torchoutil 0.3.1

Collection of functions and modules to help development in PyTorch.

torchoutil

Collection of functions and modules to help development in PyTorch.

Installation

pip install torchoutil

The only requirement is PyTorch.

To check if the package is installed and show the package version, you can use the following command:

torchoutil-info

Usage

Batch of padded sequences

import torch
from torchoutil import masked_mean

x = torch.as_tensor([1, 2, 3, 4])
mask = torch.as_tensor([True, True, False, False])
result = masked_mean(x, mask)
# result contains the mean of the values marked as True: 1.5

import torch
from torchoutil import lengths_to_non_pad_mask

x = torch.as_tensor([3, 1, 2])
mask = lengths_to_non_pad_mask(x, max_len=4)
# Each row i contains x[i] True values for non-padding mask
# tensor([[True, True, True, False],
#         [True, False, False, False],
#         [True, True, False, False]])

Multilabel conversions

import torch
from torchoutil import probs_to_names

probs = torch.as_tensor([[0.9, 0.1], [0.6, 0.9]])
names = probs_to_names(probs, threshold=0.5, idx_to_name={0: "Cat", 1: "Dog"})
# [["Cat"], ["Cat", "Dog"]]

import torch
from torchoutil import multihot_to_indices

multihot = torch.as_tensor([[1, 0, 0], [0, 1, 1], [0, 0, 0]])
indices = multihot_to_indices(multihot)
# [[0], [1, 2], []]

Easely pre-compute transforms

Here is an example of pre-computing spectrograms of torchaudio SPEECHCOMMANDS dataset, using pack_to_pickle function:

from torch import nn
from torchaudio.datasets import SPEECHCOMMANDS
from torchaudio.transforms import Spectrogram
from torchoutil.utils.pickle_dataset import pack_to_pickle

speech_commands_root = "path/to/speech_commands"
pickle_root = "path/to/pickle_dataset"

dataset = SPEECHCOMMANDS(speech_commands_root, download=True, subset="validation")

class MyTransform(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.spectrogram_extractor = Spectrogram()

    def forward(self, item):
        waveform = item[0]
        spectrogram = self.spectrogram_extractor(waveform)
        return (spectrogram,) + item[1:]

pack_to_pickle(dataset, pickle_root, MyTransform())

Then you can load the pre-computed dataset using PickleDataset:

from torchoutil.utils.pickle_dataset import PickleDataset

pickle_root = "path/to/pickle_dataset"
pickle_dataset = PickleDataset(pickle_root)
pickle_dataset[0]  # == first transformed item, i.e. transform(dataset[0])

...and more tensor manipulations!

import torch
from torchoutil import insert_at_indices

x = torch.as_tensor([1, 2, 3, 4])
result = insert_at_indices(x, indices=[0, 2], values=5)
# result contains tensor with inserted values: tensor([5, 1, 2, 5, 3, 4])

import torch
from torchoutil import get_inverse_perm

perm = torch.randperm(10)
inv_perm = get_inverse_perm(perm)

x1 = torch.rand(10)
x2 = x1[perm]
x3 = x2[inv_perm]
# inv_perm are indices that allow us to get x3 from x2, i.e. x1 == x3 here

Extras

torchoutil also provides additional modules when some specific package are already installed in your environment. All extras can be installed with pip install torchoutil[extras]

• If tensorboard is installed, the function load_event_file can be used. It is useful to load manually all data contained in an tensorboard event file.
• If numpy is installed, the classes FromNumpy and ToNumpy can be used and their related function. It is meant to be used to compose dynamic transforms into Sequential module.
• If h5py is installed, the function pack_to_hdf and class HDFDataset can be used. Can be used to pack/read dataset to HDF files, and supports variable-length sequences of data.

Contact

Maintainer:

• Étienne Labbé "Labbeti": [e-mail]