def main():
# TODO: use argv
parser = argparse.ArgumentParser(
description="Caffe2: simple video training")
parser.add_argument("--model_name",
type=str,
default='r2plus1d',
help="Name of the model")
parser.add_argument("--model_depth",
type=int,
default=18,
help="Depth of the model")
parser.add_argument("--train_data",
type=str,
default=None,
help="Path to train data",
required=True)
parser.add_argument("--test_data",
type=str,
default=None,
help="Path to test data")
parser.add_argument("--db_type",
type=str,
default="minidb",
help="Database type to save the training model")
parser.add_argument("--gpus",
type=str,
help="Comma separated list of GPU devices to use")
parser.add_argument("--num_gpus",
type=int,
default=1,
help="Number of GPU devices (instead of --gpus)")
parser.add_argument("--scale_h",
type=int,
default=128,
help="Scale image height to")
parser.add_argument("--scale_w",
type=int,
default=171,
help="Scale image width to")
parser.add_argument("--crop_size",
type=int,
default=112,
help="Input image size (to crop to)")
parser.add_argument("--num_decode_threads",
type=int,
default=4,
help="# of threads/GPU dedicated for video decoding")
parser.add_argument("--clip_length_rgb",
type=int,
default=16,
help="Length of input clips")
parser.add_argument("--sampling_rate_rgb",
type=int,
default=1,
help="Frame sampling rate")
parser.add_argument("--num_labels",
type=int,
default=101,
help="Number of labels")
parser.add_argument("--num_channels",
type=int,
default=3,
help="Number of channels")
parser.add_argument("--clip_length_of",
type=int,
default=8,
help="Frames of optical flow data")
parser.add_argument("--sampling_rate_of", type=int, default=2, help="")
parser.add_argument("--frame_gap_of", type=int, default=2, help="")
parser.add_argument("--input_type",
type=int,
default=0,
help="False=rgb, True=optical flow")
parser.add_argument(
"--flow_data_type",
type=int,
default=0,
help="0=Flow2C, 1=Flow3C, 2=FlowWithGray, 3=FlowWithRGB")
parser.add_argument("--do_flow_aggregation",
type=int,
default=0,
help="whether to aggregate optical flow across " +
"multiple frames")
parser.add_argument("--get_video_id",
type=int,
default=0,
help="Output video id")
parser.add_argument("--batch_size",
type=int,
default=32,
help="Batch size, total over all GPUs")
parser.add_argument("--epoch_size",
type=int,
default=110000,
help="Number of videos/epoch, total over all machines")
parser.add_argument("--num_epochs",
type=int,
default=50,
help="Num epochs.")
parser.add_argument("--base_learning_rate",
type=float,
default=0.003,
help="Initial learning rate.")
parser.add_argument("--step_epoch",
type=int,
default=10,
help="Reducing learning rate every step_epoch.")
parser.add_argument("--gamma",
type=float,
default=0.1,
help="Learning rate decay factor.")
parser.add_argument("--display_iter",
type=int,
default=10,
help="Display information every # of iterations.")
parser.add_argument("--weight_decay",
type=float,
default=0.005,
help="Weight decay (L2 regularization)")
parser.add_argument("--cudnn_workspace_limit_mb",
type=int,
default=64,
help="CuDNN workspace limit in MBs")
parser.add_argument("--file_store_path",
type=str,
default="/tmp",
help="Path to directory to use for saving checkpoints")
parser.add_argument("--save_model_name",
type=str,
default="simple_c3d",
help="Save the trained model to a given name")
parser.add_argument(
"--load_model_path",
type=str,
default=None,
help="Load previously saved model to continue training")
parser.add_argument("--use_cudnn", type=int, default=1, help="Use CuDNN")
parser.add_argument("--profiling",
type=int,
default=0,
help="Profile training time")
parser.add_argument("--pred_layer_name",
type=str,
default=None,
help="the prediction layer name")
parser.add_argument("--multi_label",
type=int,
default=0,
help="Multiple label training")
parser.add_argument("--channel_multiplier",
type=float,
default=1.0,
