Initial commit

preprocess/tools/note_transcription/utils/commons/base_task.py

@@ -0,0 +1,235 @@
+import logging
+import os
+import random
+import subprocess
+import sys
+from datetime import datetime
+import numpy as np
+import torch.utils.data
+from torch import nn
+from torch.utils.tensorboard import SummaryWriter
+from .dataset_utils import data_loader
+from .hparams import hparams
+from .meters import AvgrageMeter
+from .tensor_utils import tensors_to_scalars
+from .trainer import Trainer
+
+torch.multiprocessing.set_sharing_strategy(os.getenv('TORCH_SHARE_STRATEGY', 'file_system'))
+
+log_format = '%(asctime)s %(message)s'
+logging.basicConfig(stream=sys.stdout, level=logging.INFO,
+                    format=log_format, datefmt='%m/%d %I:%M:%S %p')
+
+
+class BaseTask(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super(BaseTask, self).__init__()
+        self.current_epoch = 0
+        self.global_step = 0
+        self.trainer = None
+        self.use_ddp = False
+        self.gradient_clip_norm = hparams['clip_grad_norm']
+        self.gradient_clip_val = hparams.get('clip_grad_value', 0)
+        self.model = None
+        self.training_losses_meter = None
+        self.logger: SummaryWriter = None
+
+    ######################
+    # build model, dataloaders, optimizer, scheduler and tensorboard
+    ######################
+    def build_model(self):
+        raise NotImplementedError
+
+    @data_loader
+    def train_dataloader(self):
+        raise NotImplementedError
+
+    @data_loader
+    def test_dataloader(self):
+        raise NotImplementedError
+
+    @data_loader
+    def val_dataloader(self):
+        raise NotImplementedError
+
+    def build_scheduler(self, optimizer):
+        return None
+
+    def build_optimizer(self, model):
+        raise NotImplementedError
+
+    def configure_optimizers(self):
+        optm = self.build_optimizer(self.model)
+        self.scheduler = self.build_scheduler(optm)
+        if isinstance(optm, (list, tuple)):
+            return optm
+        return [optm]
+
+    def build_tensorboard(self, save_dir, name, **kwargs):
+        log_dir = os.path.join(save_dir, name)
+        os.makedirs(log_dir, exist_ok=True)
+        self.logger = SummaryWriter(log_dir=log_dir, **kwargs)
+
+    ######################
+    # training
+    ######################
+    def on_train_start(self):
+        pass
+
+    def on_train_end(self):
+        pass
+
+    def on_epoch_start(self):
+        self.training_losses_meter = {'total_loss': AvgrageMeter()}
+
+    def on_epoch_end(self):
+        loss_outputs = {k: round(v.avg, 4) for k, v in self.training_losses_meter.items()}
+        print(f"Epoch {self.current_epoch} ended. Steps: {self.global_step}. {loss_outputs}")
+
+    def _training_step(self, sample, batch_idx, optimizer_idx):
+        """
+
+        :param sample:
+        :param batch_idx:
+        :return: total loss: torch.Tensor, loss_log: dict
+        """
+        raise NotImplementedError
+
+    def training_step(self, sample, batch_idx, optimizer_idx=-1):
+        """
+
+        :param sample:
+        :param batch_idx:
+        :param optimizer_idx:
+        :return: {'loss': torch.Tensor, 'progress_bar': dict, 'tb_log': dict}
+        """
+        loss_ret = self._training_step(sample, batch_idx, optimizer_idx)
+        if loss_ret is None:
+            return {'loss': None}
+        total_loss, log_outputs = loss_ret
+        log_outputs = tensors_to_scalars(log_outputs)
+        for k, v in log_outputs.items():
+            if k not in self.training_losses_meter:
+                self.training_losses_meter[k] = AvgrageMeter()
+            if not np.isnan(v):
+                self.training_losses_meter[k].update(v)
+        self.training_losses_meter['total_loss'].update(total_loss.item())
+
+        if optimizer_idx >= 0:
+            log_outputs[f'lr_{optimizer_idx}'] = self.trainer.optimizers[optimizer_idx].param_groups[0]['lr']
+
+        progress_bar_log = log_outputs
+        tb_log = {f'tr/{k}': v for k, v in log_outputs.items()}
+        return {
+            'loss': total_loss,
+            'progress_bar': progress_bar_log,
+            'tb_log': tb_log
+        }
+
+    def on_before_optimization(self, opt_idx):
+        if self.gradient_clip_norm > 0:
+            torch.nn.utils.clip_grad_norm_(self.parameters(), self.gradient_clip_norm)
+        if self.gradient_clip_val > 0:
+            torch.nn.utils.clip_grad_value_(self.parameters(), self.gradient_clip_val)
+
+    def on_after_optimization(self, epoch, batch_idx, optimizer, optimizer_idx):
+        if self.scheduler is not None:
+            # self.scheduler.step(self.global_step // hparams['accumulate_grad_batches'])
+            # the code above causes EPOCH_DEPRECATION_WARNING, changed it and changed the optimizer init with
+            # step_size divided by accumulate_grad_batches
+            self.scheduler.step()
+
+    ######################
+    # validation
+    ######################
+    def validation_start(self):
+        pass
+
+    def validation_step(self, sample, batch_idx):
+        """
+
+        :param sample:
+        :param batch_idx:
+        :return: output: {"losses": {...}, "total_loss": float, ...} or (total loss: torch.Tensor, loss_log: dict)
+        """
+        raise NotImplementedError
+
+    def validation_end(self, outputs):
+        """
+
+        :param outputs:
+        :return: loss_output: dict
+        """
+        all_losses_meter = {'total_loss': AvgrageMeter()}
+        for output in outputs:
+            if len(output) == 0 or output is None:
+                continue
+            if isinstance(output, dict):
+                assert 'losses' in output, 'Key "losses" should exist in validation output.'
+                n = output.pop('nsamples', 1)
+                losses = tensors_to_scalars(output['losses'])
+                total_loss = output.get('total_loss', sum(losses.values()))
+            else:
+                assert len(output) == 2, 'Validation output should only consist of two elements: (total_loss, losses)'
+                n = 1
+                total_loss, losses = output
+                losses = tensors_to_scalars(losses)
+            if isinstance(total_loss, torch.Tensor):
+                total_loss = total_loss.item()
+            for k, v in losses.items():
+                if k not in all_losses_meter:
+                    all_losses_meter[k] = AvgrageMeter()
+                all_losses_meter[k].update(v, n)
+            all_losses_meter['total_loss'].update(total_loss, n)
+        loss_output = {k: round(v.avg, 4) for k, v in all_losses_meter.items()}
+        print(f"| Validation results@{self.global_step}: {loss_output}")
+        return {
+            'tb_log': {f'val/{k}': v for k, v in loss_output.items()},
+            'val_loss': loss_output['total_loss']
+        }
+
+    ######################
+    # testing
+    ######################
+    def test_start(self):
+        pass
+
+    def test_step(self, sample, batch_idx):
+        return self.validation_step(sample, batch_idx)
+
+    def test_end(self, outputs):
+        return self.validation_end(outputs)
+
+    ######################
+    # start training/testing
+    ######################
+    @classmethod
+    def start(cls):
+        os.environ['MASTER_PORT'] = str(random.randint(15000, 30000))
+        random.seed(hparams['seed'])
+        np.random.seed(hparams['seed'])
+        work_dir = hparams['work_dir']
+        trainer = Trainer(
+            work_dir=work_dir,
+            val_check_interval=hparams['val_check_interval'],
+            tb_log_interval=hparams['tb_log_interval'],
+            max_updates=hparams['max_updates'],
+            num_sanity_val_steps=hparams['num_sanity_val_steps'] if not hparams['validate'] else 10000,
+            accumulate_grad_batches=hparams['accumulate_grad_batches'],
+            print_nan_grads=hparams['print_nan_grads'],
+            resume_from_checkpoint=hparams.get('resume_from_checkpoint', 0),
+            amp=hparams['amp'],
+            monitor_key=hparams['valid_monitor_key'],
+            monitor_mode=hparams['valid_monitor_mode'],
+            num_ckpt_keep=hparams['num_ckpt_keep'],
+            save_best=hparams['save_best'],
+            seed=hparams['seed'],
+            debug=hparams['debug']
+        )
+        if not hparams['infer']:  # train
+            trainer.fit(cls)
+        else:
+            trainer.test(cls)
+
+    def on_keyboard_interrupt(self):
+        pass

preprocess/tools/note_transcription/utils/commons/ckpt_utils.py

@@ -0,0 +1,68 @@
+import glob
+import os
+import re
+import torch
+
+
+def get_last_checkpoint(work_dir, steps=None):
+    checkpoint = None
+    last_ckpt_path = None
+    ckpt_paths = get_all_ckpts(work_dir, steps)
+    if len(ckpt_paths) > 0:
+        last_ckpt_path = ckpt_paths[0]
+        checkpoint = torch.load(last_ckpt_path, map_location='cpu')
+    return checkpoint, last_ckpt_path
+
+
+def get_all_ckpts(work_dir, steps=None):
+    if steps is None:
+        ckpt_path_pattern = f'{work_dir}/model_ckpt_steps_*.ckpt'
+    else:
+        ckpt_path_pattern = f'{work_dir}/model_ckpt_steps_{steps}.ckpt'
+    return sorted(glob.glob(ckpt_path_pattern),
+                  key=lambda x: -int(re.findall('.*steps\_(\d+)\.ckpt', x)[0]))
+
+
+def load_ckpt(cur_model, ckpt_base_dir, model_name='model', force=True, strict=True, verbose=True):
+    if os.path.isfile(ckpt_base_dir):
+        base_dir = os.path.dirname(ckpt_base_dir)
+        ckpt_path = ckpt_base_dir
+        checkpoint = torch.load(ckpt_base_dir, map_location='cpu')
+    else:
+        base_dir = ckpt_base_dir
+        checkpoint, ckpt_path = get_last_checkpoint(ckpt_base_dir)
+    if checkpoint is not None:
+        state_dict = checkpoint["state_dict"]
+        if len([k for k in state_dict.keys() if '.' in k]) > 0:
+            state_dict = {k[len(model_name) + 1:]: v for k, v in state_dict.items()
+                          if k.startswith(f'{model_name}.')}
+        else:
+            if '.' not in model_name:
+                state_dict = state_dict[model_name]
+            else:
+                base_model_name = model_name.split('.')[0]
+                rest_model_name = model_name[len(base_model_name) + 1:]
+                state_dict = {
+                    k[len(rest_model_name) + 1:]: v for k, v in state_dict[base_model_name].items()
+                    if k.startswith(f'{rest_model_name}.')}
+        if not strict:
+            cur_model_state_dict = cur_model.state_dict()
+            unmatched_keys = []
+            for key, param in state_dict.items():
+                if key in cur_model_state_dict:
+                    new_param = cur_model_state_dict[key]
+                    if new_param.shape != param.shape:
+                        unmatched_keys.append(key)
+                        print("| Unmatched keys: ", key, new_param.shape, param.shape)
+            for key in unmatched_keys:
+                del state_dict[key]
+        # print(state_dict)
+        cur_model.load_state_dict(state_dict, strict=strict)
+        if verbose:
+            print(f"| load '{model_name}' from '{ckpt_path}'.")
+    else:
+        e_msg = f"| ckpt not found in {base_dir}."
+        if force:
+            assert False, e_msg
+        else:
+            print(e_msg)

