training/detectors/utils/slowfast/config/defaults.py

#!/usr/bin/env python3
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.

"""Configs."""
import yaml
from fvcore.common.config import CfgNode as CfgNodeOri

from . import custom_config
def load_yaml_with_base(text: str, allow_unsafe: bool = False):
    """
    Just like `yaml.load(open(filename))`, but inherit attributes from its
        `_BASE_`.
    Args:
        text (str): the file name of the current config. Will be used to
            find the base config file.
        allow_unsafe (bool): whether to allow loading the config file with
            `yaml.unsafe_load`.
    Returns:
        (dict): the loaded yaml
    """
    cfg = yaml.load(text, Loader=yaml.FullLoader)
    return cfg
class CfgNode(CfgNodeOri):
    def merge_from_str(self, text, allow_unsafe=False):
        loaded_cfg = load_yaml_with_base(text, allow_unsafe=allow_unsafe)
        loaded_cfg = type(self)(loaded_cfg)
        self.merge_from_other_cfg(loaded_cfg)

# -----------------------------------------------------------------------------
# Config definition
# -----------------------------------------------------------------------------
_C = CfgNode()


# ---------------------------------------------------------------------------- #
# Batch norm options
# ---------------------------------------------------------------------------- #
_C.BN = CfgNode()

# Precise BN stats.
_C.BN.USE_PRECISE_STATS = False

# Number of samples use to compute precise bn.
_C.BN.NUM_BATCHES_PRECISE = 200

# Weight decay value that applies on BN.
_C.BN.WEIGHT_DECAY = 0.0

# Norm type, options include `batchnorm`, `sub_batchnorm`, `sync_batchnorm`
_C.BN.NORM_TYPE = "batchnorm"

# Parameter for SubBatchNorm, where it splits the batch dimension into
# NUM_SPLITS splits, and run BN on each of them separately independently.
_C.BN.NUM_SPLITS = 1

# Parameter for NaiveSyncBatchNorm3d, where the stats across `NUM_SYNC_DEVICES`
# devices will be synchronized.
_C.BN.NUM_SYNC_DEVICES = 1


# ---------------------------------------------------------------------------- #
# Training options.
# ---------------------------------------------------------------------------- #
_C.TRAIN = CfgNode()

# If True Train the model, else skip training.
_C.TRAIN.ENABLE = True

# Dataset.
_C.TRAIN.DATASET = "kinetics"

# Total mini-batch size.
_C.TRAIN.BATCH_SIZE = 64

_C.TRAIN.SPLIT = "train_subset2.pth"
# Evaluate model on test data every eval period epochs.
_C.TRAIN.EVAL_PERIOD = 1

# Save model checkpoint every checkpoint period epochs.
_C.TRAIN.CHECKPOINT_PERIOD = 1

# Save model checkpoint every checkpoint period iters.
_C.TRAIN.CHECKPOINT_PERIOD_BY_ITER = 500


# Resume training from the latest checkpoint in the output directory.
_C.TRAIN.AUTO_RESUME = True

# Path to the checkpoint to load the initial weight.
_C.TRAIN.CHECKPOINT_FILE_PATH = ""

# Checkpoint types include `caffe2` or `pytorch`.
_C.TRAIN.CHECKPOINT_TYPE = "pytorch"

# If True, perform inflation when loading checkpoint.
_C.TRAIN.CHECKPOINT_INFLATE = False


# ---------------------------------------------------------------------------- #
# Testing options
# ---------------------------------------------------------------------------- #
_C.TEST = CfgNode()

# If True test the model, else skip the testing.
_C.TEST.ENABLE = True

# Dataset for testing.
_C.TEST.DATASET = "kinetics"

_C.TEST.SPLIT = "test_subset2.pth"
# Total mini-batch size
_C.TEST.BATCH_SIZE = 8

# Path to the checkpoint to load the initial weight.
_C.TEST.CHECKPOINT_FILE_PATH = ""

# Number of clips to sample from a video uniformly for aggregating the
# prediction results.
_C.TEST.NUM_ENSEMBLE_VIEWS = 10