help="Channel multiplier")
parser.add_argument("--bottleneck_multiplier",
type=float,
default=1.0,
help="Bottleneck multiplier")
parser.add_argument("--use_dropout",
type=int,
default=0,
help="Use dropout at the prediction layer")
parser.add_argument("--conv1_temporal_stride",
type=int,
default=1,
help="Conv1 temporal striding")
parser.add_argument("--conv1_temporal_kernel",
type=int,
default=3,
help="Conv1 temporal kernel")
parser.add_argument("--video_res_type",
type=int,
default=1,
help="Video frame scaling option, 0: scaled by " +
"height x width; 1: scaled by short edge")
parser.add_argument("--use_pool1",
type=int,
default=0,
help="use pool1 layer")
parser.add_argument("--jitter_scales",
type=str,
default="128,160",
required=True,
help="spatial scales jitter, separated by commas")
parser.add_argument("--use_local_file",
type=int,
default=0,
help="use local file")
parser.add_argument("--is_checkpoint",
type=int,
default=1,
help="0: pretrained_model is used as initalization" +
"1: pretrained_model is used as a checkpoint")
parser.add_argument("--audio_input_3d",
type=int,
default=0,
help="is audio input 3d or 2d; 0 for 2d")
parser.add_argument("--g_blend",
type=int,
default=0,
help="use gradient-blending to train model")
parser.add_argument("--audio_weight",
type=float,
default=0.0,
help="g_blend weights for audio head")
parser.add_argument("--visual_weight",
type=float,
default=0.0,
help="g_blend weights for visual head")
parser.add_argument("--av_weight",
type=float,
default=1.0,
help="g_blend weights for av head")
args = parser.parse_args()
log.info(args)
assert model_builder.model_validation(
args.model_name, args.model_depth,
args.clip_length_of if args.input_type else args.clip_length_rgb,
args.crop_size)
Train(args)
def main():
parser = argparse.ArgumentParser(
description="test_net"
)
parser.add_argument("--test_data", type=str, default=None,
help="Path to test data")
parser.add_argument("--db_type", type=str, default='pickle',
help="Db type of the testing model")
parser.add_argument("--model_depth", type=int, default=18,
help="Model depth")
parser.add_argument("--model_name", type=str, default='r2plus1d',
help="Model name")
parser.add_argument("--gpus", type=str, default=None,
help="Comma separated list of GPU devices to use")
parser.add_argument("--num_gpus", type=int, default=1,
help="Number of GPU devices (instead of --gpus)")
parser.add_argument("--scale_h", type=int, default=128,
help="Scale image height to")
parser.add_argument("--scale_w", type=int, default=171,
help="Scale image width to")
parser.add_argument("--num_iter", type=int, default=0,
help="Number of test iterations; " +
"0: test the whole set")
parser.add_argument("--crop_size", type=int, default=112,
help="Input image size (to crop to)")
parser.add_argument("--clip_length_rgb", type=int, default=16,
help="Length of input clips")
parser.add_argument("--sampling_rate_rgb", type=int, default=1,
help="Frame sampling rate")
parser.add_argument("--num_labels", type=int, default=101,
help="Number of labels")
parser.add_argument("--num_channels", type=int, default=3,
help="Number of channels")
parser.add_argument("--batch_size", type=int, default=6,
help="Batch size, total over all GPUs")
parser.add_argument("--clip_per_video", type=int, default=10,
help="Number of clips to be sampled from a video")
parser.add_argument("--top_k", type=int, default=5,
help="Top k video accuracy output")
parser.add_argument("--aggregation", type=int, default=0,
help="0: avergage pool, 1: max pooling")
parser.add_argument("--load_model_path", type=str, default=None,
help="Load saved model for testing")
parser.add_argument("--use_cudnn", type=int, default=1,
help="Use CuDNN")
parser.add_argument("--pred_layer_name", type=str, default=None,
help="the prediction layer name")
parser.add_argument("--display_iter", type=int, default=10,
help="Display information every # of iterations.")