preprocess/tools/note_transcription/utils/commons/dataset_utils.py

@@ -0,0 +1,372 @@
+import os
+import sys
+import traceback
+import types
+from functools import wraps
+from itertools import chain
+import numpy as np
+import torch.utils.data
+import torch.nn.functional as F
+from torch.utils.data import ConcatDataset
+from .hparams import hparams
+
+
+def collate_1d_or_2d(values, pad_idx=0, left_pad=False, shift_right=False, max_len=None, shift_id=1):
+    if len(values[0].shape) == 1:
+        return collate_1d(values, pad_idx, left_pad, shift_right, max_len, shift_id)
+    else:
+        return collate_2d(values, pad_idx, left_pad, shift_right, max_len)
+
+
+def collate_1d(values, pad_idx=0, left_pad=False, shift_right=False, max_len=None, shift_id=1):
+    """Convert a list of 1d tensors into a padded 2d tensor."""
+    size = max(v.size(0) for v in values) if max_len is None else max_len
+    res = values[0].new(len(values), size).fill_(pad_idx)
+
+    def copy_tensor(src, dst):
+        assert dst.numel() == src.numel()
+        if shift_right:
+            dst[1:] = src[:-1]
+            dst[0] = shift_id
+        else:
+            dst.copy_(src)
+
+    for i, v in enumerate(values):
+        copy_tensor(v, res[i][size - len(v):] if left_pad else res[i][:len(v)])
+    return res
+
+
+def collate_2d(values, pad_idx=0, left_pad=False, shift_right=False, max_len=None):
+    """Convert a list of 2d tensors into a padded 3d tensor."""
+    size = max(v.size(0) for v in values) if max_len is None else max_len
+    res = values[0].new(len(values), size, values[0].shape[1]).fill_(pad_idx)
+
+    def copy_tensor(src, dst):
+        assert dst.numel() == src.numel()
+        if shift_right:
+            dst[1:] = src[:-1]
+        else:
+            dst.copy_(src)
+
+    for i, v in enumerate(values):
+        copy_tensor(v, res[i][size - len(v):] if left_pad else res[i][:len(v)])
+    return res
+
+def collate_xd(values, pad_value=0, max_len=None):
+    size = ((max(v.size(0) for v in values) if max_len is None else max_len), *values[0].shape[1:])
+    res = torch.full((len(values), *size), fill_value=pad_value, dtype=values[0].dtype, device=values[0].device)
+
+    for i, v in enumerate(values):
+        res[i, :len(v), ...] = v
+    return res
+
+def pad_or_cut_1d(values: torch.tensor, tgt_len, pad_value=0):
+    src_len = values.shape[0]
+    if src_len < tgt_len:
+        res = F.pad(values, [0, tgt_len - src_len], value=pad_value)
+    else:
+        res = values[:tgt_len]
+    return res
+
+def pad_or_cut_2d(values: torch.tensor, tgt_len, dim=-1, pad_value=0):
+    if dim == 0 or dim == -2:
+        src_len = values.shape[0]
+        if src_len < tgt_len:
+            res = F.pad(values, [0, 0, 0, tgt_len - src_len], value=pad_value)
+        else:
+            res = values[:tgt_len]
+    elif dim == 1 or dim == -1:
+        src_len = values.shape[1]
+        if src_len < tgt_len:
+            res = F.pad(values, [0, tgt_len - src_len], value=pad_value)
+        else:
+            res = values[:, :tgt_len]
+    else:
+        raise RuntimeError(f"Wrong dim number {dim} while the tensor only has {len(values.shape)} dimensions.")
+    return res
+
+def pad_or_cut_3d(values: torch.tensor, tgt_len, dim=-1, pad_value=0):
+    if dim == 0 or dim == -3:
+        src_len = values.shape[0]
+        if src_len < tgt_len:
+            res = F.pad(values, [0, 0, 0, 0, 0, tgt_len - src_len], value=pad_value)
+        else:
+            res = values[:tgt_len]
+    elif dim == 1 or dim == -2:
+        src_len = values.shape[1]
+        if src_len < tgt_len:
+            res = F.pad(values, [0, 0, 0, tgt_len - src_len], value=pad_value)
+        else:
+            res = values[:, :tgt_len]
+    elif dim == 2 or dim == -1:
+        src_len = values.shape[2]
+        if src_len < tgt_len:
+            res = F.pad(values, [0, tgt_len - src_len], value=pad_value)
+        else:
+            res = values[:, :, :tgt_len]
+    else:
+        raise RuntimeError(f"Wrong dim number {dim} while the tensor only has {len(values.shape)} dimensions.")
+    return res
+
+def pad_or_cut_xd(values, tgt_len, dim=-1, pad_value=0):
+    if len(values.shape) == 1:
+        return pad_or_cut_1d(values, tgt_len, pad_value)
+    elif len(values.shape) == 2:
+        return pad_or_cut_2d(values, tgt_len, dim, pad_value)
+    elif len(values.shape) == 3:
+        return pad_or_cut_3d(values, tgt_len, dim, pad_value)
+    else:
+        raise NotImplementedError
+
+def _is_batch_full(batch, num_tokens, max_tokens, max_sentences):
+    if len(batch) == 0:
+        return 0
+    if len(batch) == max_sentences:
+        return 1
+    if num_tokens > max_tokens:
+        return 1
+    return 0
+
+
+def batch_by_size(
+        indices, num_tokens_fn, max_tokens=None, max_sentences=None,
+        required_batch_size_multiple=1, distributed=False
+):
+    """
+    Yield mini-batches of indices bucketed by size. Batches may contain
+    sequences of different lengths.
+
+    Args:
+        indices (List[int]): ordered list of dataset indices
+        num_tokens_fn (callable): function that returns the number of tokens at
+            a given index
+        max_tokens (int, optional): max number of tokens in each batch
+            (default: None).
+        max_sentences (int, optional): max number of sentences in each
+            batch (default: None).
+        required_batch_size_multiple (int, optional): require batch size to
+            be a multiple of N (default: 1).
+    """
+    max_tokens = max_tokens if max_tokens is not None else sys.maxsize
+    max_sentences = max_sentences if max_sentences is not None else sys.maxsize
+    bsz_mult = required_batch_size_multiple
+
+    if isinstance(indices, types.GeneratorType):
+        indices = np.fromiter(indices, dtype=np.int64, count=-1)
+
+    sample_len = 0
+    sample_lens = []
+    batch = []
+    batches = []
+    for i in range(len(indices)):
+        idx = indices[i]
+        num_tokens = num_tokens_fn(idx)
+        sample_lens.append(num_tokens)
+        sample_len = max(sample_len, num_tokens)
+
+        assert sample_len <= max_tokens, (
+            "sentence at index {} of size {} exceeds max_tokens "
+            "limit of {}!".format(idx, sample_len, max_tokens)
+        )
+        num_tokens = (len(batch) + 1) * sample_len
+
+        if _is_batch_full(batch, num_tokens, max_tokens, max_sentences):
+            mod_len = max(
+                bsz_mult * (len(batch) // bsz_mult),
+                len(batch) % bsz_mult,
+            )
+            batches.append(batch[:mod_len])
+            batch = batch[mod_len:]
+            sample_lens = sample_lens[mod_len:]
+            sample_len = max(sample_lens) if len(sample_lens) > 0 else 0
+        batch.append(idx)
+    if len(batch) > 0:
+        batches.append(batch)
+    return batches
+
+
+def build_dataloader(dataset, shuffle, max_tokens=None, max_sentences=None,
+                     required_batch_size_multiple=-1, endless=False, apply_batch_by_size=True, pin_memory=False, use_ddp=False):
+    import torch.distributed as dist
+    devices_cnt = torch.cuda.device_count()
+    if devices_cnt == 0:
+        devices_cnt = 1
+    if not use_ddp:
+        devices_cnt = 1
+    if required_batch_size_multiple == -1:
+        required_batch_size_multiple = devices_cnt
+
+    def shuffle_batches(batches):
+        np.random.shuffle(batches)
+        return batches
+
+    if max_tokens is not None:
+        max_tokens *= devices_cnt
+    if max_sentences is not None:
+        max_sentences *= devices_cnt
+    indices = dataset.ordered_indices()
+    if apply_batch_by_size:
+        batch_sampler = batch_by_size(
+            indices, dataset.num_tokens, max_tokens=max_tokens, max_sentences=max_sentences,
+            required_batch_size_multiple=required_batch_size_multiple,
+        )
+    else:
+        batch_sampler = []
+        for i in range(0, len(indices), max_sentences):
+            batch_sampler.append(indices[i:i + max_sentences])
+
+    if shuffle:
+        batches = shuffle_batches(list(batch_sampler))
+        if endless:
+            batches = [b for _ in range(1000) for b in shuffle_batches(list(batch_sampler))]
+    else:
+        batches = batch_sampler
+        if endless:
+            batches = [b for _ in range(1000) for b in batches]
+    num_workers = dataset.num_workers
+    if use_ddp:
+        num_replicas = dist.get_world_size()
+        rank = dist.get_rank()
+        # batches = [x[rank::num_replicas] for x in batches if len(x) % num_replicas == 0]
+        # ensure that every sample in the dataset is covered
+        batches_ = []
+        for x in batches:
+            if len(x) % num_replicas == 0:
+                batches_.append(x[rank::num_replicas])
+            else:
+                x_ = x + [x[-1]] * (len(x) - len(x) // num_replicas * num_replicas)
+                batches_.append(x_[rank::num_replicas])
+        batches = batches_
+    return torch.utils.data.DataLoader(dataset,
+                                       collate_fn=dataset.collater,
+                                       batch_sampler=batches,
+                                       num_workers=num_workers,
+                                       pin_memory=pin_memory)
+
+
+def unpack_dict_to_list(samples):
+    samples_ = []
+    bsz = samples.get('outputs').size(0)
+    for i in range(bsz):
+        res = {}
+        for k, v in samples.items():
+            try:
+                res[k] = v[i]
+            except:
+                pass
+        samples_.append(res)
+    return samples_
+
+
+def remove_padding(x, padding_idx=0):
+    if x is None:
+        return None
+    assert len(x.shape) in [1, 2]
+    if len(x.shape) == 2:  # [T, H]
+        return x[np.abs(x).sum(-1) != padding_idx]
+    elif len(x.shape) == 1:  # [T]
+        return x[x != padding_idx]
+
+
+def data_loader(fn):
+    """
+    Decorator to make any fx with this use the lazy property
+    :param fn:
+    :return:
+    """
+
+    wraps(fn)
+    attr_name = '_lazy_' + fn.__name__
+
+    def _get_data_loader(self):
+        try:
+            value = getattr(self, attr_name)
+        except AttributeError:
+            try:
+                value = fn(self)  # Lazy evaluation, done only once.
+            except AttributeError as e:
+                # Guard against AttributeError suppression. (Issue #142)
+                traceback.print_exc()
+                error = f'{fn.__name__}: An AttributeError was encountered: ' + str(e)
+                raise RuntimeError(error) from e
+            setattr(self, attr_name, value)  # Memoize evaluation.
+        return value
+
+    return _get_data_loader
+
+
+class BaseDataset(torch.utils.data.Dataset):
+    def __init__(self, shuffle):
+        super().__init__()
+        self.hparams = hparams
+        self.shuffle = shuffle
+        self.sort_by_len = hparams['sort_by_len']
+        self.sizes = None
+
+    @property
+    def _sizes(self):
+        return self.sizes
+
+    def __getitem__(self, index):
+        raise NotImplementedError
+
+    def collater(self, samples):
+        raise NotImplementedError
+
+    def __len__(self):
+        return len(self._sizes)
+
+    def num_tokens(self, index):
+        return self.size(index)
+
+    def size(self, index):
+        """Return an example's size as a float or tuple. This value is used when
+        filtering a dataset with ``--max-positions``."""
+        return min(self._sizes[index], hparams['max_frames'])
+
+    def ordered_indices(self):
+        """Return an ordered list of indices. Batches will be constructed based
+        on this order."""
+        if self.shuffle:
+            indices = np.random.permutation(len(self))
+            if self.sort_by_len:
+                indices = indices[np.argsort(np.array(self._sizes)[indices], kind='mergesort')]
+        else:
+            indices = np.arange(len(self))
+        return indices.tolist()
+
+    @property
+    def num_workers(self):
+        return int(os.getenv('NUM_WORKERS', hparams['ds_workers']))
+
+
+class BaseConcatDataset(ConcatDataset):
+    def collater(self, samples):
+        return self.datasets[0].collater(samples)
+
+    @property
+    def _sizes(self):
+        if not hasattr(self, 'sizes'):
+            self.sizes = list(chain.from_iterable([d._sizes for d in self.datasets]))
+        return self.sizes
+
+    def size(self, index):
+        return min(self._sizes[index], hparams['max_frames'])
+
+    def num_tokens(self, index):
+        return self.size(index)
+
+    def ordered_indices(self):
+        """Return an ordered list of indices. Batches will be constructed based
+        on this order."""
+        if self.datasets[0].shuffle:
+            indices = np.random.permutation(len(self))
+            if self.datasets[0].sort_by_len:
+                indices = indices[np.argsort(np.array(self._sizes)[indices], kind='mergesort')]
+        else:
+            indices = np.arange(len(self))
+        return indices
+
+    @property
+    def num_workers(self):
+        return self.datasets[0].num_workers

