add some comments to the code

examples/voxceleb/sv0/conf/ecapa_tdnn.yaml

@@ -1,3 +1,12 @@
+###########################################
+#                Data                 #
+###########################################
+batch_size: 32
+num_workers: 2
+num_speakers: 7205 # 1211 vox1, 5994 vox2, 7205 vox1+2, test speakers: 41
+shuffle: True
+random_chunk: True
+
 ###########################################################
 #                FEATURE EXTRACTION SETTING               #
 ###########################################################
@@ -7,7 +16,6 @@ feature:
   window_size: 400     #25ms, sample rate 16000, 25 * 16000 / 1000 = 400 
   hop_length: 160     #10ms, sample rate 16000, 10 * 16000 / 1000 = 160
 
-
 ###########################################################
 #                       MODEL SETTING                     #
 ###########################################################
@@ -15,9 +23,8 @@ feature:
 # if we want use another model, please choose another configuration yaml file
 model:
   input_size: 80
-  ##"channels": [1024, 1024, 1024, 1024, 3072],
   # "channels": [512, 512, 512, 512, 1536],
-  channels: [512, 512, 512, 512, 1536]
+  channels: [1024, 1024, 1024, 1024, 3072]
   kernel_sizes: [5, 3, 3, 3, 1]
   dilations: [1, 2, 3, 4, 1]
   attention_channels: 128
@@ -26,10 +33,9 @@ model:
 ###########################################
 #                Training                 #
 ###########################################
-seed: 0
+seed: 1986 # according from speechbrain configuration
 epochs: 10
-batch_size: 32
-num_workers: 2
 save_freq: 10
-log_freq: 10
+log_interval: 10
 learning_rate: 1e-8
+

examples/voxceleb/sv0/run.sh

@@ -47,7 +47,8 @@ mkdir -p ${exp_dir}
 if [ $stage -le 0 ]; then 
      # stage 0: data prepare for vox1 and vox2, vox2 must be converted from m4a to wav
      python3 local/data_prepare.py \
-     --data-dir ${dir} --augment --vox2-base-path ${vox2_base_path}
+     --data-dir ${dir} --augment --vox2-base-path ${vox2_base_path} \
+     --config conf/ecapa_tdnn.yaml
 fi 
 
 if [ $stage -le 1 ]; then

paddleaudio/paddleaudio/metric/__init__.py

@@ -12,5 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from .dtw import dtw_distance
-from .mcd import mcd_distance
 from .eer import compute_eer
+from .eer import compute_minDCF
+from .mcd import mcd_distance

paddleaudio/paddleaudio/metric/eer.py

@@ -14,6 +14,7 @@
 from typing import List
 
 import numpy as np
+import paddle
 from sklearn.metrics import roc_curve
 
 
@@ -26,3 +27,68 @@ def compute_eer(labels: np.ndarray, scores: np.ndarray) -> List[float]:
     eer_threshold = threshold[np.nanargmin(np.absolute((fnr - fpr)))]
     eer = fpr[np.nanargmin(np.absolute((fnr - fpr)))]
     return eer, eer_threshold
+
+
+def compute_minDCF(positive_scores,
+                   negative_scores,
+                   c_miss=1.0,
+                   c_fa=1.0,
+                   p_target=0.01):
+    """
+    This is modified from SpeechBrain
+    https://github.com/speechbrain/speechbrain/blob/085be635c07f16d42cd1295045bc46c407f1e15b/speechbrain/utils/metric_stats.py#L509
+    Computes the minDCF metric normally used to evaluate speaker verification
+    systems. The min_DCF is the minimum of the following C_det function computed
+    within the defined threshold range:
+
+    C_det =  c_miss * p_miss * p_target + c_fa * p_fa * (1 -p_target)
+
+    where p_miss is the missing probability and p_fa is the probability of having
+    a false alarm.
+
+    Args:
+        positive_scores (Paddle.Tensor): The scores from entries of the same class.
+        negative_scores (Paddle.Tensor): The scores from entries of different classes.
+        c_miss (float, optional): Cost assigned to a missing error (default 1.0).
+        c_fa (float, optional): Cost assigned to a false alarm (default 1.0).
+        p_target (float, optional): Prior probability of having a target (default 0.01).
+
+    Returns:
+        _type_: min dcf 
+    """
+    # Computing candidate thresholds
+    if len(positive_scores.shape) > 1:
+        positive_scores = positive_scores.squeeze()
+
+    if len(negative_scores.shape) > 1:
+        negative_scores = negative_scores.squeeze()
+
+    thresholds = paddle.sort(paddle.concat([positive_scores, negative_scores]))
+    thresholds = paddle.unique(thresholds)
+
+    # Adding intermediate thresholds
+    interm_thresholds = (thresholds[0:-1] + thresholds[1:]) / 2
+    thresholds = paddle.sort(paddle.concat([thresholds, interm_thresholds]))
+
+    # Computing False Rejection Rate (miss detection)
+    positive_scores = paddle.concat(
+        len(thresholds) * [positive_scores.unsqueeze(0)])
+    pos_scores_threshold = positive_scores.transpose(perm=[1, 0]) <= thresholds
+    p_miss = (pos_scores_threshold.sum(0)
+              ).astype("float32") / positive_scores.shape[1]
+    del positive_scores
+    del pos_scores_threshold
+
+    # Computing False Acceptance Rate (false alarm)
+    negative_scores = paddle.concat(
+        len(thresholds) * [negative_scores.unsqueeze(0)])
+    neg_scores_threshold = negative_scores.transpose(perm=[1, 0]) > thresholds
+    p_fa = (neg_scores_threshold.sum(0)
+            ).astype("float32") / negative_scores.shape[1]
+    del negative_scores
+    del neg_scores_threshold
+
+    c_det = c_miss * p_miss * p_target + c_fa * p_fa * (1 - p_target)
+    c_min = paddle.min(c_det, axis=0)
+    min_index = paddle.argmin(c_det, axis=0)
+    return float(c_min), float(thresholds[min_index])