parser.add_argument("--clip_length_of", type=int, default=8,
help="Frames of optical flow data")
parser.add_argument("--sampling_rate_of", type=int, default=2,
help="Optical flow sampling rate (in frames)")
parser.add_argument("--frame_gap_of", type=int, default=2,
help="")
parser.add_argument("--do_flow_aggregation", type=int, default=0,
help="whether to aggregate optical flow across " +
" multiple frames")
parser.add_argument("--flow_data_type", type=int, default=0,
help="0=Flow2C, 1=Flow3C, 2=FlowWithGray, " +
"3=FlowWithRGB")
parser.add_argument("--input_type", type=int, default=0,
help="False=rgb, True=optical flow")
parser.add_argument("--get_video_id", type=int, default=0,
help="Output video id")
parser.add_argument("--use_dropout", type=int, default=0,
help="Use dropout at the prediction layer")
parser.add_argument("--use_local_file", type=int, default=0,
help="Use lmdb as a list of local filenames")
args = parser.parse_args()
log.info(args)
assert model_builder.model_validation(
args.model_name,
args.model_depth,
args.clip_length_of if args.input_type == 1 else args.clip_length_rgb,
args.crop_size
)
Test(args)
def main():
parser = argparse.ArgumentParser(description="Simple feature extraction")
parser.add_argument("--db_type",
type=str,
default='pickle',
help="Db type of the testing model")
parser.add_argument("--model_name",
type=str,
default='r2plus1d',
help="Model name")
parser.add_argument("--model_depth",
type=int,
default=18,
help="Model depth")
parser.add_argument("--gpus",
type=str,
default=None,
help="Comma separated list of GPU devices to use")
parser.add_argument("--num_gpus",
type=int,
default=1,
help="Number of GPU devices (instead of --gpus)")
parser.add_argument("--scale_h",
type=int,
default=128,
help="Scale image height to")
parser.add_argument("--scale_w",
type=int,
default=171,
help="Scale image width to")
parser.add_argument("--crop_size",
type=int,
default=112,
help="Input image size (to crop to)")
parser.add_argument("--clip_length_rgb",
type=int,
default=4,
help="Length of input clips")
parser.add_argument("--sampling_rate_rgb",
type=int,
default=1,
help="Frame sampling rate")
parser.add_argument("--num_labels",
type=int,
default=101,
help="Number of labels")
parser.add_argument("--num_channels",
type=int,
default=3,
help="Number of channels")
parser.add_argument("--batch_size",
type=int,
default=32,
help="Batch size, per-GPU")
parser.add_argument("--load_model_path",
type=str,
default='',
required=True,
help="Load saved model for testing")
parser.add_argument("--test_data",
type=str,
default="",
required=True,
help="Dataset on which we will extract features")
parser.add_argument("--output_path",
type=str,
default="",
help="Path to output pickle; defaults to " +
"features.pickle next to <test_data>")
parser.add_argument("--use_cudnn", type=int, default=1, help="Use CuDNN")
parser.add_argument("--features",
type=str,
default="final_avg",
help="Comma-separated list of blob names to fetch")
parser.add_argument("--num_iterations",
type=int,
default=-1,
help="Run only this many iterations")
parser.add_argument("--num_decode_threads", type=int, default=4, help="")
parser.add_argument("--clip_length_of",
type=int,
default=8,
help="Frames of optical flow data")
parser.add_argument("--sampling_rate_of",
type=int,
default=2,
help="Sampling rate for optial flows")
parser.add_argument("--frame_gap_of",
type=int,
default=2,
help="Frame gap of optical flows")
parser.add_argument("--input_type",
type=int,
default=0,
help="0=rgb, 1=optical flow")
parser.add_argument("--flow_data_type",
type=int,
default=0,
help="0=Flow2C, 1=Flow3C, 2=FlowWithGray, " +
"3=FlowWithRGB")
parser.add_argument("--do_flow_aggregation",
type=int,
default=0,
help="whether to aggregate optical flow across " +
"multiple frames")
parser.add_argument("--clip_per_video",
type=int,
default=1,
help="When clips_per_video > 1, sample this many " +
"clips uniformly in time")
parser.add_argument("--get_video_id",
type=int,
default=0,
help="Output video id")
