Merge branch 'main' into nmt-bottleneck

Jenkinsfile

@@ -1001,56 +1001,6 @@ pipeline {
               sh 'ls -lha examples/nlp/language_modeling'
             }
         }
-        stage('L2: Pretraining BERT pretraining from Text with char tokenizer') {
-            steps {
-              sh 'cd examples/nlp/language_modeling && \
-              python bert_pretraining.py \
-              --config-name=bert_pretraining_from_text_config.yaml \
-              trainer.gpus=[0] \
-              trainer.precision=16 \
-              trainer.amp_level=O1 \
-              +trainer.fast_dev_run=true \
-              model.train_ds.data_file=/home/TestData/nlp/wikitext-2/train.txt  \
-              model.train_ds.batch_size=32 \
-              model.validation_ds.data_file=/home/TestData/nlp/wikitext-2/valid.txt  \
-              model.validation_ds.batch_size=32 \
-              model.language_model.config_file=/home/TestData/nlp/bert_configs/bert_3200.json \
-              model.optim.lr=0.01 \
-              model.optim.sched.warmup_ratio=0.1 \
-              model.tokenizer.tokenizer_name=char \
-              model.tokenizer.vocab_file=/home/TestData/nlp/vocabs/mini_vocab.txt \
-              model.mask_prob=0.15 \
-              model.short_seq_prob=0.1 \
-              exp_manager.exp_dir=PretrainingBERTFromTextchartok \
-              '
-              sh 'rm -rf examples/nlp/language_modeling/PretrainingBERTFromTextchartok'
-            }
-        }
-        stage('L2: Pretraining BERT pretraining from Text with word tokenizer') {
-            steps {
-              sh 'cd examples/nlp/language_modeling && \
-              python bert_pretraining.py \
-              --config-name=bert_pretraining_from_text_config.yaml \
-              trainer.gpus=[1] \
-              trainer.precision=16 \
-              trainer.amp_level=O1 \
-              +trainer.fast_dev_run=true \
-              model.train_ds.data_file=/home/TestData/nlp/wikitext-2/train.txt  \
-              model.train_ds.batch_size=32 \
-              model.validation_ds.data_file=/home/TestData/nlp/wikitext-2/valid.txt  \
-              model.validation_ds.batch_size=32 \
-              model.language_model.config_file=/home/TestData/nlp/bert_configs/bert_3200.json \
-              model.optim.lr=0.01 \
-              model.optim.sched.warmup_ratio=0.1 \
-              model.tokenizer.tokenizer_name=word \
-              model.tokenizer.vocab_file=/home/TestData/nlp/vocabs/mini_vocab.txt \
-              model.mask_prob=0.15 \
-              model.short_seq_prob=0.1 \
-              exp_manager.exp_dir=PretrainingBERTFromTextwordtok \
-              '
-              sh 'rm -rf examples/nlp/language_modeling/PretrainingBERTFromTextwordtok'
-            }
-        }
       }
     }
 

examples/speaker_recognition/conf/SpeakerNet_ECAPA.yaml

@@ -0,0 +1,99 @@
+name: "SpeakerNet"
+
+model:
+
+  sample_rate: 16000
+
+  train_ds:
+    manifest_filepath: ???
+    sample_rate: ${model.sample_rate}
+    labels: null
+    batch_size: 64
+    shuffle: True
+    time_length: 3
+    augmentor:
+      noise:
+        manifest_path: null
+        prob: 0.5
+        min_snr_db: 0
+        max_snr_db: 15
+
+      speed:
+        prob: 0.5
+        sr: ${model.sample_rate}
+        resample_type: 'kaiser_fast'
+        min_speed_rate: 0.95
+        max_speed_rate: 1.05
+
+  validation_ds:
+    manifest_filepath: ???
+    sample_rate: ${model.sample_rate}
+    labels: null
+    batch_size: 128
+    shuffle: False
+    time_length: 3
+
+  preprocessor:
+    _target_: nemo.collections.asr.modules.AudioToMelSpectrogramPreprocessor
+    normalize: "per_feature"
+    window_size: 0.025
+    sample_rate: ${model.sample_rate}
+    window_stride: 0.01
+    window: "hann"
+    features: 80
+    n_fft: 512
+    frame_splicing: 1
+    dither: 0.00001
+    stft_conv: false
+
+  encoder:
+    _target_: nemo.collections.asr.modules.ECAPAEncoder
+    feat_in: ${model.preprocessor.features}
+    filters: [1024,1024,1024,1024,3072]
+    kernel_sizes: [5,3,3,3,1]
+    dilations: [1,1,1,1,1]
+    scale: 8
+
+
+  decoder:
+    _target_: nemo.collections.asr.modules.SpeakerDecoder
+    feat_in: 3072
+    num_classes: 7205
+    pool_mode: 'attention' #xvector,tap or attention
+    emb_sizes: 192
+    angular: True 
+
+  loss:
+    scale: 30
+    margin: 0.2
+
+  optim:
+    name: sgd
+    lr: 0.008
+    weight_decay: 0.0002
+
+    # scheduler setup
+    sched:
+      name: CosineAnnealing
+      warmup_ratio: 0.1
+      min_lr: 0.0001
+
+trainer:
+  gpus: 1 # number of gpus (trained on single node - 8 gpus)
+  max_epochs: 200
+  max_steps: null # computed at runtime if not set
+  num_nodes: 1
+  accelerator: ddp
+  accumulate_grad_batches: 1
+  amp_level: O0
+  deterministic: False
+  checkpoint_callback: False
+  logger: False
+  log_every_n_steps: 1  # Interval of logging.
+  val_check_interval: 1.0  # Set to 0.25 to check 4 times per epoch, or an int for number of iterations
+
+exp_manager:
+  exp_dir: null
+  name: ${name}
+  create_tensorboard_logger: True
+  create_checkpoint_callback: True

examples/speaker_recognition/conf/SpeakerNet_recognition_3x2x512.yaml

@@ -103,7 +103,7 @@ model:
     feat_in: *enc_feat_out
     num_classes: 7205
     pool_mode: 'xvector'
-    emb_sizes: 512,512
+    emb_sizes: [512,512]
     angular: False
 
   loss:

nemo/collections/asr/models/label_models.py

@@ -192,8 +192,8 @@ def forward(self, input_signal, input_signal_length):
             input_signal=input_signal, length=input_signal_length,
         )
 
-        encoded, _ = self.encoder(audio_signal=processed_signal, length=processed_signal_len)
-        logits, embs = self.decoder(encoder_output=encoded)
+        encoded, length = self.encoder(audio_signal=processed_signal, length=processed_signal_len)
+        logits, embs = self.decoder(encoder_output=encoded, length=length)
         return logits, embs
 
     # PTL-specific methods

nemo/collections/asr/modules/__init__.py

@@ -24,6 +24,7 @@
     ConvASRDecoder,
     ConvASRDecoderClassification,
     ConvASREncoder,
+    ECAPAEncoder,
     SpeakerDecoder,
 )
 from nemo.collections.asr.modules.lstm_decoder import LSTMDecoder