# Number of crops to sample from a frame spatially for aggregating the
# prediction results.
_C.TEST.NUM_SPATIAL_CROPS = 3

# Checkpoint types include `caffe2` or `pytorch`.
_C.TEST.CHECKPOINT_TYPE = "pytorch"
# Path to saving prediction results file.
_C.TEST.SAVE_RESULTS_PATH = ""
# -----------------------------------------------------------------------------
# ResNet options
# -----------------------------------------------------------------------------
_C.RESNET = CfgNode()

# Transformation function.
_C.RESNET.TRANS_FUNC = "bottleneck_transform"

# Number of groups. 1 for ResNet, and larger than 1 for ResNeXt).
_C.RESNET.NUM_GROUPS = 1

# Width of each group (64 -> ResNet; 4 -> ResNeXt).
_C.RESNET.WIDTH_PER_GROUP = 64

# Apply relu in a inplace manner.
_C.RESNET.INPLACE_RELU = True

# Apply stride to 1x1 conv.
_C.RESNET.STRIDE_1X1 = False

#  If true, initialize the gamma of the final BN of each block to zero.
_C.RESNET.ZERO_INIT_FINAL_BN = False

# Number of weight layers.
_C.RESNET.DEPTH = 50


# label of branchs
_C.RESNET.LABELS = ["continus","discontinus"]

# If the current block has more than NUM_BLOCK_TEMP_KERNEL blocks, use temporal
# kernel of 1 for the rest of the blocks.
_C.RESNET.NUM_BLOCK_TEMP_KERNEL = [[3], [4], [6], [3]]

# Size of stride on different res stages.
_C.RESNET.SPATIAL_STRIDES = [[1], [2], [2], [2]]

# Size of dilation on different res stages.
_C.RESNET.SPATIAL_DILATIONS = [[1], [1], [1], [1]]


# -----------------------------------------------------------------------------
# Nonlocal options
# -----------------------------------------------------------------------------
_C.NONLOCAL = CfgNode()

# Index of each stage and block to add nonlocal layers.
_C.NONLOCAL.LOCATION = [[[]], [[]], [[]], [[]]]

# Number of group for nonlocal for each stage.
_C.NONLOCAL.GROUP = [[1], [1], [1], [1]]

# Instatiation to use for non-local layer.
_C.NONLOCAL.INSTANTIATION = "dot_product"


# Size of pooling layers used in Non-Local.
_C.NONLOCAL.POOL = [
    # Res2
    [[1, 2, 2], [1, 2, 2]],
    # Res3
    [[1, 2, 2], [1, 2, 2]],
    # Res4
    [[1, 2, 2], [1, 2, 2]],
    # Res5
    [[1, 2, 2], [1, 2, 2]],
]

# -----------------------------------------------------------------------------
# Model options
# -----------------------------------------------------------------------------
_C.MODEL = CfgNode()

# Model architecture.
_C.MODEL.ARCH = "slowfast"

# Model name
_C.MODEL.MODEL_NAME = "SlowFast"

# The number of classes to predict for the model.
_C.MODEL.NUM_CLASSES = 400

# Loss function.
_C.MODEL.LOSS_FUNC = "cross_entropy"

_C.MODEL.MASK_WEIGHT = 100

_C.MODEL.CLASS_WEIGHT = 1

# Model architectures that has one single pathway.
_C.MODEL.SINGLE_PATHWAY_ARCH = ["c2d", "i3d", "slow"]

# Model architectures that has multiple pathways.
_C.MODEL.MULTI_PATHWAY_ARCH = ["slowfast"]

# Dropout rate before final projection in the backbone.
_C.MODEL.DROPOUT_RATE = 0.5

# The std to initialize the fc layer(s).
_C.MODEL.FC_INIT_STD = 0.01

# Activation layer for the output head.
_C.MODEL.HEAD_ACT = "softmax"


# -----------------------------------------------------------------------------
# SlowFast options
# -----------------------------------------------------------------------------
_C.SLOWFAST = CfgNode()

