### Define configuration YAML

Source: https://context7.com/facebookresearch/slowfast/llms.txt

Example configuration file structure for Kinetics dataset training.

```yaml
# Example config: configs/Kinetics/SLOWFAST_8x8_R50.yaml
TRAIN:
  ENABLE: True
  DATASET: kinetics
  BATCH_SIZE: 64
  EVAL_PERIOD: 10
  CHECKPOINT_PERIOD: 1
  AUTO_RESUME: True

DATA:
  NUM_FRAMES: 32
  SAMPLING_RATE: 2
  TRAIN_JITTER_SCALES: [256, 320]
  TRAIN_CROP_SIZE: 224
  TEST_CROP_SIZE: 256
  INPUT_CHANNEL_NUM: [3, 3]

SLOWFAST:
  ALPHA: 4        # Frame rate ratio between Slow and Fast pathways
  BETA_INV: 8     # Channel reduction ratio
  FUSION_CONV_CHANNEL_RATIO: 2
  FUSION_KERNEL_SZ: 7

MODEL:
  NUM_CLASSES: 400
  ARCH: slowfast
  MODEL_NAME: SlowFast
  LOSS_FUNC: cross_entropy
  DROPOUT_RATE: 0.5

SOLVER:
  BASE_LR: 0.1
  LR_POLICY: cosine
  MAX_EPOCH: 196
  WARMUP_EPOCHS: 34.0
  WEIGHT_DECAY: 1e-4
  OPTIMIZING_METHOD: sgd

TEST:
  ENABLE: True
  NUM_ENSEMBLE_VIEWS: 10
  NUM_SPATIAL_CROPS: 3

NUM_GPUS: 8
NUM_SHARDS: 1
OUTPUT_DIR: .
```

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

### Build and Develop SlowFast

Source: https://github.com/facebookresearch/slowfast/blob/main/INSTALL.md

Clones the SlowFast repository, navigates into its directory, and then builds and installs the package in develop mode. This step is necessary after cloning the repository and setting up dependencies.

```bash
git clone https://github.com/facebookresearch/slowfast
cd SlowFast
python setup.py build develop
```

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

### Install Core Dependencies for Detectron2

Source: https://github.com/facebookresearch/slowfast/blob/main/INSTALL.md

Installs essential packages like PyTorch, torchvision, cython, fvcore, and the COCO API. It also clones the Detectron2 repository and installs it in editable mode.

```bash
pip install -U torch torchvision cython
pip install -U 'git+https://github.com/facebookresearch/fvcore.git' 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'
git clone https://github.com/facebookresearch/detectron2 detectron2_repo
pip install -e detectron2_repo
```

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

### Run a Basic Command

Source: https://github.com/facebookresearch/slowfast/blob/main/GETTING_STARTED.md

A simplified command to run a PySlowFast job, typically used for basic execution or initial setup.

```bash
python \tools\run_net.py --cfg path/to/<pretrained_model_config_file>.yaml
```

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

### Train Video Models via CLI

Source: https://context7.com/facebookresearch/slowfast/llms.txt

Use the run_net.py script to initiate training for various architectures with specific configuration files and hardware settings.

```bash
# Train a SlowFast model on Kinetics-400
python tools/run_net.py \
  --cfg configs/Kinetics/SLOWFAST_8x8_R50.yaml \
  DATA.PATH_TO_DATA_DIR /path/to/kinetics400 \
  NUM_GPUS 8 \
  TRAIN.BATCH_SIZE 64 \
  SOLVER.BASE_LR 0.1 \
  SOLVER.MAX_EPOCH 196 \
  OUTPUT_DIR /path/to/output

# Train with mixed precision for faster training
python tools/run_net.py \
  --cfg configs/Kinetics/SLOWFAST_8x8_R50.yaml \
  DATA.PATH_TO_DATA_DIR /path/to/kinetics400 \
  TRAIN.MIXED_PRECISION True \
  NUM_GPUS 4 \
  TRAIN.BATCH_SIZE 32

# Train MViT (Multiscale Vision Transformer) model
python tools/run_net.py \
  --cfg configs/Kinetics/MVITv2_S_16x4.yaml \
  DATA.PATH_TO_DATA_DIR /path/to/kinetics400 \
  NUM_GPUS 8 \
  TRAIN.BATCH_SIZE 64

# Train X3D efficient model
python tools/run_net.py \
  --cfg configs/Kinetics/X3D_M.yaml \
  DATA.PATH_TO_DATA_DIR /path/to/kinetics400 \
  NUM_GPUS 8 \
  TRAIN.BATCH_SIZE 32
```

