PyTorch Cheat Sheet
Using PyTorch 1.2, torchaudio 0.3, torchtext 0.4, and torchvision 0.4.
General PyTorch and model I/O [Link].check_model(model)
# loading PyTorch Pre-trained models and domain-specific utils
import torch
Audio
# cuda
import [Link] as tCuda # various functions and settings import torchaudio
[Link] = True # deterministic ML? # load and save audio
[Link] = False # deterministic ML? stream, sample_rate = [Link]('file')
[Link].is_available # check if cuda is is_available [Link]('file', stream, sample_rate)
[Link]() # moving tensor to gpu # 16 bit wav files only
[Link]() # moving tensor to cpu stream, sample_rate=torchaudio.load_wav('file')
[Link](device) # copy densor to device xyz
[Link]('cuda') # or 'cuda0', 'cuda1' if multiple devices # datasets (can be used with [Link])
[Link]('cpu') # default import [Link] as aDatasets
[Link]('folder_for_storage', download=True)
# static computation graph/C++ export preparation [Link]('folder_for_storage', download=True)
[Link]()
from [Link] import script, trace
@script # transforms
import [Link] as aTransforms
# load and save a model [Link]
[Link](model, 'model_file') [Link]
model = [Link]('model_file') [Link]
[Link]() # set to inference [Link]
[Link](model.state_dict(), 'model_file') # only state dict [Link]
model = ModelCalss() [Link]
model.load_state_dict([Link]('model_file') [Link]
[Link]
# save to onnx
[Link] # kaldi support
[Link].export_to_pretty_string import [Link] as aKaldi
import torchaudio.kaldi_io as aKaldiIO
# load onnx model [Link]
import onnx [Link]
model = [Link]('[Link]') [Link]
# check model aKaldi.resample_waveform
aKaldiIO.read_vec_int_ark [Link]
aKaldiIO.read_vec_flt_scp [Link]
aKaldiIO.read_vec_flt_ark [Link]
aKaldiIO.read_mat_scp [Link]
aKaldiIO.read_mat_ark [Link]
[Link]
# functional/direct function access
import [Link] as aFunctional # question classification
[Link]
# sox effects/passing data between Python and C++
import torchaudio.sox_effects as aSox_effects # entailment
[Link]
Text [Link]
import torchtext # language modeling
# various data-related function and classes tDatasets.WikiText2
import [Link] as tData tDatasets.WikiText103
[Link] [Link]
[Link]
[Link] # machine translation
[Link] [Link] # subclass
[Link] tDatasets.Multi30k
[Link] [Link]
[Link] tDatasets.WMT14
[Link]
[Link] # sequence tagging
[Link] [Link] # subclass
[Link] [Link]
[Link] tDatasets.CoNLL2000Chunking
[Link] # similar to vTransform and sklearn's pipeline
[Link] # function # question answering
[Link] # function tDatasets.BABI20
# vocabulary and pre-trained embeddings
# datasets import [Link] as tVocab
import [Link] as tDatasets [Link] # create a vocabulary
# sentiment analysis [Link] # create subvocabulary
[Link] [Link] # word vectors
[Link] [Link] # GloVe embeddings
[Link] # subclass of all datasets below [Link] # FastText embeddings
tDatasets.AG_NEWS [Link] # character n-gram
[Link]
2
Vision # classification
[Link](pretrained=True)
import torchvision vModels.densenet121()
# datasets vModels.densenet161()
import [Link] as vDatasets vModels.densenet169()
[Link] vModels.densenet201()
[Link] [Link]()
[Link] vModels.inception_v3()
[Link] vModels.mnasnet0_5()
[Link] vModels.mnasnet0_75()
[Link] # randomly generated images vModels.mnasnet1_0()
[Link] vModels.mnasnet1_3()
[Link] vModels.mobilenet_v2()
[Link] vModels.resnet18()
