Commit
·
21e84e2
1
Parent(s):
73c48c7
Remove model codes.
Browse files- README.md +5 -5
- ced_model/__init__.py +0 -0
- ced_model/configuration_ced.py +0 -140
- ced_model/feature_extraction_ced.py +0 -121
- ced_model/modeling_ced.py +0 -575
README.md
CHANGED
@@ -28,8 +28,8 @@ Augmentation and knowledge distillation (KD) are well-established techniques emp
|
|
28 |
## Uses
|
29 |
|
30 |
```bash
|
31 |
-
git
|
32 |
-
|
33 |
```
|
34 |
|
35 |
```python
|
@@ -41,14 +41,14 @@ git clone https://huggingface.co/mispeech/ced-base
|
|
41 |
>>> model = CedForAudioClassification.from_pretrained(model_path)
|
42 |
|
43 |
>>> import torchaudio
|
44 |
-
>>> audio, sampling_rate = torchaudio.load(
|
45 |
|
46 |
>>> inputs = feature_extractor(audio, sampling_rate=sampling_rate, return_tensors="pt")
|
47 |
>>> with torch.no_grad():
|
48 |
... logits = model(**inputs).logits
|
49 |
|
|
|
50 |
>>> predicted_class_ids = torch.argmax(logits, dim=-1).item()
|
51 |
-
>>>
|
52 |
-
>>> predicted_label
|
53 |
'Finger snapping'
|
54 |
```
|
|
|
28 |
## Uses
|
29 |
|
30 |
```bash
|
31 |
+
git clone https://github.com/jimbozhang/hf_transformers_custom_model_ced.git
|
32 |
+
pip install -r requirements.txt
|
33 |
```
|
34 |
|
35 |
```python
|
|
|
41 |
>>> model = CedForAudioClassification.from_pretrained(model_path)
|
42 |
|
43 |
>>> import torchaudio
|
44 |
+
>>> audio, sampling_rate = torchaudio.load("resources/JeD5V5aaaoI_931_932.wav")
|
45 |
|
46 |
>>> inputs = feature_extractor(audio, sampling_rate=sampling_rate, return_tensors="pt")
|
47 |
>>> with torch.no_grad():
|
48 |
... logits = model(**inputs).logits
|
49 |
|
50 |
+
>>> import torch
|
51 |
>>> predicted_class_ids = torch.argmax(logits, dim=-1).item()
|
52 |
+
>>> model.config.id2label[predicted_class_ids]
|
|
|
53 |
'Finger snapping'
|
54 |
```
|
ced_model/__init__.py
DELETED
File without changes
|
ced_model/configuration_ced.py
DELETED
@@ -1,140 +0,0 @@
|
|
1 |
-
# coding=utf-8
|
2 |
-
# Copyright 2023 Xiaomi Corporation and The HuggingFace Inc. team. All rights reserved.
|
3 |
-
#
|
4 |
-
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
-
# you may not use this file except in compliance with the License.
|
6 |
-
# You may obtain a copy of the License at
|
7 |
-
#
|
8 |
-
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
-
#
|
10 |
-
# Unless required by applicable law or agreed to in writing, software
|
11 |
-
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
-
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
-
# See the License for the specific language governing permissions and
|
14 |
-
# limitations under the License.
|
15 |
-
""" CED model configuration"""
|
16 |
-
|
17 |
-
|
18 |
-
from transformers import PretrainedConfig
|
19 |
-
from transformers.utils import logging
|
20 |
-
from transformers.utils.hub import cached_file
|
21 |
-
|
22 |
-
logger = logging.get_logger(__name__)
|
23 |
-
|
24 |
-
CED_PRETRAINED_CONFIG_ARCHIVE_MAP = {
|
25 |
-
"mispeech/ced-tiny": "https://huggingface.co/mispeech/ced-tiny/resolve/main/config.json",
|
26 |
-
}
|
27 |
-
|
28 |
-
|
29 |
-
class CedConfig(PretrainedConfig):
|
30 |
-
model_type = "ced"
|
31 |
-
|
32 |
-
r"""
|
33 |
-
Configuration class for the CED model.
|
34 |
-
|
35 |
-
Args:
|
36 |
-
name (str, optional, *optional*):
|
37 |
-
Name of the pre-defined configuration. Can be "ced-tiny", "ced-mini", "ced-small" or "ced-base".
|
38 |
-
attn_drop_rate (float, *optional*, defaults to 0.0):
|
39 |
-
Dropout probability for attention weights. Default to 0.0.
|
40 |
-
depth (int, *optional*, defaults to 12): Number of transformer layers. Default to 12.
|
41 |
-
drop_path_rate (float, *optional*, defaults to 0.0): Drop path is taken from timm. Default to 0.0.
|
42 |
-
drop_rate (float, *optional*, defaults to 0.0):
|
43 |
-
Dropout probability for input embeddings. Default to 0.0.
|
44 |
-
embed_dim (int, *optional*, defaults to 768):
|
45 |
-
Dimensionality of the audio patch embeddings. Default to 768.
|
46 |
-
eval_avg (str, *optional*, defaults to `"mean"`):
|
47 |
-
Type of pooling to use for evaluation. Can be "mean", "token", "dm" or "logit". Default to "mean".
|
48 |
-
mlp_ratio (float, *optional*, defaults to 4.0):
|
49 |
-
Ratio of hidden size in the feedforward layer to the embedding size. Default to 4.0.
|
50 |
-
num_heads (int, *optional*, defaults to 12): Number of attention heads. Default to 12.
|
51 |
-
outputdim (int, *optional*, defaults to 527): Dimensionality of the output. Default to 527.