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

### Enable TensorBoard Visualization

Source: https://context7.com/facebookresearch/slowfast/llms.txt

Configure logging for training metrics or run standalone visualization tools for model analysis.

```bash
# Enable TensorBoard logging during training
python tools/run_net.py \
  --cfg configs/Kinetics/SLOWFAST_8x8_R50.yaml \
  TENSORBOARD.ENABLE True \
  TENSORBOARD.LOG_DIR /path/to/tensorboard_logs \
  TENSORBOARD.MODEL_VIS.ENABLE True \
  TENSORBOARD.MODEL_VIS.GRAD_CAM.ENABLE True \
  TENSORBOARD.MODEL_VIS.GRAD_CAM.LAYER_LIST ["s5", "s5"]

# Run visualization tool separately
python tools/visualization.py \
  --cfg configs/Kinetics/SLOWFAST_8x8_R50.yaml \
  TEST.CHECKPOINT_FILE_PATH /path/to/checkpoint.pyth \
  TENSORBOARD.ENABLE True \
  TENSORBOARD.MODEL_VIS.ENABLE True \
  TENSORBOARD.WRONG_PRED_VIS.ENABLE True
```

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

### Configure Demo for Video Inference

Source: https://github.com/facebookresearch/slowfast/blob/main/VISUALIZATION_TOOLS.md

Add this configuration to your YAML file to enable the demo mode for running inference on videos or webcam input. Specify input and output paths, and optionally enable multi-threading for video processing.

```yaml
DEMO:
  ENABLE: True
  LABEL_FILE_PATH: # Path to json file providing class_name - id mapping.
  INPUT_VIDEO: # Path to input video file.
  OUTPUT_FILE: # Path to output video file to write results to.
               # Leave an empty string if you would like to display results to a window.
  THREAD_ENABLE: # Run video reader/writer in the background with multi-threading.
  NUM_VIS_INSTANCES: # Number of CPU(s)/processes use to run video visualizer.
  NUM_CLIPS_SKIP: # Number of clips to skip prediction/visualization
                  # (mostly to smoothen/improve display quality with wecam input).
```

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

### Train MViTv2 Model from Scratch

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/mvitv2/README.md

Use this command to train a standard MViTv2 model from scratch. Ensure you replace 'path_to_your_dataset' with the actual path to your dataset.

```bash
python tools/run_net.py \
  --cfg configs/Kinetics/MVITv2_S_16x4.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_dataset \

```

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

### Train a Standard Model from Scratch

Source: https://github.com/facebookresearch/slowfast/blob/main/GETTING_STARTED.md

Use this command to train a simple C2D model. Specify the dataset path and number of GPUs. You can also configure these in the YAML file.

```bash
python tools/run_net.py \
  --cfg configs/Kinetics/C2D_8x8_R50.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_dataset \
  NUM_GPUS 2 \
  TRAIN.BATCH_SIZE 16 \

```

```yaml
DATA:
  PATH_TO_DATA_DIR: path_to_your_dataset
```

```yaml
DATA_LOADER.NUM_WORKERS 0 \
NUM_GPUS 2 \
TRAIN.BATCH_SIZE 16 \

```

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

### Run demo inference

Source: https://context7.com/facebookresearch/slowfast/llms.txt

Execute inference on video files, webcam streams, or AVA action detection tasks.

```bash
# Run demo on a video file
python tools/run_net.py \
  --cfg configs/Kinetics/SLOWFAST_8x8_R50.yaml \
  TEST.CHECKPOINT_FILE_PATH /path/to/checkpoint.pyth \
  DEMO.ENABLE True \
  DEMO.INPUT_VIDEO /path/to/video.mp4 \
  DEMO.LABEL_FILE_PATH /path/to/kinetics_labels.json \
  DEMO.OUTPUT_FILE /path/to/output.mp4 \
  NUM_GPUS 1

# Run demo from webcam
python tools/run_net.py \
  --cfg configs/Kinetics/SLOWFAST_8x8_R50.yaml \
  TEST.CHECKPOINT_FILE_PATH /path/to/checkpoint.pyth \
  DEMO.ENABLE True \
  DEMO.WEBCAM 0 \
  DEMO.LABEL_FILE_PATH /path/to/labels.json \
  NUM_GPUS 1

# AVA action detection demo with precomputed boxes
python tools/run_net.py \
  --cfg configs/AVA/SLOWFAST_32x2_R50_SHORT.yaml \
  TEST.CHECKPOINT_FILE_PATH /path/to/ava_checkpoint.pyth \
  DEMO.ENABLE True \
  DEMO.INPUT_VIDEO /path/to/video.mp4 \
  DEMO.PREDS_BOXES /path/to/precomputed_boxes.csv \
  DETECTION.ENABLE True
```