[Link] # data loader for a certain image folder structure vModels.resnet34()
[Link] # data loader for a certain folder structure vModels.resnet50()
[Link] vModels.resnet50_32x4d()
[Link] vModels.resnet101()
vDatasets.STL10 vModels.resnet101_32x8d()
[Link] vModels.resnet152()
[Link] vModels.wide_resnet50_2()
[Link] vModels.wide_resnet101_2()
[Link] vModels.shufflenet_v2_x0_5()
[Link] vModels.shufflenet_v2_x1_0()
[Link] vModels.shufflenet_v2_x1_5()
[Link] vModels.shufflenet_v2_x2_0()
[Link] vModels.squeezenet1_0()
vDatasets.Kinetics400 vModels.squeezenet1_1()
vDatasets.HMDB51 vModels.vgg11()
vDatasets.UCF101 vModels.vgg11_bn()
vModels.vgg13()
# video IO vModels.vgg13_bn()
import [Link] as vIO vModels.vgg16()
vIO.read_video('file', start_pts, end_pts) vModels.vgg16_bn()
vIO.write_video('file', video, fps, video_codec) vModels.vgg19()
[Link].save_image(image,'file') vModels.vgg19_bn()
# pretrained models/model architectures # semantic segmentation
import [Link] as vModels [Link].fcn_resnet50()
# models can be constructed with random weights () [Link].fcn_resnet101()
# or pretrained (pretrained=True) [Link].deeplabv3_resnet50()
3
[Link].deeplabv3_resnet101()
# transforms on torch tensors
# object and/or keypoint detection, instance segmentation [Link]
[Link].fasterrcnn_resnet50_fpn() [Link]
[Link].maskrcnn_resnet50_fpn() [Link]
[Link].keypointrcnn_resnet50_fpn()
# conversion
# video classification [Link]
[Link].r3d_18() [Link]
[Link].mc3_18()
[Link].r2plus1d_18() # direct access to transform functions
import [Link] as vTransformsF
# transforms
import [Link] as vTransforms # operators for computer vision
[Link](transforms) # chaining transforms # (not supported by TorchScript yet)
[Link](someLambdaFunction) import [Link] as vOps
[Link] # non-maximum suppression (NMS)
# transforms on PIL images vOps.roi_align # <=> [Link]
[Link](height,width) vOps.roi_pool # <=> [Link]
[Link](brightness=0, contrast=0,
saturation=0, hue=0)
[Link] Data loader
[Link]
[Link] # classes and functions to represent datasets
[Link](degrees, translate=None, from [Link] import Dataset, Dataloader
scale=None, shear=None,
resample=False, fillcolor=0)
[Link](transforms, p=0.5) Neural network
[Link](transforms)
[Link] import [Link] as nn
[Link]
[Link] Activation functions
[Link]
[Link] [Link]
[Link] [Link]
[Link] [Link]
[Link] [Link]
[Link] [Link]
[Link] [Link]
[Link] [Link]
[Link] [Link]
[Link]
4
[Link] optim.lr_scheduler.Scheduler
[Link]
nn.ReLU6 # optimizers
[Link](lower,upper) # sampled from uniform distribution [Link] # general optimizer classes
[Link] [Link]
[Link] [Link]
[Link] [Link]
nn.Softmax2d [Link] # adam with decoupled weight decay regularization
[Link] [Link]
[Link] [Link] # averged stochastic gradient descent
[Link] [Link]
[Link] [Link]
[Link] [Link]
[Link] [Link]
[Link] [Link] # for sparse tensors
Loss function # learning rate
optim.lr_scheduler
[Link] optim.lr_scheduler.LambdaLR
[Link] optim.lr_scheduler.StepLR
[Link] optim.lr_scheduler.MultiStepLR
[Link] optim.lr_scheduler.ExponentialLR
[Link] optim.lr_scheduler.CosineAnnealingLR
[Link] optim.lr_scheduler.ReduceLROnPlateau
[Link] optim.lr_scheduler.CyclicLR
nn.L1Loss
[Link] Pre-defined layers/deep learning
[Link]
[Link] # containers
[Link] [Link]{ ,List,Dict}
[Link] [Link]{List,Dict}
[Link] [Link]
nn.SmoothL1Loss
[Link] # linear layers
[Link] [Link]
[Link]
[Link]
