### Install CLIP Library

Source: https://github.com/open-mmlab/mmaction2/blob/main/projects/actionclip/README.md

Installs the CLIP library from its GitHub repository. This is a prerequisite for using ActionCLIP.

```shell
pip install git+https://github.com/openai/CLIP.git
```

--------------------------------

### Install and Inspect Tree Utility

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/mmaction2_tutorial.ipynb

Installs the 'tree' utility to visualize directory structures and then displays the structure of the 'kinetics400_tiny' dataset.

```bash
!apt-get -q install tree
!tree kinetics400_tiny
```

--------------------------------

### Install MMRazor from Source

Source: https://github.com/open-mmlab/mmaction2/blob/main/projects/knowledge_distillation/README.md

Clone the MMRazor repository and install it from source. This method allows for direct modifications and development.

```shell
git clone https://github.com/open-mmlab/mmrazor.git
cd mmrazor
pip install -v -e .
```

--------------------------------

### Install MMAction2 and Dependencies

Source: https://github.com/open-mmlab/mmaction2/blob/main/projects/stad_tutorial/demo_stad.ipynb

Installs necessary packages like openmim, mmengine, mmcv, and mmdet. Clones the MMAction2 repository and installs it in editable mode. Navigates to the tutorial directory.

```python
%pip install -U openmim
!mim install mmengine
!mim install mmcv
!mim install mmdet

!git clone https://github.com/open-mmlab/mmaction2.git

%cd mmaction2
%pip install -v -e .
%cd projects/stad_tutorial
```

--------------------------------

### Complete Finetuning Configuration Example

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/user_guides/finetune.md

This is a complete configuration file example for finetuning, integrating all the previously discussed modifications for model head, dataset, training schedule, and pre-trained model loading.

```python
_base_ = [
    '../../_base_/models/tsn_r50.py', '../../_base_/schedules/sgd_100e.py',
    '../../_base_/default_runtime.py'
]

# model settings
model = dict(
    cls_head=dict(
        type='TSNHead',
        num_classes=101  # change from 400 to 101
        ))


```

--------------------------------

### Compose Pipeline Configuration Example

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/guide_to_framework.md

Example configuration for an MMAction2 video processing pipeline using Compose. Demonstrates initialization and sampling transforms.

```python
import os.path as osp
from mmengine.dataset import Compose

pipeline_cfg = [
    dict(type='VideoInit'),
    dict(type='VideoSample', clip_len=16, num_clips=1, test_mode=False),

```

--------------------------------

### Start Training Runner

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/mmaction2_tutorial.ipynb

Initiates the training process using the configured runner.

```python
runner.train()
```

--------------------------------

### Example: Train TSN Model on Kinetics-400

Source: https://github.com/open-mmlab/mmaction2/blob/main/configs/recognition/tsn/README.md

Example command for training a TSN model on the Kinetics-400 dataset with deterministic settings. Ensure the configuration file path is correct.

```shell
python tools/train.py configs/recognition/tsn/tsn_imagenet-pretrained-r50_8xb32-1x1x3-100e_kinetics400-rgb.py  \
    --seed=0 --deterministic
```

--------------------------------

### Install MMENGINE

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/mmaction2_tutorial.ipynb

Installs the MMENGINE package, a foundational library for OpenMMLab projects. This command fetches the package from a specific index URL.

```bash
pip install mmengine -f https://download.openmmlab.com/mmcv/dist/cu118/torch2.0.0/index.html
```

--------------------------------

### Example Test Command for MViT on Kinetics-400

Source: https://github.com/open-mmlab/mmaction2/blob/main/configs/recognition/mvit/README.md

This example demonstrates how to test an MViT model on the Kinetics-400 dataset and save the results to a pickle file. Ensure the paths to the config and checkpoint files are correct.

```shell
python tools/test.py configs/recognition/mvit/mvit-small-p244_16x4x1_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl
```

--------------------------------

### Install Tree Utility

Source: https://github.com/open-mmlab/mmaction2/blob/main/projects/stad_tutorial/demo_stad.ipynb

Installs the 'tree' command-line utility, which is useful for visualizing directory structures. This is a prerequisite for the subsequent 'tree data' command.

```bash
!apt-get -q install tree
```

