init

.gitattributes copy

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,160 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/

README copy.md

@@ -0,0 +1,12 @@
+---
+title: SAM SLR V2
+emoji: 📉
+colorFrom: blue
+colorTo: red
+sdk: gradio
+sdk_version: 4.29.0
+app_file: app.py
+pinned: false
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

VSL_SAM_SLR_V2.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:509a2c017b6539366c59ee8a90408f62cec2484bb10078241f0af1845263ad90
+size 16709176

app.py

@@ -0,0 +1,99 @@
+import time
+import pandas as pd
+import gradio as gr
+import onnxruntime as ort
+from mediapipe.python.solutions import holistic
+from utils import get_predictions, preprocess
+
+
+title = '''
+
+'''
+
+cite_markdown = '''
+
+'''
+
+description = '''
+
+'''
+
+examples = [
+    ['000_con_cho.mp4'],
+]
+
+# Load the configuration file.
+ort_session = ort.InferenceSession('VSL_SAM_SLR_V2.onnx')
+
+# Load id-to-gloss mapping.
+id2gloss = pd.read_csv('gloss.csv', names=['id', 'gloss']).to_dict()['gloss']
+
+
+def inference(
+    video: str,
+    progress: gr.Progress = gr.Progress(),
+) -> str:
+    '''
+    Video-based inference for Vietnamese Sign Language recognition.
+
+    Parameters
+    ----------
+    video : str
+        The path to the video.
+    progress : gr.Progress, optional
+        The progress bar, by default gr.Progress()
+
+    Returns
+    -------
+    str
+        The inference message.
+    '''
+    keypoints_detector = holistic.Holistic(
+        static_image_mode=False,
+        model_complexity=2,
+        enable_segmentation=True,
+        refine_face_landmarks=True,
+    )
+
+    progress(0, desc='Preprocessing video')
+    start_time = time.time()
+    inputs = preprocess(
+        source=video,
+        keypoints_detector=keypoints_detector,
+    )
+    end_time = time.time()
+    data_time = end_time - start_time
+
+    progress(1/2, desc='Getting predictions')
+    start_time = time.time()
+    predictions = get_predictions(
+        inputs=inputs, ort_session=ort_session, id2gloss=id2gloss, k=3
+    )
+    end_time = time.time()
+    model_time = end_time - start_time
+
+    if len(predictions) == 0:
+        output_message = 'No sign language detected in the video. Please try again.'
+    else:
+        output_message = 'The top-3 predictions are:\n'
+        for i, prediction in enumerate(predictions):
+            output_message += f'\t{i+1}. {prediction["label"]} ({prediction["score"]:2f})\n'
+        output_message += f'Data processing time: {data_time:.2f} seconds\n'
+        output_message += f'Model inference time: {model_time:.2f} seconds\n'
+        output_message += f'Total time: {data_time + model_time:.2f} seconds'
+
+    progress(1/2, desc='Completed')
+
+    return output_message
+
+
+iface = gr.Interface(
+    fn=inference,
+    inputs='video',
+    outputs='text',
+    examples=examples,
+    title=title,
+    description=description,
+)
+iface.launch()
+# print(inference('000_con_cho.mp4'))

gloss.csv

@@ -0,0 +1,100 @@
+0,Con chó
+1,Con mèo
+2,Con gà
+3,Con vịt
+4,Con rùa
+5,Con thỏ
+6,Con trâu
+7,Con bò
+8,Con dê
+9,Con heo
+10,Màu đen
+11,Màu trắng
+12,Màu đỏ
+13,Màu cam
+14,Màu vàng
+15,Màu lá cây
+16,Màu da trời
+17,Màu hồng
+18,Màu tím
+19,Màu nâu
+20,Quả dâu
+21,Quả mận
+22,Quả dứa
+23,Quả đào
+24,Quả đu đủ
+25,Quả cam
+26,Quả bơ
+27,Quả chuối
+28,Quả xoài
+29,Quả dừa
+30,Bố
+31,Mẹ
+32,Con trai
+33,Con gái
+34,Vợ
+35,Chồng
+36,Ông nội
+37,Bà nội
+38,Ông ngoại
+39,Bà ngoại
+40,Ăn
+41,Uống
+42,Xem
+43,Thèm
+44,Mách
+45,Khóc
+46,Cười
+47,Học
+48,Dỗi
+49,Chết
+50,Đi
+51,Chạy
+52,Bận
+53,Hát
+54,Múa
+55,Nấu
+56,Nướng
+57,Nhầm lẫn
+58,Quan sát
+59,Cắm trại
+60,Cung cấp
+61,Bắt chước
+62,Bắt buộc
+63,Báo cáo
+64,Mua bán
+65,Không quen
+66,Không nên
+67,Không cần
+68,Không cho
+69,Không nghe lời
+70,Mặn
+71,Đắng
+72,Cay
+73,Ngọt
+74,Đậm
+75,Nhạt
+76,Ngon miệng
+77,Xấu
+78,Đẹp
+79,Chật
+80,Hẹp
+81,Rộng
+82,Dài
+83,Cao
+84,Lùn
+85,Ốm
+86,Mập
+87,Ngoan
+88,Hư
+89,Khỏe
+90,Mệt
+91,Đau
+92,Giỏi
+93,Chăm chỉ
+94,Lười biếng
+95,Tốt bụng
+96,Thú vị
+97,Hài hước
+98,Dũng cảm
+99,Sáng tạo