|
52 |
-
patch_size (int, *optional*, defaults to 16): Size of the patches. Default to 16.
|
53 |
-
patch_stride (int, *optional*, defaults to 16): Stride of the patches. Default to 16.
|
54 |
-
pooling (str, *optional*, defaults to `"mean"`):
|
55 |
-
Type of pooling to use for the output. Can be "mean", "token", "dm" or "logit". Default to "mean".
|
56 |
-
qkv_bias (bool, *optional*, defaults to `True`):
|
57 |
-
Whether to include bias terms in the query, key and value projections. Default to True.
|
58 |
-
target_length (int, *optional*, defaults to 1012): Frames of an audio chunk. Default to 1012.
|
59 |
-
"""
|
60 |
-
|
61 |
-
def __init__(
|
62 |
-
self,
|
63 |
-
name=None,
|
64 |
-
attn_drop_rate=0.0,
|
65 |
-
depth=12,
|
66 |
-
drop_path_rate=0.0,
|
67 |
-
drop_rate=0.0,
|
68 |
-
embed_dim=768,
|
69 |
-
eval_avg="mean",
|
70 |
-
mlp_ratio=4.0,
|
71 |
-
num_heads=12,
|
72 |
-
outputdim=527,
|
73 |
-
patch_size=16,
|
74 |
-
patch_stride=16,
|
75 |
-
pooling="mean",
|
76 |
-
qkv_bias=True,
|
77 |
-
target_length=1012,
|
78 |
-
**kwargs,
|
79 |
-
):
|
80 |
-
r"""
|
81 |
-
TODO: Add docstring
|
82 |
-
"""
|
83 |
-
|
84 |
-
super().__init__(**kwargs)
|
85 |
-
|
86 |
-
if name == "ced-tiny":
|
87 |
-
embed_dim = 192
|
88 |
-
num_heads = 3
|
89 |
-
elif name == "ced-mini":
|
90 |
-
embed_dim = 256
|
91 |
-
num_heads = 4
|
92 |
-
elif name == "ced-small":
|
93 |
-
embed_dim = 384
|
94 |
-
num_heads = 6
|
95 |
-
elif name == "ced-base":
|
96 |
-
embed_dim = 768
|
97 |
-
num_heads = 12
|
98 |
-
else:
|
99 |
-
logger.info("No model name specified for CedConfig, use default settings.")
|
100 |
-
|
101 |
-
assert pooling in ("mean", "token", "dm", "logit")
|
102 |
-
self.name = name
|
103 |
-
self.attn_drop_rate = attn_drop_rate
|
104 |
-
self.center = kwargs.get("center", True)
|
105 |
-
self.depth = depth
|
106 |
-
self.drop_path_rate = drop_path_rate
|
107 |
-
self.drop_rate = drop_rate
|
108 |
-
self.embed_dim = embed_dim
|
109 |
-
self.eval_avg = eval_avg
|
110 |
-
self.f_max = kwargs.get("f_max", 8000)
|
111 |
-
self.f_min = kwargs.get("f_min", 0)
|
112 |
-
self.hop_size = kwargs.get("hop_size", 160)
|
113 |
-
self.mlp_ratio = mlp_ratio
|
114 |
-
self.n_fft = kwargs.get("n_fft", 512)
|
115 |
-
self.n_mels = kwargs.get("n_mels", 64)
|
116 |
-
self.n_mels = kwargs.get("n_mels", 64)
|
117 |
-
self.num_heads = num_heads
|
118 |
-
self.outputdim = outputdim
|
119 |
-
self.pad_last = kwargs.get("pad_last", True)
|
120 |
-
self.patch_size = patch_size
|
121 |
-
self.patch_stride = patch_stride
|
122 |
-
self.pooling = pooling
|
123 |
-
self.qkv_bias = qkv_bias
|
124 |
-
self.target_length = target_length
|
125 |
-
self.win_size = kwargs.get("win_size", 512)
|
126 |
-
|
127 |
-
if self.outputdim == 527:
|
128 |
-
with open(
|
129 |
-
cached_file("topel/ConvNeXt-Tiny-AT", "class_labels_indices.csv"), "r"
|
130 |
-
) as f:
|
131 |
-
self.id2label = {
|
132 |
-
int(line.split(",", maxsplit=3)[0]): line.split(",", maxsplit=3)[2]
|
133 |
-
.replace('"', "")
|
134 |
-
.strip("\n")
|
135 |
-
for line in f.readlines()[1:]
|
136 |
-
}
|
137 |
-
self.label2id = {v: k for k, v in self.id2label.items()}
|
138 |
-
else:
|
139 |
-
self.id2label = None
|
140 |
-
self.label2id = None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
ced_model/feature_extraction_ced.py
DELETED
@@ -1,121 +0,0 @@
|
|
1 |
-
# coding=utf-8
|
2 |
-
# Copyright 2023 Xiaomi Corporation and The HuggingFace Inc. team.
|
3 |
-
#
|
4 |
-
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
-
# you may not use this file except in compliance with the License.
|
6 |
-
# You may obtain a copy of the License at
|
7 |
-
#
|
8 |
-
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
-
#
|
10 |
-
# Unless required by applicable law or agreed to in writing, software
|
11 |
-
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
-
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
-
# See the License for the specific language governing permissions and
|
14 |
-
# limitations under the License.
|
15 |
-
"""
|
16 |
-
Feature extractor class for CED.
