Update docs

.readthedocs.yml

@@ -0,0 +1,12 @@
+version: 2
+
+build:
+  image: latest
+
+sphinx:
+  configuration: docs/conf.py
+
+python:
+  version: 3.6
+  install:
+  - requirements: docs/requirements.txt

docs/_static/datasets.csv

@@ -1,8 +1,8 @@
-Dataset,Sign Language,Pose Dataset
-:class:`~openhands.datasets.isolated.AUTSLDataset`,`Turkish <https://chalearnlap.cvc.uab.cat/dataset/40/description/>`_,`Download <https://raw.github.com>`_
-:class:`~openhands.datasets.isolated.CSLDataset`,`Chinese <http://home.ustc.edu.cn/~pjh/openresources/cslr-dataset-2015/index.html>`_,`Download <https://raw.github.com>`_
-:class:`~openhands.datasets.isolated.DeviSignDataset`,`Chinese <http://vipl.ict.ac.cn/homepage/ksl/data.html>`_,`Download <https://raw.github.com>`_
-:class:`~openhands.datasets.isolated.GSLDataset`,`Greek <https://vcl.iti.gr/dataset/gsl/>`_,`Download <https://raw.github.com>`_
-:class:`~openhands.datasets.isolated.INCLUDEDataset`,`Indian <https://sign-language.ai4bharat.org/#/INCLUDE>`_,`Download <https://raw.github.com>`_
-:class:`~openhands.datasets.isolated.LSA64Dataset`,`Argentinian <http://facundoq.github.io/datasets/lsa64/>`_,`Download <https://raw.github.com>`_
-:class:`~openhands.datasets.isolated.WLASLDataset`,`American <https://dxli94.github.io/WLASL/>`_,`Download <https://raw.github.com>`_
+Dataset,Sign Language,Download Size
+:class:`~openhands.datasets.isolated.AUTSLDataset`,`Turkish <https://chalearnlap.cvc.uab.cat/dataset/40/description/>`_,`2.1GB <https://raw.github.com>`_
+:class:`~openhands.datasets.isolated.CSLDataset`,`Chinese <http://home.ustc.edu.cn/~pjh/openresources/cslr-dataset-2015/index.html>`_,`11GB <https://raw.github.com>`_
+:class:`~openhands.datasets.isolated.DeviSignDataset`,`Chinese <http://vipl.ict.ac.cn/homepage/ksl/data.html>`_,`2.3GB <https://raw.github.com>`_
+:class:`~openhands.datasets.isolated.GSLDataset`,`Greek <https://vcl.iti.gr/dataset/gsl/>`_,`0.7GB <https://raw.github.com>`_
+:class:`~openhands.datasets.isolated.INCLUDEDataset`,`Indian <https://sign-language.ai4bharat.org/#/INCLUDE>`_,`0.6GB <https://raw.github.com>`_
+:class:`~openhands.datasets.isolated.LSA64Dataset`,`Argentinian <http://facundoq.github.io/datasets/lsa64/>`_,`0.2GB <https://raw.github.com>`_
+:class:`~openhands.datasets.isolated.WLASLDataset`,`American <https://dxli94.github.io/WLASL/>`_,`1GB <https://raw.github.com>`_

docs/_static/raw_datasets.csv

@@ -1,8 +1,6 @@
-Dataset Link,Source,Size (hours)
-pose_no_sub_one_min_splits.h5,ISH News
-pose_sub_based_splits.h5,ISH News
-pose_videos_chunked_1_min.h5,MBM Vadodara
-pose_for_splitted.h5,NIOS Videos
-newshook_dataset.h5,NewsHook
-sign_library_1_min_splits.h5,Sign Library
-sign_library_sub_based.h5,Sign Library
+Source,Total Hours,Download Size
+`ISH News <https://www.youtube.com/c/ISHNews>`_,145,`15GB <>`_
+`MBM Vadodara <https://www.youtube.com/channel/UCptUQkgBpyrvDK31d16lPmA>`_,225,`23GB <>`_
+`National Institute of Open Schooling (NIOS) <https://www.youtube.com/channel/UCXBn5q8Zv4Bz-LZXWWD7Jxw>`_,115,`11GB <>`_
+`NewzHook <https://www.youtube.com/c/NewzHookNewsforDisabledinIndia>`_,615,`58GB <>`_
+`SIGN LIBRARY <https://www.youtube.com/channel/UCaTMaFNkTiwgsSQ7FYaiaTQ/>`_,29,`2.9GB <>`_