preprocess/tools/note_transcription/utils/commons/ddp_utils.py

@@ -0,0 +1,164 @@
+from torch.nn.parallel import DistributedDataParallel
+from torch.nn.parallel.distributed import _find_tensors
+import torch.optim
+import torch.utils.data
+import torch
+from packaging import version
+
+class DDP(DistributedDataParallel):
+    """
+    Override the forward call in lightning so it goes to training and validation step respectively
+    """
+
+    def forward(self, *inputs, **kwargs):  # pragma: no cover
+        # if version.parse(torch.__version__[:6]) < version.parse("1.11"):
+        if version.parse(torch.__version__) < version.parse("1.11"):    # fix the hard [:6] problem
+            self._sync_params()
+            inputs, kwargs = self.scatter(inputs, kwargs, self.device_ids)
+            assert len(self.device_ids) == 1
+            if self.module.training:
+                output = self.module.training_step(*inputs[0], **kwargs[0])
+            elif self.module.testing:
+                output = self.module.test_step(*inputs[0], **kwargs[0])
+            else:
+                output = self.module.validation_step(*inputs[0], **kwargs[0])
+            if torch.is_grad_enabled():
+                # We'll return the output object verbatim since it is a freeform
+                # object. We need to find any tensors in this object, though,
+                # because we need to figure out which parameters were used during
+                # this forward pass, to ensure we short circuit reduction for any
+                # unused parameters. Only if `find_unused_parameters` is set.
+                if self.find_unused_parameters:
+                    self.reducer.prepare_for_backward(list(_find_tensors(output)))
+                else:
+                    self.reducer.prepare_for_backward([])
+        elif version.parse("1.11") <= version.parse(torch.__version__) < version.parse("2.0"):
+            from torch.nn.parallel.distributed import \
+                logging, Join, _DDPSink, _tree_flatten_with_rref, _tree_unflatten_with_rref
+            with torch.autograd.profiler.record_function("DistributedDataParallel.forward"):
+                if torch.is_grad_enabled() and self.require_backward_grad_sync:
+                    self.logger.set_runtime_stats_and_log()
+                    self.num_iterations += 1
+                    self.reducer.prepare_for_forward()
+
+                # Notify the join context that this process has not joined, if
+                # needed
+                work = Join.notify_join_context(self)
+                if work:
+                    self.reducer._set_forward_pass_work_handle(
+                        work, self._divide_by_initial_world_size
+                    )
+
+                # Calling _rebuild_buckets before forward compuation,
+                # It may allocate new buckets before deallocating old buckets
+                # inside _rebuild_buckets. To save peak memory usage,
+                # call _rebuild_buckets before the peak memory usage increases
+                # during forward computation.
+                # This should be called only once during whole training period.
+                if torch.is_grad_enabled() and self.reducer._rebuild_buckets():
+                    logging.info("Reducer buckets have been rebuilt in this iteration.")
+                    self._has_rebuilt_buckets = True
+
+                # sync params according to location (before/after forward) user
+                # specified as part of hook, if hook was specified.
+                buffer_hook_registered = hasattr(self, 'buffer_hook')
+                if self._check_sync_bufs_pre_fwd():
+                    self._sync_buffers()
+
+                if self._join_config.enable:
+                    # Notify joined ranks whether they should sync in backwards pass or not.
+                    self._check_global_requires_backward_grad_sync(is_joined_rank=False)
+
+                # modified part
+                inputs, kwargs = self.scatter(inputs, kwargs, self.device_ids)
+                if self.module.training:
+                    output = self.module.training_step(*inputs[0], **kwargs[0])
+                elif self.module.testing:
+                    output = self.module.test_step(*inputs[0], **kwargs[0])
+                else:
+                    output = self.module.validation_step(*inputs[0], **kwargs[0])
+
+                # sync params according to location (before/after forward) user
+                # specified as part of hook, if hook was specified.
+                if self._check_sync_bufs_post_fwd():
+                    self._sync_buffers()
+
+                if torch.is_grad_enabled() and self.require_backward_grad_sync:
+                    self.require_forward_param_sync = True
+                    # We'll return the output object verbatim since it is a freeform
+                    # object. We need to find any tensors in this object, though,
+                    # because we need to figure out which parameters were used during
+                    # this forward pass, to ensure we short circuit reduction for any
+                    # unused parameters. Only if `find_unused_parameters` is set.
+                    if self.find_unused_parameters and not self.static_graph:
+                        # Do not need to populate this for static graph.
+                        self.reducer.prepare_for_backward(list(_find_tensors(output)))
+                    else:
+                        self.reducer.prepare_for_backward([])
+                else:
+                    self.require_forward_param_sync = False
+
+            # TODO: DDPSink is currently enabled for unused parameter detection and
+            # static graph training for first iteration.
+            if (self.find_unused_parameters and not self.static_graph) or (
+                    self.static_graph and self.num_iterations == 1
+            ):
+                state_dict = {
+                    'static_graph': self.static_graph,
+                    'num_iterations': self.num_iterations,
+                }
+
+                output_tensor_list, treespec, output_is_rref = _tree_flatten_with_rref(
+                    output
+                )
+                output_placeholders = [None for _ in range(len(output_tensor_list))]
+                # Do not touch tensors that have no grad_fn, which can cause issues
+                # such as https://github.com/pytorch/pytorch/issues/60733
+                for i, output in enumerate(output_tensor_list):
+                    if torch.is_tensor(output) and output.grad_fn is None:
+                        output_placeholders[i] = output
+
+                # When find_unused_parameters=True, makes tensors which require grad
+                # run through the DDPSink backward pass. When not all outputs are
+                # used in loss, this makes those corresponding tensors receive
+                # undefined gradient which the reducer then handles to ensure
+                # param.grad field is not touched and we don't error out.
+                passthrough_tensor_list = _DDPSink.apply(
+                    self.reducer,
+                    state_dict,
+                    *output_tensor_list,
+                )
+                for i in range(len(output_placeholders)):
+                    if output_placeholders[i] is None:
+                        output_placeholders[i] = passthrough_tensor_list[i]
+
+                # Reconstruct output data structure.
+                output = _tree_unflatten_with_rref(
+                    output_placeholders, treespec, output_is_rref
+                )
+        else:
+            # now pytorch version >= 2.0
+            with torch.autograd.profiler.record_function("DistributedDataParallel.forward"):
+                inputs, kwargs = self._pre_forward(*inputs, **kwargs)
+                output = (
+                    # self.module.forward(*inputs, **kwargs)
+                    # if self._delay_all_reduce_all_params
+                    # else self._run_ddp_forward(*inputs, **kwargs)
+                    # modified: delete 'delay_all_reduce_named_params' function
+                    self._run_ddp_forward(*inputs, **kwargs)
+                )
+                return self._post_forward(output)
+        return output
+
+    def _run_ddp_forward(self, *inputs, **kwargs):
+        if version.parse(torch.__version__) >= version.parse("2.0"):
+            with self._inside_ddp_forward():
+                if self.module.training:
+                    output = self.module.training_step(*inputs, **kwargs)
+                elif self.module.testing:
+                    output = self.module.test_step(*inputs, **kwargs)
+                else:
+                    output = self.module.validation_step(*inputs, **kwargs)
+                return output  # type: ignore[index]
+        else:
+            return super(DDP, self)._run_ddp_forward(*inputs, **kwargs)