paddlespeech/vector/exps/ecapa_tdnn/speaker_verification_cosine.py

@@ -45,7 +45,7 @@ def main(args, config):
 
     # stage2: build the speaker verification eval instance with backbone model
     model = SpeakerIdetification(
-        backbone=ecapa_tdnn, num_class=VoxCeleb.num_speakers)
+        backbone=ecapa_tdnn, num_class=config.num_speakers)
 
     # stage3: load the pre-trained model
     args.load_checkpoint = os.path.abspath(
@@ -93,6 +93,7 @@ def main(args, config):
     model.eval()
 
     # stage7: global embedding norm to imporve the performance
+    print("global embedding norm: {}".format(args.global_embedding_norm))
     if args.global_embedding_norm:
         global_embedding_mean = None
         global_embedding_std = None
@@ -118,6 +119,8 @@ def main(args, config):
                         -1).numpy()  # (N, emb_size, 1) -> (N, emb_size)
 
                     # Global embedding normalization.
+                    # if we use the global embedding norm
+                    # eer can reduece about relative 10%
                     if args.global_embedding_norm:
                         batch_count += 1
                         current_mean = embeddings.mean(
@@ -150,8 +153,8 @@ def main(args, config):
         for line in f.readlines():
             label, enrol_id, test_id = line.strip().split(' ')
             labels.append(int(label))
-            enrol_ids.append(enrol_id.split('.')[0].replace('/', '-'))
-            test_ids.append(test_id.split('.')[0].replace('/', '-'))
+            enrol_ids.append(enrol_id.split('.')[0].replace('/', '--'))
+            test_ids.append(test_id.split('.')[0].replace('/', '--'))
 
     cos_sim_func = paddle.nn.CosineSimilarity(axis=1)
     enrol_embeddings, test_embeddings = map(lambda ids: paddle.to_tensor(
@@ -185,11 +188,10 @@ if __name__ == "__main__":
                         default='',
                         help="Directory to load model checkpoint to contiune trainning.")
     parser.add_argument("--global-embedding-norm",
-                        type=bool,
-                        default=True,
+                        default=False,
+                        action="store_true",
                         help="Apply global normalization on speaker embeddings.")
     parser.add_argument("--embedding-mean-norm",
-                        type=bool,
                         default=True,
                         help="Apply mean normalization on speaker embeddings.")
     parser.add_argument("--embedding-std-norm",

paddlespeech/vector/exps/ecapa_tdnn/train.py

@@ -178,9 +178,9 @@ def main(args, config):
             timer.count()  # step plus one in timer
 
             # stage 9-10: print the log information only on 0-rank per log-freq batchs
-            if (batch_idx + 1) % config.log_freq == 0 and local_rank == 0:
+            if (batch_idx + 1) % config.log_interval == 0 and local_rank == 0:
                 lr = optimizer.get_lr()
-                avg_loss /= config.log_freq
+                avg_loss /= config.log_interval
                 avg_acc = num_corrects / num_samples
 
                 print_msg = 'Train Epoch={}/{}, Step={}/{}'.format(
@@ -196,7 +196,7 @@ def main(args, config):
                 num_samples = 0
 
         # stage 9-11: save the model parameters only on 0-rank per save-freq batchs
-        if epoch % config.save_freq == 0 and batch_idx + 1 == steps_per_epoch:
+        if epoch % config.save_interval == 0 and batch_idx + 1 == steps_per_epoch:
             if local_rank != 0:
                 paddle.distributed.barrier(
                 )  # Wait for valid step in main process

paddlespeech/vector/io/augment.py

@@ -11,7 +11,8 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-# this is modified from https://github.com/speechbrain/speechbrain/blob/085be635c07f16d42cd1295045bc46c407f1e15b/speechbrain/lobes/augment.py
+# this is modified from SpeechBrain
+# https://github.com/speechbrain/speechbrain/blob/085be635c07f16d42cd1295045bc46c407f1e15b/speechbrain/lobes/augment.py
 import math
 import os
 from typing import List

paddlespeech/vector/io/batch.py

@@ -75,6 +75,9 @@ def batch_feature_normalize(batch, mean_norm: bool=True, std_norm: bool=True):
             i]:].sum() == 0  # Padding valus should all be 0.
 
     # Converts into ratios.
+    # the utterance of the max length doesn't need to padding
+    # the remaining utterances need to padding and all of them will be padded to max length
+    # we convert the original length of each utterance to the ratio of the max length
     lengths = (lengths / lengths.max()).astype(np.float32)
 
     return {'ids': ids, 'feats': feats, 'lengths': lengths}