Commit
·
bf8cc62
1
Parent(s):
ea0d0f6
Add README and supporting code
Browse files- README.md +45 -3
- images/human_eval_pr_curve.png +0 -0
- inference.py +110 -0
- requirements.txt +4 -0
README.md
CHANGED
|
@@ -1,3 +1,45 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
## Model Description
|
| 2 |
+
We present a large classification model trained on a manually curated real-world dataset that can be used as a new
|
| 3 |
+
benchmark for advancing research in toxicity detection and classification.
|
| 4 |
+
Our model is fine-tuned on the [WavLM base plus](https://arxiv.org/abs/2110.13900) with 2,374 hours of audio clips from
|
| 5 |
+
voice chat for multilabel classification. The audio clips are automatically labeled using a synthetic data pipeline
|
| 6 |
+
described in [our blog post](link to blog post here). A single output can have multiple labels.
|
| 7 |
+
The model outputs a n by 6 output tensor where the inferred labels are `Profanity`, `DatingAndSexting`, `Racist`,
|
| 8 |
+
`Bullying`, `Other`, `NoViolation`. `Other` consists of policy violation categories with low prevalence such as drugs
|
| 9 |
+
and alcohol or self-harm that are combined into a single category.
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
We evaluated this model on a data set with human annotated labels that contained a total of 9,795 samples with the class
|
| 13 |
+
distribution shown below. Note that we did not include the "other" category in this evaluation data set.
|
| 14 |
+
|
| 15 |
+
|Class|Number of examples| Duration (hours)|% of dataset|
|
| 16 |
+
|---|---|---|---|
|
| 17 |
+
|Profanity | 4893| 15.38 | 49.95%|
|
| 18 |
+
|DatingAndSexting | 688 | 2.52 | 7.02% |
|
| 19 |
+
|Racist | 889 | 3.10 | 9.08% |
|
| 20 |
+
|Bullying | 1256 | 4.25 | 12.82% |
|
| 21 |
+
|NoViolation | 4185 | 9.93 | 42.73% |
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
If we set the same threshold across all classes and treat the model as a binary classifier across all 4 toxicity classes
|
| 25 |
+
(`Profanity`, `DatingAndSexting`, `Racist`, `Bullying`), we get a binarized average precision of 94.48%. The precision
|
| 26 |
+
recall curve is as shown below.
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
<p align="center">
|
| 30 |
+
<img src="images/human_eval_pr_curve.png" alt="PR Curve" width="500"/>
|
| 31 |
+
</p>
|
| 32 |
+
|
| 33 |
+
## Usage
|
| 34 |
+
The dependencies for the inference file can be installed as follows:
|
| 35 |
+
```
|
| 36 |
+
pip install -r requirements.txt
|
| 37 |
+
```
|
| 38 |
+
The inference file contains useful helper functions to preprocess the audio file for proper inference.
|
| 39 |
+
To run the inference file, please run the following command:
|
| 40 |
+
```
|
| 41 |
+
python inference.py --audio_file <your audio file path> --model_path <path to Huggingface model>
|
| 42 |
+
```
|
| 43 |
+
You can get the model weights either by downloading from the model releases page [here](https://github.com/Roblox/voice-safety-classifier/releases/tag/vs-classifier-v1), or from HuggingFace under
|
| 44 |
+
[`roblox/voice-safety-classifier`](https://huggingface.co/Roblox/voice-safety-classifier). If `model_path` isn’t
|
| 45 |
+
specified, the model will be loaded directly from HuggingFace.
|
images/human_eval_pr_curve.png
ADDED
|
inference.py
ADDED
|
@@ -0,0 +1,110 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright © 2024 Roblox Corporation
|
| 2 |
+
|
| 3 |
+
"""
|
| 4 |
+
This file gives a sample demonstration of how to use the given functions in Python, for the Voice Safety Classifier model.
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
import torch
|
| 8 |
+
import librosa
|
| 9 |
+
import numpy as np
|
| 10 |
+
import argparse
|
| 11 |
+
from transformers import WavLMForSequenceClassification
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
def feature_extract_simple(
|
| 15 |
+
wav,
|
| 16 |
+
sr=16_000,
|
| 17 |
+
win_len=15.0,
|
| 18 |
+
win_stride=15.0,
|
| 19 |
+
do_normalize=False,
|
| 20 |
+
):
|
| 21 |
+
"""simple feature extraction for wavLM
|
| 22 |
+
Parameters
|
| 23 |
+
----------
|
| 24 |
+
wav : str or array-like
|
| 25 |
+
path to the wav file, or array-like
|
| 26 |
+
sr : int, optional
|
| 27 |
+
sample rate, by default 16_000
|
| 28 |
+
win_len : float, optional
|
| 29 |
+
window length, by default 15.0
|
| 30 |
+
win_stride : float, optional
|
| 31 |
+
window stride, by default 15.0
|
| 32 |
+
do_normalize: bool, optional
|
| 33 |
+
whether to normalize the input, by default False.
