From 2dea4952d5820914125ca0c41b316e8af4408d95 Mon Sep 17 00:00:00 2001
From: James Betker <jbetker@gmail.com>
Date: Mon, 18 Apr 2022 09:22:15 -0600
Subject: [PATCH] update api constants

---
 api.py           | 10 +++++++---
 eval_multiple.py |  6 ++----
 models/cvvp.py   |  0
 3 files changed, 9 insertions(+), 7 deletions(-)
 create mode 100644 models/cvvp.py

diff --git a/api.py b/api.py
index c64918a..fa1a010 100644
--- a/api.py
+++ b/api.py
@@ -186,7 +186,9 @@ class TextToSpeech:
             'high_quality': Use if you want the absolute best. This is not really worth the compute, though.
         """
         # Use generally found best tuning knobs for generation.
-        kwargs.update({'temperature': .8, 'length_penalty': 1.0, 'repetition_penalty': 2.0, 'top_p': .8,
+        kwargs.update({'temperature': .8, 'length_penalty': 1.0, 'repetition_penalty': 2.0,
+                       #'typical_sampling': True,
+                       'top_p': .8,
                        'cond_free_k': 2.0, 'diffusion_temperature': 1.0})
         # Presets are defined here.
         presets = {
@@ -202,7 +204,8 @@ class TextToSpeech:
             # autoregressive generation parameters follow
             num_autoregressive_samples=512, temperature=.8, length_penalty=1, repetition_penalty=2.0, top_p=.8,
             # diffusion generation parameters follow
-            diffusion_iterations=100, cond_free=True, cond_free_k=2, diffusion_temperature=1.0,):
+            diffusion_iterations=100, cond_free=True, cond_free_k=2, diffusion_temperature=1.0,
+            **hf_generate_kwargs):
         text = torch.IntTensor(self.tokenizer.encode(text)).unsqueeze(0).cuda()
         text = F.pad(text, (0, 1))  # This may not be necessary.
 
@@ -228,7 +231,8 @@ class TextToSpeech:
                                                              temperature=temperature,
                                                              num_return_sequences=self.autoregressive_batch_size,
                                                              length_penalty=length_penalty,
-                                                             repetition_penalty=repetition_penalty)
+                                                             repetition_penalty=repetition_penalty,
+                                                             **hf_generate_kwargs)
                 padding_needed = self.autoregressive.max_mel_tokens - codes.shape[1]
                 codes = F.pad(codes, (0, padding_needed), value=stop_mel_token)
                 samples.append(codes)
diff --git a/eval_multiple.py b/eval_multiple.py
index 1113433..9defa52 100644
--- a/eval_multiple.py
+++ b/eval_multiple.py
@@ -16,7 +16,7 @@ if __name__ == '__main__':
         lines = [l.strip().split('\t') for l in f.readlines()]
 
     tts = TextToSpeech()
-    for k in range(4):
+    for k in range(3):
         outpath = f'{outpath_base}_{k}'
         os.makedirs(outpath, exist_ok=True)
         recorder = open(os.path.join(outpath, 'transcript.tsv'), 'w', encoding='utf-8')
@@ -27,9 +27,7 @@ if __name__ == '__main__':
             path = os.path.join(os.path.dirname(fname), line[1])
             cond_audio = load_audio(path, 22050)
             torchaudio.save(os.path.join(outpath_real, os.path.basename(line[1])), cond_audio, 22050)
-            sample = tts.tts(transcript, [cond_audio, cond_audio], num_autoregressive_samples=128, k=1,
-                                 repetition_penalty=2.0, length_penalty=2, temperature=.5, top_p=.5,
-                                 diffusion_temperature=.7, cond_free_k=2, diffusion_iterations=70)
+            sample = tts.tts_with_preset(transcript, [cond_audio, cond_audio], preset='standard')
 
             down = torchaudio.functional.resample(sample, 24000, 22050)
             fout_path = os.path.join(outpath, os.path.basename(line[1]))
diff --git a/models/cvvp.py b/models/cvvp.py
new file mode 100644
index 0000000..e69de29