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

### Enable Multigrid Training via CLI

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/multigrid/README.md

Use these flags when running the training script to activate long and short cycles. Ensure the dataset path is correctly specified.

```bash
python tools/run_net.py \
  --cfg configs/Charades/SLOWFAST_16x8_R50.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_dataset \
  MULTIGRID.LONG_CYCLE True \
  MULTIGRID.SHORT_CYCLE True \
```

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

### Run MaskFeat Training and Fine-tuning

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/maskfeat/README.md

Commands to execute pre-training and fine-tuning for MViT on Kinetics-400 and ViT on ImageNet using the run_net.py script.

```bash
python tools/run_net.py \
  --cfg configs/masked_ssl/k400_MVITv2_L_16x4_MaskFeat_PT.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_Kinetics_dataset
```

```bash
python tools/run_net.py \
  --cfg configs/masked_ssl/k400_MVITv2_L_16x4_FT.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_Kinetics_dataset \
  TRAIN.CHECKPOINT_FILE_PATH path_to_your_pretrain_checkpoint
```

```bash
python tools/run_net.py \
  --cfg configs/masked_ssl/in1k_VIT_B_MaskFeat_PT.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_ImageNet_dataset
```

```bash
python tools/run_net.py \
  --cfg configs/masked_ssl/in1k_VIT_B_FT.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_ImageNet_dataset \
  TRAIN.CHECKPOINT_FILE_PATH path_to_your_pretrain_checkpoint
```

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

### Test and Evaluate Video Models

Source: https://context7.com/facebookresearch/slowfast/llms.txt

Perform multi-view testing on pretrained models using the run_net.py script with specific checkpoint paths and evaluation parameters.

```bash
# Test a trained SlowFast model
python tools/run_net.py \
  --cfg configs/Kinetics/SLOWFAST_8x8_R50.yaml \
  DATA.PATH_TO_DATA_DIR /path/to/kinetics400 \
  TEST.CHECKPOINT_FILE_PATH /path/to/checkpoint.pyth \
  TRAIN.ENABLE False \
  TEST.ENABLE True \
  NUM_GPUS 8

# Multi-view testing with 10 temporal clips and 3 spatial crops
python tools/run_net.py \
  --cfg configs/Kinetics/SLOWFAST_8x8_R50.yaml \
  TEST.CHECKPOINT_FILE_PATH /path/to/model.pyth \
  TEST.NUM_ENSEMBLE_VIEWS 10 \
  TEST.NUM_SPATIAL_CROPS 3 \
  TRAIN.ENABLE False

# Test MViT model on Kinetics
python tools/run_net.py \
  --cfg configs/Kinetics/MVITv2_B_32x3.yaml \
  TEST.CHECKPOINT_FILE_PATH /path/to/mvitv2_model.pyth \
  TRAIN.ENABLE False \
  TEST.ENABLE True \
  NUM_GPUS 4
```

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

### Configure programmatically

Source: https://context7.com/facebookresearch/slowfast/llms.txt

Load, override, and freeze configuration settings using the YACS-based system.

```python
# Load and modify config programmatically
from slowfast.config.defaults import get_cfg
from slowfast.utils.parser import load_config, parse_args

# Get default config
cfg = get_cfg()

# Load from YAML file
cfg.merge_from_file("configs/Kinetics/SLOWFAST_8x8_R50.yaml")

# Override with command line args
cfg.merge_from_list(["NUM_GPUS", "4", "TRAIN.BATCH_SIZE", "32"])

# Freeze config
cfg.freeze()
```

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

### Build Video Models Programmatically

Source: https://context7.com/facebookresearch/slowfast/llms.txt

Construct models using the build_model function and configuration objects.

```python
from slowfast.models import build_model
from slowfast.config.defaults import get_cfg

# Build a SlowFast model
cfg = get_cfg()
cfg.MODEL.MODEL_NAME = "SlowFast"
cfg.MODEL.ARCH = "slowfast"
cfg.MODEL.NUM_CLASSES = 400
cfg.RESNET.DEPTH = 50
cfg.NUM_GPUS = 1

model = build_model(cfg)
# Model is automatically moved to GPU and wrapped in DDP if NUM_GPUS > 1

# Build an MViT model
cfg = get_cfg()
cfg.MODEL.MODEL_NAME = "MViT"
cfg.MVIT.DEPTH = 16
cfg.MVIT.EMBED_DIM = 96
cfg.MVIT.NUM_HEADS = 1
cfg.MODEL.NUM_CLASSES = 400

model = build_model(cfg)

# Build X3D model
cfg = get_cfg()
cfg.MODEL.MODEL_NAME = "X3D"
cfg.MODEL.ARCH = "x3d"
cfg.X3D.WIDTH_FACTOR = 2.0
cfg.X3D.DEPTH_FACTOR = 2.2

model = build_model(cfg)
```

