fix lint

cosyvoice/utils/common.py

@@ -102,6 +102,7 @@ def init_weights(m, mean=0.0, std=0.01):
     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)
@@ -110,6 +111,7 @@ def ras_sampling(weighted_scores, decoded_tokens, sampling, top_p=0.8, top_k=25,
         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
@@ -127,13 +129,16 @@ def nucleus_sampling(weighted_scores, top_p=0.8, top_k=25):
     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_mel, fade_out_mel, window):
     device = fade_in_mel.device
     fade_in_mel, fade_out_mel = fade_in_mel.cpu(), fade_out_mel.cpu()
     mel_overlap_len = int(window.shape[0] / 2)
-    fade_in_mel[:, :, :mel_overlap_len] = fade_in_mel[:, :, :mel_overlap_len] * window[:mel_overlap_len] + fade_out_mel[:, :, -mel_overlap_len:] * window[mel_overlap_len:]
+    fade_in_mel[:, :, :mel_overlap_len] = fade_in_mel[:, :, :mel_overlap_len] * window[:mel_overlap_len] + \
+        fade_out_mel[:, :, -mel_overlap_len:] * window[mel_overlap_len:]
     return fade_in_mel.to(device)

cosyvoice/utils/executor.py

@@ -70,7 +70,8 @@ class Executor:
                 info_dict = update_parameter_and_lr(model, optimizer, scheduler, info_dict)
                 log_per_step(writer, info_dict)
                 # NOTE specify save_per_step in cosyvoice.yaml if you want to enable step save
-                if info_dict['save_per_step'] > 0 and (self.step + 1) % info_dict['save_per_step'] == 0 and (batch_idx + 1) % info_dict["accum_grad"] == 0:
+                if info_dict['save_per_step'] > 0 and (self.step + 1) % info_dict['save_per_step'] == 0 and \
+                   (batch_idx + 1) % info_dict["accum_grad"] == 0:
                     dist.barrier()
                     self.cv(model, cv_data_loader, writer, info_dict, on_batch_end=False)
                     model.train()

cosyvoice/utils/file_utils.py

@@ -28,6 +28,7 @@ def read_lists(list_file):
             lists.append(line.strip())
     return lists
 
+
 def read_json_lists(list_file):
     lists = read_lists(list_file)
     results = {}
@@ -36,6 +37,7 @@ def read_json_lists(list_file):
             results.update(json.load(fin))
     return results
 
+
 def load_wav(wav, target_sr):
     speech, sample_rate = torchaudio.load(wav)
     speech = speech.mean(dim=0, keepdim=True)
@@ -44,6 +46,7 @@ def load_wav(wav, target_sr):
         speech = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=target_sr)(speech)
     return speech
 
+
 def speed_change(waveform, sample_rate, speed_factor: str):
     effects = [
         ["tempo", speed_factor],  # speed_factor

cosyvoice/utils/frontend_utils.py

@@ -15,6 +15,7 @@
 import re
 chinese_char_pattern = re.compile(r'[\u4e00-\u9fff]+')
 
+
 # whether contain chinese character
 def contains_chinese(text):
     return bool(chinese_char_pattern.search(text))

cosyvoice/utils/scheduler.py

@@ -567,8 +567,7 @@ class NoamAnnealing(_LRScheduler):
                  min_lr=0.0,
                  last_epoch=-1):
         self._normalize = d_model**(-0.5)
-        assert not (warmup_steps is not None
-                    and warmup_ratio is not None), \
+        assert not (warmup_steps is not None and warmup_ratio is not None), \
             "Either use particular number of step or ratio"
         assert warmup_ratio is None or max_steps is not None, \
             "If there is a ratio, there should be a total steps"

cosyvoice/utils/train_utils.py

@@ -69,7 +69,6 @@ def init_dataset_and_dataloader(args, configs):
     return train_dataset, cv_dataset, train_data_loader, cv_data_loader
 
 
-
 def check_modify_and_save_config(args, configs):
     if args.train_engine == "torch_ddp":
         configs['train_conf']["dtype"] = 'fp32'
@@ -84,7 +83,8 @@ def check_modify_and_save_config(args, configs):
             configs['train_conf']["dtype"] = "fp32"
         assert ds_configs["train_micro_batch_size_per_gpu"] == 1
         # if use deepspeed, override ddp config
-        configs['train_conf']['save_per_step'] = int(configs['train_conf']['save_per_step'] * configs['train_conf']['accum_grad'] / ds_configs["gradient_accumulation_steps"])
+        configs['train_conf']['save_per_step'] = int(configs['train_conf']['save_per_step'] *
+                                                     configs['train_conf']['accum_grad'] / ds_configs["gradient_accumulation_steps"])
         configs['train_conf']['accum_grad'] = ds_configs["gradient_accumulation_steps"]
         configs['train_conf']['grad_clip'] = ds_configs["gradient_clipping"]
         configs['train_conf']['log_interval'] = ds_configs["steps_per_print"]