|
17 |
-
"""
|
18 |
-
|
19 |
-
from typing import Optional, Union
|
20 |
-
|
21 |
-
import numpy as np
|
22 |
-
import torch
|
23 |
-
import torchaudio.transforms as audio_transforms
|
24 |
-
|
25 |
-
from transformers.feature_extraction_sequence_utils import SequenceFeatureExtractor
|
26 |
-
from transformers.feature_extraction_utils import BatchFeature
|
27 |
-
from transformers.utils import logging
|
28 |
-
|
29 |
-
|
30 |
-
logger = logging.get_logger(__name__)
|
31 |
-
|
32 |
-
|
33 |
-
class CedFeatureExtractor(SequenceFeatureExtractor):
|
34 |
-
r"""
|
35 |
-
CedFeatureExtractor extracts Mel spectrogram features from audio signals.
|
36 |
-
|
37 |
-
Args:
|
38 |
-
f_min (int, *optional*, defaults to 0): Minimum frequency for the Mel filterbank.
|
39 |
-
sampling_rate (int, *optional*, defaults to 16000):
|
40 |
-
Sampling rate of the input audio signal.
|
41 |
-
win_size (int, *optional*, defaults to 512): Window size for the STFT.
|
42 |
-
center (bool, *optional*, defaults to `True`):
|
43 |
-
Whether to pad the signal on both sides to center it.
|
44 |
-
n_fft (int, *optional*, defaults to 512): Number of FFT points for the STFT.
|
45 |
-
f_max (int, optional, *optional*): Maximum frequency for the Mel filterbank.
|
46 |
-
hop_size (int, *optional*, defaults to 160): Hop size for the STFT.
|
47 |
-
feature_size (int, *optional*, defaults to 64): Number of Mel bands to generate.
|
48 |
-
padding_value (float, *optional*, defaults to 0.0): Value for padding.
|
49 |
-
|
50 |
-
Returns:
|
51 |
-
BatchFeature: A BatchFeature object containing the extracted features.
|
52 |
-
"""
|
53 |
-
|
54 |
-
def __init__(
|
55 |
-
self,
|
56 |
-
f_min: int = 0,
|
57 |
-
sampling_rate: int = 16000,
|
58 |
-
win_size: int = 512,
|
59 |
-
center: bool = True,
|
60 |
-
n_fft: int = 512,
|
61 |
-
f_max: Optional[int] = None,
|
62 |
-
hop_size: int = 160,
|
63 |
-
feature_size: int = 64,
|
64 |
-
padding_value: float = 0.0,
|
65 |
-
**kwargs,
|
66 |
-
):
|
67 |
-
super().__init__(
|
68 |
-
feature_size=feature_size,
|
69 |
-
sampling_rate=sampling_rate,
|
70 |
-
padding_value=padding_value,
|
71 |
-
**kwargs,
|
72 |
-
)
|
73 |
-
self.f_min = f_min
|
74 |
-
self.win_size = win_size
|
75 |
-
self.center = center
|
76 |
-
self.n_fft = n_fft
|
77 |
-
self.f_max = f_max
|
78 |
-
self.hop_size = hop_size
|
79 |
-
|
80 |
-
def __call__(
|
81 |
-
self,
|
82 |
-
x: Union[np.ndarray, torch.Tensor],
|
83 |
-
sampling_rate: Optional[int] = None,
|
84 |
-
return_tensors="pt",
|
85 |
-
) -> BatchFeature:
|
86 |
-
r"""
|
87 |
-
Extracts Mel spectrogram features from an audio signal tensor.
|
88 |
-
|
89 |
-
Args:
|
90 |
-
x: Input audio signal tensor.
|
91 |
-
|
92 |
-
Returns:
|
93 |
-
BatchFeature: A dictionary containing the extracted features.
|
94 |
-
"""
|
95 |
-
if sampling_rate is None:
|
96 |
-
sampling_rate = self.sampling_rate
|
97 |
-
|
98 |
-
if return_tensors != "pt":
|
99 |
-
raise NotImplementedError(
|
100 |
-
"Only return_tensors='pt' is currently supported."
|
101 |
-
)
|
102 |
-
|
103 |
-
mel_spectrogram = audio_transforms.MelSpectrogram(
|
104 |
-
f_min=self.f_min,
|
105 |
-
sample_rate=sampling_rate,
|
106 |
-
win_length=self.win_size,
|
107 |
-
center=self.center,
|
108 |
-
n_fft=self.n_fft,
|
109 |
-
f_max=self.f_max,
|
110 |
-
hop_length=self.hop_size,
|
111 |
-
n_mels=self.feature_size,
|
112 |
-
)
|
113 |
-
amplitude_to_db = audio_transforms.AmplitudeToDB(top_db=120)
|
114 |
-
|
115 |
-
x = torch.from_numpy(x).float() if isinstance(x, np.ndarray) else x.float()
|
116 |
-
if x.dim() == 1:
|
117 |
-
x = x.unsqueeze(0)
|
118 |
-
|
119 |
-
x = mel_spectrogram(x)
|
120 |
-
x = amplitude_to_db(x)
|
121 |
-
return BatchFeature({"input_values": x})
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
ced_model/modeling_ced.py
DELETED
@@ -1,575 +0,0 @@
|
|
1 |
-
# coding=utf-8
|
2 |
-
# Copyright 2023 Xiaomi Corporation and The HuggingFace Inc. team. All rights reserved.
|
3 |
-
#
|
4 |
-
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
-
# you may not use this file except in compliance with the License.
|
6 |
-
# You may obtain a copy of the License at
|
7 |
-
#
|
8 |
-
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
-
#
|
10 |
-
# Unless required by applicable law or agreed to in writing, software
|
11 |
-
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
-
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
-
# See the License for the specific language governing permissions and
|
14 |
-
# limitations under the License.
|
15 |
-
""" PyTorch CED (Ced) model."""