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

### Enable Tensorboard Support in Config

Source: https://github.com/facebookresearch/slowfast/blob/main/VISUALIZATION_TOOLS.md

Add this configuration to your YAML file to enable Tensorboard support for live monitoring of metrics and class-level performance during training, evaluation, and testing.

```yaml
TENSORBOARD:
  ENABLE: True
  LOG_DIR: # Leave empty to use cfg.OUTPUT_DIR/runs-{cfg.TRAIN.DATASET} as path.
  CLASS_NAMES_PATH: # Path to json file providing class_name - id mapping.
  CONFUSION_MATRIX:
    ENABLE: True
    SUBSET_PATH: # Path to txt file contains class names separated by newline characters.
                 # Only classes in this file will be visualized in the confusion matrix.
  HISTOGRAM:
    ENABLE: True
    TOP_K: 10   # Top-k most frequently predicted classes for each class in the dataset.
    SUBSET_PATH: # Path to txt file contains class names separated by newline characters.
                 # Only classes in this file will be visualized with histograms.
```

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

### Configure Multigrid Training

Source: https://context7.com/facebookresearch/slowfast/llms.txt

Train models using multigrid schedules to accelerate convergence by varying mini-batch shapes.

```bash
# Train with multigrid (short cycle only)
python tools/run_net.py \
  --cfg configs/Kinetics/SLOWFAST_8x8_R50_stepwise_multigrid.yaml \
  DATA.PATH_TO_DATA_DIR /path/to/kinetics400 \
  MULTIGRID.SHORT_CYCLE True \
  NUM_GPUS 8

# Train with multigrid long cycle
python tools/run_net.py \
  --cfg configs/Kinetics/SLOWFAST_8x8_R50_stepwise_multigrid.yaml \
  MULTIGRID.LONG_CYCLE True \
  MULTIGRID.SHORT_CYCLE True \
  MULTIGRID.EPOCH_FACTOR 1.5 \
  NUM_GPUS 8
```

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

### Download AVA Annotations

Source: https://github.com/facebookresearch/slowfast/blob/main/slowfast/datasets/DATASET.md

Downloads official AVA annotation files, including training/validation CSVs and action lists, into the specified annotations directory. Creates the directory if it does not exist.

```bash
DATA_DIR="../../data/ava/annotations"

if [[ ! -d "${DATA_DIR}" ]]; then
  echo "${DATA_DIR} doesn't exist. Creating it.";
  mkdir -p ${DATA_DIR}
fi

wget https://research.google.com/ava/download/ava_train_v2.1.csv -P ${DATA_DIR}
wget https://research.google.com/ava/download/ava_val_v2.1.csv -P ${DATA_DIR}
wget https://research.google.com/ava/download/ava_action_list_v2.1_for_activitynet_2018.pbtxt -P ${DATA_DIR}
wget https://research.google.com/ava/download/ava_train_excluded_timestamps_v2.1.csv -P ${DATA_DIR}
wget https://research.google.com/ava/download/ava_val_excluded_timestamps_v2.1.csv -P ${DATA_DIR}
```

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

### Configure AVA Video Demo

Source: https://github.com/facebookresearch/slowfast/blob/main/VISUALIZATION_TOOLS.md

Defines the input paths and output settings for visualizing prediction results on AVA-format videos.

```yaml
DEMO:
  ENABLE: True
  OUTPUT_FILE: yourPath/output.mp4
  LABEL_FILE_PATH:  yourPath/ava_classnames.json
  INPUT_VIDEO: yourPath/frames/HVAmkvLrthQ  # Path to a video file or image folder
  PREDS_BOXES: yourPath/ava_detection_train_boxes_and_labels_include_negative.csv # Path to pre-computed bouding boxes in AVA format.
  GT_BOXES: yourPath/ava_train_v2.2.csv # Path to ground-truth boxes and labels in AVA format (optional).
```

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

### Train and test Rev-MViT video model

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/rev/README.md

Command to execute training and testing for the Rev-MViT video model on Kinetics using the specified configuration.

```bash

python tools/run_net.py \

--cfg configs/Kinetics/REV_MVIT_B_16x4_CONV.yaml \

DATA.PATH_TO_DATA_DIR path_to_your_dataset \
NUM_GPUS 1 \
```

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

### Perform Testing with PySlowFast

Source: https://github.com/facebookresearch/slowfast/blob/main/GETTING_STARTED.md

Run this command for testing. Set TRAIN.ENABLE to False and provide the path to the model checkpoint for testing.

```bash
python tools/run_net.py \
  --cfg configs/Kinetics/C2D_8x8_R50.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_dataset \
  TEST.CHECKPOINT_FILE_PATH path_to_your_checkpoint \
  TRAIN.ENABLE False \