parser.add_argument("--sanity_check",
type=int,
default=0,
help="Sanity check on the accuracy/auc")
parser.add_argument("--decode_type",
type=int,
default=2,
help="0: random, 1: uniform sampling, " +
"2: use starting frame")
parser.add_argument("--use_local_file",
type=int,
default=0,
help="Use lmdb as a list of local filenames")
args = parser.parse_args()
log.info(args)
assert model_builder.model_validation(
args.model_name, args.model_depth,
args.clip_length_of if args.input_type == 1 else args.clip_length_rgb,
args.crop_size)
ExtractFeatures(args)
def main():
# TODO: use argv
parser = argparse.ArgumentParser(
description="Caffe2: simple video training")
parser.add_argument("--model_name",
type=str,
default='r2plus1d',
help="Name of the model")
parser.add_argument("--model_depth",
type=int,
default=18,
help="Depth of the model")
parser.add_argument("--train_data",
type=str,
default=None,
help="Path to train data",
required=True)
parser.add_argument("--test_data",
type=str,
default=None,
help="Path to test data")
parser.add_argument("--db_type",
type=str,
default="minidb",
help="Database type to save the training model")
parser.add_argument("--gpus",
type=str,
help="Comma separated list of GPU devices to use")
parser.add_argument("--num_gpus",
type=int,
default=1,
help="Number of GPU devices (instead of --gpus)")
parser.add_argument("--scale_h",
type=int,
default=128,
help="Scale image height to")
parser.add_argument("--scale_w",
type=int,
default=171,
help="Scale image width to")
parser.add_argument("--crop_size",
type=int,
default=112,
help="Input image size (to crop to)")
parser.add_argument("--num_decode_threads",
type=int,
default=4,
help="# of threads/GPU dedicated for video decoding")
parser.add_argument("--clip_length_rgb",
type=int,
default=16,
help="Length of input clips")
parser.add_argument("--sampling_rate_rgb",
type=int,
default=1,
help="Frame sampling rate")
parser.add_argument("--num_labels",
type=int,
default=101,
help="Number of labels")
parser.add_argument("--num_channels",
type=int,
default=3,
help="Number of channels")
parser.add_argument("--clip_length_of",
type=int,
default=8,
help="Frames of optical flow data")
parser.add_argument("--sampling_rate_of", type=int, default=2, help="")
parser.add_argument("--frame_gap_of", type=int, default=2, help="")
parser.add_argument("--input_type",
type=int,
default=0,
help="0: rgb, 1: optical flow")
parser.add_argument("--flow_data_type",
type=int,
default=0,
help="0: Flow2C, 1: Flow3C, 2: FlowWithGray, " +
"3: FlowWithRGB")
parser.add_argument("--do_flow_aggregation",
type=int,
default=0,
help="whether to aggregate optical flow across " +
"multiple frames")
parser.add_argument("--get_video_id",
type=int,
default=0,
help="Output video id")
parser.add_argument("--batch_size",
type=int,
default=32,
help="Batch size, total over all GPUs")
parser.add_argument("--epoch_size",
type=int,
default=110000,
help="Number of videos/epoch, total over all machines")
parser.add_argument("--num_epochs",
type=int,
default=50,
help="Num epochs.")
parser.add_argument("--base_learning_rate",
type=float,
default=0.003,
help="Initial learning rate.")
parser.add_argument("--step_epoch",
type=int,
default=10,
help="Reducing learning rate every step_epoch.")
parser.add_argument("--gamma",
type=float,
default=0.1,
help="Learning rate decay factor.")
parser.add_argument("--display_iter",
type=int,
default=10,
help="Display information every # of iterations.")
parser.add_argument("--weight_decay",
type=float,
default=0.005,
help="Weight decay (L2 regularization)")
parser.add_argument("--cudnn_workspace_limit_mb",
type=int,
default=64,
help="CuDNN workspace limit in MBs")
parser.add_argument("--file_store_path",
type=str,
default=".",
help="Path to directory to use for saving checkpoints")
parser.add_argument(
"--pretrained_model",
type=str,
default=None,
help="Load saved model to continue training" + "if is_checkpoint = 1" +
"Load pretrained model for finetuning" + "if is_checkpoint = 0.")