# Corresponds to the inverse of the channel reduction ratio, $\beta$ between
# the Slow and Fast pathways.
_C.SLOWFAST.BETA_INV = 8

# Corresponds to the frame rate reduction ratio, $\alpha$ between the Slow and
# Fast pathways.
_C.SLOWFAST.ALPHA = 8

# Ratio of channel dimensions between the Slow and Fast pathways.
_C.SLOWFAST.FUSION_CONV_CHANNEL_RATIO = 2

# Kernel dimension used for fusing information from Fast pathway to Slow
# pathway.
_C.SLOWFAST.FUSION_KERNEL_SZ = 5


# -----------------------------------------------------------------------------
# Data options
# -----------------------------------------------------------------------------
_C.DATA = CfgNode()

# The path to the data directory.
_C.DATA.PATH_TO_DATA_DIR = ""

_C.DATA.DATASET = "faceforensics"

_C.DATA.MODE = ""

_C.DATA.ADAPTIVE = False

_C.DATA.SCALE = 1.0
# The separator used between path and label.
_C.DATA.PATH_LABEL_SEPARATOR = " "

# Video path prefix if any.
_C.DATA.PATH_PREFIX = ""

# The spatial crop size of the input clip.
_C.DATA.CROP_SIZE = 224

# The number of frames of the input clip.
_C.DATA.NUM_FRAMES = 8

_C.DATA.NUM_FRAMES_RANGE = [1,2,3,4,5,6,7,8]

# The video sampling rate of the input clip.
_C.DATA.SAMPLING_RATE = 8

# The mean value of the video raw pixels across the R G B channels.
_C.DATA.MEAN = [0.45, 0.45, 0.45]
# List of input frame channel dimensions.

_C.DATA.INPUT_CHANNEL_NUM = [3, 3]

# The std value of the video raw pixels across the R G B channels.
_C.DATA.STD = [0.225, 0.225, 0.225]

# The spatial augmentation jitter scales for training.
_C.DATA.TRAIN_JITTER_SCALES = [256, 320]

# The spatial crop size for training.
_C.DATA.TRAIN_CROP_SIZE = 224

# The spatial crop size for testing.
_C.DATA.TEST_CROP_SIZE = 256

# Input videos may has different fps, convert it to the target video fps before
# frame sampling.
_C.DATA.TARGET_FPS = 30

# Decoding backend, options include `pyav` or `torchvision`
_C.DATA.DECODING_BACKEND = "pyav"

# if True, sample uniformly in [1 / max_scale, 1 / min_scale] and take a
# reciprocal to get the scale. If False, take a uniform sample from
# [min_scale, max_scale].
_C.DATA.INV_UNIFORM_SAMPLE = False

# If True, perform random horizontal flip on the video frames during training.
_C.DATA.RANDOM_FLIP = True

# If True, calculdate the map as metric.
_C.DATA.MULTI_LABEL = False

# Method to perform the ensemble, options include "sum" and "max".
_C.DATA.ENSEMBLE_METHOD = "sum"

# If True, revert the default input channel (RBG <-> BGR).
_C.DATA.REVERSE_INPUT_CHANNEL = False


# ---------------------------------------------------------------------------- #
# Optimizer options
# ---------------------------------------------------------------------------- #
_C.SOLVER = CfgNode()

