"""
The setting of Superb IC
Authors
* Wei-Cheng Tseng 2021
* Leo 2021
* Leo 2022
"""
import logging
import pickle
from pathlib import Path
import pandas as pd
import torch
from omegaconf import MISSING
from torch.utils.data import Dataset
from s3prl.dataio.corpus.fluent_speech_commands import FluentSpeechCommands
from s3prl.dataio.dataset import EncodeCategories, LoadAudio
from s3prl.dataio.encoder.category import CategoryEncoders
from s3prl.dataio.sampler import FixedBatchSizeBatchSampler
from s3prl.nn.linear import MeanPoolingLinear
from s3prl.task.utterance_classification_task import (
UtteranceMultiClassClassificationTask,
)
from .run import Common
logger = logging.getLogger(__name__)
__all__ = [
"fsc_for_multi_classification",
"SuperbIC",
]
[docs]def fsc_for_multi_classification(
target_dir: str,
cache_dir: str,
dataset_root: str,
n_jobs: int = 6,
get_path_only: bool = False,
):
"""
Prepare Fluent Speech Command for multi-class classfication
following :obj:`SuperbIC.prepare_data` format. The standard usage
is to use three labels jointly: action, object, and location.
Args:
dataset_root (str): The root path of Fluent Speech Command
n_jobs (int): to speed up the corpus parsing procedure
"""
target_dir = Path(target_dir)
train_path = target_dir / f"train.csv"
valid_path = target_dir / f"valid.csv"
test_paths = [target_dir / f"test.csv"]
if get_path_only:
return train_path, valid_path, test_paths
def format_fields(data_points: dict):
return {
key: dict(
wav_path=value["path"],
labels=f"{value['action']} ; {value['object']} ; {value['location']}",
)
for key, value in data_points.items()
}
corpus = FluentSpeechCommands(dataset_root, n_jobs)
train_data, valid_data, test_data = corpus.data_split
train_data = format_fields(train_data)
valid_data = format_fields(valid_data)
test_data = format_fields(test_data)
def dict_to_csv(data_dict, csv_path):
keys = sorted(list(data_dict.keys()))
fields = sorted(data_dict[keys[0]].keys())
data = dict()
for field in fields:
data[field] = []
for key in keys:
data[field].append(data_dict[key][field])
data["id"] = keys
df = pd.DataFrame(data)
df.to_csv(csv_path, index=False)
dict_to_csv(train_data, train_path)
dict_to_csv(valid_data, valid_path)
dict_to_csv(test_data, test_paths[0])
return train_path, valid_path, test_paths
[docs]class SuperbIC(Common):
[docs] def default_config(self) -> dict:
return dict(
start=0,
stop=None,
target_dir=MISSING,
cache_dir=None,
remove_all_cache=False,
prepare_data=dict(
dataset_root=MISSING,
),
build_encoder=dict(),
build_dataset=dict(),
build_batch_sampler=dict(
train=dict(
batch_size=32,
shuffle=True,
),
valid=dict(
batch_size=32,
),
test=dict(
batch_size=32,
),
),
build_upstream=dict(
name=MISSING,
),
build_featurizer=dict(
layer_selections=None,
normalize=False,
),
build_downstream=dict(
hidden_size=256,
),
build_model=dict(
upstream_trainable=False,
),
build_task=dict(),
build_optimizer=dict(
name="Adam",
conf=dict(
lr=1.0e-4,
),
),
build_scheduler=dict(
name="ExponentialLR",
gamma=0.9,
),
save_model=dict(),
save_task=dict(),
train=dict(
total_steps=200000,
log_step=100,
eval_step=5000,
save_step=250,
gradient_clipping=1.0,
gradient_accumulate=1,
valid_metric="accuracy",
valid_higher_better=True,
auto_resume=True,
resume_ckpt_dir=None,
),
)
[docs] def prepare_data(
self,
prepare_data: dict,
target_dir: str,
cache_dir: str,
get_path_only: bool = False,
):
"""
Prepare the task-specific data metadata (path, labels...).
