diff

configs_template/diffusion_template.yaml

@@ -0,0 +1,48 @@
+data:
+  sampling_rate: 44100
+  block_size: 512 # Equal to hop_length
+  duration: 2 # Audio duration during training, must be less than the duration of the shortest audio clip
+  encoder: 'vec768l12' # 'hubertsoft', 'vec256l9', 'vec768l12'
+  cnhubertsoft_gate: 10
+  encoder_sample_rate: 16000
+  encoder_hop_size: 320
+  encoder_out_channels: 768 # 256 if using 'hubertsoft'
+  train_path: dataset/44k # Create a folder named "audio" under this path and put the audio clip in it
+  filelists_path: filelists/ # FileLists path
+  extensions: # List of extension included in the data collection
+    - wav
+model:
+  type: 'Diffusion'
+  n_layers: 20
+  n_chans: 512
+  n_hidden: 256
+  use_pitch_aug: true  
+  n_spk: 1 # max number of different speakers
+device: cuda
+vocoder:
+  type: 'nsf-hifigan'
+  ckpt: 'pretrain/nsf_hifigan/model'
+infer:
+  speedup: 10
+  method: 'dpm-solver' # 'pndm' or 'dpm-solver'
+env:
+  expdir: exp/diffusion-test
+  gpu_id: 0
+train:
+  num_workers: 2 # If your cpu and gpu are both very strong, set to 0 may be faster!
+  amp_dtype: fp32 # fp32, fp16 or bf16 (fp16 or bf16 may be faster if it is supported by your gpu)
+  batch_size: 48
+  cache_all_data: true # Save Internal-Memory or Graphics-Memory if it is false, but may be slow
+  cache_device: 'cpu' # Set to 'cuda' to cache the data into the Graphics-Memory, fastest speed for strong gpu
+  cache_fp16: true
+  epochs: 100000
+  interval_log: 10
+  interval_val: 2000
+  interval_force_save: 10000
+  lr: 0.0002
+  decay_step: 100000
+  gamma: 0.5
+  weight_decay: 0
+  save_opt: false
+spk:
+  'nyaru': 0

diffusion/data_loaders.py

@@ -51,7 +51,7 @@ def traverse_dir(
 
 def get_data_loaders(args, whole_audio=False):
     data_train = AudioDataset(
-        args.data.train_path,
+        filelists_path = args.filelists_path,
         waveform_sec=args.data.duration,
         hop_size=args.data.block_size,
         sample_rate=args.data.sampling_rate,
@@ -71,7 +71,7 @@ def get_data_loaders(args, whole_audio=False):
         pin_memory=True if args.train.cache_device=='cpu' else False
     )
     data_valid = AudioDataset(
-        args.data.valid_path,
+        filelists_path = args.filelists_path,
         waveform_sec=args.data.duration,
         hop_size=args.data.block_size,
         sample_rate=args.data.sampling_rate,
@@ -92,7 +92,7 @@ def get_data_loaders(args, whole_audio=False):
 class AudioDataset(Dataset):
     def __init__(
         self,
-        path_root,
+        filelists,
         waveform_sec,
         hop_size,
         sample_rate,
@@ -109,7 +109,7 @@ def __init__(
         self.waveform_sec = waveform_sec
         self.sample_rate = sample_rate
         self.hop_size = hop_size
-        self.path_root = path_root
+        self.filelists = filelists
         self.paths = traverse_dir(
             os.path.join(path_root, 'audio'),
             extensions=extensions,