# Base learning rate.
_C.SOLVER.BASE_LR = 0.1

# Learning rate policy (see utils/lr_policy.py for options and examples).
_C.SOLVER.LR_POLICY = "cosine"

# Exponential decay factor.
_C.SOLVER.GAMMA = 0.1

# Step size for 'exp' and 'cos' policies (in epochs).
_C.SOLVER.STEP_SIZE = 1

# Steps for 'steps_' policies (in epochs).
_C.SOLVER.STEPS = []

# Learning rates for 'steps_' policies.
_C.SOLVER.LRS = []

# Maximal number of epochs.
_C.SOLVER.MAX_EPOCH = 300

# Momentum.
_C.SOLVER.MOMENTUM = 0.9

# Momentum dampening.
_C.SOLVER.DAMPENING = 0.0

# Nesterov momentum.
_C.SOLVER.NESTEROV = True

# L2 regularization.
_C.SOLVER.WEIGHT_DECAY = 1e-4

# Start the warm up from SOLVER.BASE_LR * SOLVER.WARMUP_FACTOR.
_C.SOLVER.WARMUP_FACTOR = 0.1

# Gradually warm up the SOLVER.BASE_LR over this number of epochs.
_C.SOLVER.WARMUP_EPOCHS = 0.0

# The start learning rate of the warm up.
_C.SOLVER.WARMUP_START_LR = 0.01

# Optimization method.
_C.SOLVER.OPTIMIZING_METHOD = "sgd"

_C.SOLVER.LR_STEP = 50000

_C.SOLVER.TOTAL_STEP = 200000

_C.SOLVER.FREEZE_STEP = 10000


# ---------------------------------------------------------------------------- #
# Misc options
# ---------------------------------------------------------------------------- #

# Number of GPUs to use (applies to both training and testing).
_C.NUM_GPUS = 1

# Number of machine to use for the job.
_C.NUM_SHARDS = 1

# The index of the current machine.
_C.SHARD_ID = 0

# Output basedir.
_C.OUTPUT_DIR = "./tmp"

# train module
_C.TRAIN_MODULE= "train_unet_by_iter"

# Note that non-determinism may still be present due to non-deterministic
# operator implementations in GPU operator libraries.
_C.RNG_SEED = 1

# Log period in iters.
_C.LOG_PERIOD = 10

# If True, log the model info.
_C.LOG_MODEL_INFO = True

# Distributed backend.
_C.DIST_BACKEND = "nccl"

# ---------------------------------------------------------------------------- #
# Benchmark options
# ---------------------------------------------------------------------------- #
_C.BENCHMARK = CfgNode()

# Number of epochs for data loading benchmark.
_C.BENCHMARK.NUM_EPOCHS = 5

# Log period in iters for data loading benchmark.
_C.BENCHMARK.LOG_PERIOD = 100

# If True, shuffle dataloader for epoch during benchmark.
_C.BENCHMARK.SHUFFLE = True


# ---------------------------------------------------------------------------- #
# Common train/test data loader options
# ---------------------------------------------------------------------------- #
_C.DATA_LOADER = CfgNode()

# Number of data loader workers per training process.
_C.DATA_LOADER.NUM_WORKERS = 8

# Load data to pinned host memory.
_C.DATA_LOADER.PIN_MEMORY = True

# Enable multi thread decoding.
_C.DATA_LOADER.ENABLE_MULTI_THREAD_DECODE = False


# ---------------------------------------------------------------------------- #
# Detection options.
# ---------------------------------------------------------------------------- #
_C.DETECTION = CfgNode()

# Whether enable video detection.
_C.DETECTION.ENABLE = False

# Aligned version of RoI. More details can be found at slowfast/models/head_helper.py
_C.DETECTION.ALIGNED = True

# Spatial scale factor.
_C.DETECTION.SPATIAL_SCALE_FACTOR = 16

# RoI tranformation resolution.
_C.DETECTION.ROI_XFORM_RESOLUTION = 7


# -----------------------------------------------------------------------------
# AVA Dataset options
# -----------------------------------------------------------------------------
_C.AVA = CfgNode()

# Directory path of frames.
_C.AVA.FRAME_DIR = "/mnt/fair-flash3-east/ava_trainval_frames.img/"

# Directory path for files of frame lists.
_C.AVA.FRAME_LIST_DIR = (
    "/mnt/vol/gfsai-flash3-east/ai-group/users/haoqifan/ava/frame_list/"
)