preprocess/tools/note_transcription/utils/commons/gpu_mem_track.py

@@ -0,0 +1,113 @@
+import gc
+import datetime
+import inspect
+
+import torch
+import numpy as np
+
+dtype_memory_size_dict = {
+    torch.float64: 64/8,
+    torch.double: 64/8,
+    torch.float32: 32/8,
+    torch.float: 32/8,
+    torch.float16: 16/8,
+    torch.half: 16/8,
+    torch.int64: 64/8,
+    torch.long: 64/8,
+    torch.int32: 32/8,
+    torch.int: 32/8,
+    torch.int16: 16/8,
+    torch.short: 16/6,
+    torch.uint8: 8/8,
+    torch.int8: 8/8,
+}
+# compatibility of torch1.0
+if getattr(torch, "bfloat16", None) is not None:
+    dtype_memory_size_dict[torch.bfloat16] = 16/8
+if getattr(torch, "bool", None) is not None:
+    dtype_memory_size_dict[torch.bool] = 8/8 # pytorch use 1 byte for a bool, see https://github.com/pytorch/pytorch/issues/41571
+
+def get_mem_space(x):
+    try:
+        ret = dtype_memory_size_dict[x]
+    except KeyError:
+        print(f"dtype {x} is not supported!")
+    return ret
+
+class MemTracker(object):
+    """
+    Class used to track pytorch memory usage
+    Arguments:
+        detail(bool, default True): whether the function shows the detail gpu memory usage
+        path(str): where to save log file
+        verbose(bool, default False): whether show the trivial exception
+        device(int): GPU number, default is 0
+    """
+    def __init__(self, detail=True, path='', verbose=False, device=0):
+        self.print_detail = detail
+        self.last_tensor_sizes = set()
+        self.gpu_profile_fn = path + f'{datetime.datetime.now():%d-%b-%y-%H:%M:%S}-gpu_mem_track.txt'
+        self.verbose = verbose
+        self.begin = True
+        self.device = device
+
+    def get_tensors(self):
+        for obj in gc.get_objects():
+            try:
+                if torch.is_tensor(obj) or (hasattr(obj, 'data') and torch.is_tensor(obj.data)):
+                    tensor = obj
+                else:
+                    continue
+                if tensor.is_cuda:
+                    yield tensor
+            except Exception as e:
+                if self.verbose:
+                    print('A trivial exception occured: {}'.format(e))
+
+    def get_tensor_usage(self):
+        sizes = [np.prod(np.array(tensor.size())) * get_mem_space(tensor.dtype) for tensor in self.get_tensors()]
+        return np.sum(sizes) / 1024**2
+
+    def get_allocate_usage(self):
+        return torch.cuda.memory_allocated() / 1024**2
+
+    def clear_cache(self):
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def print_all_gpu_tensor(self, file=None):
+        for x in self.get_tensors():
+            print(x.size(), x.dtype, np.prod(np.array(x.size()))*get_mem_space(x.dtype)/1024**2, file=file)
+
+    def track(self):
+        """
+        Track the GPU memory usage
+        """
+        frameinfo = inspect.stack()[1]
+        where_str = frameinfo.filename + ' line ' + str(frameinfo.lineno) + ': ' + frameinfo.function
+
+        with open(self.gpu_profile_fn, 'a+') as f:
+
+            if self.begin:
+                f.write(f"GPU Memory Track | {datetime.datetime.now():%d-%b-%y-%H:%M:%S} |"
+                        f" Total Tensor Used Memory:{self.get_tensor_usage():<7.1f}Mb"
+                        f" Total Allocated Memory:{self.get_allocate_usage():<7.1f}Mb\n\n")
+                self.begin = False
+
+            if self.print_detail is True:
+                ts_list = [(tensor.size(), tensor.dtype) for tensor in self.get_tensors()]
+                new_tensor_sizes = {(type(x),
+                                    tuple(x.size()),
+                                    ts_list.count((x.size(), x.dtype)),
+                                    np.prod(np.array(x.size()))*get_mem_space(x.dtype)/1024**2,
+                                    x.dtype) for x in self.get_tensors()}
+                for t, s, n, m, data_type in new_tensor_sizes - self.last_tensor_sizes:
+                    f.write(f'+ | {str(n)} * Size:{str(s):<20} | Memory: {str(m*n)[:6]} M | {str(t):<20} | {data_type}\n')
+                for t, s, n, m, data_type in self.last_tensor_sizes - new_tensor_sizes:
+                    f.write(f'- | {str(n)} * Size:{str(s):<20} | Memory: {str(m*n)[:6]} M | {str(t):<20} | {data_type}\n')
+
+                self.last_tensor_sizes = new_tensor_sizes
+
+            f.write(f"\nAt {where_str:<50}"
+                    f" Total Tensor Used Memory:{self.get_tensor_usage():<7.1f}Mb"
+                    f" Total Allocated Memory:{self.get_allocate_usage():<7.1f}Mb\n\n")

preprocess/tools/note_transcription/utils/commons/hparams.py

@@ -0,0 +1,131 @@
+import argparse
+import os
+import yaml
+
+global_print_hparams = True
+hparams = {}
+
+
+class Args:
+    def __init__(self, **kwargs):
+        for k, v in kwargs.items():
+            self.__setattr__(k, v)
+
+
+def override_config(old_config: dict, new_config: dict):
+    for k, v in new_config.items():
+        if isinstance(v, dict) and k in old_config:
+            override_config(old_config[k], new_config[k])
+        else:
+            old_config[k] = v
+
+
+def set_hparams(config='', exp_name='', hparams_str='', print_hparams=True, global_hparams=True, root_dir=''):
+    if config == '' and exp_name == '':
+        parser = argparse.ArgumentParser(description='')
+        parser.add_argument('--config', type=str, default='',
+                            help='location of the data corpus')
+        parser.add_argument('--exp_name', type=str, default='', help='exp_name')
+        parser.add_argument('-hp', '--hparams', type=str, default='',
+                            help='location of the data corpus')
+        parser.add_argument('--infer', action='store_true', help='infer')
+        parser.add_argument('--validate', action='store_true', help='validate')
+        parser.add_argument('--reset', action='store_true', help='reset hparams')
+        parser.add_argument('--remove', action='store_true', help='remove old ckpt')
+        parser.add_argument('--debug', action='store_true', help='debug')
+        parser.add_argument('--root_dir', type=str, default='', help='root directory of the project.')
+        args, unknown = parser.parse_known_args()
+        print("| Unknow hparams: ", unknown)
+    else:
+        args = Args(config=config, exp_name=exp_name, hparams=hparams_str,
+                    infer=False, validate=False, reset=False, debug=False, remove=False, root_dir=root_dir)
+    global hparams
+    assert args.config != '' or args.exp_name != ''
+    root_dir = args.root_dir
+    if args.config != '':
+        assert os.path.exists(os.path.join(root_dir, args.config)), f'| Wrong config path! root_dir: {root_dir}, config_path: {args.config}'
+
+    config_chains = []
+    loaded_config = set()
+
+    def load_config(config_fn):
+        # deep first inheritance and avoid the second visit of one node
+        if not os.path.exists(os.path.join(root_dir, config_fn)):
+            return {}
+        with open(os.path.join(root_dir, config_fn)) as f:
+            hparams_ = yaml.safe_load(f)
+        loaded_config.add(config_fn)
+        if 'base_config' in hparams_:
+            ret_hparams = {}
+            if not isinstance(hparams_['base_config'], list):
+                hparams_['base_config'] = [hparams_['base_config']]
+            for c in hparams_['base_config']:
+                if c.startswith('.'):
+                    c = f'{os.path.dirname(config_fn)}/{c}'
+                    c = os.path.normpath(c)
+                if c not in loaded_config:
+                    override_config(ret_hparams, load_config(c))
+            override_config(ret_hparams, hparams_)
+        else:
+            ret_hparams = hparams_
+        config_chains.append(config_fn)
+        return ret_hparams
+
+    saved_hparams = {}
+    args_work_dir = ''
+    if args.exp_name != '':
+        args_work_dir = os.path.join(root_dir, f'checkpoints/{args.exp_name}')
+        ckpt_config_path = f'{args_work_dir}/config.yaml'
+        if os.path.exists(ckpt_config_path):
+            with open(ckpt_config_path) as f:
+                saved_hparams_ = yaml.safe_load(f)
+                if saved_hparams_ is not None:
+                    saved_hparams.update(saved_hparams_)
+    hparams_ = {}
+    if args.config != '':
+        hparams_.update(load_config(args.config))
+    if not args.reset:
+        hparams_.update(saved_hparams)
+    hparams_['work_dir'] = args_work_dir
+
+    # Support config overriding in command line. Support list type config overriding.
+    # Examples: --hparams="a=1,b.c=2,d=[1 1 1]"
+    if args.hparams != "":
+        for new_hparam in args.hparams.split(","):
+            k, v = new_hparam.split("=")
+            v = v.strip("\'\" ")
+            config_node = hparams_
+            for k_ in k.split(".")[:-1]:
+                config_node = config_node[k_]
+            k = k.split(".")[-1]
+            if v in ['True', 'False'] or type(config_node[k]) in [bool, list, dict]:
+                if type(config_node[k]) == list:
+                    v = v.replace(" ", ",")
+                config_node[k] = eval(v)
+            else:
+                config_node[k] = type(config_node[k])(v)
+    if args_work_dir != '' and args.remove:
+        answer = input("REMOVE old checkpoint? Y/N [Default: N]: ")
+        if answer.lower() == "y":
+            pass
+    if args_work_dir != '' and (not os.path.exists(ckpt_config_path) or args.reset) and not args.infer:
+        os.makedirs(hparams_['work_dir'], exist_ok=True)
+        with open(ckpt_config_path, 'w') as f:
+            yaml.safe_dump(hparams_, f)
+
+    hparams_['infer'] = args.infer
+    hparams_['debug'] = args.debug
+    hparams_['validate'] = args.validate
+    hparams_['exp_name'] = args.exp_name
+    global global_print_hparams
+    if global_hparams:
+        hparams.clear()
+        hparams.update(hparams_)
+    if print_hparams and global_print_hparams and global_hparams:
+        # print('| Hparams chains: ', config_chains)
+        # print('| Hparams: ')
+        # for i, (k, v) in enumerate(sorted(hparams_.items())):
+        #     print(f"\033[;33;m{k}\033[0m: {v}, ", end="\n" if i % 5 == 4 else "")
+        # print("")
+        global_print_hparams = False
+    return hparams_