|
16 |
-
|
17 |
-
import collections
|
18 |
-
import math
|
19 |
-
from functools import partial
|
20 |
-
from typing import Any, Callable, Optional, Tuple, Union
|
21 |
-
|
22 |
-
import torch
|
23 |
-
import torch.utils.checkpoint
|
24 |
-
from torch import nn
|
25 |
-
|
26 |
-
from transformers.modeling_outputs import SequenceClassifierOutput
|
27 |
-
from transformers.modeling_utils import PreTrainedModel
|
28 |
-
from transformers.utils import (
|
29 |
-
add_code_sample_docstrings,
|
30 |
-
add_start_docstrings,
|
31 |
-
add_start_docstrings_to_model_forward,
|
32 |
-
logging,
|
33 |
-
)
|
34 |
-
from .configuration_ced import CedConfig
|
35 |
-
|
36 |
-
|
37 |
-
logger = logging.get_logger(__name__)
|
38 |
-
|
39 |
-
_CONFIG_FOR_DOC = "CedConfig"
|
40 |
-
_SEQ_CLASS_EXPECTED_OUTPUT = "'Speech synthesizer'"
|
41 |
-
_SEQ_CLASS_EXPECTED_LOSS = 0.69
|
42 |
-
|
43 |
-
# Audio classification docstring
|
44 |
-
_SEQ_CLASS_CHECKPOINT = "mispeech/ced-tiny"
|
45 |
-
|
46 |
-
|
47 |
-
CED_PRETRAINED_MODEL_ARCHIVE_LIST = [
|
48 |
-
"mispeech/ced-tiny",
|
49 |
-
"mispeech/ced-mini",
|
50 |
-
"mispeech/ced-small",
|
51 |
-
"mispeech/ced-base",
|
52 |
-
# See all CED models at https://huggingface.co/models?search=mispeech%2Fced
|
53 |
-
]
|
54 |
-
|
55 |
-
|
56 |
-
class CedPreTrainedModel(PreTrainedModel):
|
57 |
-
"""
|
58 |
-
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
|
59 |
-
models.
|
60 |
-
"""
|
61 |
-
|
62 |
-
config_class = CedConfig
|
63 |
-
base_model_prefix = "ced"
|
64 |
-
main_input_name = "input_values"
|
65 |
-
supports_gradient_checkpointing = True
|
66 |
-
|
67 |
-
def _init_weights(self, module):
|
68 |
-
"""Initialize the weights"""
|
69 |
-
if isinstance(module, nn.Linear):
|
70 |
-
trunc_normal_(module.weight, std=0.02)
|
71 |
-
if module.bias is not None:
|
72 |
-
nn.init.zeros_(module.bias)
|
73 |
-
elif isinstance(module, nn.LayerNorm):
|
74 |
-
nn.init.constant_(module.bias, 0)
|
75 |
-
nn.init.constant_(module.weight, 1.0)
|
76 |
-
|
77 |
-
|
78 |
-
Conv_Kernel = Union[int, Tuple[int, int]]
|
79 |
-
|
80 |
-
|
81 |
-
def to_2tuple(x: Any) -> Tuple[Any, Any]:
|
82 |
-
if isinstance(x, collections.abc.Iterable):
|
83 |
-
return x
|
84 |
-
return (x, x)
|
85 |
-
|
86 |
-
|
87 |
-
class CedAudioPatchEmbed(nn.Module):
|
88 |
-
def __init__(
|
89 |
-
self,
|
90 |
-
input_size: Conv_Kernel = 224,
|
91 |
-
patch_size: Conv_Kernel = 16,
|
92 |
-
patch_stride: Conv_Kernel = 16,
|
93 |
-
in_chans: int = 1,
|
94 |
-
embed_dim: int = 768,
|
95 |
-
norm_layer: Optional[Callable] = None,
|
96 |
-
flatten: bool = False,
|
97 |
-
):
|
98 |
-
super().__init__()
|
99 |
-
self.input_size = to_2tuple(input_size)
|
100 |
-
self.patch_size = to_2tuple(patch_size)
|
101 |
-
self.patch_stride = to_2tuple(patch_stride)
|
102 |
-
self.grid_size = (
|
103 |
-
self.input_size[0] // self.patch_stride[0],
|
104 |
-
self.input_size[1] // self.patch_stride[1],
|
105 |
-
)
|
106 |
-
self.num_patches = self.grid_size[0] * self.grid_size[1]
|
107 |
-
self.flatten = flatten
|
108 |
-
|
109 |
-
self.proj = nn.Conv2d(
|
110 |
-
in_chans, embed_dim, kernel_size=patch_size, stride=patch_stride
|
111 |
-
)
|
112 |
-
self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
|
113 |
-
|
114 |
-
def forward(self, x):
|
115 |
-
x = self.proj(x)
|
116 |
-
if self.flatten:
|
117 |
-
x = torch.permute(torch.flatten(x, 2, 3), (0, 2, 1))
|
118 |
-
x = self.norm(x)
|
119 |
-
return x
|
120 |
-
|
121 |
-
|
122 |
-
class CedAttention(nn.Module):
|
123 |
-
def __init__(
|
124 |
-
self,
|
125 |
-
dim,
|
126 |
-
num_heads=8,
|
127 |
-
qkv_bias=False,
|
128 |
-
attn_drop=0.0,
|
129 |
-
proj_drop=0.0,
|
130 |
-
causal: bool = False,
|
131 |
-
):
|
132 |
-
super().__init__()
|
133 |
-
assert dim % num_heads == 0, "dim should be divisible by num_heads"
|
134 |
-
self.num_heads = num_heads
|
135 |
-
head_dim = dim // num_heads
|
136 |
-
self.scale = head_dim**-0.5
|
137 |
-
|
138 |
-
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