# Directory path for annotation files.
_C.AVA.ANNOTATION_DIR = (
    "/mnt/vol/gfsai-flash3-east/ai-group/users/haoqifan/ava/frame_list/"
)

# Filenames of training samples list files.
_C.AVA.TRAIN_LISTS = ["train.csv"]

# Filenames of test samples list files.
_C.AVA.TEST_LISTS = ["val.csv"]

# Filenames of box list files for training. Note that we assume files which
# contains predicted boxes will have a suffix "predicted_boxes" in the
# filename.
_C.AVA.TRAIN_GT_BOX_LISTS = ["ava_train_v2.2.csv"]
_C.AVA.TRAIN_PREDICT_BOX_LISTS = []

# Filenames of box list files for test.
_C.AVA.TEST_PREDICT_BOX_LISTS = ["ava_val_predicted_boxes.csv"]

# This option controls the score threshold for the predicted boxes to use.
_C.AVA.DETECTION_SCORE_THRESH = 0.9

# If use BGR as the format of input frames.
_C.AVA.BGR = False

# Training augmentation parameters
# Whether to use color augmentation method.
_C.AVA.TRAIN_USE_COLOR_AUGMENTATION = False

# Whether to only use PCA jitter augmentation when using color augmentation
# method (otherwise combine with color jitter method).
_C.AVA.TRAIN_PCA_JITTER_ONLY = True

# Eigenvalues for PCA jittering. Note PCA is RGB based.
_C.AVA.TRAIN_PCA_EIGVAL = [0.225, 0.224, 0.229]

# Eigenvectors for PCA jittering.
_C.AVA.TRAIN_PCA_EIGVEC = [
    [-0.5675, 0.7192, 0.4009],
    [-0.5808, -0.0045, -0.8140],
    [-0.5836, -0.6948, 0.4203],
]

# Whether to do horizontal flipping during test.
_C.AVA.TEST_FORCE_FLIP = False

# Whether to use full test set for validation split.
_C.AVA.FULL_TEST_ON_VAL = False

# The name of the file to the ava label map.
_C.AVA.LABEL_MAP_FILE = "ava_action_list_v2.2_for_activitynet_2019.pbtxt"

# The name of the file to the ava exclusion.
_C.AVA.EXCLUSION_FILE = "ava_val_excluded_timestamps_v2.2.csv"

# The name of the file to the ava groundtruth.
_C.AVA.GROUNDTRUTH_FILE = "ava_val_v2.2.csv"

# Backend to process image, includes `pytorch` and `cv2`.
_C.AVA.IMG_PROC_BACKEND = "cv2"

# ---------------------------------------------------------------------------- #
# Multigrid training options
# See https://arxiv.org/abs/1912.00998 for details about multigrid training.
# ---------------------------------------------------------------------------- #
_C.MULTIGRID = CfgNode()

# Multigrid training allows us to train for more epochs with fewer iterations.
# This hyperparameter specifies how many times more epochs to train.
# The default setting in paper trains for 1.5x more epochs than baseline.
_C.MULTIGRID.EPOCH_FACTOR = 1.5

# Enable short cycles.
_C.MULTIGRID.SHORT_CYCLE = False
# Short cycle additional spatial dimensions relative to the default crop size.
_C.MULTIGRID.SHORT_CYCLE_FACTORS = [0.5, 0.5 ** 0.5]

_C.MULTIGRID.LONG_CYCLE = False
# (Temporal, Spatial) dimensions relative to the default shape.
_C.MULTIGRID.LONG_CYCLE_FACTORS = [
    (0.25, 0.5 ** 0.5),
    (0.5, 0.5 ** 0.5),
    (0.5, 1),
    (1, 1),
]

# While a standard BN computes stats across all examples in a GPU,
# for multigrid training we fix the number of clips to compute BN stats on.
# See https://arxiv.org/abs/1912.00998 for details.
_C.MULTIGRID.BN_BASE_SIZE = 8

