config.yaml

expt_name: "default"
# Just set to multiple values to run the same config multiple times. Just there
# to take into account random variation
run_id: 0
seed: 42
# A common place, so can be overriden in notebooks, which don't support ":"
# interpolation
cwd: ${hydra:runtime.cwd}

sync_bn: false

test_only: false

# Set this to force data parallel training. Num nodes should be 1.
data_parallel: false

dist_backend: nccl

pytorch:
  # This only works with the compiled version of torchvision, and might have
  # some memory issues?
  video_backend: "video_reader"

train:
  fn: 'train'  # Which file in func/ directory to use for training
  batch_size: 2
  # This can have structure as follows:
  # <module name in model>:<module name in ckpt>:<path to ckpt> <>...
  # By default also supports just the <path to ckpt>
  # But the more complex structure can be used to init separate parts of model
  # using diff checkpoints. By default if only 2 elements are specified with :,
  # module_name_in_ckpt is assumed to be null
  init_from_model: null
  # Total epochs to train for
  num_epochs: 45
  # Evaluate within training, every these many epochs
  eval_freq: 1
  # Shuffle data at train time
  shuffle_data: true
  # Store the best performing checkpoint
  store_best: false
  train_one_epoch_fn:
    _target_: func.train.train_one_epoch
    print_freq: 10
    print_large_freq: 1000  # How often to write images/videos summary
    grad_clip_params: ${opt.grad_clip}  # DO NOT CHANGE HERE, change in opt
    # Set the following to store models every so many epochs. By default
    # will only store the last checkpoint and the best checkpoint.
    save_freq: null
    # Num of minutes to save at, same as above -- must set save_intermediate
    # true to save like this
    save_freq_min: 60  # At least save every 60 mins
    # Whether or not to save the intermediate models
    save_intermediates: false
    loss_wts:
      cls_action: 1.0
      cls_verb: 1.0
      cls_noun: 1.0
      pred: 1.0
      feat: 1.0
      distill: 150.0
      ens_distill: 0.0
      distill_feat: 0.0
      distill_feat_mse: 0.0
      # Past predictions, default 0 to be backward compatible
      past_cls_action: 0.0
      past_cls_verb: 0.0
      past_cls_noun: 0.0


eval:
  batch_size: null  # Will automatically figure out from train if null
  eval_fn:
    _target_: func.train.evaluate
    store: true
    store_endpoint: logits
    only_run_featext: false

model:
  backbone_dim: 2048
  # Use the backbone dim if null. Don't use the interpolation since the
  # backbone dim might be updated in the code
  intermediate_featdim: null
  backbone_last_n_modules_to_drop: 2  # Avg pool and linear layer
  dropout: 0.0
  # Set to a number to project the temp_agg and future features to this
  # dimension using a MLP before applying the NCE loss.
  # Note this is also applied when doing L2 regression loss, so the name is a
  # bit of a misnomer.
  project_dim_for_nce: null
  # Set to true to also add a regression head -- that is used for dense
  # anticipation when predicting the duration of an action
  add_regression_head: False
  bn:
    eps: 0.001
    mom: 0.1
  # Set this to true if you want to have the same temporal aggregated feat
  # dim as from the original backbone (backbone_dim). This will add a linear
  # layer to get that. It's useful when training future predictive models,
  # with future feat avg as the target.
  same_temp_agg_dim: false
  # Set this to true to use the class mappings to get the other predictions
  # eg, verb/noun from action, instead of adding additional linear layers
  # Only applicable when predicting multiple output classes
  use_cls_mappings: false
  # Apply the classifier on the past predictions too
  classifier_on_past: false