```

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

### Enable Model Analysis on Tensorboard

Source: https://github.com/facebookresearch/slowfast/blob/main/VISUALIZATION_TOOLS.md

Add this configuration to your YAML file to enable Tensorboard support for model analysis, including visualization of weights, activations, and Grad-CAM.

```yaml
TENSORBOARD:
  ENABLE: True
  MODEL_VIS:
    ENABLE: True
    MODEL_WEIGHTS: # Set to True to visualize model weights.
    ACTIVATIONS: # Set to True to visualize feature maps.
    INPUT_VIDEO: # Set to True to visualize the input video(s) for the corresponding feature maps.
    LAYER_LIST: # List of layer names to visualize weights and activations for.
    GRAD_CAM:
      ENABLE: True
      LAYER_LIST: # List of CNN layers to use for Grad-CAM visualization method.
                  # The number of layer must be equal to the number of pathway(s).
```

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

### Train AVA action detection

Source: https://context7.com/facebookresearch/slowfast/llms.txt

Command to initiate training for spatio-temporal action detection on the AVA dataset.

```bash
# Train AVA action detection model
python tools/run_net.py \
  --cfg configs/AVA/SLOWFAST_32x2_R50_SHORT.yaml \
  AVA.FRAME_DIR /path/to/ava_frames \
  AVA.FRAME_LIST_DIR /path/to/frame_lists \
  AVA.ANNOTATION_DIR /path/to/annotations \
  TRAIN.CHECKPOINT_FILE_PATH /path/to/kinetics_pretrained.pyth \
  NUM_GPUS 8 \
  DETECTION.ENABLE True
```

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

### Download AVA Videos

Source: https://github.com/facebookresearch/slowfast/blob/main/slowfast/datasets/DATASET.md

Downloads the AVA dataset videos and organizes them into a specified directory. Ensures the directory exists before downloading.

```bash
DATA_DIR="../../data/ava/videos"

if [[ ! -d "${DATA_DIR}" ]]; then
  echo "${DATA_DIR} doesn't exist. Creating it.";
  mkdir -p ${DATA_DIR}
fi

wget https://s3.amazonaws.com/ava-dataset/annotations/ava_file_names_trainval_v2.1.txt

for line in $(cat ava_file_names_trainval_v2.1.txt)
do
  wget https://s3.amazonaws.com/ava-dataset/trainval/$line -P ${DATA_DIR}
done
```

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

### Train and Test MViT-B Model

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/mvit/README.md

Use this command to train and test an MViT-B model on your dataset. Ensure you have the correct configuration file and dataset path.

```bash
python tools/run_net.py \
  --cfg configs/Kinetics/MVIT-B.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_dataset \

```

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

### Run SlowFast Pipeline

Source: https://github.com/facebookresearch/slowfast/blob/main/INSTALL.md

Executes the SlowFast pipeline using a specified configuration file. This command requires the number of GPUs, batch size, learning rate, and the path to the data directory.

```bash
python tools/run_net.py --cfg configs/Kinetics/C2D_8x8_R50.yaml NUM_GPUS 1 TRAIN.BATCH_SIZE 8 SOLVER.BASE_LR 0.0125 DATA.PATH_TO_DATA_DIR path_to_your_data_folder
```

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

### Train and Test X3D Model

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/x3d/README.md

Use this command to train and test an extra small (XS) X3D model on your dataset. Ensure you have the correct configuration file and dataset path.

```bash
python tools/run_net.py \
  --cfg configs/Kinetics/X3D-XS.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_dataset \

```

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

### Perform Self-Supervised Learning

Source: https://context7.com/facebookresearch/slowfast/llms.txt

Execute contrastive pretraining or fine-tune models on downstream tasks using SSL workflows.

```bash
# Pretrain with MoCo
python tools/run_net.py \
  --cfg configs/contrastive_ssl/MoCo_SlowR50_8x8.yaml \
  DATA.PATH_TO_DATA_DIR /path/to/kinetics400 \
  TASK ssl \
  CONTRASTIVE.TYPE moco \
  CONTRASTIVE.T 0.07 \
  CONTRASTIVE.QUEUE_LEN 65536 \
  NUM_GPUS 8

# Fine-tune on downstream task
python tools/run_net.py \
  --cfg configs/contrastive_ssl/finetune_SSv2_Slow_R50_syn8.yaml \
  DATA.PATH_TO_DATA_DIR /path/to/ssv2 \
  TRAIN.CHECKPOINT_FILE_PATH /path/to/ssl_pretrained.pyth \
  TASK ssl_eval \
  NUM_GPUS 8
```