preprocess/tools/note_transcription/utils/commons/indexed_datasets.py

@@ -0,0 +1,71 @@
+import pickle
+from copy import deepcopy
+
+import numpy as np
+
+
+class IndexedDataset:
+    def __init__(self, path, num_cache=1):
+        super().__init__()
+        self.path = path
+        self.data_file = None
+        self.data_offsets = np.load(f"{path}.idx", allow_pickle=True).item()['offsets']
+        self.data_file = open(f"{path}.data", 'rb', buffering=-1)
+        self.cache = []
+        self.num_cache = num_cache
+
+    def check_index(self, i):
+        if i < 0 or i >= len(self.data_offsets) - 1:
+            raise IndexError('index out of range')
+
+    def __del__(self):
+        if self.data_file:
+            self.data_file.close()
+
+    def __getitem__(self, i):
+        self.check_index(i)
+        if self.num_cache > 0:
+            for c in self.cache:
+                if c[0] == i:
+                    return c[1]
+        self.data_file.seek(self.data_offsets[i])
+        b = self.data_file.read(self.data_offsets[i + 1] - self.data_offsets[i])
+        item = pickle.loads(b)
+        if self.num_cache > 0:
+            self.cache = [(i, deepcopy(item))] + self.cache[:-1]
+        return item
+
+    def __len__(self):
+        return len(self.data_offsets) - 1
+
+class IndexedDatasetBuilder:
+    def __init__(self, path):
+        self.path = path
+        self.out_file = open(f"{path}.data", 'wb')
+        self.byte_offsets = [0]
+
+    def add_item(self, item):
+        s = pickle.dumps(item)
+        bytes = self.out_file.write(s)
+        self.byte_offsets.append(self.byte_offsets[-1] + bytes)
+
+    def finalize(self):
+        self.out_file.close()
+        np.save(open(f"{self.path}.idx", 'wb'), {'offsets': self.byte_offsets})
+
+
+if __name__ == "__main__":
+    import random
+    from tqdm import tqdm
+    ds_path = '/tmp/indexed_ds_example'
+    size = 100
+    items = [{"a": np.random.normal(size=[10000, 10]),
+              "b": np.random.normal(size=[10000, 10])} for i in range(size)]
+    builder = IndexedDatasetBuilder(ds_path)
+    for i in tqdm(range(size)):
+        builder.add_item(items[i])
+    builder.finalize()
+    ds = IndexedDataset(ds_path)
+    for i in tqdm(range(10000)):
+        idx = random.randint(0, size - 1)
+        assert (ds[idx]['a'] == items[idx]['a']).all()

preprocess/tools/note_transcription/utils/commons/losses.py

@@ -0,0 +1,53 @@
+import torch
+import torch.nn.functional as F
+
+def sigmoid_focal_loss(
+    inputs: torch.Tensor,
+    targets: torch.Tensor,
+    alpha: float = 0.25,
+    gamma: float = 2,
+    reduction: str = "none",
+) -> torch.Tensor:
+    """
+    Loss used in RetinaNet for dense detection: https://arxiv.org/abs/1708.02002.
+
+    Args:
+        inputs (Tensor): A float tensor of arbitrary shape.
+                The predictions for each example.
+        targets (Tensor): A float tensor with the same shape as inputs. Stores the binary
+                classification label for each element in inputs
+                (0 for the negative class and 1 for the positive class).
+        alpha (float): Weighting factor in range (0,1) to balance
+                positive vs negative examples or -1 for ignore. Default: ``0.25``.
+        gamma (float): Exponent of the modulating factor (1 - p_t) to
+                balance easy vs hard examples. Default: ``2``.
+        reduction (string): ``'none'`` | ``'mean'`` | ``'sum'``
+                ``'none'``: No reduction will be applied to the output.
+                ``'mean'``: The output will be averaged.
+                ``'sum'``: The output will be summed. Default: ``'none'``.
+    Returns:
+        Loss tensor with the reduction option applied.
+    """
+    # Original implementation from https://github.com/facebookresearch/fvcore/blob/master/fvcore/nn/focal_loss.py
+    p = torch.sigmoid(inputs)
+    ce_loss = F.binary_cross_entropy_with_logits(inputs, targets, reduction="none")
+    p_t = p * targets + (1 - p) * (1 - targets)
+    loss = ce_loss * ((1 - p_t) ** gamma)
+
+    if alpha >= 0:  # decrease the importance of negative samples
+        alpha_t = alpha * targets + (1 - alpha) * (1 - targets)
+        loss = alpha_t * loss
+
+    # Check reduction option and return loss accordingly
+    if reduction == "none":
+        pass
+    elif reduction == "mean":
+        loss = loss.mean()
+    elif reduction == "sum":
+        loss = loss.sum()
+    else:
+        raise ValueError(
+            f"Invalid Value for arg 'reduction': '{reduction} \n Supported reduction modes: 'none', 'mean', 'sum'"
+        )
+    return loss
+

preprocess/tools/note_transcription/utils/commons/meters.py

@@ -0,0 +1,42 @@
+import time
+import torch
+
+
+class AvgrageMeter(object):
+
+    def __init__(self):
+        self.reset()
+
+    def reset(self):
+        self.avg = 0
+        self.sum = 0
+        self.cnt = 0
+
+    def update(self, val, n=1):
+        self.sum += val * n
+        self.cnt += n
+        self.avg = self.sum / self.cnt
+
+
+class Timer:
+    timer_map = {}
+
+    def __init__(self, name, enable=False):
+        if name not in Timer.timer_map:
+            Timer.timer_map[name] = 0
+        self.name = name
+        self.enable = enable
+
+    def __enter__(self):
+        if self.enable:
+            if torch.cuda.is_available():
+                torch.cuda.synchronize()
+            self.t = time.time()
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if self.enable:
+            if torch.cuda.is_available():
+                torch.cuda.synchronize()
+            Timer.timer_map[self.name] += time.time() - self.t
+            if self.enable:
+                print(f'[Timer] {self.name}: {Timer.timer_map[self.name]}')