--------------------------------

### Install OpenMIM

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/mmaction2_tutorial.ipynb

Installs the OpenMIM package, a dependency for managing models and configurations. Ensure you have a stable internet connection.

```bash
pip install -r requirements/optional.txt
```

--------------------------------

### Install MMAction2 from Source

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/installation.md

Clone the MMAction2 repository and install it in editable mode using pip. This method is recommended for development and customization.

```shell
git clone https://github.com/open-mmlab/mmaction2.git
cd mmaction2
pip install -v -e .
```

--------------------------------

### Example: Test LFB Model on AVA and Dump Results

Source: https://github.com/open-mmlab/mmaction2/blob/main/configs/detection/lfb/README.md

Example command for testing an LFB model on the AVA dataset with half-precision long-term feature banks. Results are dumped to a pkl file.

```shell
python tools/test.py configs/detection/lfb/slowonly-lfb-nl_kinetics400-pretrained-r50_8xb12-4x16x1-20e_ava21-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl
```

--------------------------------

### Configure SGD Optimizer

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/advanced_guides/customize_optimizer.md

Use 'optim_wrapper' to configure optimization strategies. This example sets up a SGD optimizer with a specific learning rate and weight decay.

```python
optim_wrapper = dict(
    type='OptimWrapper',
    optimizer=dict(type='SGD', lr=0.0003, weight_decay=0.0001)
)
```

--------------------------------

### Example: Train LFB Model on AVA with Half-Precision Features

Source: https://github.com/open-mmlab/mmaction2/blob/main/configs/detection/lfb/README.md

Example command for training an LFB model on the AVA dataset using half-precision long-term feature banks. Includes setting a random seed and enabling deterministic operations.

```shell
python tools/train.py configs/detection/lfb/slowonly-lfb-nl_kinetics400-pretrained-r50_8xb12-4x16x1-20e_ava21-rgb.py \
  --seed 0 --deterministic
```

--------------------------------

### Example: Test TSN Model and Dump Results

Source: https://github.com/open-mmlab/mmaction2/blob/main/configs/recognition/tsn/README.md

Example command for testing a TSN model on the Kinetics-400 dataset and saving the results to a pkl file. Replace 'SOME_CHECKPOINT.pth' with your actual checkpoint file.

```shell
python tools/test.py configs/recognition/tsn/tsn_imagenet-pretrained-r50_8xb32-1x1x3-100e_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl
```

--------------------------------

### Train a model with MMAction2

Source: https://github.com/open-mmlab/mmaction2/blob/main/configs/detection/slowfast/README.md

Use this command to start the training process. Specify the configuration file and any optional arguments for customization.

```shell
python tools/train.py ${CONFIG_FILE} [optional arguments]
```

```shell
python tools/train.py configs/detection/slowfast/slowfast_kinetics400-pretrained-r50_8xb16-4x16x1-20e_ava21-rgb.py \
    --seed 0 --deterministic
```

--------------------------------

### Initialize and Train Recognizer

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/mmaction2_tutorial.ipynb

Initializes the work directory and builds the runner from the configuration object to start training.

```python
import os.path as osp
import mmengine
from mmengine.runner import Runner

# Create work_dir
mmengine.mkdir_or_exist(osp.abspath(cfg.work_dir))

# build the runner from config
runner = Runner.from_cfg(cfg)
```

--------------------------------

### Initialize Pre-commit Hook

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/contribution_guide.md

Install the pre-commit hook in the repository. This ensures that code linters and formatters are automatically enforced on every commit.

```shell
pre-commit install
```

--------------------------------

### Demo with STGCN and Faster R-CNN

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/README.md

Example command to run the skeleton-based action recognition demo using STGCN as the action recognizer and Faster R-CNN for human detection. This configuration uses different action recognition checkpoints and configs compared to the PoseC3D example.

```shell
python demo/demo_skeleton.py demo/demo_skeleton.mp4 demo/demo_skeleton_out.mp4 \
    --config configs/skeleton/stgcn/stgcn_8xb16-joint-u100-80e_ntu60-xsub-keypoint-2d.py \
    --checkpoint https://download.openmmlab.com/mmaction/v1.0/skeleton/stgcn/stgcn_8xb16-joint-u100-80e_ntu60-xsub-keypoint-2d/stgcn_8xb16-joint-u100-80e_ntu60-xsub-keypoint-2d_20221129-484a394a.pth \
    --det-config demo/demo_configs/faster-rcnn_r50_fpn_2x_coco_infer.py \
    --det-checkpoint http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth \
    --det-score-thr 0.9 \
    --det-cat-id 0 \
    --pose-config demo/demo_configs/td-hm_hrnet-w32_8xb64-210e_coco-256x192_infer.py \
    --pose-checkpoint https://download.openmmlab.com/mmpose/top_down/hrnet/hrnet_w32_coco_256x192-c78dce93_20200708.pth \
    --label-map tools/data/skeleton/label_map_ntu60.txt
```