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

### Resume Training from a Checkpoint

Source: https://github.com/facebookresearch/slowfast/blob/main/GETTING_STARTED.md

Specify the path to your PyTorch or Caffe2 checkpoint to resume training. For Caffe2 checkpoints, also set the checkpoint type.

```bash
TRAIN.CHECKPOINT_FILE_PATH path_to_your_PyTorch_checkpoint
```

```bash
TRAIN.CHECKPOINT_FILE_PATH path_to_your_Caffe2_checkpoint \
TRAIN.CHECKPOINT_TYPE caffe2
```

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

### Cut AVA Videos to 15-Minute Segments

Source: https://github.com/facebookresearch/slowfast/blob/main/slowfast/datasets/DATASET.md

Processes downloaded AVA videos, cutting each to a 15-minute duration (900 seconds) and saving to an output directory. Skips processing if the output file already exists.

```bash
IN_DATA_DIR="../../data/ava/videos"
OUT_DATA_DIR="../../data/ava/videos_15min"

if [[ ! -d "${OUT_DATA_DIR}" ]]; then
  echo "${OUT_DATA_DIR} doesn't exist. Creating it.";
  mkdir -p ${OUT_DATA_DIR}
fi

for video in $(ls -A1 -U ${IN_DATA_DIR}/*)
do
  out_name="${OUT_DATA_DIR}/${video##*/}"
  if [ ! -f "${out_name}" ]; then
    ffmpeg -ss 900 -t 901 -i "${video}" "${out_name}"
  fi
done
```

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

### Test AVA model

Source: https://context7.com/facebookresearch/slowfast/llms.txt

Execute detection testing on the AVA dataset using a specified configuration and checkpoint.

```bash
python tools/run_net.py \
  --cfg configs/AVA/SLOWFAST_32x2_R50_SHORT.yaml \
  AVA.FRAME_DIR /path/to/ava_frames \
  TEST.CHECKPOINT_FILE_PATH /path/to/ava_checkpoint.pyth \
  TRAIN.ENABLE False \
  TEST.ENABLE True \
  DETECTION.ENABLE True
```

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

### Enable Confusion Matrix on Tensorboard

Source: https://github.com/facebookresearch/slowfast/blob/main/VISUALIZATION_TOOLS.md

Configure Tensorboard to display confusion matrices by setting ENABLE to True and providing a path to a JSON file for category-to-class mapping.

```yaml
TENSORBOARD:
  ENABLE: True
  CATEGORIES_PATH: # Path to a json file for categories -> classes mapping
                   # in the format {"parent_class": ["child_class1", "child_class2",...], ...}.
  CONFUSION_MATRIX:
    ENABLE: True
```

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

### Manage model checkpoints

Source: https://context7.com/facebookresearch/slowfast/llms.txt

Utilities for saving, loading, and verifying model checkpoints, including support for Caffe2 formats and weight inflation.

```python
from slowfast.utils import checkpoint as cu
from slowfast.models import build_model
from slowfast.config.defaults import get_cfg

cfg = get_cfg()
model = build_model(cfg)

# Save checkpoint
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
cu.save_checkpoint(
    path_to_job="/path/to/output",
    model=model,
    optimizer=optimizer,
    epoch=10,
    cfg=cfg,
    scaler=None  # GradScaler for mixed precision
)
# Saves to: /path/to/output/checkpoints/checkpoint_epoch_00011.pyth

# Load checkpoint for training
epoch = cu.load_checkpoint(
    path_to_checkpoint="/path/to/checkpoint.pyth",
    model=model,
    data_parallel=True,  # True if using DistributedDataParallel
    optimizer=optimizer,
    inflation=False,  # Set True to inflate 2D weights to 3D
    convert_from_caffe2=False
)

# Load from Caffe2 checkpoint
cu.load_checkpoint(
    path_to_checkpoint="/path/to/caffe2_model.pkl",
    model=model,
    data_parallel=True,
    convert_from_caffe2=True
)

# Load test checkpoint (simplified)
cu.load_test_checkpoint(cfg, model)

# Check if checkpoint exists
has_ckpt = cu.has_checkpoint("/path/to/output")
last_ckpt = cu.get_last_checkpoint("/path/to/output", task="")
```

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

### Construct Data Loaders

Source: https://context7.com/facebookresearch/slowfast/llms.txt

Initialize data loaders for training, validation, and testing phases using the construct_loader function.

```python
from slowfast.datasets import loader
from slowfast.config.defaults import get_cfg

cfg = get_cfg()
cfg.TRAIN.DATASET = "kinetics"
cfg.DATA.PATH_TO_DATA_DIR = "/path/to/kinetics400"
cfg.DATA.NUM_FRAMES = 8
cfg.DATA.SAMPLING_RATE = 8
cfg.TRAIN.BATCH_SIZE = 32
cfg.DATA_LOADER.NUM_WORKERS = 8

# Construct training loader
train_loader = loader.construct_loader(cfg, "train")

# Construct validation loader
val_loader = loader.construct_loader(cfg, "val")

# Construct test loader
cfg.TEST.DATASET = "kinetics"
cfg.TEST.BATCH_SIZE = 64
test_loader = loader.construct_loader(cfg, "test")
```