requirements.txt

@@ -0,0 +1,9 @@
+torch
+torchvision
+numpy
+opencv-python
+mediapipe
+timm
+einops
+yacs
+onnxruntime

utils/data.py

@@ -0,0 +1,325 @@
+import cv2
+import random
+import numpy as np
+from mediapipe.python.solutions import pose
+
+
+SELECTED_JOINTS = {
+    27: {
+        'pose': [0, 11, 12, 13, 14, 15, 16],
+        'hand': [0, 4, 5, 8, 9, 12, 13, 16, 17, 20],
+    },  # 27
+}
+
+
+def pad(joints: np.ndarray, num_frames: int = 150) -> np.ndarray:
+    '''
+    Add padding to the joints.
+
+    Parameters
+    ----------
+    joints : np.ndarray
+        The joints to pad.
+    num_frames : int, default=150
+        The number of frames to pad.
+
+    Returns
+    -------
+    np.ndarray
+        The padded joints.
+    '''
+    if joints.shape[0] < num_frames:
+        L = joints.shape[0]
+        padded_joints = np.zeros((num_frames, joints.shape[1], joints.shape[2]))
+        padded_joints[:L, :, :] = joints
+        rest = num_frames - L
+        num = int(np.ceil(rest / L))
+        pad = np.concatenate([joints for _ in range(num)], 0)[:rest]
+        padded_joints[L:, :, :] = pad
+    else:
+        padded_joints = joints[:num_frames]
+    return padded_joints
+
+
+def extract_joints(
+    source: str,
+    keypoints_detector,
+    resize_to: tuple = (256, 256),
+    num_joints: int = 27,
+    num_frames: int = 150,
+    num_bodies: int = 1,
+    num_channels: int = 3,
+) -> np.ndarray:
+    '''
+    Extract the joints from the video.
+
+    Parameters
+    ----------
+    source : str
+        The path to the video.
+    keypoints_detector : mediapipe.solutions.holistic.Holistic
+        The keypoints detector.
+    resize_to : tuple, default=(256, 256)
+        The size to resize the image.
+    num_joints : int, default=27
+        The number of joints.
+    num_frames : int, default=150
+        The number of frames.
+    num_bodies : int, default=1
+        The number of bodies.
+    num_channels : int, default=3
+        The number of channels.
+
+    Returns
+    -------
+    np.ndarray
+        The extracted joints.
+    '''
+    cap = cv2.VideoCapture(source)
+
+    extracted_joints = []
+    while cap.isOpened():
+        success, image = cap.read()
+        if not success:
+            break
+        image = cv2.resize(image, resize_to)
+        image = cv2.flip(image, flipCode=1)
+        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+        frame_joints = []
+
+        results = keypoints_detector.process(image)
+
+        pose = [(0.0, 0.0, 0.0)] * len(SELECTED_JOINTS[num_joints]['pose'])
+        if results.pose_landmarks is not None:
+            pose = [
+                (landmark.x * resize_to[0], landmark.y * resize_to[1], landmark.visibility)
+                for i, landmark in enumerate(results.pose_landmarks.landmark)
+                if i in SELECTED_JOINTS[num_joints]['pose']
+            ]
+        frame_joints.extend(pose)
+
+        left_hand = [(0.0, 0.0, 0.0)] * len(SELECTED_JOINTS[num_joints]['hand'])
+        if results.left_hand_landmarks is not None:
+            left_hand = [
+                (landmark.x * resize_to[0], landmark.y * resize_to[1], landmark.visibility)
+                for i, landmark in enumerate(results.left_hand_landmarks.landmark)
+                if i in SELECTED_JOINTS[num_joints]['hand']
+            ]
+        frame_joints.extend(left_hand)
+
+        right_hand = [(0.0, 0.0, 0.0)] * len(SELECTED_JOINTS[num_joints]['hand'])
+        if results.right_hand_landmarks is not None:
+            right_hand = [
+                (landmark.x * resize_to[0], landmark.y * resize_to[1], landmark.visibility)
+                for i, landmark in enumerate(results.right_hand_landmarks.landmark)
+                if i in SELECTED_JOINTS[num_joints]['hand']
+            ]
+        frame_joints.extend(right_hand)
+
+        assert len(frame_joints) == num_joints, \
+            f'Expected {num_joints} joints, got {len(frame_joints)} joints.'
+        extracted_joints.append(frame_joints)
+
+    extracted_joints = np.array(extracted_joints)
+    extracted_joints = pad(extracted_joints, num_frames=num_frames)
+
+    fp = np.zeros(
+        (num_frames, num_joints, num_channels, num_bodies),
+        dtype=np.float32,
+    )
+    fp[:, :, :, 0] = extracted_joints
+
+    return np.transpose(fp, [2, 0, 1, 3])
+
+
+def preprocess(
+    source: str,
+    data_args: dict,
+    keypoints_detector,
+    device: str = 'cpu',
+) -> torch.Tensor:
+    '''
+    Preprocess the video.
+
+    Parameters
+    ----------
+    source : str
+        The path to the video.
+
+    Returns
+    -------
+    dict
+        The model inputs.
+    '''
+    print('Extracting joints from pose...')