--------------------------------

### Start Multi-GPU Training

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/user_guides/train_test.md

Use this script for multi-GPU training on a single machine. Specify the config file, number of GPUs, and optional Python arguments.

```shell
bash tools/dist_train.sh ${CONFIG} ${GPUS} [PY_ARGS]
```

--------------------------------

### Download and Preprocess Something-Something V2 with MIM

Source: https://github.com/open-mmlab/mmaction2/blob/main/tools/data/sthv2/README.md

Use OpenXlab CLI tools to download and preprocess the Something-Something V2 dataset. Ensure you have installed openxlab and logged in.

```bash
# install OpenXlab CLI tools
pip install -U openxlab
# log in OpenXLab
openxlab login
# download and preprocess by MIM
mim download mmaction2 --dataset sthv2
```

--------------------------------

### Install MMCV

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/faq.md

Instructions for uninstalling existing mmcv and installing the correct version. Refer to the official installation guide for detailed steps.

```bash
pip uninstall mmcv

```

--------------------------------

### RecognizerZelda Initialization and Usage Example

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/guide_to_framework.md

Demonstrates how to build, train, and test the RecognizerZelda model using a configuration dictionary. Includes data preprocessing and weight initialization.

```python
import torch
import copy
from mmaction.registry import MODELS

model_cfg = dict(
    type='RecognizerZelda',
    backbone=dict(type='BackBoneZelda'),
    cls_head=dict(
        type='ClsHeadZelda',
        num_classes=2,
        in_channels=128,
        average_clips='prob'),
    data_preprocessor = dict(
        type='DataPreprocessorZelda',
        mean=[123.675, 116.28, 103.53],
        std=[58.395, 57.12, 57.375]))

model = MODELS.build(model_cfg)

# Train
model.train()
model.init_weights()
data_batch_train = copy.deepcopy(batched_packed_results)
data = model.data_preprocessor(data_batch_train, training=True)
loss = model(**data, mode='loss')
print('loss dict: ', loss)

# Test
with torch.no_grad():
    model.eval()
    data_batch_test = copy.deepcopy(batched_packed_results)
    data = model.data_preprocessor(data_batch_test, training=False)
    predictions = model(**data, mode='predict')
print('Label of Sample[0]', predictions[0].gt_label)
print('Scores of Sample[0]', predictions[0].pred_score)
```

--------------------------------

### Train with Single GPU

Source: https://github.com/open-mmlab/mmaction2/blob/main/projects/knowledge_distillation/README.md

Command to start training a knowledge distillation model using a single GPU. Specify the configuration file for the desired distillation setup.

```bash
mim train mmrazor configs/kd_logits_tsm-res50_tsm-mobilenetv2_8xb16_k400.py
```

--------------------------------

### Visualizer Configuration

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/mmaction2_tutorial.ipynb

Sets up the visualizer, which utilizes the configured visualization backends.

```python
visualizer = dict(
```

--------------------------------

### Define Training Pipeline with Decord - MMAction2

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/advanced_guides/customize_pipeline.md

Example of a typical training data pipeline using Decord for video loading and various transformations. Ensure Decord is installed for this pipeline.

```python
train_pipeline = [
    dict(type='DecordInit',),
    dict(type='SampleFrames', clip_len=32, frame_interval=2, num_clips=1),
    dict(type='DecordDecode'),
    dict(type='Resize', scale=(-1, 256)),
    dict(type='RandomResizedCrop'),
    dict(type='Resize', scale=(224, 224), keep_ratio=False),
    dict(type='Flip', flip_ratio=0.5),
    dict(type='FormatShape', input_format='NCTHW'),
    dict(type='PackActionInputs')
]
```