Optimizer
import [Link] as optim # dropout layers
# general useage [Link]
scheduler = [Link](....) [Link]{ ,2d,3d}
[Link]() # step-wise
5
# convolutional layers Functional
[Link]{1,2,3}d
import [Link] as F
[Link]{1,2,3}d
# direct function access and not via classes ([Link]) ???
[Link]
[Link]
NumPy-like functions
# pooling
[Link]{1,2,3}d Loading PyTorch and tensor basics
[Link]{1,2,3}d
[Link]{1,2,3}d # loading PyTorch
[Link]{1,2,3}d import torch
[Link]{1,2,3}d
# defining a tensor
# recurrent layers [Link]((values))
[Link]
[Link] # define data type
[Link] [Link]((values), dtype=torch.int16)
# padding layers # converting a NumPy array to a PyTorch tensor
[Link]{1,2}d torch.from_numpy(numpyArray)
[Link]{1,2,3}d
[Link]{1,2,3}d # create a tensor of zeros
[Link]((shape))
# normalization layers torch.zeros_like(other_tensor)
[Link]{1,2,3}d
[Link]{1,2,3}d # create a tensor of ones
[Link]((shape))
# transformer layers torch.ones_like(other_tensor)
[Link]
[Link] # create an idenity matrix
[Link] [Link](numberOfRows)
[Link]
[Link] # create tensor with same values
[Link]((shape), value)
torch.full_like(other_tensor,value)
# create an empty tensor
Computational graph [Link]((shape))
torch.empty_like(other_tensor)
# various functions and classes to use and manipulate
# automaic differentiation and the computational graph # create sequences
import [Link] as autograd [Link](startNumber, endNumber, stepSize)
6
[Link](startNumber, endNumber, stepSize) torch.randn_like(other_tensor)
[Link](startNUmber, endNumber, stepSize)
# random permuation of integers
# concatenate tensors # range [0,n-1)
[Link]((tensors), axis) [Link]()
# split tensors into sub-tensors Math (element-wise)
[Link](tensor, splitSize)
# basic operations
# (un)squeeze tensor [Link](tensor)
[Link](tensor, dimension) [Link](tensor, tensor2) # or tensor+scalar
[Link](tensor, dim) [Link](tensor, tensor2) # or tensor/scalar
[Link](tensor,tensor2) # or tensor*scalar
# reshape tensor [Link](tensor, tensor2) # or tensor-scalar
[Link](tensor, shape) [Link](tensor)
[Link](tensor)
# transpose tensor [Link](tensor, devisor) #or [Link]()
torch.t(tensor) # 1D and 2D tensors [Link](tensor)
[Link](tensor, dim0, dim1)
# trigonometric functions
[Link](tensor)
Random numbers
[Link](tensor)
# set seed [Link](tensor)
torch.manual_seed(seed) torch.atan2(tensor)
[Link](tensor)
# generate a tensor with random numbers [Link](tensor)
# of interval [0,1) [Link](tensor)
[Link](size) [Link](tensor)
torch.rand_like(other_tensor) [Link](tensor)
[Link](tensor)
# generate a tensor with random integer numbers
# of interval [lowerInt,higherInt] # exponentials and logarithms
[Link](lowerInt, [Link](tensor)
higherInt, torch.expm1(tensor) # exp(input-1)
(tensor_shape)) [Link](tensor)
torch.randint_like(other_tensor, torch.log10(tensor)
lowerInt, torch.log1p(tensor) # log(1+input)
higherInt) torch.log2(tensor)
# generate a tensor of random numbers drawn # other
# from a normal distribution (mean=0, var=1) [Link](tensor) # error function
[Link]((size)) [Link](tensor) # inverse error function
7
[Link](tensor) # round to full integer [Link](tensor) # pseudo-inverse
[Link](tensor, power)
Other
Math (not element-wise) [Link](tensor)
[Link](tensor) [Link](tensor)
[Link](tensor) [Link](tensor, signal_dim)
[Link](tensor) [Link](tensor, signal_dim)
[Link](tensor) [Link](tensor, signal_dim)
[Link](tensor) [Link](tensor, signal_dim)
[Link](tensor) [Link](tensor, n_fft)
[Link](tensor, norm) [Link](tensor)
[Link](tensor) # product of all elements [Link](tensor)
[Link](tensor) [Link](tensor, start_dim)
[Link](tensor) torch.rot90(tensor)
[Link](tensor) [Link](tensor)
[Link](tensor) [Link](tensor)
[Link](tensor1,tensor2) [Link](tensor)
torch.cartesian_prod(tensor1, tensor2, ...)
[Link](equation,tensor)
[Link](tensor1,tensor2)
PyTorch C++
[Link](tensor) (aka libtorch)
torch.cholesky_torch(tensor)
[Link](tensor1, tensor2) // PyTorch header file(s)
[Link](tensor) #import <torch/script.h>
[Link](tensor)
[Link](tensor) torch::jit::script::Module module;
©Simon Wenkel ([Link]
This pdf is licensed under the CC BY-SA 4.0 license.
