Args

Streamer parser

usage: streamer.arguments.base_arguments [-h] [--dataset DATASET] [--output OUTPUT] [--name NAME] [--p_name P_NAME] [--p_device {gpu,slurm,cpu,mps}]
                                         [--p_partition P_PARTITION] [--p_n_nodes P_N_NODES] [--p_n_gpus P_N_GPUS] [--p_n_cpus P_N_CPUS] [--p_ram P_RAM]
                                         [--p_backend {nccl,gloo}] [--p_verbose P_VERBOSE] [--p_logs P_LOGS] [--frame_size FRAME_SIZE [FRAME_SIZE ...]]
                                         [--num_workers NUM_WORKERS] [--dataset_split {train,test}] [--dataset_percent DATASET_PERCENT] [--feature_dim FEATURE_DIM]
                                         [--max_layers MAX_LAYERS] [--evolve_every EVOLVE_EVERY] [--init_layers INIT_LAYERS] [--init_ckpt INIT_CKPT]
                                         [--buffer_size BUFFER_SIZE] [--force_fixed_buffer FORCE_FIXED_BUFFER] [--demarcation_mode {fixed,accum,average}]
                                         [--distance_mode {similarity,distance}] [--force_base_dist FORCE_BASE_DIST] [--normalize_imgs NORMALIZE_IMGS]
                                         [--window_size WINDOW_SIZE] [--modifier_type {add,multiply}] [--modifier MODIFIER] [--loss_threshold LOSS_THRESHOLD]
                                         [--lr LR] [--alpha ALPHA] [--optimize_every OPTIMIZE_EVERY] [--average_every AVERAGE_EVERY] [--optimize OPTIMIZE]
                                         [--save_every SAVE_EVERY] [--hgn_timescale HGN_TIMESCALE] [--hgn_reach HGN_REACH] [--dbg] [--tb] [--log_prefix LOG_PREFIX]
                                         [--log_postfix LOG_POSTFIX] [--log_base_every LOG_BASE_EVERY]

Named Arguments

--dataset

Path to dataset directory

Default: “data/epic”

--output

Path to output directory. Will be created if does not exist

Default: “out”

--name

Name of the experiment for logging purposes

Default: “my_model”

--p_name

Platform job name for SLURM

Default: “job”

--p_device

Possible choices: gpu, slurm, cpu, mps

Platform device

Default: “gpu”

--p_partition

Platform partition for SLURM

Default: “general”

--p_n_nodes

Platform number of nodes for SLURM

Default: 1

--p_n_gpus

Platform number of GPUs per node

Default: 1

--p_n_cpus

Platform number of total CPUs per process/GPU

Default: 2

--p_ram

Platform total RAM in GB

Default: 10

--p_backend

Possible choices: nccl, gloo

Platform backend for IPC

Default: “nccl”

--p_verbose

Platform verbose

Default: True

--p_logs

Platform console logs path. Will be added to output folder automatically

Default: “./logs”

--frame_size

Frame size of images

Default: [128, 128]

--num_workers

Dataloader number of workers

Default: 1

--dataset_split

Possible choices: train, test

Dataset split

Default: “train”

--dataset_percent

Dataset portion to use in percentage

Default: 100

--feature_dim

Feature dimension of the representation at all layers

Default: 1024

--max_layers

Maximum number of layers to stack

Default: 3

--evolve_every

Add new layer every evolve_every steps

Default: 100000.0

--init_layers

Number of layers to initialize the model

Default: 1

--init_ckpt

Checkpoint path to load. Overrides init_layers with layers count

Default: “”

--buffer_size

Buffer size for TemporalEncoding function

Default: 10

--force_fixed_buffer

Force fixed buffer size. If True, a boundary is forced after buffer is full

Default: False

--demarcation_mode

Possible choices: fixed, accum, average

Demarcation type. Only average is supported

Default: “average”

--distance_mode

Possible choices: similarity, distance

Distance type. Only similarity is supported

Default: “similarity”

--force_base_dist

Force base layer distance mode to be distance. Not recommended

Default: False

--normalize_imgs

Normalize input images along channel dimension

Default: False

--window_size

Window size for average demarcation type

Default: 50

--modifier_type

Possible choices: add, multiply

Modifier type for average demarcation type

Default: “multiply”

--modifier

Modifier value for average demarcation type

Default: 1.0

--loss_threshold

Loss threshold for fixed and accum demarcation types.

Default: 0.1

--lr

Learning rate of all modules and layers

Default: 0.0001

--alpha

Reach hyperparameter for gradient normalization

Default: 3.0

--optimize_every

Optimize model every optimize_every steps of highest layer

Default: 100

--average_every

Average weights of models every average_every steps of lowest layer

Default: 1000

--optimize

Optimize the model. Set to False during inference or testing

Default: True

--save_every

Save model every save_every steps

Default: 25000

--hgn_timescale

Optimize higher layers with bigger gradients

Default: True

--hgn_reach

Optimize with reach of influence

Default: True

--dbg

Flag for debugging and development. Overrides log files.

Default: False

--tb

Flag for tb logging. If False, does not save tensorboard files.

Default: False

--log_prefix

Prefix to write in json file

Default: “/data/D2/datasets/epic_kitchen/videos/”

--log_postfix

Postfix extension to write in json file

Default: “MP4”

--log_base_every

Log base (e.g., images) signal to tensorboard every log_base_every

Default: 1000