# Multigrid training epochs are not proportional to actual training time or
# computations, so _C.TRAIN.EVAL_PERIOD leads to too frequent or rare
# evaluation. We use a multigrid-specific rule to determine when to evaluate:
# This hyperparameter defines how many times to evaluate a model per long
# cycle shape.
_C.MULTIGRID.EVAL_FREQ = 3

# No need to specify; Set automatically and used as global variables.
_C.MULTIGRID.LONG_CYCLE_SAMPLING_RATE = 0
_C.MULTIGRID.DEFAULT_B = 0
_C.MULTIGRID.DEFAULT_T = 0
_C.MULTIGRID.DEFAULT_S = 0

# -----------------------------------------------------------------------------
# Tensorboard Visualization Options
# -----------------------------------------------------------------------------
_C.TENSORBOARD = CfgNode()

# Log to summary writer, this will automatically.
# log loss, lr and metrics during train/eval.
_C.TENSORBOARD.ENABLE = False
# Provide path to prediction results for visualization.
# This is a pickle file of [prediction_tensor, label_tensor]
_C.TENSORBOARD.PREDICTIONS_PATH = ""
# Path to directory for tensorboard logs.
# Default to to cfg.OUTPUT_DIR/runs-{cfg.TRAIN.DATASET}.
_C.TENSORBOARD.LOG_DIR = ""
# Path to a json file providing class_name - id mapping
# in the format {"class_name1": id1, "class_name2": id2, ...}.
# This file must be provided to enable plotting confusion matrix
# by a subset or parent categories.
_C.TENSORBOARD.CLASS_NAMES_PATH = ""

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

# Config for confusion matrices visualization.
_C.TENSORBOARD.CONFUSION_MATRIX = CfgNode()
# Visualize confusion matrix.
_C.TENSORBOARD.CONFUSION_MATRIX.ENABLE = False
# Figure size of the confusion matrices plotted.
_C.TENSORBOARD.CONFUSION_MATRIX.FIGSIZE = [8, 8]
# Path to a subset of categories to visualize.
# File contains class names separated by newline characters.
_C.TENSORBOARD.CONFUSION_MATRIX.SUBSET_PATH = ""

# Config for histogram visualization.
_C.TENSORBOARD.HISTOGRAM = CfgNode()
# Visualize histograms.
_C.TENSORBOARD.HISTOGRAM.ENABLE = False
# Path to a subset of classes to plot histograms.
# Class names must be separated by newline characters.
_C.TENSORBOARD.HISTOGRAM.SUBSET_PATH = ""
# Visualize top-k most predicted classes on histograms for each
# chosen true label.
_C.TENSORBOARD.HISTOGRAM.TOPK = 10
# Figure size of the histograms plotted.
_C.TENSORBOARD.HISTOGRAM.FIGSIZE = [8, 8]

# Config for layers' weights and activations visualization.
# _C.TENSORBOARD.ENABLE must be True.
_C.TENSORBOARD.MODEL_VIS = CfgNode()