--------------------------------

### Building and Indexing Training Dataset

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/guide_to_framework.md

Demonstrates how to build the training dataset using the configured dictionary and access the first data sample by index. It then prints the shape of the inputs and metadata.

```python
from mmaction.registry import DATASETS

train_dataset = DATASETS.build(train_dataset_cfg)

packed_results = train_dataset[0]

inputs = packed_results['inputs']
data_sample = packed_results['data_samples']

print('shape of the inputs: ', inputs.shape)

# Get metainfo of the inputs
print('image_shape: ', data_sample.img_shape)
print('num_clips: ', data_sample.num_clips)
print('clip_len: ', data_sample.clip_len)
```

--------------------------------

### Install Core Dependencies with MIM

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/installation.md

Install MMEngine, MMCV, MMDetection, and MMPose using the MIM package manager. This is a prerequisite for installing MMAction2.

```shell
pip install -U openmim
mim install mmengine
mim install mmcv
mim install mmdet
mim install mmpose
```

--------------------------------

### Recognize Video and Output to File

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/README.md

Example showing how to run the video demo script and save the output to a specified MP4 file using the `--out-filename` argument.

```shell
# The demo.mp4 and label_map_k400.txt are both from Kinetics-400
python demo/demo.py demo/demo_configs/tsn_r50_1x1x8_video_infer.py \
    checkpoints/tsn_r50_8xb32-1x1x8-100e_kinetics400-rgb_20220818-2692d16c.pth \
    demo/demo.mp4 tools/data/kinetics/label_map_k400.txt --out-filename demo/demo_out.mp4
```

--------------------------------

### Install MMAction2 Dependencies

Source: https://github.com/open-mmlab/mmaction2/blob/main/projects/stad_tutorial/demo_stad_zh_CN.ipynb

Installs necessary packages including openmim, mmengine, mmcv, and mmdet using mim install. This is a prerequisite for setting up MMAction2.

```bash
%pip install -U openmim
!mim install mmengine
!mim install mmcv
!mim install mmdet
```

--------------------------------

### Demo with PoseC3D and Faster R-CNN

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/README.md

Example command to run the skeleton-based action recognition demo using PoseC3D as the action recognizer and Faster R-CNN for human detection. Ensure all specified config and checkpoint paths are correct.

```shell
python demo/demo_skeleton.py demo/demo_skeleton.mp4 demo/demo_skeleton_out.mp4 \
    --config configs/skeleton/posec3d/slowonly_r50_8xb16-u48-240e_ntu60-xsub-keypoint.py \
    --checkpoint https://download.openmmlab.com/mmaction/skeleton/posec3d/slowonly_r50_u48_240e_ntu60_xsub_keypoint/slowonly_r50_u48_240e_ntu60_xsub_keypoint-f3adabf1.pth \
    --det-config demo/demo_configs/faster-rcnn_r50_fpn_2x_coco_infer.py \
    --det-checkpoint http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth \
    --det-score-thr 0.9 \
    --det-cat-id 0 \
    --pose-config demo/demo_configs/td-hm_hrnet-w32_8xb64-210e_coco-256x192_infer.py \
    --pose-checkpoint https://download.openmmlab.com/mmpose/top_down/hrnet/hrnet_w32_coco_256x192-c78dce93_20200708.pth \
    --label-map tools/data/skeleton/label_map_ntu60.txt
```

--------------------------------

### Troubleshoot 'Please install XXCODEBASE' Error

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/zh_cn/get_started/faq.md

This error indicates a missing dependency. Ensure mmcv and mmengine are installed before other OpenMMLAB codebases, following the official installation tutorial.

--------------------------------

### Run Webcam Demo

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/README.md

Execute the webcam demo script with specified configuration, checkpoint, and label files. Optional arguments allow customization of device, camera ID, and prediction thresholds.

```shell
python demo/webcam_demo.py ${CONFIG_FILE} ${CHECKPOINT_FILE} ${LABEL_FILE} \
    [--device ${DEVICE_TYPE}] [--camera-id ${CAMERA_ID}] [--threshold ${THRESHOLD}] \
    [--average-size ${AVERAGE_SIZE}] [--drawing-fps ${DRAWING_FPS}] [--inference-fps ${INFERENCE_FPS}]
```

--------------------------------

### Install PyTorch on GPU Platforms

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/installation.md

Install PyTorch and torchvision using Conda. This command automatically installs the latest compatible versions of PyTorch and cudatoolkit. Verify compatibility with your environment.

```shell
conda install pytorch torchvision -c pytorch
```

