Adding Hybrid RNNT-CTC model

Jenkinsfile

@@ -562,6 +562,24 @@ pipeline {
     //         sh 'rm -rf examples/asr/speech_to_text_rnnt_wpe_results'
     //       }
     //     }
+    //     stage('L3: Speech to Text Hybrid Transducer-CTC WPE') {
+    //       steps {
+    //         sh 'STRICT_NUMBA_COMPAT_CHECK=false python examples/asr/asr_hybrid_transducer_ctc/speech_to_text_hybrid_rnnt_ctc_bpe.py \
+    //         --config-path="../conf/conformer/hybrid_transducer_ctc/conformer_hybrid_transducer_ctc/" --config-name="conformer_hybrid_transducer_ctc_bpe.yaml" \
+    //         model.train_ds.manifest_filepath=/home/TestData/an4_dataset/an4_train.json \
+    //         model.validation_ds.manifest_filepath=/home/TestData/an4_dataset/an4_val.json \
+    //         model.encoder.n_layers= 2 \
+    //         model.train_ds.batch_size=2 \
+    //         model.validation_ds.batch_size=2 \
+    //         model.tokenizer.dir="/home/TestData/asr_tokenizers/an4_wpe_128/" \
+    //         model.tokenizer.type="wpe" \
+    //         trainer.devices=[0] \
+    //         trainer.accelerator="gpu" \
+    //         +trainer.fast_dev_run=True \
+    //         exp_manager.exp_dir=examples/asr/speech_to_text_hybrid_transducer_ctc_wpe_results'
+    //         sh 'rm -rf examples/asr/speech_to_text_hybrid_transducer_ctc_wpe_results'
+    //       }
+    //     }
     //   }
     // }
 

docs/source/asr/models.rst

@@ -236,6 +236,27 @@ You may find the example config files of Squeezeformer-CTC model with character-
 ``<NeMo_git_root>/examples/asr/conf/squeezeformer/squeezeformer_ctc_char.yaml`` and
 with sub-word encoding at ``<NeMo_git_root>/examples/asr/conf/squeezeformer/squeezeformer_ctc_bpe.yaml``.
 
+.. _Hybrid-Transducer_CTC_model:
+
+Hybrid-Transducer-CTC
+---------------------
+
+Hybrid RNNT-CTC models is a group of models with both the RNNT and CTC decoders. Training a unified model would speedup the convergence for the CTC models and would enable
+the user to use a single model which works as both a CTC and RNNT model. This category can be used with any of the ASR models.
+Hybrid models uses two decoders of CTC and RNNT on the top of the encoder. The default decoding strategy after the training is done is RNNT.
+User may use the ``asr_model.change_decoding_strategy(decoder_type='ctc' or 'rnnt')`` to change the default decoding.
+
+The variant with sub-word encoding is a BPE-based model
+which can be instantiated using the :class:`~nemo.collections.asr.models.EncDecHybridRNNTCTCBPEModel` class, while the
+character-based variant is based on :class:`~nemo.collections.asr.models.EncDecHybridRNNTCTCModel`.
+
+You may use the example scripts under ``<NeMo_git_root>/examples/asr/asr_hybrid_transducer_ctc`` for both the char-based encoding and sub-word encoding.
+These examples can be used to train any Hybrid ASR model like Conformer, Citrinet, QuartzNet, etc.
+
+You may find the example config files of Conformer variant of such hybrid models with character-based encoding at
+``<NeMo_git_root>/examples/asr/conf/conformer/hybrid_transducer_ctc/conformer_hybrid_transducer_ctc_char.yaml`` and
+with sub-word encoding at ``<NeMo_git_root>/examples/asr/conf/conformer/hybrid_transducer_ctc/conformer_hybrid_transducer_ctc_bpe.yaml``.
+
 