# If False, skip model visualization.
_C.TENSORBOARD.MODEL_VIS.ENABLE = False

# If False, skip visualizing model weights.
_C.TENSORBOARD.MODEL_VIS.MODEL_WEIGHTS = False

# If False, skip visualizing model activations.
_C.TENSORBOARD.MODEL_VIS.ACTIVATIONS = False

# If False, skip visualizing input videos.
_C.TENSORBOARD.MODEL_VIS.INPUT_VIDEO = False


# List of strings containing data about layer names and their indexing to
# visualize weights and activations for. The indexing is meant for
# choosing a subset of activations outputed by a layer for visualization.
# If indexing is not specified, visualize all activations outputed by the layer.
# For each string, layer name and indexing is separated by whitespaces.
# e.g.: [layer1 1,2;1,2, layer2, layer3 150,151;3,4]; this means for each array `arr`
# along the batch dimension in `layer1`, we take arr[[1, 2], [1, 2]]
_C.TENSORBOARD.MODEL_VIS.LAYER_LIST = []
# Top-k predictions to plot on videos
_C.TENSORBOARD.MODEL_VIS.TOPK_PREDS = 1
# Colormap to for text boxes and bounding boxes colors
_C.TENSORBOARD.MODEL_VIS.COLORMAP = "Pastel2"
# Config for visualization video inputs with Grad-CAM.
# _C.TENSORBOARD.ENABLE must be True.
_C.TENSORBOARD.MODEL_VIS.GRAD_CAM = CfgNode()
# Whether to run visualization using Grad-CAM technique.
_C.TENSORBOARD.MODEL_VIS.GRAD_CAM.ENABLE = True
# CNN layers to use for Grad-CAM. The number of layers must be equal to
# number of pathway(s).
_C.TENSORBOARD.MODEL_VIS.GRAD_CAM.LAYER_LIST = []
# If True, visualize Grad-CAM using true labels for each instances.
# If False, use the highest predicted class.
_C.TENSORBOARD.MODEL_VIS.GRAD_CAM.USE_TRUE_LABEL = False
# Colormap to for text boxes and bounding boxes colors
_C.TENSORBOARD.MODEL_VIS.GRAD_CAM.COLORMAP = "viridis"

# Config for visualization for wrong prediction visualization.
# _C.TENSORBOARD.ENABLE must be True.
_C.TENSORBOARD.WRONG_PRED_VIS = CfgNode()
_C.TENSORBOARD.WRONG_PRED_VIS.ENABLE = False
# Folder tag to origanize model eval videos under.
_C.TENSORBOARD.WRONG_PRED_VIS.TAG = "Incorrectly classified videos."
# Subset of labels to visualize. Only wrong predictions with true labels
# within this subset is visualized.
_C.TENSORBOARD.WRONG_PRED_VIS.SUBSET_PATH = ""



###############
_C.JITTER = CfgNode()

_C.JITTER.ENABLE = False

_C.JITTER.CONTINUS_METHODS=["blend_diff_person","blend_downsampled","blend_same_person"]
_C.JITTER.DISCONTINUS_METHODS=["light", "rotate", "skip"]

_C.JITTER.STRONG_INNER_CLIP_MASK_JITTER= False

# ---------------------------------------------------------------------------- #
# Demo options
# ---------------------------------------------------------------------------- #
_C.DEMO = CfgNode()

# Run model in DEMO mode.
_C.DEMO.ENABLE = False

# Path to a json file providing class_name - id mapping
# in the format {"class_name1": id1, "class_name2": id2, ...}.
_C.DEMO.LABEL_FILE_PATH = ""

# Specify a camera device as input. This will be prioritized
# over input video if set.
# If -1, use input video instead.
_C.DEMO.WEBCAM = -1

# Path to input video for demo.
_C.DEMO.INPUT_VIDEO = ""
# Custom width for reading input video data.
_C.DEMO.DISPLAY_WIDTH = 0
# Custom height for reading input video data.
_C.DEMO.DISPLAY_HEIGHT = 0
# Path to Detectron2 object detection model configuration,
# only used for detection tasks.
_C.DEMO.DETECTRON2_CFG = "COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml"
# Path to Detectron2 object detection model pre-trained weights.
_C.DEMO.DETECTRON2_WEIGHTS = "detectron2://COCO-Detection/faster_rcnn_R_50_FPN_3x/137849458/model_final_280758.pkl"
# Threshold for choosing predicted bounding boxes by Detectron2.
_C.DEMO.DETECTRON2_THRESH = 0.9
# Number of overlapping frames between 2 consecutive clips.
# Increase this number for more frequent action predictions.
# The number of overlapping frames cannot be larger than
# half of the sequence length `cfg.DATA.NUM_FRAMES * cfg.DATA.SAMPLING_RATE`
_C.DEMO.BUFFER_SIZE = 0
# If specified, the visualized outputs will be written this a video file of
# this path. Otherwise, the visualized outputs will be displayed in a window.
_C.DEMO.OUTPUT_FILE = ""
# Frames per second rate for writing to output video file.
# If not set (-1), use fps rate from input file.
_C.DEMO.OUTPUT_FPS = -1
# Input format from demo video reader ("RGB" or "BGR").
_C.DEMO.INPUT_FORMAT = "BGR"
# Draw visualization frames in [keyframe_idx - CLIP_VIS_SIZE, keyframe_idx + CLIP_VIS_SIZE] inclusively.
_C.DEMO.CLIP_VIS_SIZE = 10
# Number of processes to run video visualizer.
_C.DEMO.NUM_VIS_INSTANCES = 2