By default call :obj:`fsc_for_multi_classification` with :code:`**prepare_data`
Args:
prepare_data (dict): same in :obj:`default_config`,
arguments for :obj:`fsc_for_multi_classification`
target_dir (str): Parse your corpus and save the csv file into this directory
cache_dir (str): If the parsing or preprocessing takes too long time, you can save
the temporary files into this directory. This directory is expected to be shared
across different training sessions (different hypers and :code:`target_dir`)
get_path_only (str): Directly return the filepaths no matter they exist or not.
Returns:
tuple
1. train_path (str)
2. valid_path (str)
3. test_paths (List[str])
Each path (str) should be a csv file containing the following columns:
==================== ====================
column description
==================== ====================
id (str) - the unique id for this data point
wav_path (str) - the absolute path of the waveform file
labels (str) - the string labels of the waveform, separated by a ';'
==================== ====================
The number of the label columns can be arbitrary.
"""
return fsc_for_multi_classification(
**self._get_current_arguments(flatten_dict="prepare_data")
)
[docs] def build_encoder(
self,
build_encoder: dict,
target_dir: str,
cache_dir: str,
train_csv_path: str,
valid_csv_path: str,
test_csv_paths: list,
get_path_only: bool = False,
):
"""
Build the encoder (for the labels) given the data metadata, and return the saved encoder path.
By default generate and save a :obj:`s3prl.dataio.encoder.CategoryEncoders` from all the columns
prefixing :code:`label` from all the csv files.
Args:
build_encoder (dict): same in :obj:`default_config`, no argument supported for now
target_dir (str): Save your encoder into this directory
cache_dir (str): If the preprocessing takes too long time, you can save
the temporary files into this directory. This directory is expected to be shared
across different training sessions (different hypers and :code:`target_dir`)
train_csv_path (str): the train path from :obj:`prepare_data`
valid_csv_path (str): the valid path from :obj:`prepare_data`
test_csv_paths (List[str]): the test paths from :obj:`prepare_data`
get_path_only (bool): Directly return the filepaths no matter they exist or not.
Returns:
str
tokenizer_path: The tokenizer should be saved in the pickle format
"""
encoder_path = Path(target_dir) / "encoder.pkl"
if get_path_only:
return encoder_path
train_csv = pd.read_csv(train_csv_path)
valid_csv = pd.read_csv(valid_csv_path)
test_csvs = [pd.read_csv(path) for path in test_csv_paths]
all_csv = pd.concat([train_csv, valid_csv, *test_csvs])
multilabels = [
[label.strip() for label in multilabel.split(";")]
for multilabel in all_csv["labels"].tolist()
]
encoder = CategoryEncoders(
[single_category_labels for single_category_labels in zip(*multilabels)]
)
with open(encoder_path, "wb") as f:
pickle.dump(encoder, f)
return encoder
[docs] def build_dataset(
self,
build_dataset: dict,
target_dir: str,
cache_dir: str,
mode: str,
data_csv: str,
encoder_path: str,
frame_shift: int,
):
"""
Build the dataset for train/valid/test.