diffusion/logger/saver.py

@@ -0,0 +1,150 @@
+'''
+author: wayn391@mastertones
+'''
+
+import os
+import json
+import time
+import yaml
+import datetime
+import torch
+import matplotlib.pyplot as plt
+from . import utils
+from torch.utils.tensorboard import SummaryWriter
+
+class Saver(object):
+    def __init__(
+            self, 
+            args,
+            initial_global_step=-1):
+
+        self.expdir = args.env.expdir
+        self.sample_rate = args.data.sampling_rate
+
+        # cold start
+        self.global_step = initial_global_step
+        self.init_time = time.time()
+        self.last_time = time.time()
+
+        # makedirs
+        os.makedirs(self.expdir, exist_ok=True)       
+
+        # path
+        self.path_log_info = os.path.join(self.expdir, 'log_info.txt')
+
+        # ckpt
+        os.makedirs(self.expdir, exist_ok=True)       
+
+        # writer
+        self.writer = SummaryWriter(os.path.join(self.expdir, 'logs'))
+
+        # save config
+        path_config = os.path.join(self.expdir, 'config.yaml')
+        with open(path_config, "w") as out_config:
+            yaml.dump(dict(args), out_config)
+
+
+    def log_info(self, msg):
+        '''log method'''
+        if isinstance(msg, dict):
+            msg_list = []
+            for k, v in msg.items():
+                tmp_str = ''
+                if isinstance(v, int):
+                    tmp_str = '{}: {:,}'.format(k, v)
+                else:
+                    tmp_str = '{}: {}'.format(k, v)
+
+                msg_list.append(tmp_str)
+            msg_str = '\n'.join(msg_list)
+        else:
+            msg_str = msg
+
+        # dsplay
+        print(msg_str)
+
+        # save
+        with open(self.path_log_info, 'a') as fp:
+            fp.write(msg_str+'\n')
+
+    def log_value(self, dict):
+        for k, v in dict.items():
+            self.writer.add_scalar(k, v, self.global_step)
+
+    def log_spec(self, name, spec, spec_out, vmin=-14, vmax=3.5):  
+        spec_cat = torch.cat([(spec_out - spec).abs() + vmin, spec, spec_out], -1)
+        spec = spec_cat[0]
+        if isinstance(spec, torch.Tensor):
+            spec = spec.cpu().numpy()
+        fig = plt.figure(figsize=(12, 9))
+        plt.pcolor(spec.T, vmin=vmin, vmax=vmax)
+        plt.tight_layout()
+        self.writer.add_figure(name, fig, self.global_step)
+
+    def log_audio(self, dict):
+        for k, v in dict.items():
+            self.writer.add_audio(k, v, global_step=self.global_step, sample_rate=self.sample_rate)
+
+    def get_interval_time(self, update=True):
+        cur_time = time.time()
+        time_interval = cur_time - self.last_time
+        if update:
+            self.last_time = cur_time
+        return time_interval
+
+    def get_total_time(self, to_str=True):
+        total_time = time.time() - self.init_time
+        if to_str:
+            total_time = str(datetime.timedelta(
+                seconds=total_time))[:-5]
+        return total_time
+
+    def save_model(
+            self,
+            model, 
+            optimizer,
+            name='model',
+            postfix='',
+            to_json=False):
+        # path
+        if postfix:
+            postfix = '_' + postfix
+        path_pt = os.path.join(
+            self.expdir , name+postfix+'.pt')
+
+        # check
+        print(' [*] model checkpoint saved: {}'.format(path_pt))
+
+        # save
+        if optimizer is not None:
+            torch.save({
+                'global_step': self.global_step,
+                'model': model.state_dict(),
+                'optimizer': optimizer.state_dict()}, path_pt)
+        else:
+            torch.save({
+                'global_step': self.global_step,
+                'model': model.state_dict()}, path_pt)
+
+        # to json
+        if to_json:
+            path_json = os.path.join(
+                self.expdir , name+'.json')
+            utils.to_json(path_params, path_json)
+
+    def delete_model(self, name='model', postfix=''):
+        # path
+        if postfix:
+            postfix = '_' + postfix
+        path_pt = os.path.join(
+            self.expdir , name+postfix+'.pt')
+
+        # delete
+        if os.path.exists(path_pt):
+            os.remove(path_pt)
+            print(' [*] model checkpoint deleted: {}'.format(path_pt))
+
+    def global_step_increment(self):
+        self.global_step += 1
+
+