# Path to pre-computed predicted boxes
_C.DEMO.PREDS_BOXES = ""
# Whether to run in with multi-threaded video reader.
_C.DEMO.THREAD_ENABLE = False
# Take one clip for every `DEMO.NUM_CLIPS_SKIP` + 1 for prediction and visualization.
# This is used for fast demo speed by reducing the prediction/visualiztion frequency.
# If -1, take the most recent read clip for visualization. This mode is only supported
# if `DEMO.THREAD_ENABLE` is set to True.
_C.DEMO.NUM_CLIPS_SKIP = 0
# Path to ground-truth boxes and labels (optional)
_C.DEMO.GT_BOXES = ""
# The starting second of the video w.r.t bounding boxes file.
_C.DEMO.STARTING_SECOND = 900
# Frames per second of the input video/folder of images.
_C.DEMO.FPS = 30
# Visualize with top-k predictions or predictions above certain threshold(s).
# Option: {"thres", "top-k"}
_C.DEMO.VIS_MODE = "thres"
# Threshold for common class names.
_C.DEMO.COMMON_CLASS_THRES = 0.7
# Theshold for uncommon class names. This will not be
# used if `_C.DEMO.COMMON_CLASS_NAMES` is empty.
_C.DEMO.UNCOMMON_CLASS_THRES = 0.3
# This is chosen based on distribution of examples in
# each classes in AVA dataset.
_C.DEMO.COMMON_CLASS_NAMES = [
    "watch (a person)",
    "talk to (e.g., self, a person, a group)",
    "listen to (a person)",
    "touch (an object)",
    "carry/hold (an object)",
    "walk",
    "sit",
    "lie/sleep",
    "bend/bow (at the waist)",
]
# Slow-motion rate for the visualization. The visualized portions of the
# video will be played `_C.DEMO.SLOWMO` times slower than usual speed.
_C.DEMO.SLOWMO = 1

# Add custom config with default values.
custom_config.add_custom_config(_C)


def _assert_and_infer_cfg(cfg):
    # BN assertions.
    if cfg.BN.USE_PRECISE_STATS:
        assert cfg.BN.NUM_BATCHES_PRECISE >= 0
    # TRAIN assertions.
    assert cfg.TRAIN.CHECKPOINT_TYPE in ["pytorch", "caffe2"]
    assert cfg.TRAIN.BATCH_SIZE % cfg.NUM_GPUS == 0

    # TEST assertions.
    assert cfg.TEST.CHECKPOINT_TYPE in ["pytorch", "caffe2"]
    assert cfg.TEST.BATCH_SIZE % cfg.NUM_GPUS == 0
    assert cfg.TEST.NUM_SPATIAL_CROPS == 3

    # RESNET assertions.
    assert cfg.RESNET.NUM_GROUPS > 0
    assert cfg.RESNET.WIDTH_PER_GROUP > 0
    assert cfg.RESNET.WIDTH_PER_GROUP % cfg.RESNET.NUM_GROUPS == 0

    # General assertions.
    assert cfg.SHARD_ID < cfg.NUM_SHARDS
    return cfg


def get_cfg():
    """
    Get a copy of the default config.
    """
    return _assert_and_infer_cfg(_C.clone())