opt:
  # Not using an overall LR anymore, since everything is now defined per
  # module.
  # Use a list format to specify per-layer LRs and WD. The first element is
  # module_name ("__all__" => all params), LR and WD.
  # Note that if there is any overlap between parameters, those params
  # will get updated that many number of times as they appear in the list.
  # It WILL NOT take the last options as highest precedence. (TODO future)
  # The first term can also be a list, to give it a bunch of modules to set
  # the same LR and WD for.
  lr_wd: [[__all__, 0.1, 0.0001]]
  # Set this to true to also scale the LR by the batch size (normally it will
  # be scaled by the #replicas, so the LR is specified per given batch size).
  # This allows to further specify a LR per batch element (useful when doing
  # sweeps over batch size).
  scale_lr_by_bs: false
  # Set this to true to only train the last classifier layer.
  # Also, will set all BN layers to not compute mean/var at runtime.
  classifier_only: false
  bias_bn_wd_scale: 1.0  # Scale the WD for bias and BN layers by this amount
  grad_clip:
    max_norm: null  # By default, no clipping
    norm_type: 2
  warmup:
    _target_: common.scheduler.Warmup
    init_lr_ratio: 0.0  # Warmup from this ratio of the orig LRs
    num_epochs: 0  # Warmup for this many epochs (will take out of total epochs)

moco:
  _target_: moco.moco.builder.MoCo
  dim: 128
  K: 65536
  m: 0.999
  T: 0.2  # From moco-v2
  mlp: true  # From moco-v2

defaults:
  - train_eval_op: basic
  - train_eval_op/cls_loss_acc_fn: basic
  - train_eval_op/reg_criterion: mse
  - opt/optimizer: sgd
  - model/backbone: r2plus1d_34
  - model/temporal_aggregator: mean
  - model/future_predictor: identity
  - model/temporal_aggregator_after_future_pred: identity
  - model/classifier: linear
  - opt/scheduler: warmup_multi_step
  # Any keys with dataset_train prefix, like dataset_train2, etc, will all
  # be used for training by concatentating all those datasets. So you can
  # use multiple datasets in training by adding
  # +dataseset_train2=hmdb51/train to the command line config.
  # Note that this only works with standard datasets, ConcatDataset can't
  # handle overly customized datasets as we use in EpicKitchens
  - dataset@dataset_train: dundee50salads/anticipation_train
  # Any keys with the dataset_eval prefix, will all be evaluated on separately.
  # The postfix will be used to identify which dataset the results are on.
  # So, you can use > 1 evaluation datasets that way, by adding it in the
  # command line config, like +dataset_eval2=hmdb51/val
  - dataset@dataset_eval: dundee50salads/anticipation_val
  - data@data_train: default
  - data@data_eval: default
  # Load any common dataset files, that will be used to create other dataset
  # elements.
  - dataset/epic_kitchens/common
  - dataset/epic_kitchens100/common
  - dataset/dundee50salads/common
  - dataset/dundee50salads/annot_reader_fn: orig
  - dataset/egtea/common
  # Overrides
  - override hydra/launcher: submitit_slurm
  - override hydra/job_logging: colorlog
  - override hydra/hydra_logging: colorlog

hydra:
  job:
    name: "AVT"
  launcher:
    # All params in https://github.com/facebookresearch/hydra/blob/master/plugins/hydra_submitit_launcher/hydra_plugins/hydra_submitit_launcher/config.py
    timeout_min: 2880
    cpus_per_task: 10
    gpus_per_node: 8
    tasks_per_node: ${hydra.launcher.gpus_per_node}
    # This is the memory requested per node. So all GPUs on a given
    # node will share this memory
    mem_gb: 450
    nodes: 2
    # Use these parameters through + options in hydra
    # partition: learnfair
    # max_num_timeout: 3
    # constraint: ${hydra.launcher.gpu_type} # Any, or could say [volta|pascal]
    # comment: ""
  run:
    dir: ./outputs/  # Specified in the launch script
  sweep:
    dir: ${hydra.run.dir}
    # Output sub directory for sweep runs.
    subdir: ${hydra.job.num}  # ${hydra.job.override_dirname}