--------------------------------

### Inference Result Example

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/quick_run.md

This is an example of the expected output format for inference results.

```json
{'predictions': [{'rec_labels': [[6]], 'rec_scores': [[...]]}]}
```

--------------------------------

### Visualizer Configuration

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/mmaction2_tutorial.ipynb

Sets up the visualizer backend and type for action visualization.

```python
vis_backends = [dict(type='LocalVisBackend')]
visualizer = dict(
    type='ActionVisualizer',
    vis_backends=[dict(type='LocalVisBackend')])
```

--------------------------------

### Example Metric Output

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/guide_to_framework.md

This is an example of the output format for the AccuracyMetric, showing top-1 and top-5 accuracy values.

```shell
OrderedDict([('topk1', 0.5), ('topk5', 1.0)])
```

--------------------------------

### Visualizer Configuration with Multiple Backends

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/migration.md

Sets up the visualizer with a specified type and a list of visualization backends, including local and TensorBoard options.

```python
visualizer = dict(
    type='ActionVisualizer',
    vis_backends=[
        dict(type='LocalVisBackend'),
        # Uncomment the below line to save the log and visualization results to TensorBoard.
        # dict(type='TensorboardVisBackend')
    ]
)
```

--------------------------------

### Verify PyTorch, MMAction2, MMCV, and MMEngine Installation

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/mmaction2_tutorial.ipynb

This script checks the installation and versions of PyTorch (including CUDA availability), MMAction2, MMCV (compiling CUDA version and compiler version), and MMEngine. Ensure these libraries are installed correctly before proceeding.

```python
# Check Pytorch installation
import torch, torchvision
print(torch.__version__, torch.cuda.is_available())

# Check MMAction2 installation
import mmaction
print(mmaction.__version__)

# Check MMCV installation
from mmcv.ops import get_compiling_cuda_version, get_compiler_version
print(get_compiling_cuda_version())
print(get_compiler_version())

# Check MMEngine installation
from mmengine.utils.dl_utils import collect_env
print(collect_env())
```

--------------------------------

### Download Demo Files

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/installation.md

Download the necessary configuration and checkpoint files for running the inference demo using MIM. Specify the config name and destination directory.

```shell
mim download mmaction2 --config tsn_imagenet-pretrained-r50_8xb32-1x1x8-100e_kinetics400-rgb --dest .
```

--------------------------------

### Install Pre-commit Hook

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/contribution_guide.md

Install the pre-commit package to manage hooks. This is a prerequisite for enforcing code style and formatting.

```shell
pip install -U pre-commit
```

--------------------------------

### Video Demo with Output and Customization

Source: https://context7.com/open-mmlab/mmaction2/llms.txt

Performs a video demo, optionally saving the output video and customizing visualization parameters like font size and color.

```bash
python demo/demo.py \
    demo/demo_configs/tsn_r50_1x1x8_video_infer.py \
    checkpoints/tsn_r50.pth \
    demo/demo.mp4 \
    tools/data/kinetics/label_map_k400.txt \
    --out-filename demo/output.mp4
```

```bash
python demo/demo.py \
    demo/demo_configs/tsn_r50_1x1x8_video_infer.py \
    checkpoints/tsn_r50.pth \
    demo/demo.mp4 \
    tools/data/kinetics/label_map_k400.txt \
    --out-filename demo/output.mp4 \
    --font-scale 1.0 \
    --font-color white \
    --target-resolution 1280 720
```

--------------------------------

### Install MMAction2 with Conda

Source: https://github.com/open-mmlab/mmaction2/blob/main/README.md

Follow these steps to install MMAction2 using Conda. Ensure your PyTorch and CUDA versions are compatible.

```shell
conda create --name openmmlab python=3.8 -y
conda activate openmmlab
conda install pytorch torchvision -c pytorch  # This command will automatically install the latest version PyTorch and cudatoolkit, please check whether they match your environment.
pip install -U openmim
mim install mmengine
mim install mmcv
mim install mmdet  # optional
mim install mmpose  # optional
git clone https://github.com/open-mmlab/mmaction2.git
cd mmaction2
pip install -v -e .
```

--------------------------------

### Clone and Install MMAction2

Source: https://github.com/open-mmlab/mmaction2/blob/main/projects/stad_tutorial/demo_stad_zh_CN.ipynb