docs/api.rst

@@ -0,0 +1,47 @@
+API Documentation
+=================
+
+Tasks
+-----
+
+.. automodule:: openhands.apis
+    :members:
+
+Models
+------
+
+Encoders
+^^^^^^^^
+
+.. automodule:: openhands.models.encoder
+    :members:
+
+Decoders
+^^^^^^^^
+
+.. automodule:: openhands.models.decoder
+    :members:
+
+Datasets
+--------
+
+Augmentations
+^^^^^^^^^^^^^
+
+.. automodule:: openhands.datasets.pose_transforms
+    :members:
+
+ISLR Datasets
+^^^^^^^^^^^^^^^^^^^^^^^
+
+Base class
+++++++++++
+
+.. automodule:: openhands.datasets.isolated.base
+    :members:
+
+Labelled datasets
++++++++++++++++++
+
+.. automodule:: openhands.datasets.isolated
+    :members:

docs/index.rst

@@ -6,35 +6,23 @@
 Welcome to 👐OpenHands's documentation!
 =======================================
 
-**👐OpenHands** is an open source toolkit to democratize sign language research by making pose-based Sign Language Recognition (SLR) more accessible to everyone.
-
-Features
---------
-
-- Easily train pose-based isolated sign language recognizers (ISLR)
-   - Currently supports 4 different models
-- Support for many run-time augmentations for pose keypoints
-   - Check for complete list of pose transforms
-- Support for efficient inference
-   - Check for more details: 
-- Easily pre-train models using monolingual SL data
-   - And fine-tune on any small ISLR dataset
-- All Training and Inference is completely config-based
-   - No-code required
-   - Check for example configs: 
-- Supports 6 sign languages out-of-the-box
-   - Check for supported datasets: 
-
-For detailed explanation, please check out our paper: `👐OpenHands: Making ... <https://arxiv.org>`_
+**👐OpenHands** is an `open source <https://github.com/AI4Bharat/OpenHands>`_ toolkit to democratize sign language research by making pose-based Sign Language Recognition (SLR) more accessible to everyone.
 
 .. toctree::
    :maxdepth: 2
    :caption: Getting Started
 
+   instructions/features
    instructions/installation
    instructions/datasets
    instructions/models
    instructions/training
    instructions/self_supervised
    instructions/inference
    instructions/support
+
+.. toctree::
+   :maxdepth: 2
+   :caption: Code Documentation
+
+   api

docs/instructions/datasets.rst

@@ -28,5 +28,22 @@ You can now directly proceed to train!
 Custom Datasets
 ---------------
 
-To add support for your own dataset, follow the steps below:
+To add support for your own dataset, create a class of the following structure:
 
+.. code:: python
+
+    from .base import BaseIsolatedDataset
+
+    class MyDatasetDataset(BaseIsolatedDataset):
+        def read_glosses(self):
+            self.glosses = ... # Populate the list of all glosses
+
+        def read_original_dataset(self):
+            self.data = ... # Populate the list of all video files and gloss IDs as tuples
+
+        def read_video_data(self, index):
+            # Read the following ...
+            return imgs, label, video_name
+
+- For implementation examples, check `this folder in the source code <https://github.com/AI4Bharat/OpenHands/tree/main/openhands/datasets/isolated>`_
+- This class can now be referenced in your config file appropriately, and used for training or inference.

docs/instructions/features.rst

@@ -0,0 +1,29 @@
+Features
+========
+
+For detailed explanation, please check out our paper: `👐OpenHands: Making ... <https://arxiv.org>`_
+
+Currently supported
+-------------------
+
+- Easily train pose-based isolated sign language recognizers (ISLR)
+   - Currently supports 4 different models
+- Support for many run-time augmentations for pose keypoints
+   - Check for complete list of pose transforms
+- Support for efficient inference
+   - Check for more details: 
+- Easily pre-train models using monolingual SL data
+   - And fine-tune on any small ISLR dataset
+- All Training and Inference is completely config-based
+   - No-code required
+   - Check for example configs: 
+- Supports 6 sign languages out-of-the-box
+   - Check for supported datasets: 
+
+Future Plan
+-----------
+
+- [ ] Add more sign langauges
+- [ ] Explore more SoTA models
+- [ ] Support for sign language detection
+- [ ] Support for continuous sign langauge recognition (CSLR)

docs/instructions/self_supervised.rst

@@ -1,4 +1,4 @@
-Self-supervised Learning
+Self-Supervised Learning
 ========================
 
 This section is dedicated to explaining how to perform pre-training and fine-tuning using any required datasets for ISLR.
@@ -18,8 +18,8 @@ HDF5 format
 Generating HDF5 datasets
 ^^^^^^^^^^^^^^^^^^^^^^^^
 
-- See `this script <https://github.com/AI4Bharat/OpenHands/blob/main/scripts/mediapipe_extract.py>`_ to extract pose for all the given videos using MediaPipe Holistic
-- See `this script <https://github.com/AI4Bharat/OpenHands/blob/main/scripts/pkl_to_h5.py>`_ to convert all the above individual pose files (in ``.pkl``) to HDF5 format.
+- `See this script <https://github.com/AI4Bharat/OpenHands/blob/main/scripts/mediapipe_extract.py>`_ to extract pose for all the given videos using MediaPipe Holistic
+- Use `this script <https://github.com/AI4Bharat/OpenHands/blob/main/scripts/pkl_to_h5.py>`_ to convert all the above individual pose files (in ``.pkl``) to HDF5 format.
 
