Log a new value to the meter Args: value: Next restult to include.

Finds a function handle with the name given as 'type' in config, and returns the instance initialized with corresponding keyword args given as

Load the text features and raw captions to be used in the challenge.

Canonical path lookup for visual features

Produce a common collection of feature names shared amongst datasets.

Load the text features and raw captions to be used in the challenge.

save the current content to the HMTL file

Insert a header to the HTML file Parameters: text (str) -- the header text

Provide a context for managing temporary, cloned copies of retrieval sample tensors. The rationale here is that to use nan-checking in the mo

Generate a zip file for a submission in the challenge format.

Use assertions that will be written to the logs. This is a recipe from: http://code.activestate.com/recipes/577074-logging-asserts/

Extract performace statistics from experiment log files. Args: logger (logger): reference to primary logging instance log_path (P

Get the value of the meter in the current state.

Saving checkpoints :param epoch: current epoch number :param log: logging information of the epoch :param save_best: if True,

create a single empty directory if it didn't exist Parameters: path (str) -- a single directory path

Compute a similarity matrix from sharded vectors. Args: embds1 (dict[str:th.Tensor]): the set of sub-embeddings that, when co

Compute a similarity matrix from sharded vectors. Args: embds (dict[str:th.Tensor]): the set of sub-embeddings that, when concatenated,

Full training logic. Responsible for iterating over epochs, early stopping, checkpointing and logging metrics.

Provide backwards compatible support for web directories Args: config: a configuration object containing experiment paths

Access a nested object in tree by sequence of keys.

setup GPU device if available, move model into configured device

Resume from saved checkpoints :param resume_path: Checkpoint path to be resumed

Set a value in a nested object in tree by sequence of keys.

3D space to depth trick for TPU optimization.

Training logic for an epoch :param epoch: Current epoch number

Validate model after an epoch of training and store results to disk. Args: epoch (int): the current epoch Returns:

Add a new observation Args: output (Tensor): NxK tensor that for each of the N examples indicates the probability

add images to the HTML file Parameters: ims (str list) -- a list of image paths txts (str list) -- a list of image

add images to the HTML file Parameters: vids (str list) -- a list of image paths txts (str list) -- a list of imag

We store the indices of the top-k predictions, rather than the full similarity matrix, to reduce storage requirements. NOTE: The similarity m

Partition the datset into train/val/test splits. Args: split_name (str): the name of the split

Generates a datastructure containing all the paths required to load features

Display current results on visdom; save current results to an HTML file. Parameters: visuals (OrderedDict) - - dictionary of imag

Remove nans, which we expect to find at missing indices. Args: x (th.Tensor): features ind (th.Tensor): binary values denoting wh