|
139 |
-
self.attn_drop = nn.Dropout(attn_drop)
|
140 |
-
self.proj = nn.Linear(dim, dim)
|
141 |
-
self.proj_drop = nn.Dropout(proj_drop)
|
142 |
-
self.causal = causal
|
143 |
-
|
144 |
-
def forward(self, x):
|
145 |
-
B, N, C = x.shape
|
146 |
-
qkv = (
|
147 |
-
self.qkv(x)
|
148 |
-
.reshape(B, N, 3, self.num_heads, C // self.num_heads)
|
149 |
-
.permute(2, 0, 3, 1, 4)
|
150 |
-
)
|
151 |
-
q, k, v = qkv.unbind(0) # make torchscript happy (cannot use tensor as tuple)
|
152 |
-
|
153 |
-
attn = (q @ k.transpose(-2, -1)) * self.scale
|
154 |
-
# if mask is not None:
|
155 |
-
# # Mask is a tensor of shape [B, T, T]
|
156 |
-
# # Different from self.causal == True, the mask might be something like:
|
157 |
-
# # [False, False, True]
|
158 |
-
# # [False, False, True]
|
159 |
-
# # [True, True, True]
|
160 |
-
# # We use -inf to pad here, since if we would pad by any number, the entries at rows only containing
|
161 |
-
# # [True, True, True] would lead to weights such as: [0.33,0.33,0.33], which is not correct
|
162 |
-
# mask_value = torch.as_tensor(-float('inf'))
|
163 |
-
# print(mask.shape, attn.shape)
|
164 |
-
# attn = attn.masked_fill(mask, mask_value)
|
165 |
-
if self.causal:
|
166 |
-
mask_value = -torch.finfo(attn.dtype).max
|
167 |
-
i, j = attn.shape[-2:]
|
168 |
-
mask = torch.ones(i, j, device=q.device, dtype=torch.bool).triu(j - i + 1)
|
169 |
-
attn = attn.masked_fill(mask, mask_value)
|
170 |
-
attn = attn.softmax(dim=-1)
|
171 |
-
# Only for the case that a mask with all True entries on a row is passed.
|
172 |
-
# attn = torch.nan_to_num(attn)
|
173 |
-
attn = self.attn_drop(attn)
|
174 |
-
|
175 |
-
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
|
176 |
-
x = self.proj(x)
|
177 |
-
x = self.proj_drop(x)
|
178 |
-
return x
|
179 |
-
|
180 |
-
|
181 |
-
class CedMlp(nn.Module):
|
182 |
-
def __init__(
|
183 |
-
self,
|
184 |
-
in_features: int,
|
185 |
-
hidden_features: Optional[int] = None,
|
186 |
-
out_features: Optional[int] = None,
|
187 |
-
act_layer: Callable = nn.GELU,
|
188 |
-
drop: float = 0.0,
|
189 |
-
):
|
190 |
-
super().__init__()
|
191 |
-
out_features = out_features or in_features
|
192 |
-
hidden_features = hidden_features or in_features
|
193 |
-
self.fc1 = nn.Linear(in_features, hidden_features)
|
194 |
-
self.act = act_layer()
|
195 |
-
self.fc2 = nn.Linear(hidden_features, out_features)
|
196 |
-
self.drop = nn.Dropout(drop)
|
197 |
-
|
198 |
-
def forward(self, x):
|
199 |
-
x = self.fc1(x)
|
200 |
-
x = self.act(x)
|
201 |
-
x = self.drop(x)
|
202 |
-
x = self.fc2(x)
|
203 |
-
x = self.drop(x)
|
204 |
-
return x
|
205 |
-
|
206 |
-
|
207 |
-
# Drop path is taken from Timm
|
208 |
-
# https://github.com/huggingface/pytorch-image-models/blob/7c67d6aca992f039eece0af5f7c29a43d48c00e4/timm/models/layers/drop.py#L155
|
209 |
-
class DropPath(nn.Module):
|
210 |
-
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""
|
211 |
-
|
212 |
-
def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):
|
213 |
-
super(DropPath, self).__init__()
|
214 |
-
self.drop_prob = drop_prob
|
215 |
-
self.scale_by_keep = scale_by_keep
|
216 |
-
|
217 |
-
def forward(self, x):
|
218 |
-
return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)
|
219 |
-
|
220 |
-
def extra_repr(self):
|
221 |
-
return f"drop_prob={round(self.drop_prob,3):0.3f}"
|
222 |
-
|
223 |
-
|
224 |
-
def drop_path(
|
225 |
-
x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True
|
226 |
-
):
|
227 |
-
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
|
228 |
-
|
229 |
-
This is the same as the DropConnect impl I (https://github.com/rwightman) created for EfficientNet, etc networks,
|
230 |
-
however, the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
|
231 |
-
See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for changing the
|
232 |
-
layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use 'survival rate' as the
|
233 |
-
argument.