 Download datasets
 ^^^^^^^^^^^^^^^^^

docs/instructions/support.rst

@@ -1,14 +1,10 @@
 Support
 =======
 
-👐OpenHands is developed by AI4Bharat, an non-profit organization based out of IIT-Madras.
+👐OpenHands is developed by AI4Bharat, a non-profit organization based out of IIT-Madras in India.
 
 For contributions, feedback or bug reports, please create an issue here: `GitHub Issues <https://github.com/AI4Bharat/OpenHands/issues>`_.
 
-Future Plan
------------
-
-- [ ] Add more sign langauges
-- [ ] Explore more SoTA models
-- [ ] Support for sign language detection
-- [ ] Support for continuous sign langauge recognition (CSLR)
+- Primary developers:
+   - Gokul NC (gokulnc at ai4bharat dot org)
+   - Prem Kumar SelvaRaj (prem at ai4bharat dot org)

docs/requirements.txt

@@ -0,0 +1,3 @@
+sphinx
+myst-parser
+sphinx_rtd_theme

openhands/apis/__init__.py

@@ -1 +1,5 @@
 from .classification_model import ClassificationModel
+from .dpc import PretrainingModelDPC
+from .inference import InferenceModel
+
+__all__ = ["ClassificationModel", "PretrainingModelDPC", "InferenceModel"]

openhands/apis/classification_model.py

@@ -2,8 +2,8 @@
 import torch.nn.functional as F
 import torchmetrics
 from ..models.loader import get_model
-from .losses import CrossEntropyLoss, SmoothedCrossEntropyLoss
-from .data import DataModule
+from ..core.losses import CrossEntropyLoss, SmoothedCrossEntropyLoss
+from ..core.data import DataModule
 from .inference import InferenceModel
 
 class ClassificationModel(InferenceModel):

openhands/apis/inference.py

@@ -3,7 +3,7 @@
 from tqdm import tqdm
 import time
 
-from .data import DataModule
+from ..core.data import DataModule
 from ..models.loader import get_model
 
 class InferenceModel(pl.LightningModule):

openhands/core/exp_utils.py

@@ -12,7 +12,7 @@ def experiment_manager(trainer, cfg=None):
     if cfg is None:
         return
 
-    trainer._default_root_dir = os.getcwd()
+    # trainer._default_root_dir = cfg.output_path if cfg.output_path else os.getcwd()
     if cfg.create_tensorboard_logger or cfg.create_wandb_logger:
         configure_loggers(
             trainer,
@@ -66,7 +66,7 @@ def configure_checkpointing(trainer, cfg):
     """
     Creates ModelCheckpoint callback and and attach it to the trainer.
     """
-    checkpoint_callback = ModelCheckpoint(**cfg, dirpath=trainer._default_root_dir)
+    checkpoint_callback = ModelCheckpoint(**cfg)#, dirpath=trainer._default_root_dir)
     trainer.callbacks.append(checkpoint_callback)
 
 

openhands/datasets/isolated/__init__.py

@@ -5,3 +5,5 @@
 from .include import INCLUDEDataset
 from .lsa64 import LSA64Dataset
 from .wlasl import WLASLDataset
+
+__all__ = ["AUTSLDataset", "CSLDataset", "DeviSignDataset", "GSLDataset", "INCLUDEDataset", "LSA64Dataset", "WLASLDataset"]

openhands/models/decoder/__init__.py

@@ -0,0 +1,5 @@
+from .fc import FC as FullyConnectedClassifier
+from .rnn import RNNClassifier
+from .bert_hf import BERT
+
+__all__ = ["FullyConnectedClassifier", "RNNClassifier", "BERT"]

openhands/models/encoder/__init__.py

@@ -0,0 +1,5 @@
+from .graph.pose_flattener import PoseFlattener
+from .graph.decoupled_gcn import DecoupledGCN
+from .graph.st_gcn import STGCN
+
+__all__ = ["PoseFlattener", "DecoupledGCN", "STGCN"]

openhands/models/loader.py

@@ -94,7 +94,7 @@ def load_ssl_backbone(cfg, in_channels, num_class):
         raise ValueError(f"SSL Type '{cfg.type}' not supported.")
 
 def get_model(config, in_channels, num_class):
-    if config.pretrained:
+    if "pretrained" in config:
         # Load self-supervised backbone
         return load_ssl_backbone(config.pretrained, in_channels, num_class)