 References
 ----------

examples/asr/asr_hybrid_transducer_ctc/README.md

@@ -0,0 +1,32 @@
+# ASR with Hybrid Transducer/CTC Models
+
+This directory contains example scripts to train ASR models with two decoders of Transducer and CTC Loss. 
+
+Currently supported models are - 
+
+* Character based Hybrid RNNT/CTC model
+* Subword based Hybrid RNNT/CTC model
+
+# Model execution overview
+
+The training scripts in this directory execute in the following order. When preparing your own training-from-scratch / fine-tuning scripts, please follow this order for correct training/inference.
+
+```mermaid
+
+graph TD
+    A[Hydra Overrides + Yaml Config] --> B{Config}
+    B --> |Init| C[Trainer]
+    C --> D[ExpManager]
+    B --> D[ExpManager]
+    C --> E[Model]
+    B --> |Init| E[Model]
+    E --> |Constructor| F1(Change Vocabulary)
+    F1 --> F2(Setup Adapters if available)
+    F2 --> G(Setup Train + Validation + Test Data loaders)
+    G --> H1(Setup Optimization)
+    H1 --> H2(Change Transducer Decoding Strategy)
+    H2 --> I[Maybe init from pretrained]
+    I --> J["trainer.fit(model)"]
+```
+
+During restoration of the model, you may pass the Trainer to the restore_from / from_pretrained call, or set it after the model has been initialized by using `model.set_trainer(Trainer)`.

examples/asr/asr_hybrid_transducer_ctc/speech_to_text_hybrid_rnnt_ctc_bpe.py

@@ -0,0 +1,91 @@
+# Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+
+"""
+# Preparing the Tokenizer for the dataset
+Use the `process_asr_text_tokenizer.py` script under <NEMO_ROOT>/scripts/tokenizers/ in order to prepare the tokenizer.
+
+```sh
+python <NEMO_ROOT>/scripts/tokenizers/process_asr_text_tokenizer.py \
+        --manifest=<path to train manifest files, seperated by commas>
+        OR
+        --data_file=<path to text data, seperated by commas> \
+        --data_root="<output directory>" \
+        --vocab_size=<number of tokens in vocabulary> \
+        --tokenizer=<"spe" or "wpe"> \
+        --no_lower_case \
+        --spe_type=<"unigram", "bpe", "char" or "word"> \
+        --spe_character_coverage=1.0 \
+        --log
+```
+
+# Training the model
+```sh
+python speech_to_text_hybrid_rnnt_ctc_bpe.py \
+    # (Optional: --config-path=<path to dir of configs> --config-name=<name of config without .yaml>) \
+    model.train_ds.manifest_filepath=<path to train manifest> \
+    model.validation_ds.manifest_filepath=<path to val/test manifest> \
+    model.tokenizer.dir=<path to directory of tokenizer (not full path to the vocab file!)> \
+    model.tokenizer.type=<either bpe or wpe> \
+    model.aux_ctc.ctc_loss_weight=0.3 \
+    trainer.devices=-1 \
+    trainer.max_epochs=100 \
+    model.optim.name="adamw" \
+    model.optim.lr=0.001 \
+    model.optim.betas=[0.9,0.999] \
+    model.optim.weight_decay=0.0001 \
+    model.optim.sched.warmup_steps=2000
+    exp_manager.create_wandb_logger=True \
+    exp_manager.wandb_logger_kwargs.name="<Name of experiment>" \
+    exp_manager.wandb_logger_kwargs.project="<Name of project>"
+```
+
+# Fine-tune a model
+
+For documentation on fine-tuning this model, please visit -
+https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/configs.html#fine-tuning-configurations
+
+"""
+
+import pytorch_lightning as pl
+from omegaconf import OmegaConf
+
+from nemo.collections.asr.models import EncDecHybridRNNTCTCBPEModel
+from nemo.core.config import hydra_runner
+from nemo.utils import logging
+from nemo.utils.exp_manager import exp_manager
+
+
+@hydra_runner(
+    config_path="../conf/conformer/hybrid_transducer_ctc/", config_name="conformer_hybrid_transducer_ctc_bpe"
+)
+def main(cfg):
+    logging.info(f'Hydra config: {OmegaConf.to_yaml(cfg)}')
+
+    trainer = pl.Trainer(**cfg.trainer)
+    exp_manager(trainer, cfg.get("exp_manager", None))
+    asr_model = EncDecHybridRNNTCTCBPEModel(cfg=cfg.model, trainer=trainer)
+
+    # Initialize the weights of the model from another model, if provided via config
+    asr_model.maybe_init_from_pretrained_checkpoint(cfg)
+
+    trainer.fit(asr_model)
+
+    if hasattr(cfg.model, 'test_ds') and cfg.model.test_ds.manifest_filepath is not None:
+        if asr_model.prepare_test(trainer):
+            trainer.test(asr_model)
+
+
+if __name__ == '__main__':
+    main()  # noqa pylint: disable=no-value-for-parameter