Clones the MMAction2 repository from GitHub and installs it in editable mode. This command should be run after navigating into the cloned directory.

```bash
!git clone https://github.com/open-mmlab/mmaction2.git

%cd mmaction2
%pip install -v -e .
%cd projects/stad_tutorial
```

--------------------------------

### Optimizer and Learning Rate Scheduler

Source: https://github.com/open-mmlab/mmaction2/blob/main/projects/stad_tutorial/demo_stad.ipynb

Sets up the learning rate scheduler and the optimizer wrapper. This example uses a ConstantLR scheduler and an SGD optimizer with specified learning rate, momentum, and weight decay.

```python
param_scheduler = [
    dict(type='ConstantLR', factor=1.0, by_epoch=False, begin=0, end=500)
]
```

```python
optim_wrapper = dict(
    type='OptimWrapper',
    optimizer=dict(type='SGD', lr=0.005, momentum=0.9, weight_decay=0.0001))
```

--------------------------------

### Action Visualizer Configuration

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/user_guides/config.md

Sets up the action visualizer, specifying the type and the visualization backends to be employed.

```python
visualizer = dict(
    type='ActionVisualizer',
    vis_backends=vis_backends)
```

--------------------------------

### Basic Video Demo

Source: https://context7.com/open-mmlab/mmaction2/llms.txt

Runs a basic video demo using a specified configuration, checkpoint, video file, and label map.

```bash
python demo/demo.py \
    demo/demo_configs/tsn_r50_1x1x8_video_infer.py \
    https://download.openmmlab.com/mmaction/v1.0/recognition/tsn/tsn_r50_8xb32-1x1x8-100e_kinetics400-rgb/tsn_r50_8xb32-1x1x8-100e_kinetics400-rgb_20220818-2692d16c.pth \
    demo/demo.mp4 \
    tools/data/kinetics/label_map_k400.txt
```

--------------------------------

### Install Optional Requirements for MMAction2

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/mmaction2_tutorial.ipynb

Use this command to install additional requirements specified in the optional.txt file. This is useful for enabling extended functionalities.

```bash
!pip install -r requirements/optional.txt
```

--------------------------------

### Install PyTorch with CUDA 11.8 Support

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/mmaction2_tutorial.ipynb

Installs PyTorch, TorchVision, and TorchAudio specifically compiled for CUDA 11.8. This is crucial for GPU acceleration.

```python
# install dependencies: (if your colab has CUDA 11.8)
%pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
```

--------------------------------

### Building and Indexing a Dataset

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/zh_cn/get_started/guide_to_framework.md

Builds the training dataset using the configuration and then accesses the first sample by index. Prints information about the input data.

```python
from mmaction.registry import DATASETS

train_dataset = DATASETS.build(train_dataset_cfg)

packed_results = train_dataset[0]

inputs = packed_results['inputs']
data_sample = packed_results['data_samples']

print('shape of the inputs: ', inputs.shape)

# 获取输入的信息
print('image_shape: ', data_sample.img_shape)
print('num_clips: ', data_sample.num_clips)
print('clip_len: ', data_sample.clip_len)

# 获取输入的标签
print('label: ', data_sample.gt_label)
```

--------------------------------

### Install MMRazor as a Python Package

Source: https://github.com/open-mmlab/mmaction2/blob/main/projects/knowledge_distillation/README.md

Install MMRazor version 1.0.0 or higher using pip. This is a prerequisite for using the knowledge distillation features.

```shell
mim install "mmrazor>=1.0.0"
```

--------------------------------

### Install PyTorch on CPU Platforms

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/installation.md

Install PyTorch and torchvision for CPU-only usage using Conda. This is suitable for environments without a compatible GPU.

```shell
conda install pytorch torchvision cpuonly -c pytorch
```

