base_configs/pretrain.py

# coding=utf-8
# Copyright 2024 The Google Research Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""Default pretraining config values."""

import ml_collections


def get_config():
    """Returns default config."""
    config = ml_collections.ConfigDict()

    # ============================================== #
    # General experiment params.
    # ============================================== #
    # The root directory where experiments will be saved.
    config.root_dir = "/tmp/xirl/pretrain_runs/"
    # Rng seed. Set this to `none` to disable seeding.
    config.seed = 1
    # cudnn-related parameters that affect reproducibility.
    config.cudnn_deterministic = False
    config.cudnn_benchmark = True
    # Pretraining algorithm to use.
    config.algorithm = "tcc"
    # Number of steps between tensorboard logging.
    config.logging_frequency = 100
    # Number of steps between consecutive checkpoints.
    config.checkpointing_frequency = 200

    # ============================================== #
    # Dataset params.
    # ============================================== #
    config.data = ml_collections.ConfigDict()

    # Absolute path to the dataset root.
    config.data.root = "/tmp/xirl/datasets/xmagical/"
    # The mini-batch size. Note this only specifies the number of videos to
    # load frames from in a single batch. The effective batch size is actually
    # larger since we sample multiple frame sequences per video.
    config.data.batch_size = 4
    # Which action classes to select for creating the pretraining dataset. Leave
    # it empty to load all action classes.
    config.data.pretrain_action_class = ()
    # Which action classes to select for creating the dowstream dataset. Leave
    # it empty to load all action classes.
    config.data.downstream_action_class = ()
    # Restrict the number of videos per class. This is useful for experiments
    # that test sample complexity based on the number of pretraining
    # demonstrations.
    config.data.max_vids_per_class = -1
    # This controls how a video batch is created. If set to 'random', videos
    # are sampled randomly from different classes. If set to 'same_class', only
    # videos belonging to the same class folder are sampled within a batch.
    config.data.pretraining_video_sampler = "random"

    # ============================================== #
    # Frame sampling params.
    # ============================================== #
    config.frame_sampler = ml_collections.ConfigDict()

    # A wildcard specifying the file extension for images in each video folder.
    # This will usually be either "*.jpg" or "*.png".
    config.frame_sampler.image_ext = "*.png"
    # This controls the type of sampling we perform on video frames.
    config.frame_sampler.strategy = "uniform"
    # The number of frames to sample per video.
    config.frame_sampler.num_frames_per_sequence = 15
    # The number of context frames to sample per frame. This is useful for
    # models that use 3D convolutions.
    config.frame_sampler.num_context_frames = 1
    # The stride between sampled context frames.
    config.frame_sampler.context_stride = 3

    config.frame_sampler.all_sampler = ml_collections.ConfigDict()
    config.frame_sampler.all_sampler.stride = 1

    config.frame_sampler.strided_sampler = ml_collections.ConfigDict()
    config.frame_sampler.strided_sampler.stride = 3
    config.frame_sampler.strided_sampler.offset = True

    config.frame_sampler.uniform_sampler = ml_collections.ConfigDict()
    config.frame_sampler.uniform_sampler.offset = 0

    # Currently, this frame sampler has no additional kwargs.
    config.frame_sampler.window_sampler = ml_collections.ConfigDict()

    # ============================================== #
    # Data augmentation params.
    # ============================================== #
    config.data_augmentation = ml_collections.ConfigDict()

    # The image resolution to train on.
    config.data_augmentation.image_size = (112, 112)
    # A list of image augmentations to apply to the training dataset. note that
    # the order matters, e.g. normalize should be done last if you decide to
    # turn it on.
    config.data_augmentation.train_transforms = [
        "random_resized_crop",
        "color_jitter",
        "grayscale",
        "gaussian_blur",
        # "normalize",
    ]
    # A list of image augmentations to apply to the evaluation dataset.
    config.data_augmentation.eval_transforms = [
        "global_resize",
        # "normalize",
    ]

    # ============================================== #
    # Evaluator params.
    # ============================================== #
    config.eval = ml_collections.ConfigDict()

    # How many iterations of the downstream dataloaders to run. Set to None to
    # evaluate the entire dataloader.
    config.eval.val_iters = 20
    # The number of steps in between every evaluation.
    config.eval.eval_frequency = 500
    # A list of downstream task evaluators that will be run sequentially every
    # EVAL_FREQUENCY steps.
    config.eval.downstream_task_evaluators = [
        "reward_visualizer",
        "kendalls_tau",
    ]
    # What distance metric to use in the embedding space. Should match what was
    # used in the loss computation.
    # Can be one of ['cosine', 'sqeuclidean'].
    config.eval.distance = "sqeuclidean"

    config.eval.kendalls_tau = ml_collections.ConfigDict()
    config.eval.kendalls_tau.stride = 3

    config.eval.reward_visualizer = ml_collections.ConfigDict()
    config.eval.reward_visualizer.num_plots = 2

    config.eval.cycle_consistency = ml_collections.ConfigDict()
    config.eval.cycle_consistency.stride = 1

    config.eval.nearest_neighbour_visualizer = ml_collections.ConfigDict()
    config.eval.nearest_neighbour_visualizer.num_videos = 4

    config.eval.embedding_visualizer = ml_collections.ConfigDict()
    config.eval.embedding_visualizer.num_seqs = 2

    config.eval.reconstruction_visualizer = ml_collections.ConfigDict()
    config.eval.reconstruction_visualizer.num_frames = 2

    # ============================================== #
    # Model params.
    # ============================================== #
    config.model = ml_collections.ConfigDict()

    config.model.model_type = "resnet18_linear"
    config.model.embedding_size = 32
    config.model.normalize_embeddings = False
    config.model.learnable_temp = False

    # ============================================== #
    # Loss params.
    # ============================================== #
    config.loss = ml_collections.ConfigDict()

    ## TCC loss.
    config.loss.tcc = ml_collections.ConfigDict()
    config.loss.tcc.stochastic_matching = False
    config.loss.tcc.loss_type = "regression_mse"
    config.loss.tcc.cycle_length = 2
    config.loss.tcc.label_smoothing = 0.1
    config.loss.tcc.softmax_temperature = 0.1
    config.loss.tcc.normalize_indices = True
    config.loss.tcc.variance_lambda = 0.001
    config.loss.tcc.huber_delta = 0.1
    config.loss.tcc.similarity_type = "l2"  # cosine

    ## TCN loss.
    config.loss.tcn = ml_collections.ConfigDict()
    config.loss.tcn.pos_radius = 1
    config.loss.tcn.neg_radius = 4
    config.loss.tcn.num_pairs = 2
    config.loss.tcn.margin = 1.0
    config.loss.tcn.temperature = 0.1

    ## LIFS loss.
    config.loss.lifs = ml_collections.ConfigDict()
    config.loss.lifs.temperature = 1.0

    # ============================================== #
    # Optimizer params
    # ============================================== #
    config.optim = ml_collections.ConfigDict()

    config.optim.train_max_iters = 4_000
    # L2 regularization.
    config.optim.weight_decay = 1e-4
    # Learning rate.
    config.optim.lr = 1e-5

    # ============================================== #
    # End of config file
    # ============================================== #

    return config