examples/asr/asr_hybrid_transducer_ctc/speech_to_text_hybrid_rnnt_ctc_char.py

@@ -0,0 +1,100 @@
+# Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+
+"""
+# Training the model
+
+Basic run (on CPU for 50 epochs):
+    python examples/asr/asr_transducer/speech_to_text_hybrid_rnnt_ctc.py \
+        # (Optional: --config-path=<path to dir of configs> --config-name=<name of config without .yaml>) \
+        model.train_ds.manifest_filepath="<path to manifest file>" \
+        model.validation_ds.manifest_filepath="<path to manifest file>" \
+        trainer.devices=1 \
+        trainer.accelerator='cpu' \
+        trainer.max_epochs=50
+
+
+Add PyTorch Lightning Trainer arguments from CLI:
+    python speech_to_text_rnnt.py \
+        ... \
+        +trainer.fast_dev_run=true
+
+Hydra logs will be found in "$(./outputs/$(date +"%y-%m-%d")/$(date +"%H-%M-%S")/.hydra)"
+PTL logs will be found in "$(./outputs/$(date +"%y-%m-%d")/$(date +"%H-%M-%S")/lightning_logs)"
+
+Override some args of optimizer:
+    python speech_to_text_hybrid_rnnt_ctc.py \
+    --config-path="../conf/conformer/hybrid_transducer_ctc/conformer_hybrid_transducer_ctc" \
+    --config-name="config_rnnt" \
+    model.train_ds.manifest_filepath="./an4/train_manifest.json" \
+    model.validation_ds.manifest_filepath="./an4/test_manifest.json" \
+    trainer.devices=2 \
+    model.aux_ctc.ctc_loss_weight=0.3 \
+    trainer.precision=16 \
+    trainer.max_epochs=2 \
+    model.optim.betas=[0.8,0.5] \
+    model.optim.weight_decay=0.0001
+
+Override optimizer entirely
+    python speech_to_text_hybrid_rnnt_ctc.py \
+    --config-path="../conf/conformer/hybrid_transducer_ctc/conformer_hybrid_transducer_ctc" \
+    --config-name="config_rnnt" \
+    model.train_ds.manifest_filepath="./an4/train_manifest.json" \
+    model.validation_ds.manifest_filepath="./an4/test_manifest.json" \
+    model.aux_ctc.ctc_loss_weight=0.3 \
+    trainer.devices=2 \
+    trainer.precision=16 \
+    trainer.max_epochs=2 \
+    model.optim.name=adamw \
+    model.optim.lr=0.001 \
+    ~model.optim.args \
+    +model.optim.args.betas=[0.8,0.5]\
+    +model.optim.args.weight_decay=0.0005
+
+# Fine-tune a model
+
+For documentation on fine-tuning this model, please visit -
+https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/configs.html#fine-tuning-configurations
+
+"""
+
+import pytorch_lightning as pl
+from omegaconf import OmegaConf
+
+from nemo.collections.asr.models import EncDecHybridRNNTCTCModel
+from nemo.core.config import hydra_runner
+from nemo.utils import logging
+from nemo.utils.exp_manager import exp_manager
+
+
+@hydra_runner(config_path="../conf/conformer/hybrid_transducer_ctc/", config_name="conformer_hybrid_transducer_ctc")
+def main(cfg):
+    logging.info(f'Hydra config: {OmegaConf.to_yaml(cfg)}')
+
+    trainer = pl.Trainer(**cfg.trainer)
+    exp_manager(trainer, cfg.get("exp_manager", None))
+    asr_model = EncDecHybridRNNTCTCModel(cfg=cfg.model, trainer=trainer)
+
+    # Initialize the weights of the model from another model, if provided via config
+    asr_model.maybe_init_from_pretrained_checkpoint(cfg)
+
+    trainer.fit(asr_model)
+
+    if hasattr(cfg.model, 'test_ds') and cfg.model.test_ds.manifest_filepath is not None:
+        if asr_model.prepare_test(trainer):
+            trainer.test(asr_model)
+
+
+if __name__ == '__main__':
+    main()  # noqa pylint: disable=no-value-for-parameter