+    inputs = extract_joints_new(source=source, keypoints_detector=keypoints_detector)
+    T = inputs.shape[1]
+    print('Sampling video...')
+    if data_args['random_choose']:
+        inputs = random_sample_np(inputs, data_args['window_size'])
+    else:
+        inputs = uniform_sample_np(inputs, data_args['window_size'])
+
+    print('Normalizing video...')
+    print(inputs.shape, inputs)
+    return np.squeeze(inputs).transpose(1, 2, 0).astype(np.float32)
+
+
+def random_sample_np(data: np.ndarray, size: int) -> np.ndarray:
+    '''
+    Sample the data randomly.
+
+    Parameters
+    ----------
+    data : np.ndarray
+        The data to sample.
+    size : int
+        The size of the data to sample.
+
+    Returns
+    -------
+    np.ndarray
+        The sampled data.
+    '''
+    C, T, V, M = data.shape
+    if T == size:
+        return data
+    interval = int(np.ceil(size / T))
+    random_list = sorted(random.sample(list(range(T))*interval, size))
+    return data[:, random_list]
+
+
+def uniform_sample_np(data: np.ndarray, size: int) -> np.ndarray:
+    '''
+    Sample the data uniformly.
+
+    Parameters
+    ----------
+    data : np.ndarray
+        The data to sample.
+    size : int
+        The size of the data to sample.
+
+    Returns
+    -------
+    np.ndarray
+        The sampled data.
+    '''
+    C, T, V, M = data.shape
+    if T == size:
+        return data
+    interval = T / size
+    uniform_list = [int(i * interval) for i in range(size)]
+    return data[:, uniform_list]
+
+
+def calculate_angle(
+    shoulder: list,
+    elbow: list,
+    wrist: list,
+) -> float:
+    '''
+    Calculate the angle between the shoulder, elbow, and wrist.
+
+    Parameters
+    ----------
+    shoulder : list
+        Shoulder coordinates.
+    elbow : list
+        Elbow coordinates.
+    wrist : list
+        Wrist coordinates.
+
+    Returns
+    -------
+    float
+        Angle in degree between the shoulder, elbow, and wrist.
+    '''
+    shoulder = np.array(shoulder)
+    elbow = np.array(elbow)
+    wrist = np.array(wrist)
+
+    radians = np.arctan2(wrist[1] - elbow[1], wrist[0] - elbow[0]) \
+        - np.arctan2(shoulder[1] - elbow[1], shoulder[0] - elbow[0])
+    angle = np.abs(radians * 180.0 / np.pi)
+
+    if angle > 180.0:
+        angle = 360 - angle
+    return angle
+
+
+def do_hands_relax(
+    pose_landmarks: list,
+    angle_threshold: float = 160.0,
+) -> bool:
+    '''
+    Check if the hand is down.
+
+    Parameters
+    ----------
+    hand_landmarks : list
+        Hand landmarks.
+    angle_threshold : float, optional
+        Angle threshold, by default 160.0.
+
+    Returns
+    -------
+    bool
+        True if the hand is down, False otherwise.
+    '''
+    if pose_landmarks is None:
+        return True
+
+    landmarks = pose_landmarks.landmark
+    left_shoulder = [
+        landmarks[pose.PoseLandmark.LEFT_SHOULDER.value].x,
+        landmarks[pose.PoseLandmark.LEFT_SHOULDER.value].y,
+        landmarks[pose.PoseLandmark.LEFT_SHOULDER.value].visibility,
+    ]
+    left_elbow = [
+        landmarks[pose.PoseLandmark.LEFT_ELBOW.value].x,
+        landmarks[pose.PoseLandmark.LEFT_ELBOW.value].y,
+        landmarks[pose.PoseLandmark.LEFT_SHOULDER.value].visibility,
+    ]
+    left_wrist = [
+        landmarks[pose.PoseLandmark.LEFT_WRIST.value].x,
+        landmarks[pose.PoseLandmark.LEFT_WRIST.value].y,
+        landmarks[pose.PoseLandmark.LEFT_SHOULDER.value].visibility,
+    ]
+    left_angle = calculate_angle(left_shoulder, left_elbow, left_wrist)
+
+    right_shoulder = [
+        landmarks[pose.PoseLandmark.RIGHT_SHOULDER.value].x,
+        landmarks[pose.PoseLandmark.RIGHT_SHOULDER.value].y,
+        landmarks[pose.PoseLandmark.RIGHT_SHOULDER.value].visibility,
+    ]
+    right_elbow = [
+        landmarks[pose.PoseLandmark.RIGHT_ELBOW.value].x,
+        landmarks[pose.PoseLandmark.RIGHT_ELBOW.value].y,
+        landmarks[pose.PoseLandmark.RIGHT_SHOULDER.value].visibility,
+    ]
+    right_wrist = [
+        landmarks[pose.PoseLandmark.RIGHT_WRIST.value].x,
+        landmarks[pose.PoseLandmark.RIGHT_WRIST.value].y,
+        landmarks[pose.PoseLandmark.RIGHT_SHOULDER.value].visibility,
+    ]
+    right_angle = calculate_angle(right_shoulder, right_elbow, right_wrist)
+
+    is_visible = all(
+        [
+            left_shoulder[2] > 0,
+            left_elbow[2] > 0,
+            left_wrist[2] > 0,
+            right_shoulder[2] > 0,
+            right_elbow[2] > 0,
+            right_wrist[2] > 0,
+        ]
+    )
+
+    return all(
+        [
+            is_visible,
+            left_angle < angle_threshold,
+            right_angle < angle_threshold,
+        ]
+    )