--------------------------------

### Weight Initialization Logs for RecognizerZelda

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/guide_to_framework.md

Logs detailing the weight initialization process for different layers of the RecognizerZelda model, including Kaiming and Normal initialization strategies.

```shell
04/03 23:28:01 - mmengine - INFO -
backbone.conv1.0.weight - torch.Size([64, 3, 3, 7, 7]):
KaimingInit: a=0, mode=fan_out, nonlinearity=relu, distribution =normal, bias=0

04/03 23:28:01 - mmengine - INFO -
backbone.conv1.0.bias - torch.Size([64]):
KaimingInit: a=0, mode=fan_out, nonlinearity=relu, distribution =normal, bias=0

04/03 23:28:01 - mmengine - INFO -
backbone.conv1.1.weight - torch.Size([64]):
The value is the same before and after calling `init_weights` of RecognizerZelda

04/03 23:28:01 - mmengine - INFO -
backbone.conv1.1.bias - torch.Size([64]):
The value is the same before and after calling `init_weights` of RecognizerZelda

04/03 23:28:01 - mmengine - INFO -
backbone.conv.0.weight - torch.Size([128, 64, 3, 3, 3]):
KaimingInit: a=0, mode=fan_out, nonlinearity=relu, distribution =normal, bias=0

04/03 23:28:01 - mmengine - INFO -
backbone.conv.0.bias - torch.Size([128]):
KaimingInit: a=0, mode=fan_out, nonlinearity=relu, distribution =normal, bias=0

04/03 23:28:01 - mmengine - INFO -
backbone.conv.1.weight - torch.Size([128]):
The value is the same before and after calling `init_weights` of RecognizerZelda

04/03 23:28:01 - mmengine - INFO -
backbone.conv.1.bias - torch.Size([128]):
The value is the same before and after calling `init_weights` of RecognizerZelda

04/03 23:28:01 - mmengine - INFO -
cls_head.fc.weight - torch.Size([2, 128]):
NormalInit: mean=0, std=0.01, bias=0

04/03 23:28:01 - mmengine - INFO -
cls_head.fc.bias - torch.Size([2]):
NormalInit: mean=0, std=0.01, bias=0

loss dict:  {'loss_cls': tensor(0.6853, grad_fn=<NllLossBackward0>)}
Label of Sample[0] tensor([0])
Scores of Sample[0] tensor([0.5240, 0.4760])
```

--------------------------------

### Print Complete Config

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/user_guides/config.md

Use this command to inspect the full configuration of a given config file.

```bash
python tools/analysis_tools/print_config.py /PATH/TO/CONFIG
```

--------------------------------

### Initialize Recognizer and Perform Inference

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/user_guides/inference.md

Initializes a recognizer with a configuration and checkpoint, then performs inference on a video. Ensure the config and demo video are downloaded before running. The device can be set to 'cpu' or 'cuda:0'.

```python
from mmaction.apis import inference_recognizer, init_recognizer

config_path = 'configs/recognition/tsn/tsn_imagenet-pretrained-r50_8xb32-1x1x8-100e_kinetics400-rgb.py'
checkpoint_path = 'https://download.openmmlab.com/mmaction/v1.0/recognition/tsn/tsn_imagenet-pretrained-r50_8xb32-1x1x8-100e_kinetics400-rgb/tsn_imagenet-pretrained-r50_8xb32-1x1x8-100e_kinetics400-rgb_20220906-2692d16c.pth' # can be a local path
img_path = 'demo/demo.mp4'   # you can specify your own picture path

# build the model from a config file and a checkpoint file
model = init_recognizer(config_path, checkpoint_path, device="cpu")  # device can be 'cuda:0'
# test a single image
result = inference_recognizer(model, img_path)

```

--------------------------------

### Install MMAction2 and Dependencies

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/mmaction2_tutorial.ipynb

Installs MMEngine, MMCV, and MMAction2 using MIM and pip. This includes cloning the repository and setting up the package in editable mode.

```python
# install MMEngine, MMCV and MMDetection using MIM
%pip install -U openmim
!mim install mmengine
!mim install "mmcv>=2.0.0"

# Install mmaction2
!rm -rf mmaction2
!git clone https://github.com/open-mmlab/mmaction2.git -b main
%cd mmaction2

!pip install -e .
```

--------------------------------

### Run Video Demo Script

Source: https://github.com/open-mmlab/mmaction2/blob/main/demo/README.md

Use this command to run the demo script for video recognition. Specify configuration, checkpoint, video, and label files. Optional arguments allow customization of device, FPS, text styling, resolution, and output file.

```shell
python demo/demo.py ${CONFIG_FILE} ${CHECKPOINT_FILE} ${VIDEO_FILE} ${LABEL_FILE} \
    [--device ${DEVICE_TYPE}] [--fps ${FPS}] [--font-scale ${FONT_SCALE}] [--font-color ${FONT_COLOR}] \
    [--target-resolution ${TARGET_RESOLUTION}] [--out-filename ${OUT_FILE}]
```

