Update Nvidia Waveglow/Tacotron2 example (#1905)

* example.tts: Remove old create_mar.py

Remove create_mar.py as there has been updated create_mar.sh file.

* example.tts: Update tacotron2 inference interface

There are two updates in Nvidia Tacotron2 interface:
1. infer() receives input_lengths argument
2. infer() returns mel, mel_lengths, alightments

Co-authored-by: Aaqib <maaquib@gmail.com>

examples/text_to_speech_synthesizer/README.md

@@ -2,7 +2,7 @@
 
 **This example works only on NVIDIA CUDA device and not on CPU**
 
-We have used the following Waveglow/Tacotron2 model for this example: 
+We have used the following Waveglow/Tacotron2 model for this example:
 
 https://pytorch.org/hub/nvidia_deeplearningexamples_waveglow/
 
@@ -20,40 +20,40 @@ pip install librosa --user
 # Serve the WaveGlow speech synthesis model on TorchServe
 
  * Generate the model archive for waveglow speech synthesis model using following command
- 
+
     ```bash
-    python create_mar.py
+    ./create_mar.sh
     ```
-   
+
  * Register the model on TorchServe using the above model archive file
-   
+
     ```bash
     mkdir model_store
     mv waveglow_synthesizer.mar model_store/
     torchserve --start --model-store model_store --models waveglow_synthesizer.mar
     ```
-  * Run inference and download audio output using curl command : 
+  * Run inference and download audio output using curl command :
     ```bash
     curl http://127.0.0.1:8080/predictions/waveglow_synthesizer -T sample_text.txt -o audio.wav
     ```
-    
+
   * Run inference and download audio output using python script :
-  
+
     ```python
     import requests
-    
+
     files = {'data': open('sample_text.txt','rb')}
     response = requests.post('http://localhost:8080/predictions/waveglow_synthesizer', files=files)
     data = response.content
-    
+
     with open('audio.wav', 'wb') as audio_file:
         audio_file.write(data)
     ```
-  
+
   * Change the host and port in above samples as per your server configuration.
-  
+
   * Response :
     An audio.wav file gets downloaded.
-    
+
   **Note :** The above example works only for smaller text size. Refer following NVidia/DeepLearningExamples ticket for more details :
   https://github.com/NVIDIA/DeepLearningExamples/issues/497

examples/text_to_speech_synthesizer/waveglow_handler.py

@@ -1,18 +1,19 @@
 import logging
-import numpy as np
 import os
-import torch
 import uuid
 import zipfile
+
+import numpy as np
+import torch
+from scipy.io.wavfile import write
 from waveglow_model import WaveGlow
-from scipy.io.wavfile import write, read
+
 from ts.torch_handler.base_handler import BaseHandler
 
 logger = logging.getLogger(__name__)
 
 
 class WaveGlowSpeechSynthesizer(BaseHandler):
-
     def __init__(self):
         self.waveglow_model = None
         self.tacotron2_model = None
@@ -31,16 +32,21 @@ def _unwrap_distributed(self, state_dict):
         """
         new_state_dict = {}
         for key, value in state_dict.items():
-            new_key = key.replace('module.', '')
+            new_key = key.replace("module.", "")
             new_state_dict[new_key] = value
         return new_state_dict
 
     def _load_tacotron2_model(self, model_dir):
         from PyTorch.SpeechSynthesis.Tacotron2.tacotron2 import model as tacotron2
         from PyTorch.SpeechSynthesis.Tacotron2.tacotron2.text import text_to_sequence
-        tacotron2_checkpoint = torch.load(os.path.join(model_dir, 'nvidia_tacotron2pyt_fp32_20190427.pth'))
-        tacotron2_state_dict = self._unwrap_distributed(tacotron2_checkpoint['state_dict'])
-        tacotron2_config = tacotron2_checkpoint['config']
+
+        tacotron2_checkpoint = torch.load(
+            os.path.join(model_dir, "nvidia_tacotron2pyt_fp32_20190427.pth")
+        )
+        tacotron2_state_dict = self._unwrap_distributed(
