update stream code

cosyvoice/utils/common.py

@@ -101,3 +101,37 @@ def init_weights(m, mean=0.0, std=0.01):
     classname = m.__class__.__name__
     if classname.find("Conv") != -1:
         m.weight.data.normal_(mean, std)
+
+# Repetition Aware Sampling in VALL-E 2
+def ras_sampling(weighted_scores, decoded_tokens, sampling, top_p=0.8, top_k=25, win_size=10, tau_r=0.1):
+    top_ids = nucleus_sampling(weighted_scores, top_p=top_p, top_k=top_k)
+    rep_num = (torch.tensor(decoded_tokens[-win_size:]).to(weighted_scores.device) == top_ids).sum().item()
+    if rep_num >= win_size * tau_r:
+        top_ids = random_sampling(weighted_scores, decoded_tokens, sampling)
+    return top_ids
+
+def nucleus_sampling(weighted_scores, top_p=0.8, top_k=25):
+    prob, indices = [], []
+    cum_prob = 0.0
+    sorted_value, sorted_idx = weighted_scores.softmax(dim=0).sort(descending=True, stable=True)
+    for i in range(len(sorted_idx)):
+        # sampling both top-p and numbers.
+        if cum_prob < top_p and len(prob) < top_k:
+            cum_prob += sorted_value[i]
+            prob.append(sorted_value[i])
+            indices.append(sorted_idx[i])
+        else:
+            break
+    prob = torch.tensor(prob).to(weighted_scores)
+    indices = torch.tensor(indices, dtype=torch.long).to(weighted_scores.device)
+    top_ids = indices[prob.multinomial(1, replacement=True)]
+    return top_ids
+
+def random_sampling(weighted_scores, decoded_tokens, sampling):
+    top_ids = weighted_scores.softmax(dim=0).multinomial(1, replacement=True)
+    return top_ids
+
+def fade_in_out(fade_in_speech, fade_out_speech, window):
+    speech_overlap_len = int(window.shape[0] / 2)
+    fade_in_speech[:, :speech_overlap_len] = fade_in_speech[:, :speech_overlap_len] * window[:speech_overlap_len] + fade_out_speech[:, -speech_overlap_len:] * window[speech_overlap_len:]
+    return fade_in_speech