|
| 34 |
+
Returns
|
| 35 |
+
-------
|
| 36 |
+
np.ndarray
|
| 37 |
+
batched input to wavLM
|
| 38 |
+
"""
|
| 39 |
+
if type(wav) == str:
|
| 40 |
+
signal, _ = librosa.core.load(wav, sr=sr)
|
| 41 |
+
else:
|
| 42 |
+
try:
|
| 43 |
+
signal = np.array(wav).squeeze()
|
| 44 |
+
except Exception as e:
|
| 45 |
+
print(e)
|
| 46 |
+
raise RuntimeError
|
| 47 |
+
batched_input = []
|
| 48 |
+
stride = int(win_stride * sr)
|
| 49 |
+
l = int(win_len * sr)
|
| 50 |
+
if len(signal) / sr > win_len:
|
| 51 |
+
for i in range(0, len(signal), stride):
|
| 52 |
+
if i + int(win_len * sr) > len(signal):
|
| 53 |
+
# padding the last chunk to make it the same length as others
|
| 54 |
+
chunked = np.pad(signal[i:], (0, l - len(signal[i:])))
|
| 55 |
+
else:
|
| 56 |
+
chunked = signal[i : i + l]
|
| 57 |
+
if do_normalize:
|
| 58 |
+
chunked = (chunked - np.mean(chunked)) / (np.std(chunked) + 1e-7)
|
| 59 |
+
batched_input.append(chunked)
|
| 60 |
+
if i + int(win_len * sr) > len(signal):
|
| 61 |
+
break
|
| 62 |
+
else:
|
| 63 |
+
if do_normalize:
|
| 64 |
+
signal = (signal - np.mean(signal)) / (np.std(signal) + 1e-7)
|
| 65 |
+
batched_input.append(signal)
|
| 66 |
+
return np.stack(batched_input) # [N, T]
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
def infer(model, inputs):
|
| 70 |
+
output = model(inputs)
|
| 71 |
+
probs = torch.sigmoid(torch.Tensor(output.logits))
|
| 72 |
+
return probs
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
if __name__ == "__main__":
|
| 76 |
+
parser = argparse.ArgumentParser()
|
| 77 |
+
parser.add_argument(
|
| 78 |
+
"--audio_file",
|
| 79 |
+
type=str,
|
| 80 |
+
help="File to run inference",
|
| 81 |
+
)
|
| 82 |
+
parser.add_argument(
|
| 83 |
+
"--model_path",
|
| 84 |
+
type=str,
|
| 85 |
+
default="roblox/voice-safety-classifier",
|
| 86 |
+
help="checkpoint file of model",
|
| 87 |
+
)
|
| 88 |
+
args = parser.parse_args()
|
| 89 |
+
labels_name_list = [
|
| 90 |
+
"Profanity",
|
| 91 |
+
"DatingAndSexting",
|
| 92 |
+
"Racist",
|
| 93 |
+
"Bullying",
|
| 94 |
+
"Other",
|
| 95 |
+
"NoViolation",
|
| 96 |
+
]
|
| 97 |
+
# Model is trained on only 16kHz audio
|
| 98 |
+
audio, _ = librosa.core.load(args.audio_file, sr=16000)
|
| 99 |
+
input_np = feature_extract_simple(audio, sr=16000)
|
| 100 |
+
input_pt = torch.Tensor(input_np)
|
| 101 |
+
model = WavLMForSequenceClassification.from_pretrained(
|
| 102 |
+
args.model_path, num_labels=len(labels_name_list)
|
| 103 |
+
)
|
| 104 |
+
probs = infer(model, input_pt)
|
| 105 |
+
probs = probs.reshape(-1, 6).detach().tolist()
|
| 106 |
+
print(f"Probabilities for {args.audio_file} is:")
|
| 107 |
+
for chunk_idx in range(len(probs)):
|
| 108 |
+
print(f"\nSegment {chunk_idx}:")
|
| 109 |
+
for label_idx, label in enumerate(labels_name_list):
|
| 110 |
+
print(f"{label} : {probs[chunk_idx][label_idx]}")
|
requirements.txt
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
torch
|
| 2 |
+
transformers
|
| 3 |
+
librosa
|
| 4 |
+
numpy
|