|
234 |
-
|
235 |
-
"""
|
236 |
-
if drop_prob == 0.0 or not training:
|
237 |
-
return x
|
238 |
-
keep_prob = 1 - drop_prob
|
239 |
-
shape = (x.shape[0],) + (1,) * (
|
240 |
-
x.ndim - 1
|
241 |
-
) # work with diff dim tensors, not just 2D ConvNets
|
242 |
-
random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
|
243 |
-
if keep_prob > 0.0 and scale_by_keep:
|
244 |
-
random_tensor.div_(keep_prob)
|
245 |
-
return x * random_tensor
|
246 |
-
|
247 |
-
|
248 |
-
class CedBlock(nn.Module):
|
249 |
-
def __init__(
|
250 |
-
self,
|
251 |
-
dim,
|
252 |
-
num_heads,
|
253 |
-
mlp_ratio=4.0,
|
254 |
-
qkv_bias=False,
|
255 |
-
drop=0.0,
|
256 |
-
attn_drop=0.0,
|
257 |
-
drop_path=0.0,
|
258 |
-
act_layer: Callable = nn.GELU,
|
259 |
-
norm_layer: Callable = nn.LayerNorm,
|
260 |
-
attention_type: Callable = CedAttention,
|
261 |
-
attention_kwargs={},
|
262 |
-
**kwargs,
|
263 |
-
):
|
264 |
-
super().__init__()
|
265 |
-
self.norm1 = norm_layer(dim)
|
266 |
-
self.attn = attention_type(
|
267 |
-
dim,
|
268 |
-
num_heads=num_heads,
|
269 |
-
qkv_bias=qkv_bias,
|
270 |
-
attn_drop=attn_drop,
|
271 |
-
proj_drop=drop,
|
272 |
-
**attention_kwargs,
|
273 |
-
)
|
274 |
-
self.ls1 = nn.Identity()
|
275 |
-
self.drop_path1 = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
|
276 |
-
|
277 |
-
self.norm2 = norm_layer(dim)
|
278 |
-
self.mlp = CedMlp(
|
279 |
-
in_features=dim,
|
280 |
-
hidden_features=int(dim * mlp_ratio),
|
281 |
-
act_layer=act_layer,
|
282 |
-
drop=drop,
|
283 |
-
)
|
284 |
-
self.ls2 = nn.Identity()
|
285 |
-
self.drop_path2 = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
|
286 |
-
|
287 |
-
def forward(self, x):
|
288 |
-
x = x + self.drop_path1(self.ls1(self.attn(self.norm1(x))))
|
289 |
-
x = x + self.drop_path2(self.ls2(self.mlp(self.norm2(x))))
|
290 |
-
return x
|
291 |
-
|
292 |
-
|
293 |
-
# Taken from timm
|
294 |
-
def trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0):
|
295 |
-
return _no_grad_trunc_normal_(tensor, mean, std, a, b)
|
296 |
-
|
297 |
-
|
298 |
-
def _no_grad_trunc_normal_(tensor, mean, std, a, b):
|
299 |
-
# Cut & paste from PyTorch official master until it's in a few official releases - RW
|
300 |
-
# Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
|
301 |
-
def norm_cdf(x):
|
302 |
-
# Computes standard normal cumulative distribution function
|
303 |
-
return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
|
304 |
-
|
305 |
-
with torch.no_grad():
|
306 |
-
# Values are generated by using a truncated uniform distribution and
|
307 |
-
# then using the inverse CDF for the normal distribution.
|
308 |
-
# Get upper and lower cdf values
|
309 |
-
l = norm_cdf((a - mean) / std)
|
310 |
-
u = norm_cdf((b - mean) / std)
|
311 |
-
|
312 |
-
# Uniformly fill tensor with values from [l, u], then translate to
|
313 |
-
# [2l-1, 2u-1].
|
314 |
-
tensor.uniform_(2 * l - 1, 2 * u - 1)
|
315 |
-
|
316 |
-
# Use inverse cdf transform for normal distribution to get truncated
|
317 |
-
# standard normal
|
318 |
-
tensor.erfinv_()
|
319 |
-
|
320 |
-
# Transform to proper mean, std
|
321 |
-
tensor.mul_(std * math.sqrt(2.0))
|
322 |
-
tensor.add_(mean)
|
323 |
-
|
324 |
-
# Clamp to ensure it's in the proper range
|
325 |
-
tensor.clamp_(min=a, max=b)
|
326 |
-
return tensor
|
327 |
-
|
328 |
-
|
329 |
-
CED_START_DOCSTRING = r"""
|
330 |
-
|
331 |
-
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
332 |
-
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
333 |
-
etc.)
|
334 |
-
|
335 |
-
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
336 |
-
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
337 |
-
and behavior.
|
338 |
-
|
339 |
-
Parameters:
|
340 |
-
config ([`CedConfig`]): Model configuration class with all the parameters of the model.
|
341 |
-
Initializing with a config file does not load the weights associated with the model, only the
|
342 |
-
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
343 |
-
"""
|
344 |
-
|
345 |
-
CED_INPUTS_DOCSTRING = r"""
|
346 |
-
Args:
|
347 |
-
input_values (`torch.FloatTensor` of shape `(batch_size, n_mels, sequence_length)`):
|
348 |
-
The sequence of audio features extracted from the audio signal. Can be obtained from a raw audio waveform
|
349 |
-
using `~transformers.CedFeatureExtractor.__call__`.