Args:
build_dataset (dict): same in :obj:`default_config`, no argument supported for now
target_dir (str): Current experiment directory
cache_dir (str): If the preprocessing takes too long time, you can save
the temporary files into this directory. This directory is expected to be shared
across different training sessions (different hypers and :code:`target_dir`)
mode (str): train/valid/test
data_csv (str): The metadata csv file for the specific :code:`mode`
encoder_path (str): The pickled encoder path for encoding the labels
Returns:
torch Dataset
For all train/valid/test mode, the dataset should return each item as a dictionary
containing the following keys:
==================== ====================
key description
==================== ====================
x (torch.FloatTensor) - the waveform in (seq_len, 1)
x_len (int) - the waveform length :code:`seq_len`
class_ids (torch.LongTensor) - the encoded class ids. shape: (num_class, )
labels (List[str]) - the class name. length: num_class
unique_name (str) - the unique id for this datapoint
==================== ====================
"""
csv = pd.read_csv(data_csv)
ids = csv["id"].tolist()
audio_loader = LoadAudio(csv["wav_path"].tolist())
with open(encoder_path, "rb") as f:
encoder = pickle.load(f)
label_encoder = EncodeCategories(
[
[label.strip() for label in multilabel.split(";")]
for multilabel in csv["labels"].tolist()
],
encoder,
)
class Dataset:
def __len__(self):
return len(audio_loader)
def __getitem__(self, index: int):
audio = audio_loader[index]
label = label_encoder[index]
return {
"x": audio["wav"],
"x_len": audio["wav_len"],
"class_ids": label["class_ids"],
"labels": label["labels"],
"unique_name": ids[index],
}
dataset = Dataset()
return dataset
[docs] def build_batch_sampler(
self,
build_batch_sampler: dict,
target_dir: str,
cache_dir: str,
mode: str,
data_csv: str,
dataset: Dataset,
):
"""
Return the batch sampler for torch DataLoader.
By default call :obj:`superb_sid_batch_sampler` with :code:`**build_batch_sampler`.
Args:
build_batch_sampler (dict): same in :obj:`default_config`
==================== ====================
key description
==================== ====================
train (dict) - arguments for :obj:`FixedBatchSizeBatchSampler`
valid (dict) - arguments for :obj:`FixedBatchSizeBatchSampler`
test (dict) - arguments for :obj:`FixedBatchSizeBatchSampler`
==================== ====================
target_dir (str): Current experiment directory
cache_dir (str): If the preprocessing takes too long time, save
the temporary files into this directory. This directory is expected to be shared
across different training sessions (different hypers and :code:`target_dir`)
mode (str): train/valid/test
data_csv (str): the :code:`mode` specific csv from :obj:`prepare_data`
dataset: the dataset from :obj:`build_dataset`
Returns:
batch sampler for torch DataLoader
"""
def _build_batch_sampler(
train: dict = None, valid: dict = None, test: dict = None
):
if mode == "train":
return FixedBatchSizeBatchSampler(dataset, **train)
elif mode == "valid":
return FixedBatchSizeBatchSampler(dataset, **valid)
elif mode == "test":
return FixedBatchSizeBatchSampler(dataset, **test)
return _build_batch_sampler(**build_batch_sampler)
[docs] def build_downstream(
self,
build_downstream: dict,
downstream_input_size: int,
downstream_output_size: int,
downstream_input_stride: int,
):
"""
Return the task-specific downstream model.
By default build the :obj:`MeanPoolingLinear` model
Args:
build_downstream (dict): same in :obj:`default_config`,
support arguments of :obj:`MeanPoolingLinear`
downstream_input_size (int): the required input size of the model
downstream_output_size (int): the required output size of the model
downstream_input_stride (int): the input feature's stride (from 16 KHz)
Returns:
:obj:`AbsUtteranceModel`
"""
model = MeanPoolingLinear(
downstream_input_size, downstream_output_size, **build_downstream
)
return model
[docs] def build_task(
self,
build_task: dict,
model: torch.nn.Module,
encoder,
valid_df: pd.DataFrame = None,
test_df: pd.DataFrame = None,
):
"""
Build the task, which defines the logics for every train/valid/test forward step for the :code:`model`,
and the logics for how to reduce all the batch results from multiple train/valid/test steps into metrics
By default build :obj:`UtteranceMultiClassClassificationTask`
Args:
build_task (dict): same in :obj:`default_config`, no argument supported for now
model (torch.nn.Module): the model built by :obj:`build_model`
encoder: the encoder built by :obj:`build_encoder`
valid_df (pd.DataFrame): metadata of the valid set
test_df (pd.DataFrame): metadata of the test set
Returns:
Task
"""
return UtteranceMultiClassClassificationTask(model, encoder)