Get attributes >>> d = EasyDict({'foo':3}) >>> d['foo'] 3 >>> d.foo 3 >>> d.bar Traceback (most recent call last):

Streaming randomized quantile computation for torch. Add any amount of data repeatedly via add(data). At any time, quantile estimates (

Instantiate this object (passing a WorkerBase subclass type as its first argument) to create a worker pool. Then call pool.add(*args) to

A multicategory segmentation data set. Returns three streams: (1) The image (3, h, w). (2) The multicategory segmentation (labelcoun

Samples a fixed number of samples from the dataset, deterministically. Arguments: data_source, sample_size, seed (optional

A class to keep a running tally of the the top k values (and indexes) of any number of torch feature components. Will work on the GPU if

This is a wrapper for a more complicated multi-class segmenter, as described in https://arxiv.org/pdf/1807.10221.pdf, and as released in

Represents a fixed sequence of data set indices. Subsets can be created by specifying a subset of output indexes.

A wrapper for hooking, probing and intervening in pytorch Modules. Example usage: ``` model = load_my_model() with inst as Instr

Just like ClevrMulticlassDataset, but the second stream is a one-hot segmentation tensor rather than a flat one-hot presence vector. Mul

A data loader that looks for parallel image filenames, for example photo1/park/004234.jpg photo1/park/004236.jpg photo1/park/004237.

Equivalent to a map from conditions (any python hashable type) to RunningQuantiles. The reason for the type is to allow limited GPU memo

Running computation. Use this when an off-diagonal block of the covariance matrix is needed (e.g., when the whole covariance matrix does

DissectionProject understand how to drive a GanTester within a dissection project directory structure: it caches data in files, creates i

GanTester holds on to a specific model to test. (1) loads and instantiates the GAN; (2) instruments it at every layer so that units can

Provides a view of the subset of a dict where string keys begin with the given prefix. The prefix is stripped from all keys of the view.

Applies the reverse of torchvision.transforms.Normalize.

Utility for scaling segmentations, using nearest-neighbor zooming.

Provides a view of a dict of numpy values as torch tensors, where the types are converted to match the types and devices in the given dic