New entry point for subprocess/CLI-based inference

tts/tacotron2_v1_early_fakeyou/vocodes_common.py

@@ -284,16 +284,21 @@ class FakeYouTt2Pipeline:
             len(self.tacotron_model_cache)))
 
     def infer(self, args):
+        # Model parameters
         assert('synthesizer_checkpoint_path' in args)
         assert('vocoder_type' in args)
+        assert('text_pipeline_type' in args)
+        # Optional mel scaling before vocodinng
         assert('use_default_mel_multiply_factor' in args)
         assert('maybe_custom_mel_multiply_factor' in args)
-        assert('text_pipeline_type' in args)
+        # Premium features
+        assert('max_decoder_steps' in args)
+        # User input
         assert('raw_text' in args)
+        # Output files
         assert('output_audio_filename' in args)
         assert('output_spectrogram_filename' in args)
         assert('output_metadata_filename' in args)
-        assert('max_decoder_steps' in args)
 
         LOGGER.info(f"raw inference text: {args['raw_text']}")
 

tts/tacotron2_v1_early_fakeyou/vocodes_inference_updated.py

@@ -0,0 +1,113 @@
+#!/usr/bin/env python
+
+"""
+Vocodes Tacotron Inference Engine
+This is a work in progress and is likely going to be slow for now.
+"""
+
+# ========== ARGUMENTS (FAIL EARLY IF ABSENT) ==========
+
+import argparse
+
+parser = argparse.ArgumentParser(description='Run TTS inference')
+
+# Model parameters
+parser.add_argument('--synthesizer_checkpoint_path', type=str, help='path the TTS synthesizer model', required=True)
+parser.add_argument('--text_pipeline_type', type=str, help='', required=True)
+parser.add_argument('--vocoder_type', type=str, help='', required=True)
+parser.add_argument('--waveglow_vocoder_checkpoint_path', type=str, help='path the TTS vocoder model')
+parser.add_argument('--hifigan_vocoder_checkpoint_path', type=str, help='path the TTS vocoder model')
+parser.add_argument('--hifigan_superres_vocoder_checkpoint_path', type=str, help='path the TTS vocoder model')
+
+# Optional mel scaling before vocoding
+parser.add_argument('--use_default_mel_multiply_factor', type=bool, help='', action='store_true')
+parser.add_argument('--maybe_custom_mel_multiply_factor', type=int, help='')
+
+# Premium features
+parser.add_argument('--max_decoder_steps', type=int, help='')
+
+# User input 
+parser.add_argument('--input_text_filename', type=str, help='path the file containing text to run', required=True)
+
+# Output files
+parser.add_argument('--output_audio_filename', type=str, help='where to save result audio', required=True)
+parser.add_argument('--output_spectrogram_filename', type=str, help='where to save result spectrogram', required=True)
+parser.add_argument('--output_metadata_filename', type=str, help='where to save extra metadata', required=True)
+
+args = parser.parse_args()
+
+
+# ========== IMPORTS ==========
+
+import sys
+sys.path.append('waveglow/')
+import shutil
+import json
+
+import numpy as np
+import torch
+
+from hparams import create_hparams
+from model import Tacotron2
+from text import text_to_sequence
+from denoiser import Denoiser
+
+from vocodes_common import FakeYouTt2Pipeline
+from vocodes_common import print_gpu_info
+from vocodes_common import parse_int_or_default
+
+# For metadata
+import subprocess
+import magic
+import os
+
+from logger import LOGGER
+
+
+# ========== PRINT GPU INFO ==========
+
+import tensorflow as tf
+print("TensorFlow version: {}".format(tf.version.VERSION))
+print_gpu_info()
+
+
+# ========== LOAD MODEL + SETUP ==========
+
+pipeline = FakeYouTt2Pipeline()
+
+# NB: Not sure if there was a historical decision to load the models first outside of the pipeline. 
+# Probably not, though. Seems like this could be cleaned up.
+if args.vocoder_type == 'hifigan-superres':
+    pipeline.maybe_load_waveglow_model(args.waveglow_vocoder_checkpoint_path)
+elif args.vocoder_type == 'waveglow':
+    pipeline.maybe_load_waveglow_model(args.hifigan_vocoder_checkpoint_path)
+    pipeline.maybe_load_waveglow_model(args.hifigan_superres_vocoder_checkpoint_path)
+else:
+    raise Exception('Wrong vocoder type: {}'.format(args.vocoder_type))
+
+
+pipeline.maybe_load_tacotron_model_to_cache(args.synthesizer_checkpoint_path)
+
+# NB: 1000 is the default and is roughly 12 seconds.
+max_decoder_steps = parse_int_or_default(args.max_decoder_steps, 1000) 
+
+inference_args = {
+    'raw_text': args.inference_text,
+    'use_default_mel_multiply_factor': args.use_default_mel_multiply_factor,
+    'maybe_custom_mel_multiply_factor': args.maybe_custom_mel_multiply_factor,
+    'synthesizer_checkpoint_path': args.synthesizer_checkpoint_path,
+    'vocoder_type': args.vocoder_type,
+    'text_pipeline_type': args.text_pipeline_type,
+    'output_audio_filename': args.output_audio_filename,
+    'output_spectrogram_filename': args.output_spectrogram_filename,
+    'output_metadata_filename': args.output_metadata_filename,
+    'max_decoder_steps': max_decoder_steps,
+}
+
+LOGGER.info(f"running inference with args: {inference_args}")
+
+
+# ========== INFERENCE ==========
+
+pipeline.infer(inference_args)
+

tts/tacotron2_v1_early_fakeyou/vocodes_server.py

@@ -222,25 +222,31 @@ class ApiHandler():
 
         LOGGER.info(f"TTS Request: {raw_data}")
 
-        # ========== RUST SERVICE ARGUMENTS ==========
+        # ========== RUST SERVICE ARGUMENTS (REQUEST PARAMETERS) ==========
 
-        # Request parameters
+        # Model parameters
         synthesizer_checkpoint_path = raw_data.get('synthesizer_checkpoint_path')
         text_pipeline_type = raw_data.get('text_pipeline_type')
-        use_default_mel_multiply_factor = raw_data.get('use_default_mel_multiply_factor')
-        maybe_custom_mel_multiply_factor = raw_data.get('maybe_custom_mel_multiply_factor')
         vocoder_type = raw_data.get('vocoder_type')
         waveglow_vocoder_checkpoint_path = raw_data.get('waveglow_vocoder_checkpoint_path')
         hifigan_vocoder_checkpoint_path = raw_data.get('hifigan_vocoder_checkpoint_path')
         hifigan_superres_vocoder_checkpoint_path = raw_data.get('hifigan_superres_vocoder_checkpoint_path')
+
+        #  Optional mel scaling before vocoding
+        use_default_mel_multiply_factor = raw_data.get('use_default_mel_multiply_factor')
+        maybe_custom_mel_multiply_factor = raw_data.get('maybe_custom_mel_multiply_factor')
+
+        # Premium features
+        maybe_max_decoder_steps = raw_data.get('max_decoder_steps')
+
+        # User input
         inference_text = raw_data.get('inference_text')
+
+        # Output files
         output_audio_filename = raw_data.get('output_audio_filename')
         output_spectrogram_filename = raw_data.get('output_spectrogram_filename')
         output_metadata_filename = raw_data.get('output_metadata_filename')
 
-        # Temporarily optional request parameters
-        maybe_max_decoder_steps = raw_data.get('max_decoder_steps')
-
         # Optional request parameters
         maybe_clear_synthesizer_checkpoint_path = \
             raw_data.get('maybe_clear_synthesizer_checkpoint_path')