|
350 |
-
"""
|
351 |
-
|
352 |
-
|
353 |
-
@add_start_docstrings(
|
354 |
-
"The bare Ced Model transformer outputting raw hidden-states without any specific head on top.",
|
355 |
-
CED_START_DOCSTRING,
|
356 |
-
)
|
357 |
-
class CedModel(CedPreTrainedModel):
|
358 |
-
def __init__(self, config: CedConfig) -> None:
|
359 |
-
super().__init__(config)
|
360 |
-
self.config = config
|
361 |
-
self.name = config.name
|
362 |
-
|
363 |
-
# Allowed length in number of frames, otherwise the positional embedding will throw an error
|
364 |
-
self.maximal_allowed_length = self.config.target_length
|
365 |
-
|
366 |
-
self.init_bn = torch.nn.BatchNorm2d(config.n_mels, momentum=0.01)
|
367 |
-
|
368 |
-
self.patch_embed = CedAudioPatchEmbed(
|
369 |
-
input_size=(config.n_mels, config.target_length),
|
370 |
-
embed_dim=config.embed_dim,
|
371 |
-
patch_size=config.patch_size,
|
372 |
-
flatten=False,
|
373 |
-
patch_stride=config.patch_stride,
|
374 |
-
)
|
375 |
-
|
376 |
-
self.time_pos_embed = nn.Parameter(
|
377 |
-
torch.randn(1, config.embed_dim, 1, self.patch_embed.grid_size[1]) * 0.02
|
378 |
-
)
|
379 |
-
self.freq_pos_embed = nn.Parameter(
|
380 |
-
torch.randn(1, config.embed_dim, self.patch_embed.grid_size[0], 1) * 0.02
|
381 |
-
)
|
382 |
-
norm_layer = partial(nn.LayerNorm, eps=1e-6)
|
383 |
-
act_layer = nn.GELU
|
384 |
-
dpr = [
|
385 |
-
x.item() for x in torch.linspace(0, config.drop_path_rate, config.depth)
|
386 |
-
] # stochastic depth decay rule
|
387 |
-
self.pos_drop = nn.Dropout(p=config.drop_rate)
|
388 |
-
self.blocks = nn.Sequential(
|
389 |
-
*[
|
390 |
-
CedBlock(
|
391 |
-
dim=config.embed_dim,
|
392 |
-
num_heads=config.num_heads,
|
393 |
-
mlp_ratio=config.mlp_ratio,
|
394 |
-
qkv_bias=config.qkv_bias,
|
395 |
-
drop=config.drop_rate,
|
396 |
-
attn_drop=config.attn_drop_rate,
|
397 |
-
drop_path=dpr[i],
|
398 |
-
norm_layer=norm_layer,
|
399 |
-
act_layer=act_layer,
|
400 |
-
attention_type=CedAttention,
|
401 |
-
)
|
402 |
-
for i in range(config.depth)
|
403 |
-
]
|
404 |
-
)
|
405 |
-
self.norm = norm_layer(config.embed_dim)
|
406 |
-
|
407 |
-
# Initialize weights and apply final processing
|
408 |
-
self.post_init()
|
409 |
-
|
410 |
-
def forward_features(self, x: torch.Tensor) -> torch.Tensor:
|
411 |
-
x = self.patch_embed(x)
|
412 |
-
_, _, _, t = x.shape
|
413 |
-
x = x + self.time_pos_embed[:, :, :, :t]
|
414 |
-
x = (
|
415 |
-
x + self.freq_pos_embed[:, :, :, :]
|
416 |
-
) # Just to support __getitem__ in posembed
|
417 |
-
|
418 |
-
# x = rearrange(x, 'b c f t -> b (f t) c')
|
419 |
-
x = torch.permute(torch.flatten(x, 2, 3), (0, 2, 1))
|
420 |
-
|
421 |
-
if self.config.pooling == "token":
|
422 |
-
cls_token = self.cls_token.expand(x.shape[0], -1, -1)
|
423 |
-
cls_token = cls_token + self.token_pos_embed
|
424 |
-
x = torch.cat((cls_token, x), dim=1)
|
425 |
-
x = self.pos_drop(x)
|
426 |
-
x = self.blocks(x)
|
427 |
-
x = self.norm(x)
|
428 |
-
return x
|
429 |
-
|
430 |
-
def forward(self, input_values: torch.Tensor):
|
431 |
-
r"""
|
432 |
-
Runs a forward pass of the CED model as an audio encoder.
|
433 |
-
"""
|
434 |
-
x = torch.unsqueeze(input_values, 1)
|
435 |
-
|
436 |
-
x = torch.permute(x, (0, 2, 1, 3))
|
437 |
-
x = self.init_bn(x)
|
438 |
-
x = torch.permute(x, (0, 2, 1, 3))
|
439 |
-
|
440 |
-
if x.shape[-1] > self.maximal_allowed_length:
|
441 |
-
splits = x.split(self.maximal_allowed_length, -1)
|
442 |
-
|
443 |
-
if splits[-1].shape[-1] < self.maximal_allowed_length:
|
444 |
-
if self.config.pad_last:
|
445 |
-
pad = torch.zeros(
|
446 |
-
*x.shape[:-1], self.maximal_allowed_length, device=x.device
|
447 |
-
)
|
448 |
-
pad[..., : splits[-1].shape[-1]] = splits[-1]
|
449 |
-
splits = torch.stack((*splits[:-1], pad), dim=0)
|
450 |
-
else:
|
451 |
-
splits = torch.stack(splits[:-1], dim=0)
|
452 |
-
else:
|
453 |
-
splits = torch.stack(splits[:-1], dim=0)
|
454 |
-
n_splits = len(splits)
|
455 |
-
x = torch.flatten(splits, 0, 1) # spl b c f t-> (spl b) c f t
|
456 |
-
x = self.forward_head(self.ced(x))
|
457 |
-
x = torch.reshape(
|
458 |
-
x, (n_splits, -1, self.outputdim)
|
459 |
-
) # (spl b) d -> spl b d, spl=n_splits
|
460 |
-
|
461 |
-
if self.config.eval_avg == "mean":
|
462 |
-
x = x.mean(0)
|
463 |
-
elif self.config.eval_avg == "max":
|
464 |
-
x = x.max(0)[0]
|
465 |
-
else:
|
466 |
-
raise ValueError(f"Unknown Eval average function ({self.eval_avg})")
|
467 |
-
else:
|
468 |
-
x = self.forward_features(x)
|
469 |
-
|
470 |
-
return SequenceClassifierOutput(logits=x)
|
471 |
-
|
472 |
-
|
473 |
-
@add_start_docstrings(
|
474 |
-
"""
|
475 |
-
Ced model with an audio classification head on top (a linear layer on top of the pooled output).