diffusion/logger/utils.py

@@ -0,0 +1,126 @@
+import os
+import yaml
+import json
+import pickle
+import torch
+
+def traverse_dir(
+        root_dir,
+        extensions,
+        amount=None,
+        str_include=None,
+        str_exclude=None,
+        is_pure=False,
+        is_sort=False,
+        is_ext=True):
+
+    file_list = []
+    cnt = 0
+    for root, _, files in os.walk(root_dir):
+        for file in files:
+            if any([file.endswith(f".{ext}") for ext in extensions]):
+                # path
+                mix_path = os.path.join(root, file)
+                pure_path = mix_path[len(root_dir)+1:] if is_pure else mix_path
+
+                # amount
+                if (amount is not None) and (cnt == amount):
+                    if is_sort:
+                        file_list.sort()
+                    return file_list
+
+                # check string
+                if (str_include is not None) and (str_include not in pure_path):
+                    continue
+                if (str_exclude is not None) and (str_exclude in pure_path):
+                    continue
+
+                if not is_ext:
+                    ext = pure_path.split('.')[-1]
+                    pure_path = pure_path[:-(len(ext)+1)]
+                file_list.append(pure_path)
+                cnt += 1
+    if is_sort:
+        file_list.sort()
+    return file_list
+
+
+
+class DotDict(dict):
+    def __getattr__(*args):         
+        val = dict.get(*args)         
+        return DotDict(val) if type(val) is dict else val   
+
+    __setattr__ = dict.__setitem__    
+    __delattr__ = dict.__delitem__
+
+
+def get_network_paras_amount(model_dict):
+    info = dict()
+    for model_name, model in model_dict.items():
+        # all_params = sum(p.numel() for p in model.parameters())
+        trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
+
+        info[model_name] = trainable_params
+    return info
+
+
+def load_config(path_config):
+    with open(path_config, "r") as config:
+        args = yaml.safe_load(config)
+    args = DotDict(args)
+    # print(args)
+    return args
+
+def save_config(path_config,config):
+    config = dict(config)
+    with open(path_config, "w") as f:
+        yaml.dump(config, f)
+
+def to_json(path_params, path_json):
+    params = torch.load(path_params, map_location=torch.device('cpu'))
+    raw_state_dict = {}
+    for k, v in params.items():
+        val = v.flatten().numpy().tolist()
+        raw_state_dict[k] = val
+
+    with open(path_json, 'w') as outfile:
+        json.dump(raw_state_dict, outfile,indent= "\t")
+
+
+def convert_tensor_to_numpy(tensor, is_squeeze=True):
+    if is_squeeze:
+        tensor = tensor.squeeze()
+    if tensor.requires_grad:
+        tensor = tensor.detach()
+    if tensor.is_cuda:
+        tensor = tensor.cpu()
+    return tensor.numpy()
+
+
+def load_model(
+        expdir, 
+        model,
+        optimizer,
+        name='model',
+        postfix='',
+        device='cpu'):
+    if postfix == '':
+        postfix = '_' + postfix
+    path = os.path.join(expdir, name+postfix)
+    path_pt = traverse_dir(expdir, ['pt'], is_ext=False)
+    global_step = 0
+    if len(path_pt) > 0:
+        steps = [s[len(path):] for s in path_pt]
+        maxstep = max([int(s) if s.isdigit() else 0 for s in steps])
+        if maxstep >= 0:
+            path_pt = path+str(maxstep)+'.pt'
+        else:
+            path_pt = path+'best.pt'
+        print(' [*] restoring model from', path_pt)
+        ckpt = torch.load(path_pt, map_location=torch.device(device))
+        global_step = ckpt['global_step']
+        model.load_state_dict(ckpt['model'], strict=False)
+        if ckpt.get('optimizer') != None:
+            optimizer.load_state_dict(ckpt['optimizer'])
+    return global_step, model, optimizer