preprocess/tools/note_transcription/utils/commons/multiprocess_utils.py

@@ -0,0 +1,180 @@
+import os
+import traceback
+from functools import partial
+from tqdm import tqdm
+import torch
+
+def chunked_worker(worker_id, args_queue=None, results_queue=None, init_ctx_func=None):
+    ctx = init_ctx_func(worker_id) if init_ctx_func is not None else None
+    while True:
+        args = args_queue.get()
+        if args == '<KILL>':
+            return
+        job_idx, map_func, arg = args
+        try:
+            map_func_ = partial(map_func, ctx=ctx) if ctx is not None else map_func
+            if isinstance(arg, dict):
+                res = map_func_(**arg)
+            elif isinstance(arg, (list, tuple)):
+                res = map_func_(*arg)
+            else:
+                res = map_func_(arg)
+            results_queue.put((job_idx, res))
+        except:
+            traceback.print_exc()
+            results_queue.put((job_idx, None))
+
+
+class MultiprocessManager:
+    def __init__(self, num_workers=None, init_ctx_func=None, multithread=False, queue_max=-1):
+        if multithread:
+            from multiprocessing.dummy import Queue, Process
+        else:
+            from multiprocessing import Queue, Process
+        if num_workers is None:
+            num_workers = int(os.getenv('N_PROC', os.cpu_count()))
+        self.num_workers = num_workers
+        self.results_queue = Queue(maxsize=-1)
+        self.jobs_pending = []
+        self.args_queue = Queue(maxsize=queue_max)
+        self.workers = []
+        self.total_jobs = 0
+        self.multithread = multithread
+        for i in range(num_workers):
+            if multithread:
+                p = Process(target=chunked_worker,
+                            args=(i, self.args_queue, self.results_queue, init_ctx_func))
+            else:
+                p = Process(target=chunked_worker,
+                            args=(i, self.args_queue, self.results_queue, init_ctx_func),
+                            daemon=True)
+            self.workers.append(p)
+            p.start()
+
+    def add_job(self, func, args):
+        if not self.args_queue.full():
+            self.args_queue.put((self.total_jobs, func, args))
+        else:
+            self.jobs_pending.append((self.total_jobs, func, args))
+        self.total_jobs += 1
+
+    def get_results(self):
+        self.n_finished = 0
+        while self.n_finished < self.total_jobs:
+            while len(self.jobs_pending) > 0 and not self.args_queue.full():
+                self.args_queue.put(self.jobs_pending[0])
+                self.jobs_pending = self.jobs_pending[1:]
+            job_id, res = self.results_queue.get()
+            yield job_id, res
+            self.n_finished += 1
+        for w in range(self.num_workers):
+            self.args_queue.put("<KILL>")
+        for w in self.workers:
+            w.join()
+
+    def close(self):
+        if not self.multithread:
+            for w in self.workers:
+                w.terminate()
+
+    def __len__(self):
+        return self.total_jobs
+
+
+def multiprocess_run_tqdm(map_func, args, num_workers=None, ordered=True, init_ctx_func=None,
+                          multithread=False, queue_max=-1, desc=None):
+    for i, res in tqdm(
+            multiprocess_run(map_func, args, num_workers, ordered, init_ctx_func, multithread,
+                             queue_max=queue_max),
+            total=len(args), desc=desc):
+        yield i, res
+
+
+def multiprocess_run(map_func, args, num_workers=None, ordered=True, init_ctx_func=None, multithread=False,
+                     queue_max=-1):
+    """
+    Multiprocessing running chunked jobs.
+
+    Examples:
+    >>> for res in tqdm(multiprocess_run(job_func, args):
+    >>>     print(res)
+
+    :param map_func:
+    :param args:
+    :param num_workers:
+    :param ordered:
+    :param init_ctx_func:
+    :param q_max_size:
+    :param multithread:
+    :return:
+    """
+    if num_workers is None:
+        num_workers = int(os.getenv('N_PROC', os.cpu_count()))
+    manager = MultiprocessManager(num_workers, init_ctx_func, multithread, queue_max=queue_max)
+    for arg in args:
+        manager.add_job(map_func, arg)
+    if ordered:
+        n_jobs = len(args)
+        results = ['<WAIT>' for _ in range(n_jobs)]
+        i_now = 0
+        for job_i, res in manager.get_results():
+            results[job_i] = res
+            while i_now < n_jobs and (not isinstance(results[i_now], str) or results[i_now] != '<WAIT>'):
+                yield i_now, results[i_now]
+                results[i_now] = None
+                i_now += 1
+    else:
+        for job_i, res in manager.get_results():
+            yield job_i, res
+    manager.close()
+
+# #### this is the old version of chunked_multiprocess_run
+def chunked_worker_old(worker_id, map_func, args, results_queue=None, init_ctx_func=None):
+    ctx = init_ctx_func(worker_id) if init_ctx_func is not None else None
+    for job_idx, arg in args:
+        try:
+            if not isinstance(arg, tuple) and not isinstance(arg, list):
+                arg = [arg]
+            if ctx is not None:
+                res = map_func(*arg, ctx=ctx)
+            else:
+                res = map_func(*arg)
+            results_queue.put((job_idx, res))
+        except:
+            traceback.print_exc()
+            results_queue.put((job_idx, None))
+
+def chunked_multiprocess_run(
+        map_func, args, num_workers=None, ordered=True,
+        init_ctx_func=None, q_max_size=1000, multithread=False):
+    if multithread:
+        from multiprocessing.dummy import Queue, Process
+    else:
+        from multiprocessing import Queue, Process
+    args = zip(range(len(args)), args)
+    args = list(args)
+    n_jobs = len(args)
+    if num_workers is None:
+        num_workers = int(os.getenv('N_PROC', os.cpu_count()))
+    results_queues = []
+    if ordered:
+        for i in range(num_workers):
+            results_queues.append(Queue(maxsize=q_max_size // num_workers))
+    else:
+        results_queue = Queue(maxsize=q_max_size)
+        for i in range(num_workers):
+            results_queues.append(results_queue)
+    workers = []
+    for i in range(num_workers):
+        args_worker = args[i::num_workers]
+        p = Process(target=chunked_worker_old, args=(
+            i, map_func, args_worker, results_queues[i], init_ctx_func), daemon=True)
+        workers.append(p)
+        p.start()
+    for n_finished in range(n_jobs):
+        results_queue = results_queues[n_finished % num_workers]
+        job_idx, res = results_queue.get()
+        assert job_idx == n_finished or not ordered, (job_idx, n_finished)
+        yield res
+    for w in workers:
+        w.join()

preprocess/tools/note_transcription/utils/commons/signal.py

@@ -0,0 +1,74 @@
+import numpy as np
+import torch
+import torch.nn as nn
+
+def get_filter_2d(kernel, kernel_size, channels, no_grad=True):
+    # Reshape to 2d depthwise convolutional weight
+    kernel = kernel.view(1, 1, kernel_size, kernel_size)
+    kernel = kernel.repeat(channels, 1, 1, 1)
+
+    filter = nn.Conv2d(in_channels=channels, out_channels=channels, kernel_size=kernel_size, groups=channels,
+                       bias=False, padding=kernel_size // 2)
+
+    filter.weight.data = kernel
+    if no_grad:
+        filter.weight.requires_grad = False
+
+    return filter
+
+def get_filter_1d(kernel, kernel_size, channels, no_grad=True):
+    kernel = kernel.view(1, 1, kernel_size)
+    kernel = kernel.repeat(channels, 1, 1)
+
+    filter = nn.Conv1d(in_channels=channels, out_channels=channels, kernel_size=kernel_size, groups=channels,
+                       bias=False, padding=kernel_size // 2)
+
+    filter.weight.data = kernel
+    if no_grad:
+        filter.weight.requires_grad = False
+
+    return filter
+
+def get_gaussian_kernel_2d(kernel_size, sigma):
+    # Create a x, y coordinate grid of shape (kernel_size, kernel_size, 2)
+    x_coord = torch.arange(kernel_size)
+    x_grid = x_coord.repeat(kernel_size).view(kernel_size, kernel_size)
+    y_grid = x_grid.t()
+    xy_grid = torch.stack([x_grid, y_grid], dim=-1).float()
+
+    mean = (kernel_size - 1) / 2.
+    variance = sigma ** 2.
+
+    # Calculate the 2-dimensional gaussian kernel which is
+    # the product of two gaussian distributions for two different
+    # variables (in this case called x and y)
+    gaussian_kernel = (1. / (2. * np.pi * variance)) * torch.exp(
+        -torch.sum((xy_grid - mean) ** 2., dim=-1) / (2 * variance))
+
+    # Make sure sum of values in gaussian kernel equals 1.
+    gaussian_kernel = gaussian_kernel / torch.sum(gaussian_kernel)
+
+    return gaussian_kernel
+
+def get_gaussian_kernel_1d(kernel_size, sigma):
+    x_grid = torch.arange(kernel_size)
+    mean = (kernel_size - 1) / 2.
+    variance = sigma ** 2.
+    gaussian_kernel = (1. / ((2. * np.pi) ** 0.5 * sigma)) * torch.exp(-(x_grid - mean) ** 2. / (2 * variance))
+    gaussian_kernel = gaussian_kernel / torch.sum(gaussian_kernel)
+    return gaussian_kernel
+
+def get_hann_kernel_1d(kernel_size, periodic=False):
+    # periodic=False gives symmetric kernel, otherwise equivalent to hann(kernel_size + 1)
+    return torch.hann_window(kernel_size, periodic)
+
+def get_triangle_kernel_1d(kernel_size):
+    kernel = torch.zeros(kernel_size)
+    for idx in range(kernel_size):
+        kernel[idx] = 1 - abs((idx - (kernel_size - 1) / 2) / ((kernel_size - 1) / 2))
+    return kernel
+
+def add_gaussian_noise(tensor, mean=0, std=1):
+    noise = torch.randn(tensor.size()) * std + mean
+    noisy_tensor = tensor + noise
+    return noisy_tensor

preprocess/tools/note_transcription/utils/commons/single_thread_env.py

@@ -0,0 +1,5 @@
+import os
+
+os.environ["OMP_NUM_THREADS"] = "1"
+os.environ['TF_NUM_INTEROP_THREADS'] = '1'
+os.environ['TF_NUM_INTRAOP_THREADS'] = '1'

preprocess/tools/note_transcription/utils/commons/tensor_utils.py

@@ -0,0 +1,92 @@
+import torch
+import torch.distributed as dist
+
+
+def reduce_tensors(metrics):
+    new_metrics = {}
+    for k, v in metrics.items():
+        if isinstance(v, torch.Tensor):
+            dist.all_reduce(v)
+            v = v / dist.get_world_size()
+        if type(v) is dict:
+            v = reduce_tensors(v)
+        new_metrics[k] = v
+    return new_metrics
+
+
+def tensors_to_scalars(tensors):
+    if isinstance(tensors, torch.Tensor):
+        tensors = tensors.item()
+        return tensors
+    elif isinstance(tensors, dict):
+        new_tensors = {}
+        for k, v in tensors.items():
+            v = tensors_to_scalars(v)
+            new_tensors[k] = v
+        return new_tensors
+    elif isinstance(tensors, list):
+        return [tensors_to_scalars(v) for v in tensors]
+    else:
+        return tensors
+
+
+def tensors_to_np(tensors):
+    if isinstance(tensors, dict):
+        new_np = {}
+        for k, v in tensors.items():
+            if isinstance(v, torch.Tensor):
+                v = v.cpu().numpy()
+            if type(v) is dict:
+                v = tensors_to_np(v)
+            new_np[k] = v
+    elif isinstance(tensors, list):
+        new_np = []
+        for v in tensors:
+            if isinstance(v, torch.Tensor):
+                v = v.cpu().numpy()
+            if type(v) is dict:
+                v = tensors_to_np(v)
+            new_np.append(v)
+    elif isinstance(tensors, torch.Tensor):
+        v = tensors
+        if isinstance(v, torch.Tensor):
+            v = v.cpu().numpy()
+        if type(v) is dict:
+            v = tensors_to_np(v)
+        new_np = v
+    else:
+        raise Exception(f'tensors_to_np does not support type {type(tensors)}.')
+    return new_np
+
+
+def move_to_cpu(tensors):
+    ret = {}
+    for k, v in tensors.items():
+        if isinstance(v, torch.Tensor):
+            v = v.cpu()
+        if type(v) is dict:
+            v = move_to_cpu(v)
+        ret[k] = v
+    return ret
+
+
+def move_to_cuda(batch, gpu_id=0):
+    # base case: object can be directly moved using `cuda` or `to`
+    if callable(getattr(batch, 'cuda', None)):
+        return batch.cuda(gpu_id, non_blocking=True)
+    elif callable(getattr(batch, 'to', None)):
+        return batch.to(torch.device('cuda', gpu_id), non_blocking=True)
+    elif isinstance(batch, list):
+        for i, x in enumerate(batch):
+            batch[i] = move_to_cuda(x, gpu_id)
+        return batch
+    elif isinstance(batch, tuple):
+        batch = list(batch)
+        for i, x in enumerate(batch):
+            batch[i] = move_to_cuda(x, gpu_id)
+        return tuple(batch)
+    elif isinstance(batch, dict):
+        for k, v in batch.items():
+            batch[k] = move_to_cuda(v, gpu_id)
+        return batch
+    return batch