|
476 |
-
""",
|
477 |
-
CED_START_DOCSTRING,
|
478 |
-
)
|
479 |
-
class CedForAudioClassification(CedPreTrainedModel):
|
480 |
-
def __init__(self, config: CedConfig) -> None:
|
481 |
-
super().__init__(config)
|
482 |
-
self.config = config
|
483 |
-
|
484 |
-
self.encoder = CedModel(config)
|
485 |
-
|
486 |
-
# Classifier head
|
487 |
-
self.outputlayer = nn.Sequential(
|
488 |
-
nn.LayerNorm(config.embed_dim),
|
489 |
-
nn.Linear(config.embed_dim, config.outputdim),
|
490 |
-
)
|
491 |
-
|
492 |
-
# Initialize weights and apply final processing
|
493 |
-
self.post_init()
|
494 |
-
|
495 |
-
def forward_head(self, x: torch.Tensor) -> torch.Tensor:
|
496 |
-
if self.config.pooling == "token":
|
497 |
-
x = x[:, 0]
|
498 |
-
return self.outputlayer(x).sigmoid()
|
499 |
-
elif self.config.pooling == "mean":
|
500 |
-
x = x.mean(1)
|
501 |
-
return self.outputlayer(x).sigmoid()
|
502 |
-
elif self.config.pooling == "logit":
|
503 |
-
x = x.mean(1)
|
504 |
-
return self.outputlayer(x)
|
505 |
-
elif self.config.pooling == "dm":
|
506 |
-
# Unpack using the frequency dimension, which is constant
|
507 |
-
# 'b (f t) d -> b f t d', f=self.patch_embed.grid_size[0])
|
508 |
-
x = torch.reshape(
|
509 |
-
x, (x.shape[0], self.patch_embed.grid_size[0], -1, x.shape[3])
|
510 |
-
)
|
511 |
-
|
512 |
-
# First poolin frequency, then sigmoid the (B T D) output
|
513 |
-
x = self.outputlayer(x.mean(1)).sigmoid()
|
514 |
-
return x.mean(1)
|
515 |
-
else:
|
516 |
-
return x.mean(1)
|
517 |
-
|
518 |
-
@add_start_docstrings_to_model_forward(
|
519 |
-
CED_INPUTS_DOCSTRING.format("batch_size, sequence_length")
|
520 |
-
)
|
521 |
-
@add_code_sample_docstrings(
|
522 |
-
checkpoint=_SEQ_CLASS_CHECKPOINT,
|
523 |
-
output_type=SequenceClassifierOutput,
|
524 |
-
config_class=_CONFIG_FOR_DOC,
|
525 |
-
modality="audio",
|
526 |
-
model_cls="CedForAudioClassification",
|
527 |
-
expected_output=_SEQ_CLASS_EXPECTED_OUTPUT,
|
528 |
-
expected_loss=_SEQ_CLASS_EXPECTED_LOSS,
|
529 |
-
)
|
530 |
-
def forward(
|
531 |
-
self, input_values: torch.Tensor, labels: Optional[torch.Tensor] = None
|
532 |
-
):
|
533 |
-
"""
|
534 |
-
Runs a forward pass of the CED model for audio classification task.
|
535 |
-
|
536 |
-
Examples:
|
537 |
-
|
538 |
-
```python
|
539 |
-
>>> from transformers import AutoFeatureExtractor, AutoModelForAudioClassification
|
540 |
-
>>> from datasets import load_dataset
|
541 |
-
>>> import torch
|
542 |
-
|
543 |
-
>>> dataset = load_dataset("hf-internal-testing/librispeech_asr_demo", "clean", split="validation")
|
544 |
-
>>> dataset = dataset.sort("id")
|
545 |
-
>>> sampling_rate = dataset.features["audio"].sampling_rate
|
546 |
-
|
547 |
-
>>> feature_extractor = AutoFeatureExtractor.from_pretrained("mispeech/ced-tiny")
|
548 |
-
>>> model = AutoModelForAudioClassification.from_pretrained("mispeech/ced-tiny")
|
549 |
-
|
550 |
-
>>> inputs = feature_extractor(dataset[0]["audio"]["array"], sampling_rate=sampling_rate, return_tensors="pt")
|
551 |
-
|
552 |
-
>>> with torch.no_grad():
|
553 |
-
... logits = model(**inputs).logits
|
554 |
-
|
555 |
-
>>> predicted_class_ids = torch.argmax(logits, dim=-1).item()
|
556 |
-
>>> predicted_label = model.config.id2label[predicted_class_ids]
|
557 |
-
>>> predicted_label
|
558 |
-
'Speech synthesizer'
|
559 |
-
```
|
560 |
-
"""
|
561 |
-
last_hidden_states = self.encoder(input_values).logits
|
562 |
-
logits = self.forward_head(last_hidden_states)
|
563 |
-
|
564 |
-
if labels is not None:
|
565 |
-
loss_fct = nn.BCEWithLogitsLoss()
|
566 |
-
labels = nn.functional.one_hot(
|
567 |
-
labels, num_classes=self.config.outputdim
|
568 |
-
).float()
|
569 |
-
loss = loss_fct(logits, labels)
|
570 |
-
else:
|
571 |
-
loss = None
|
572 |
-
|
573 |
-
return SequenceClassifierOutput(
|
574 |
-
logits=logits, loss=loss, hidden_states=last_hidden_states
|
575 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|