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

### Clone SlowFast Repository

Source: https://github.com/facebookresearch/slowfast/blob/main/INSTALL.md

Clones the SlowFast video understanding repository. This is the main repository for the project.

```bash
git clone https://github.com/facebookresearch/slowfast
```

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

### Fine-tune MAE ViT-L on Kinetics-400

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/mae/README.md

This command is used to fine-tune a pre-trained MAE ViT-L model. You need to provide the path to your Kinetics dataset and the pre-trained checkpoint file.

```bash
python tools/run_net.py \
  --cfg configs/masked_ssl/k400_VIT_L_16x4_FT.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_Kinetics_dataset \
  TRAIN.CHECKPOINT_FILE_PATH path_to_your_pretrain_checkpoint
```

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

### Iterate through training batches

Source: https://context7.com/facebookresearch/slowfast/llms.txt

Standard loop for processing video data batches in a training loader.

```python
# Iterate through batches
for inputs, labels, video_idx, time, meta in train_loader:
    # inputs: list of tensors [slow_pathway, fast_pathway] for SlowFast
    #         or single tensor for other models
    # labels: class labels
    # video_idx: video indices
    # meta: additional metadata (boxes for detection)
    pass
```

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

### Train and test Rev-ViT Base Image model

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/rev/README.md

Command to execute training and testing for the Rev-ViT Base model on ImageNet using the specified configuration.

```bash
python tools/run_net.py \

--cfg configs/ImageNet/REV_VIT_B.yaml \

DATA.PATH_TO_DATA_DIR path_to_your_dataset \
NUM_GPUS 1 \
```

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

### Extract frames with ffmpeg

Source: https://github.com/facebookresearch/slowfast/blob/main/slowfast/datasets/DATASET.md

Use this command to extract video frames at 30 FPS for dataset preparation.

```bash
ffmpeg -i "${video}" -r 30 -q:v 1 "${out_name}"
```

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

### Cite Multigrid Training

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/multigrid/README.md

BibTeX entry for referencing the Multigrid method in academic research.

```BibTeX
@inproceedings{multigrid2020,
  Author    = {Chao-Yuan Wu and Ross Girshick and Kaiming He and Christoph Feichtenhofer
               and Philipp Kr\"{a}henb\"{u}hl},
  Title     = {{A Multigrid Method for Efficiently Training Video Models}},
  Booktitle = {{CVPR}},
  Year      = {2020}}
```

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

### Clone PyTorch Repository

Source: https://github.com/facebookresearch/slowfast/blob/main/INSTALL.md

Clones the PyTorch repository with its submodules. This is a prerequisite for building PyTorch from source.

```bash
git clone --recursive https://github.com/pytorch/pytorch
```

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

### Extract Frames from AVA Videos

Source: https://github.com/facebookresearch/slowfast/blob/main/slowfast/datasets/DATASET.md

Extracts frames from the processed AVA videos at 30 FPS and saves them as JPEG images in a structured directory. Handles both .webm and other video extensions.

```bash
IN_DATA_DIR="../../data/ava/videos_15min"
OUT_DATA_DIR="../../data/ava/frames"

if [[ ! -d "${OUT_DATA_DIR}" ]]; then
  echo "${OUT_DATA_DIR} doesn't exist. Creating it.";
  mkdir -p ${OUT_DATA_DIR}
fi

for video in $(ls -A1 -U ${IN_DATA_DIR}/*)
do
  video_name=${video##*/}

  if [[ $video_name = *".webm" ]]; then
    video_name=${video_name::-5}
  else
    video_name=${video_name::-4}
  fi

  out_video_dir=${OUT_DATA_DIR}/${video_name}/
  mkdir -p "${out_video_dir}"

  out_name="${out_video_dir}/${video_name}_%06d.jpg"

  ffmpeg -i "${video}" -r 30 -q:v 1 "${out_name}"
done
```

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

### BibTeX citation for Reversible Vision Transformers

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/rev/README.md

BibTeX entry for citing the Reversible Vision Transformers paper.

```BibTeX

@inproceedings{mangalam2022,

  title = {Reversible Vision Transformers},

  author = {Mangalam, Karttikeya and Fan, Haoqi and Li, Yanghao and Wu, Chao-Yuan and Xiong, Bo and Feichtenhofer, Christoph and Malik, Jitendra},

  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},

  year = {2022},
}
```