preprocess/tools/note_transcription/utils/commons/trainer.py

@@ -0,0 +1,557 @@
+import random
+import subprocess
+import traceback
+from datetime import datetime
+
+from torch.cuda.amp import GradScaler, autocast
+import numpy as np
+import torch.optim
+import torch.utils.data
+import copy
+import logging
+import os
+import re
+import sys
+import torch
+import torch.distributed as dist
+import torch.multiprocessing as mp
+import tqdm
+
+from .ckpt_utils import get_last_checkpoint, get_all_ckpts
+from .ddp_utils import DDP
+from .hparams import hparams
+from .tensor_utils import move_to_cuda
+
+
+class Tee(object):
+    def __init__(self, name, mode):
+        self.file = open(name, mode)
+        self.stdout = sys.stdout
+        sys.stdout = self
+
+    def __del__(self):
+        sys.stdout = self.stdout
+        self.file.close()
+
+    def write(self, data):
+        self.file.write(data)
+        self.stdout.write(data)
+
+    def flush(self):
+        self.file.flush()
+
+
+class Trainer:
+    def __init__(
+            self,
+            work_dir,
+            default_save_path=None,
+            accumulate_grad_batches=1,
+            max_updates=160000,
+            print_nan_grads=False,
+            val_check_interval=2000,
+            num_sanity_val_steps=5,
+            amp=False,
+            # tb logger
+            log_save_interval=100,
+            tb_log_interval=10,
+            # checkpoint
+            monitor_key='val_loss',
+            monitor_mode='min',
+            num_ckpt_keep=5,
+            save_best=True,
+            resume_from_checkpoint=0,
+            seed=1234,
+            debug=False,
+    ):
+        os.makedirs(work_dir, exist_ok=True)
+        self.work_dir = work_dir
+        self.accumulate_grad_batches = accumulate_grad_batches
+        self.max_updates = max_updates
+        self.num_sanity_val_steps = num_sanity_val_steps
+        self.print_nan_grads = print_nan_grads
+        self.default_save_path = default_save_path
+        self.resume_from_checkpoint = resume_from_checkpoint if resume_from_checkpoint > 0 else None
+        self.seed = seed
+        self.debug = debug
+        # model and optm
+        self.task = None
+        self.optimizers = []
+
+        # trainer state
+        self.testing = False
+        self.global_step = 0
+        self.current_epoch = 0
+        self.total_batches = 0
+
+        # configure checkpoint
+        self.monitor_key = monitor_key
+        self.num_ckpt_keep = num_ckpt_keep
+        self.save_best = save_best
+        self.monitor_op = np.less if monitor_mode == 'min' else np.greater
+        self.best_val_results = np.Inf if monitor_mode == 'min' else -np.Inf
+        self.mode = 'min'
+
+        # allow int, string and gpu list
+        self.all_gpu_ids = [
+            int(x) for x in os.environ.get("CUDA_VISIBLE_DEVICES", "").split(",") if x != '']
+        self.num_gpus = len(self.all_gpu_ids)
+        self.on_gpu = self.num_gpus > 0
+        self.root_gpu = 0
+        logging.info(f'GPU available: {torch.cuda.is_available()}, GPU used: {self.all_gpu_ids}')
+        self.use_ddp = self.num_gpus > 1
+        self.proc_rank = 0
+        # Tensorboard logging
+        self.log_save_interval = log_save_interval
+        self.val_check_interval = val_check_interval
+        self.tb_log_interval = tb_log_interval
+        self.amp = amp
+        self.amp_scalar = GradScaler()
+
+    def test(self, task_cls):
+        self.testing = True
+        self.fit(task_cls)
+
+    def fit(self, task_cls):
+        if len(self.all_gpu_ids) > 1:
+            mp.spawn(self.ddp_run, nprocs=self.num_gpus, args=(task_cls, copy.deepcopy(hparams)))
+        else:
+            self.task = task_cls()
+            self.task.trainer = self
+            self.run_single_process(self.task)
+        return 1
+
+    def ddp_run(self, gpu_idx, task_cls, hparams_):
+        hparams.update(hparams_)
+        self.proc_rank = gpu_idx
+        self.init_ddp_connection(self.proc_rank, self.num_gpus)
+        if dist.get_rank() != 0 and not self.debug:
+            sys.stdout = open(os.devnull, "w")
+            sys.stderr = open(os.devnull, "w")
+        task = task_cls()
+        task.trainer = self
+        torch.cuda.set_device(gpu_idx)
+        self.root_gpu = gpu_idx
+        self.task = task
+        self.run_single_process(task)
+
+    def run_single_process(self, task):
+        """Sanity check a few things before starting actual training.
+
+        :param task:
+        """
+        # build model, optm and load checkpoint
+        if self.proc_rank == 0:
+            self.save_terminal_logs()
+            if not self.testing:
+                self.save_codes()
+
+        model = task.build_model()
+        if model is not None:
+            task.model = model
+        checkpoint, _ = get_last_checkpoint(self.work_dir, self.resume_from_checkpoint)
+        if checkpoint is not None:
+            self.restore_weights(checkpoint)
+        elif self.on_gpu:
+            task.cuda(self.root_gpu)
+        if not self.testing:
+            self.optimizers = task.configure_optimizers()
+            self.fisrt_epoch = True
+        if checkpoint is not None:
+            self.restore_opt_state(checkpoint)
+        del checkpoint
+        # clear cache after restore
+        if self.on_gpu:
+            torch.cuda.empty_cache()
+
+        if self.use_ddp:
+            self.task = self.configure_ddp(self.task)
+            dist.barrier()
+
+        task_ref = self.get_task_ref()
+        task_ref.trainer = self
+        task_ref.testing = self.testing
+        # link up experiment object
+        if self.proc_rank == 0:
+            task_ref.build_tensorboard(save_dir=self.work_dir, name='tb_logs')
+        else:
+            os.makedirs('tmp', exist_ok=True)
+            task_ref.build_tensorboard(save_dir='tmp', name='tb_tmp')
+        self.logger = task_ref.logger
+        try:
+            if self.testing:
+                self.run_evaluation(test=True)
+            else:
+                self.train()
+        except KeyboardInterrupt as e:
+            traceback.print_exc()
+            task_ref.on_keyboard_interrupt()
+
+    ####################
+    # valid and test
+    ####################
+    def run_evaluation(self, test=False):
+        eval_results = self.evaluate(self.task, test, tqdm_desc='Valid' if not test else 'test',
+                                     max_batches=hparams['eval_max_batches'])
+        if eval_results is not None and 'tb_log' in eval_results:
+            tb_log_output = eval_results['tb_log']
+            self.log_metrics_to_tb(tb_log_output)
+        if self.proc_rank == 0 and not test:
+            self.save_checkpoint(epoch=self.current_epoch, logs=eval_results)
+
+    def evaluate(self, task, test=False, tqdm_desc='Valid', max_batches=None):
+        if max_batches == -1:
+            max_batches = None
+        # enable eval mode
+        task.zero_grad()
+        task.eval()
+        torch.set_grad_enabled(False)
+
+        task_ref = self.get_task_ref()
+        if test:
+            ret = task_ref.test_start()
+            if ret == 'EXIT':
+                return
+        else:
+            task_ref.validation_start()
+        outputs = []
+        dataloader = task_ref.test_dataloader() if test else task_ref.val_dataloader()
+        pbar = tqdm.tqdm(dataloader, desc=tqdm_desc, total=max_batches, dynamic_ncols=True, unit='step',
+                         disable=self.root_gpu > 0)
+        # give model a chance to do something with the outputs (and method defined)
+        for batch_idx, batch in enumerate(pbar):
+            if batch is None:  # pragma: no cover
+                continue
+            # stop short when on fast_dev_run (sets max_batch=1)
+            if max_batches is not None and batch_idx >= max_batches:
+                break
+
+            # make dataloader_idx arg in validation_step optional
+            if self.on_gpu:
+                batch = move_to_cuda(batch, self.root_gpu)
+            args = [batch, batch_idx]
+            if self.use_ddp:
+                output = task(*args)
+            else:
+                if test:
+                    output = task_ref.test_step(*args)
+                else:
+                    output = task_ref.validation_step(*args)
+            # track outputs for collation
+            outputs.append(output)
+        # give model a chance to do something with the outputs (and method defined)
+        if test:
+            eval_results = task_ref.test_end(outputs)
+        else:
+            eval_results = task_ref.validation_end(outputs)
+        # enable train mode again
+        task.train()
+        torch.set_grad_enabled(True)
+        return eval_results
+
+    ####################
+    # train
+    ####################
+    def train(self):
+        task_ref = self.get_task_ref()
+        task_ref.on_train_start()
+        if self.num_sanity_val_steps > 0:
+            # run tiny validation (if validation defined) to make sure program won't crash during val
+            self.evaluate(self.task, False, 'Sanity Val', max_batches=self.num_sanity_val_steps)
+        # clear cache before training
+        if self.on_gpu:
+            torch.cuda.empty_cache()
+        dataloader = task_ref.train_dataloader()
+        epoch = self.current_epoch
+        # run all epochs
+        while True:
+            # set seed for distributed sampler (enables shuffling for each epoch)
+            if self.use_ddp and hasattr(dataloader.sampler, 'set_epoch'):
+                dataloader.sampler.set_epoch(epoch)
+            # update training progress in trainer and model
+            task_ref.current_epoch = epoch
+            self.current_epoch = epoch
+            # total batches includes multiple val checks
+            self.batch_loss_value = 0  # accumulated grads
+            # before epoch hook
+            task_ref.on_epoch_start()
+
+            # run epoch
+            train_pbar = tqdm.tqdm(dataloader, initial=self.global_step, total=float('inf'),
+                                   dynamic_ncols=True, unit='step', disable=self.root_gpu > 0)
+            for batch_idx, batch in enumerate(train_pbar):
+                if self.global_step % self.val_check_interval == 0 and not self.fisrt_epoch:
+                    self.run_evaluation()
+                pbar_metrics, tb_metrics = self.run_training_batch(batch_idx, batch)
+                train_pbar.set_postfix(**pbar_metrics)
+                self.fisrt_epoch = False
+                # when metrics should be logged
+                if (self.global_step + 1) % self.tb_log_interval == 0:
+                    # logs user requested information to logger