utils/model.py

@@ -0,0 +1,47 @@
+import torch
+import numpy as np
+import onnxruntime as ort
+
+
+def get_predictions(
+    inputs: np.ndarray,
+    ort_session: ort.InferenceSession,
+    id2gloss: dict,
+    k: int = 3,
+) -> list:
+    '''
+    Get the top-k predictions.
+
+    Parameters
+    ----------
+    inputs : dict
+        Model inputs.
+    ort_session : ort.InferenceSession
+        ONNX Runtime session.
+    id2gloss : dict
+        Mapping from class index to class label.
+    k : int, optional
+        Number of predictions to return, by default 3.
+
+    Returns
+    -------
+    list
+        Top-k predictions.
+    '''
+    if inputs is None:
+        return []
+
+    logits = torch.from_numpy(ort_session.run(None, {'x': inputs})[0])
+
+    # Get top-3 predictions
+    topk_scores, topk_indices = torch.topk(logits, k, dim=1)
+    topk_scores = torch.nn.functional.softmax(topk_scores, dim=1).squeeze().detach().numpy()
+    topk_indices = topk_indices.squeeze().detach().numpy()
+
+    return [
+        {
+            'label': id2gloss[topk_indices[i]],
+            'score': topk_scores[i],
+        }
+        for i in range(k)
+    ]