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

### BibTeX Citation

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/maskfeat/README.md

BibTeX entry for citing the MaskFeat research paper.

```BibTeX
@InProceedings{wei2022masked,
    author    = {Wei, Chen and Fan, Haoqi and Xie, Saining and Wu, Chao-Yuan and Yuille, Alan and Feichtenhofer, Christoph},
    title     = {Masked Feature Prediction for Self-Supervised Visual Pre-Training},
    booktitle = {CVPR},
    year      = {2022},
}
```

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

### Train MAE ViT-L on Kinetics-400

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/mae/README.md

Use this command to train a Masked Autoencoder Vision Transformer Large model on the Kinetics-400 dataset. Ensure the path to your dataset is correctly specified.

```bash
python tools/run_net.py \
  --cfg configs/masked_ssl/k400_VIT_L_16x4_MAE_PT.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_Kinetics_dataset
```

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

### Train MoCo R50 Slow-only Model

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/contrastive_ssl/README.md

Use this command to train a MoCo R50 Slow-only model with 8x8 sampling on your dataset. Ensure you replace 'path_to_your_dataset' with the actual path.

```bash
python tools/run_net.py \
  --cfg configs/Kinetics/contrastive_ssl/MoCo_SlowR50_8x8.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_dataset \

```

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

### Multiscale Vision Transformers Citation (ICCV 2021)

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/mvitv2/README.md

Cite this paper when referencing the original multiscale vision transformers concept. This entry is for the ICCV 2021 publication.

```bibtex
@inproceedings{fan2021multiscale,
  title={Multiscale vision transformers},
  author={Fan, Haoqi and Xiong, Bo and Mangalam, Karttikeya and Li, Yanghao and Yan, Zhicheng and Malik, Jitendra and Feichtenhofer, Christoph},
  booktitle={ICCV},
  year={2021}
}
```

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

### MViTv2 Citation (CVPR 2022)

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/mvitv2/README.md

Cite this paper when using MViTv2 for your research. This entry is for the CVPR 2022 publication.

```bibtex
@inproceedings{li2021improved,
  title={MViTv2: Improved multiscale vision transformers for classification and detection},
  author={Li, Yanghao and Wu, Chao-Yuan and Fan, Haoqi and Mangalam, Karttikeya and Xiong, Bo and Malik, Jitendra and Feichtenhofer, Christoph},
  booktitle={CVPR},
  year={2022}
}
```

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

### MViT Paper Citation

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/mvit/README.md

BibTeX entry for citing the Multiscale Vision Transformers paper. Include this in your LaTeX documents for academic referencing.

```bibtex
@Article{mvit2021,
  author = {Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, Christoph Feichtenhofer},
  title = {Multiscale Vision Transformers},
  journal = {arXiv:2104.11227},
  Year = {2021},
}

```

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

### Add SlowFast to PYTHONPATH

Source: https://github.com/facebookresearch/slowfast/blob/main/INSTALL.md

Appends the SlowFast directory to the PYTHONPATH environment variable, making it accessible for Python imports.

```bash
export PYTHONPATH=/path/to/SlowFast/slowfast:$PYTHONPATH
```

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

### Kinetics CSV File Format

Source: https://github.com/facebookresearch/slowfast/blob/main/slowfast/datasets/DATASET.md

The CSV file format required for the Kinetics dataset. Each line should contain the path to a video file and its corresponding label.

```text
path_to_video_1 label_1
path_to_video_2 label_2
path_to_video_3 label_3
...
path_to_video_N label_N
```

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

### X3D Paper Citation

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/x3d/README.md

BibTeX entry for citing the X3D paper in academic work. Include this in your LaTeX document's bibliography.

```bibtex
@inproceedings{x3d2020,
  Author    = {Christoph Feichtenhofer},
  Title     = {{X3D}: Progressive Network Expansion for Efficient Video Recognition},
  Booktitle = {{CVPR}},
  Year      = {2020}}

```

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

### MAE Paper Citation

Source: https://github.com/facebookresearch/slowfast/blob/main/projects/mae/README.md

BibTeX entry for citing the 'Masked Autoencoders As Spatiotemporal Learners' paper. Use this in your research publications if you find the work useful.

```bibtex
@article{feichtenhofer2022masked,
  title={Masked Autoencoders As Spatiotemporal Learners},
  author={Feichtenhofer, Christoph and Fan, Haoqi and Li, Yanghao and He, Kaiming},
  journal={arXiv preprint arXiv:2205.09113},
  year={2022}
}
```

=== COMPLETE CONTENT === This response contains all available snippets from this library. No additional content exists. Do not make further requests.