+            tacotron2_checkpoint["state_dict"]
+        )
+        tacotron2_config = tacotron2_checkpoint["config"]
         self.tacotron2_model = tacotron2.Tacotron2(**tacotron2_config)
         self.tacotron2_model.load_state_dict(tacotron2_state_dict)
         self.tacotron2_model.text_to_sequence = text_to_sequence
@@ -51,16 +57,20 @@ def initialize(self, ctx):
 
         properties = ctx.system_properties
         model_dir = properties.get("model_dir")
-        if not torch.cuda.is_available() or properties.get("gpu_id") is None :
+        if not torch.cuda.is_available() or properties.get("gpu_id") is None:
             raise RuntimeError("This model is not supported on CPU machines.")
         self.device = torch.device("cuda:" + str(properties.get("gpu_id")))
 
-        with zipfile.ZipFile(model_dir + '/tacotron.zip', 'r') as zip_ref:
+        with zipfile.ZipFile(model_dir + "/tacotron.zip", "r") as zip_ref:
             zip_ref.extractall(model_dir)
 
-        waveglow_checkpoint = torch.load(os.path.join(model_dir, "nvidia_waveglowpyt_fp32_20190427.pth"))
-        waveglow_state_dict = self._unwrap_distributed(waveglow_checkpoint['state_dict'])
-        waveglow_config = waveglow_checkpoint['config']
+        waveglow_checkpoint = torch.load(
+            os.path.join(model_dir, "nvidia_waveglowpyt_fp32_20190427.pth")
+        )
+        waveglow_state_dict = self._unwrap_distributed(
+            waveglow_checkpoint["state_dict"]
+        )
+        waveglow_config = waveglow_checkpoint["config"]
         self.waveglow_model = WaveGlow(**waveglow_config)
         self.waveglow_model.load_state_dict(waveglow_state_dict)
         self.waveglow_model = self.waveglow_model.remove_weightnorm(self.waveglow_model)
@@ -69,29 +79,34 @@ def initialize(self, ctx):
 
         self._load_tacotron2_model(model_dir)
 
-        logger.debug('WaveGlow model file loaded successfully')
+        logger.debug("WaveGlow model file loaded successfully")
         self.initialized = True
 
     def preprocess(self, data):
         """
-         converts text to sequence of IDs using tacatron2 text_to_sequence
-         with english cleaners to transform text and standardize input
-         (ex: lowercasing, expanding abbreviations and numbers, etc.)
-         returns an Numpy array
+        converts text to sequence of IDs using tacatron2 text_to_sequence
+        with english cleaners to transform text and standardize input
+        (ex: lowercasing, expanding abbreviations and numbers, etc.)
+        returns an Numpy array
         """
         text = data[0].get("data")
         if text is None:
             text = data[0].get("body")
-        text = text.decode('utf-8')
+        text = text.decode("utf-8")
 
-        sequence = np.array(self.tacotron2_model.text_to_sequence(text, ['english_cleaners']))[None, :]
+        sequence = np.array(
+            self.tacotron2_model.text_to_sequence(text, ["english_cleaners"])
+        )[None, :]
         sequence = torch.from_numpy(sequence).to(device=self.device, dtype=torch.int64)
 
         return sequence
 
     def inference(self, data):
         with torch.no_grad():
-            _, mel, _, _ = self.tacotron2_model.infer(data)
+            input_lengths = torch.IntTensor([data.size(1)]).to(
+                device=self.device, dtype=torch.int64
+            )
+            mel, _, _ = self.tacotron2_model.infer(data, input_lengths)
             audio = self.waveglow_model.infer(mel)
 
             return audio
@@ -100,7 +115,7 @@ def postprocess(self, inference_output):
         audio_numpy = inference_output[0].data.cpu().numpy()
         path = "/tmp/{}.wav".format(uuid.uuid4().hex)
         write(path, 22050, audio_numpy)
-        with open(path, 'rb') as output:
+        with open(path, "rb") as output:
             data = output.read()
         os.remove(path)
         return [data]