
+from torch.autograd import Variable
+
+import pandas as pd
+import math
+import os
+import configure as c
+
+from DB_wav_reader import read_feats_structure
+from SR_Dataset import read_MFB, ToTensorTestInput
+from model.model2 import background_resnet
+
+def load_model(use_cuda, log_dir, cp_num, embedding_size, n_classes):
+    model = background_resnet(embedding_size=embedding_size, num_classes=n_classes)
+    if use_cuda:
+        model.cuda()
+    print('=> loading checkpoint')
+    # original saved file with DataParallel
+    checkpoint = torch.load(log_dir + '/checkpoint_' + str(cp_num) + '.pth')
+    # create new OrderedDict that does not contain `module.`
+    model.load_state_dict(checkpoint['state_dict'])
+    model.eval()
+    return model
+
+def split_enroll_and_test(dataroot_dir):
+    DB_all = read_feats_structure(dataroot_dir)
+    enroll_DB = pd.DataFrame()
+    test_DB = pd.DataFrame()
+    
+    enroll_DB = DB_all[DB_all['filename'].str.contains('enroll.p')]
+    test_DB = DB_all[DB_all['filename'].str.contains('test.p')]
+    
+    # Reset the index
+    enroll_DB = enroll_DB.reset_index(drop=True)
+    test_DB = test_DB.reset_index(drop=True)
+    return enroll_DB, test_DB
+
+def load_enroll_embeddings(embedding_dir):
+    embeddings = {}
+    for f in os.listdir(embedding_dir):
+        spk = f.replace('.pth','')
+        # Select the speakers who are in the 'enroll_spk_list'
+        embedding_path = os.path.join(embedding_dir, f)
+        tmp_embeddings = torch.load(embedding_path)
+        embeddings[spk] = tmp_embeddings
+        
+    return embeddings
+
+def get_embeddings(use_cuda, filename, model, test_frames):
+    input, label = read_MFB(filename) # input size:(n_frames, n_dims)
+    
+    tot_segments = math.ceil(len(input)/test_frames) # total number of segments with 'test_frames' 
+    activation = 0
+    with torch.no_grad():
+        for i in range(tot_segments):
+            temp_input = input[i*test_frames:i*test_frames+test_frames]
+            
+            TT = ToTensorTestInput()
+            temp_input = TT(temp_input) # size:(1, 1, n_dims, n_frames)
+    
+            if use_cuda:
+                temp_input = temp_input.cuda()
+            temp_activation,_ = model(temp_input)
+            activation += torch.sum(temp_activation, dim=0, keepdim=True)
+    
+    activation = l2_norm(activation, 1)
+                
+    return activation
+
+def l2_norm(input, alpha):
+    input_size = input.size()  # size:(n_frames, dim)
+    buffer = torch.pow(input, 2)  # 2 denotes a squared operation. size:(n_frames, dim)
+    normp = torch.sum(buffer, 1).add_(1e-10)  # size:(n_frames)
+    norm = torch.sqrt(normp)  # size:(n_frames)
+    _output = torch.div(input, norm.view(-1, 1).expand_as(input))
+    output = _output.view(input_size)
+    # Multiply by alpha = 10 as suggested in https://arxiv.org/pdf/1703.09507.pdf
+    output = output * alpha
+    return output
+
+def perform_verification(use_cuda, model, embeddings, enroll_speaker, test_filename, test_frames, thres):
+    enroll_embedding = embeddings[enroll_speaker]
+    test_embedding = get_embeddings(use_cuda, test_filename, model, test_frames)
+
+    score = F.cosine_similarity(test_embedding, enroll_embedding)
+    score = score.data.cpu().numpy() 
+        
+    if score > thres:
+        result = 'Accept'
+    else:
+        result = 'Reject'
+        
+    test_spk = test_filename.split('/')[-2].split('_')[0]
+    print("\n=== Speaker verification ===")
+    print("True speaker: %s\nClaimed speaker : %s\n\nResult : %s\n" %(enroll_speaker, test_spk, result))
+    print("Score : %0.4f\nThreshold : %0.2f\n" %(score, thres))
+
+def main():
+    
+    log_dir = 'new_model2' # Where the checkpoints are saved
+    embedding_dir = 'enroll_embeddings2' # Where embeddings are saved
+    test_dir = 'feat_logfbank_nfilt40/test/' # Where test features are saved
+    
+    # Settings
+    use_cuda = True # Use cuda or not
+    embedding_size = 128 # Dimension of speaker embeddings
+    cp_num = 30  # Which checkpoint to use?
+    n_classes = 241 # How many speakers in training data?
+    test_frames = 100 # Split the test utterance 
+
+    # Load model from checkpoint
+    model = load_model(use_cuda, log_dir, cp_num, embedding_size, n_classes)
+    
+    # Get the dataframe for test DB
+    enroll_DB, test_DB = split_enroll_and_test(c.TEST_FEAT_DIR)
+    
+    # Load enroll embeddings
+    embeddings = load_enroll_embeddings(embedding_dir)
+    
+    """ Test speaker list
+    '103F3021', '207F2088', '213F5100', '217F3038', '225M4062', 
+    '229M2031', '230M4087', '233F4013', '236M3043', '240M3063'
+    """ 
+    
+    # Set the true speaker
+    enroll_speaker = '103F3021'
+    
+    # Set the claimed speaker
+    test_speaker = '207F2088' 
+    
+    # Threshold
+    thres = 0.95
+    
+    test_path = os.path.join(test_dir, test_speaker, 'test.p')
+    
+    # Perform the test 
+    perform_verification(use_cuda, model, embeddings, enroll_speaker, test_path, test_frames, thres)
+
+if __name__ == '__main__':
+    main()