parser.add_argument("--is_checkpoint",
type=int,
default=1,
help="0: pretrained_model is used as initalization" +
"1: pretrained_model is used as a checkpoint")
parser.add_argument("--use_cudnn", type=int, default=1, help="Use CuDNN")
parser.add_argument("--profiling",
type=int,
default=0,
help="Profile training time")
parser.add_argument("--pred_layer_name",
type=str,
default=None,
help="the prediction layer name")
parser.add_argument("--use_dropout",
type=int,
default=0,
help="Use dropout at the prediction layer")
parser.add_argument("--use_local_file",
type=int,
default=0,
help="Use lmdb as a list of local filenames")
args = parser.parse_args()
log.info(args)
assert model_builder.model_validation(
args.model_name, args.model_depth,
args.clip_length_of if args.input_type else args.clip_length_rgb,
args.crop_size)
Train(args)
def main():
parser = argparse.ArgumentParser(
description="Tool for testing large networks")
parser.add_argument("--test_data",
type=str,
default=None,
help="Path to test data")
parser.add_argument("--db_type",
type=str,
default='pickle',
help="Db type of the testing model")
parser.add_argument("--model_depth",
type=int,
default=18,
help="Model depth")
parser.add_argument("--model_name",
type=str,
default='r2plus1d',
help="Model name")
parser.add_argument("--gpus",
type=str,
default=None,
help="Comma separated list of GPU devices to use")
parser.add_argument("--num_gpus",
type=int,
default=1,
help="Number of GPU devices (instead of --gpus)")
parser.add_argument("--scale_h",
type=int,
default=128,
help="Scale image height to")
parser.add_argument("--scale_w",
type=int,
default=171,
help="Scale image width to")
parser.add_argument(
"--num_iter",
type=int,
default=0,
help="Number of test iterations; 0: test the whole set")
parser.add_argument("--crop_size",
type=int,
default=112,
help="Input image size (to crop to)")
parser.add_argument("--clip_length_rgb",
type=int,
default=16,
help="Length of input clips")
parser.add_argument("--sampling_rate_rgb",
type=int,
default=1,
help="Frame sampling rate")
parser.add_argument("--num_labels",
type=int,
default=101,
help="Number of labels")
parser.add_argument("--num_channels",
type=int,
default=3,
help="Number of channels")
parser.add_argument("--batch_size",
type=int,
default=1,
help="Batch size, total over all GPUs")
parser.add_argument("--clip_per_video",
type=int,
default=10,
help="Number of clips to be sampled from a video")
parser.add_argument("--top_k",
type=int,
default=5,
help="Top k video accuracy output")
parser.add_argument("--aggregation",
type=int,
default=0,
help="0: avergage pool, 1: max pooling")
parser.add_argument("--load_model_path",
type=str,
default=None,
help="Load saved model for testing")
parser.add_argument("--use_cudnn", type=int, default=1, help="Use CuDNN")
parser.add_argument("--print_per_class_metrics",
type=int,
default=0,
help="Log per class accuracy for multi-class setting")
parser.add_argument("--pred_layer_name",
type=str,
default=None,
help="the prediction layer name")
parser.add_argument("--multi_label",
type=int,
default=0,
help="Multiple label testing")
parser.add_argument("--display_iter",
type=int,
default=10,
help="Display information every # of iterations.")
parser.add_argument("--clip_length_of",
type=int,
default=8,
help="Frames of optical flow data")
parser.add_argument("--sampling_rate_of", type=int, default=2, help="")
parser.add_argument("--frame_gap_of", type=int, default=2, help="")
parser.add_argument("--do_flow_aggregation",
type=int,
default=0,
help="whether to aggregate optical flow across" +
" multiple frames")
parser.add_argument(
"--flow_data_type",
type=int,
default=0,
help="0=Flow2C, 1=Flow3C, 2=FlowWithGray, 3=FlowWithRGB")
parser.add_argument("--input_type",
type=int,
default=0,
help="False=rgb, True=optical flow")
parser.add_argument("--num_decode_threads",
type=int,
default=4,
help="number of decoding threads")
parser.add_argument("--channel_multiplier",
type=float,
default=1.0,
help="Channel multiplier")
parser.add_argument("--bottleneck_multiplier",
type=float,
default=1.0,
help="Bottleneck multiplier")
parser.add_argument("--use_dropout",
type=int,
default=0,
help="Use dropout at the prediction layer")
parser.add_argument("--conv1_temporal_stride",
type=int,
default=1,
help="Conv1 temporal striding")
parser.add_argument("--conv1_temporal_kernel",
type=int,
default=3,
help="Conv1 temporal kernel")
parser.add_argument("--use_convolutional_pred",
type=int,
default=0,
help="using convolutional predictions")
parser.add_argument("--video_res_type",
type=int,
default=0,
help="Video frame scaling option, 0: scaled by " +
"height x width; 1: scaled by shorter edge")
parser.add_argument("--use_pool1",
type=int,
default=0,
help="use pool1 layer")
parser.add_argument("--use_local_file",
type=int,
default=0,
help="use local file")
parser.add_argument("--crop_per_clip",
type=int,
default=1,
help="number of spatial crops per clip")
parser.add_argument("--crop_per_inference",
type=int,
default=1,
help="number of spatial crops GPU memory can handle" +
"per one pass of inference")
args = parser.parse_args()
log.info(args)
assert model_builder.model_validation(
args.model_name, args.model_depth,
args.clip_length_of if args.input_type == 1 else args.clip_length_rgb,
args.crop_size if not args.use_convolutional_pred else 112)
Test(args)