This is a playground for pytorch beginners, which contains predefined models on popular dataset. Currently we support

• mnist, svhn
• cifar10, cifar100
• stl10
• alexnet
• vgg16, vgg16_bn, vgg19, vgg19_bn
• resnet18, resnet34, resnet50, resnet101, resnet152
• squeezenet_v0, squeezenet_v1
• inception_v3

Here is an example for MNIST dataset. This will download the dataset and pre-trained model automatically.

import torch
from torch.autograd import Variable
from utee import selector
model_raw, ds_fetcher, is_imagenet = selector.select('mnist')
ds_val = ds_fetcher(batch_size=10, train=False, val=True)
for idx, (data, target) in enumerate(ds_val):
    data =  Variable(torch.FloatTensor(data)).cuda()
    output = model_raw(data)

Also, if want to train the MLP model on mnist, simply run python mnist/train.py

Install

python3 setup.py develop --user

ImageNet dataset

We provide precomputed imagenet validation dataset with 224x224x3 size. We first resize the shorter size of image to 256, then we crop 224x224 image in the center. Then we encode the cropped images to jpg string and dump to pickle.

• cd script
• Download the val224_compressed.pkl (Tsinghua / Google Drive)
• python convert.py (needs 48G memory, thanks @jnorwood )

Quantization

We also provide a simple demo to quantize these models to specified bit-width with several methods, including linear method, minmax method and non-linear method.

quantize --type cifar10 --quant_method linear --param_bits 8 --fwd_bits 8 --bn_bits 8 --ngpu 1

Top1 Accuracy

We evaluate the performance of popular dataset and models with linear quantized method. The bit-width of running mean and running variance in BN are 10 bits for all results. (except for 32-float)

Model	32-float	12-bit	10-bit	8-bit	6-bit
MNIST	98.42	98.43	98.44	98.44	98.32
SVHN	96.03	96.03	96.04	96.02	95.46
CIFAR10	93.78	93.79	93.80	93.58	90.86
CIFAR100	74.27	74.21	74.19	73.70	66.32
STL10	77.59	77.65	77.70	77.59	73.40
AlexNet	55.70/78.42	55.66/78.41	55.54/78.39	54.17/77.29	18.19/36.25
VGG16	70.44/89.43	70.45/89.43	70.44/89.33	69.99/89.17	53.33/76.32
VGG19	71.36/89.94	71.35/89.93	71.34/89.88	70.88/89.62	56.00/78.62
ResNet18	68.63/88.31	68.62/88.33	68.49/88.25	66.80/87.20	19.14/36.49
ResNet34	72.50/90.86	72.46/90.82	72.45/90.85	71.47/90.00	32.25/55.71
ResNet50	74.98/92.17	74.94/92.12	74.91/92.09	72.54/90.44	2.43/5.36
ResNet101	76.69/93.30	76.66/93.25	76.22/92.90	65.69/79.54	1.41/1.18
ResNet152	77.55/93.59	77.51/93.62	77.40/93.54	74.95/92.46	9.29/16.75
SqueezeNetV0	56.73/79.39	56.75/79.40	56.70/79.27	53.93/77.04	14.21/29.74
SqueezeNetV1	56.52/79.13	56.52/79.15	56.24/79.03	54.56/77.33	17.10/32.46
InceptionV3	76.41/92.78	76.43/92.71	76.44/92.73	73.67/91.34	1.50/4.82

Note: ImageNet 32-float models are directly from torchvision

Selected Arguments

Here we give an overview of selected arguments of quantize.py

Flag	Default value	Description & Options
type	cifar10	mnist,svhn,cifar10,cifar100,stl10,alexnet,vgg16,vgg16_bn,vgg19,vgg19_bn,resent18,resent34,resnet50,resnet101,resnet152,squeezenet_v0,squeezenet_v1,inception_v3
quant_method	linear	quantization method:linear,minmax,log,tanh
param_bits	8	bit-width of weights and bias
fwd_bits	8	bit-width of activation
bn_bits	32	bit-width of running mean and running vairance
overflow_rate	0.0	overflow rate threshold for linear quantization method
n_samples	20	number of samples to make statistics for activation