+                    self.log_metrics_to_tb(tb_metrics)
+
+                self.global_step += 1
+                task_ref.global_step = self.global_step
+                if self.global_step > self.max_updates:
+                    print("| Training end..")
+                    break
+            # epoch end hook
+            task_ref.on_epoch_end()
+            epoch += 1
+            if self.global_step > self.max_updates:
+                break
+        task_ref.on_train_end()
+
+    def run_training_batch(self, batch_idx, batch):
+        if batch is None:
+            return {}
+        all_progress_bar_metrics = []
+        all_log_metrics = []
+        task_ref = self.get_task_ref()
+        for opt_idx, optimizer in enumerate(self.optimizers):
+            if optimizer is None:
+                continue
+            # make sure only the gradients of the current optimizer's paramaters are calculated
+            # in the training step to prevent dangling gradients in multiple-optimizer setup.
+            if len(self.optimizers) > 1:
+                for param in task_ref.parameters():
+                    param.requires_grad = False
+                for group in optimizer.param_groups:
+                    for param in group['params']:
+                        param.requires_grad = True
+
+            # forward pass
+            with autocast(enabled=self.amp):
+                if self.on_gpu:
+                    batch = move_to_cuda(copy.copy(batch), self.root_gpu)
+                args = [batch, batch_idx, opt_idx]
+                if self.use_ddp:
+                    output = self.task(*args)
+                else:
+                    output = task_ref.training_step(*args)
+                loss = output['loss']
+                if loss is None:
+                    continue
+                progress_bar_metrics = output['progress_bar']
+                log_metrics = output['tb_log']
+                # accumulate loss
+                loss = loss / self.accumulate_grad_batches
+
+            # backward pass
+            if loss.requires_grad:
+                if self.amp:
+                    self.amp_scalar.scale(loss).backward()
+                else:
+                    loss.backward()
+
+            # track progress bar metrics
+            all_log_metrics.append(log_metrics)
+            all_progress_bar_metrics.append(progress_bar_metrics)
+
+            if loss is None:
+                continue
+
+            # nan grads
+            if self.print_nan_grads:
+                has_nan_grad = False
+                for name, param in task_ref.named_parameters():
+                    if (param.grad is not None) and torch.isnan(param.grad.float()).any():
+                        print("| NaN params: ", name, param, param.grad)
+                        has_nan_grad = True
+                if has_nan_grad:
+                    exit(0)
+
+            # gradient update with accumulated gradients
+            if (self.global_step + 1) % self.accumulate_grad_batches == 0:
+                task_ref.on_before_optimization(opt_idx)
+                if self.amp:
+                    self.amp_scalar.step(optimizer)
+                    self.amp_scalar.update()
+                else:
+                    optimizer.step()
+                optimizer.zero_grad()
+                task_ref.on_after_optimization(self.current_epoch, batch_idx, optimizer, opt_idx)
+
+        # collapse all metrics into one dict
+        all_progress_bar_metrics = {k: v for d in all_progress_bar_metrics for k, v in d.items()}
+        all_log_metrics = {k: v for d in all_log_metrics for k, v in d.items()}
+        return all_progress_bar_metrics, all_log_metrics
+
+    ####################
+    # load and save checkpoint
+    ####################
+    def restore_weights(self, checkpoint):
+        # load model state
+        task_ref = self.get_task_ref()
+
+        for k, v in checkpoint['state_dict'].items():
+            getattr(task_ref, k).load_state_dict(v)
+
+        if self.on_gpu:
+            task_ref.cuda(self.root_gpu)
+        # load training state (affects trainer only)
+        self.best_val_results = checkpoint['checkpoint_callback_best']
+        self.global_step = checkpoint['global_step']
+        self.current_epoch = checkpoint['epoch']
+        task_ref.global_step = self.global_step
+
+        # wait for all models to restore weights
+        if self.use_ddp:
+            # wait for all processes to catch up
+            dist.barrier()
+
+    def restore_opt_state(self, checkpoint):
+        if self.testing:
+            return
+        # restore the optimizers
+        optimizer_states = checkpoint['optimizer_states']
+        for optimizer, opt_state in zip(self.optimizers, optimizer_states):
+            if optimizer is None:
+                return
+            try:
+                optimizer.load_state_dict(opt_state)
+                # move optimizer to GPU 1 weight at a time
+                if self.on_gpu:
+                    for state in optimizer.state.values():
+                        for k, v in state.items():
+                            if isinstance(v, torch.Tensor):
+                                state[k] = v.cuda(self.root_gpu)
+            except ValueError:
+                print("| WARMING: optimizer parameters not match !!!")
+        try:
+            if dist.is_initialized() and dist.get_rank() > 0:
+                return
+        except Exception as e:
+            print(e)
+            return
+        did_restore = True
+        return did_restore
+
+    def save_checkpoint(self, epoch, logs=None):
+        monitor_op = np.less
+        ckpt_path = f'{self.work_dir}/model_ckpt_steps_{self.global_step}.ckpt'
+        logging.info(f'Epoch {epoch:05d}@{self.global_step}: saving model to {ckpt_path}')
+        self._atomic_save(ckpt_path)
+        for old_ckpt in get_all_ckpts(self.work_dir)[self.num_ckpt_keep:]:
+            pass
+        current = None
+        if logs is not None and self.monitor_key in logs:
+            current = logs[self.monitor_key]
+        if current is not None and self.save_best:
+            if monitor_op(current, self.best_val_results):
+                best_filepath = f'{self.work_dir}/model_ckpt_best.pt'
+                self.best_val_results = current
+                logging.info(
+                    f'Epoch {epoch:05d}@{self.global_step}: {self.monitor_key} reached {current:0.5f}. '
+                    f'Saving model to {best_filepath}')
+                self._atomic_save(best_filepath)
+
+    def _atomic_save(self, filepath):
+        checkpoint = self.dump_checkpoint()
+        tmp_path = str(filepath) + ".part"
+        torch.save(checkpoint, tmp_path, _use_new_zipfile_serialization=False)
+        os.replace(tmp_path, filepath)
+
+    def dump_checkpoint(self):
+        checkpoint = {'epoch': self.current_epoch, 'global_step': self.global_step,
+                      'checkpoint_callback_best': self.best_val_results}
+        # save optimizers
+        optimizer_states = []
+        for i, optimizer in enumerate(self.optimizers):
+            if optimizer is not None:
+                optimizer_states.append(optimizer.state_dict())
+
+        checkpoint['optimizer_states'] = optimizer_states
+        task_ref = self.get_task_ref()
+        checkpoint['state_dict'] = {
+            k: v.state_dict() for k, v in task_ref.named_children() if len(list(v.parameters())) > 0}
+        return checkpoint
+
+    ####################
+    # DDP
+    ####################
+    def configure_ddp(self, task):
+        task = DDP(task, device_ids=[self.root_gpu], find_unused_parameters=hparams.get('find_unused_parameters', True))
+        random.seed(self.seed)
+        np.random.seed(self.seed)
+        return task
+
+    def init_ddp_connection(self, proc_rank, world_size):
+        root_node = '127.0.0.1'
+        root_node = self.resolve_root_node_address(root_node)
+        os.environ['MASTER_ADDR'] = root_node
+        dist.init_process_group('nccl', rank=proc_rank, world_size=world_size)
+
+    def resolve_root_node_address(self, root_node):
+        if '[' in root_node:
+            name = root_node.split('[')[0]
+            number = root_node.split(',')[0]
+            if '-' in number:
+                number = number.split('-')[0]
+            number = re.sub('[^0-9]', '', number)
+            root_node = name + number
+        return root_node
+
+    ####################
+    # utils
+    ####################
+    def get_task_ref(self):
+        from .base_task import BaseTask
+        task: BaseTask = self.task.module if isinstance(self.task, DDP) else self.task
+        return task
+
+    def log_metrics_to_tb(self, metrics, step=None):
+        """Logs the metric dict passed in.
+
+        :param metrics:
+        """
+        # turn all tensors to scalars
+        scalar_metrics = self.metrics_to_scalars(metrics)
+
+        step = step if step is not None else self.global_step
+        # log actual metrics
+        if self.proc_rank == 0:
+            self.log_metrics(self.logger, scalar_metrics, step=step)
+
+    @staticmethod
+    def log_metrics(logger, metrics, step=None):
+        for k, v in metrics.items():
+            if isinstance(v, torch.Tensor):
+                v = v.item()
+            logger.add_scalar(k, v, step)
+
+    def metrics_to_scalars(self, metrics):
+        new_metrics = {}
+        for k, v in metrics.items():
+            if isinstance(v, torch.Tensor):
+                v = v.item()
+
+            if type(v) is dict:
+                v = self.metrics_to_scalars(v)
+
+            new_metrics[k] = v
+
+        return new_metrics
+
+    def save_terminal_logs(self):
+        t = datetime.now().strftime('%Y%m%d%H%M%S')
+        os.makedirs(f'{self.work_dir}/terminal_logs', exist_ok=True)
+        Tee(f'{self.work_dir}/terminal_logs/log_{t}.txt', 'w')
+
+    def save_codes(self):
+        if len(hparams['save_codes']) > 0:
+            t = datetime.now().strftime('%Y%m%d%H%M%S')
+            code_dir = f'{self.work_dir}/codes/{t}'
+            subprocess.check_call(f'mkdir -p "{code_dir}"', shell=True)
+            for c in hparams['save_codes']:
+                if os.path.exists(c):
+                    subprocess.check_call(
+                        f'rsync -aR '
+                        f'--include="*.py" '
+                        f'--include="*.yaml" '
+                        f'--exclude="__pycache__" '
+                        f'--include="*/" '
+                        f'--exclude="*" '
+                        f'"./{c}" "{code_dir}/"',
+                        shell=True)
+            print(f"| Copied codes to {code_dir}.")