--------------------------------

### Install MoviePy for Video Processing

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/zh_cn/get_started/faq.md

Install MoviePy using pip. For Windows, ensure ImageMagick is correctly configured. For Linux, adjust the ImageMagick policy file if necessary.

```bash
pip install moviepy
```

--------------------------------

### Optimizer Wrapper Configuration

Source: https://github.com/open-mmlab/mmaction2/blob/main/projects/stad_tutorial/demo_stad.ipynb

Sets up the optimizer wrapper using SGD with a learning rate of 0.005, momentum of 0.9, and weight decay of 0.0001.

```python
optim_wrapper = dict(
    type='OptimWrapper',
    optimizer=dict(type='SGD', lr=0.005, momentum=0.9, weight_decay=0.0001))

```

--------------------------------

### Install MMCV with Pip for PyTorch 1.10.x and CUDA 11.3

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/installation.md

Use this command to install MMCV when not using MIM. It specifies the find-url based on the PyTorch and CUDA versions.

```shell
pip install mmcv -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.10/index.html
```

--------------------------------

### Real-time Spatio-temporal Action Detection Webcam Demo

Source: https://context7.com/open-mmlab/mmaction2/llms.txt

Run a real-time spatio-temporal action detection demo using a webcam. This requires specifying the input video source (0 for webcam), model configuration, pre-trained checkpoint, and detection configurations.

```bash
python demo/webcam_demo_spatiotemporal_det.py \
    --input-video 0 \
    --config configs/detection/slowonly/slowonly_kinetics400-pretrained-r101_8xb16-8x8x1-20e_ava21-rgb.py \
    --checkpoint https://download.openmmlab.com/mmaction/detection/ava/slowonly_omnisource_pretrained_r101_8x8x1_20e_ava_rgb/slowonly_omnisource_pretrained_r101_8x8x1_20e_ava_rgb_20201217-16378594.pth \
    --det-config demo/demo_configs/faster-rcnn_r50_fpn_2x_coco_infer.py \
    --det-checkpoint http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth \
    --det-score-thr 0.9 \
    --action-score-thr 0.5 \
    --label-map tools/data/ava/label_map.txt \
    --predict-stepsize 40 \
    --output-fps 20 \
    --show
```

--------------------------------

### Install MMAction2 as a Python Package

Source: https://github.com/open-mmlab/mmaction2/blob/main/docs/en/get_started/installation.md

Install MMAction2 directly as a Python package using pip. This is suitable for users who only need to import and use MMAction2's APIs.

```shell
pip install mmaction2
```

--------------------------------

### Download AVA Videos with GNU Parallel

Source: https://github.com/open-mmlab/mmaction2/blob/main/tools/data/ava/README.md

Speeds up AVA video downloading using GNU Parallel, which may require installation via `sudo apt-get install parallel`.

```shell
# sudo apt-get install parallel
bash download_videos_gnu_parallel.sh
```

--------------------------------

### Webcam Demo with I3D

Source: https://context7.com/open-mmlab/mmaction2/llms.txt

Run a webcam demo using the I3D model for accurate action recognition. Ensure you have the necessary configuration, checkpoint, and label map files.

```bash
python demo/webcam_demo.py \
    demo/demo_configs/i3d_r50_32x2x1_video_infer.py \
    checkpoints/i3d_imagenet-pretrained-r50_8xb8-32x2x1-100e_kinetics400-rgb.pth \
    tools/data/kinetics/label_map_k400.txt \
    --average-size 5 \
    --threshold 0.2
```

--------------------------------

### Create Directories and Symbolic Links for Rawframes

Source: https://github.com/open-mmlab/mmaction2/blob/main/tools/data/kinetics/README.md

Execute these commands to create directories for extracted rawframes and symbolic links to the data directory. Assumes the SSD is mounted at /mnt/SSD/.

```shell
mkdir /mnt/SSD/${DATASET}_extracted_train/
ln -s /mnt/SSD/${DATASET}_extracted_train/ ../../../data/${DATASET}/rawframes_train/
mkdir /mnt/SSD/${DATASET}_extracted_val/
ln -s /mnt/SSD/${DATASET}_extracted_val/ ../../../data/${DATASET}/rawframes_val/
```