/caffe-model

Caffe models (imagenet pretrain) and prototxt generator scripts for inception_v3 \ inception_v4 \ inception_resnet \ fractalnet \ resnext

Primary LanguagePython

Caffe-model

Python script to generate prototxt on Caffe, specially the inception_v3\inception_v4\inception_resnet\fractalnet

Generator scripts

The prototxts can be visualized by ethereon.

Every model has a bn (batch normalization) version (maybe only bn version), the paper is Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Classificaiton (imagenet)

Introduction

This folder contains the deploy files(include generator scripts) and pre-train models of resnet-v1, resnet-v2, inception-v3, inception-resnet-v2 and densenet(coming soon).

We didn't train any model from scratch, some of them are converted from other deep learning framworks (inception-v3 from mxnet, inception-resnet-v2 from tensorflow), some of them are converted from other modified caffe (resnet-v2). But to achieve the original performance, finetuning is performed on imagenet for several epochs.

The main contribution belongs to the authors and model trainers.

Performance on imagenet

  1. Top-1/5 accuracy of pre-train models in this repository.
Network 224/299(single-crop) 224/299(12-crop) 320/395(single-crop) 320/395(12-crop)
resnet101-v2 78.05/93.88 80.01/94.96 79.63/94.84 80.71/95.43
resnet152-v2 79.15/94.58 80.76/95.32 80.34/95.26 81.16/95.68
resnet269-v2 80.29/95.00 81.75/95.80 81.30/95.67 82.13/96.15
inception-v3 78.33/94.25 80.40/95.27 79.90/95.18 80.75/95.76
inception-v4 79.97/94.91 81.40/95.70 81.32/95.68 81.88/96.08
inception-resnet-v2 80.14/95.17 81.54/95.92 81.25/95.98 81.85/96.29
resnext50_32x4d 77.63/93.69 79.47/94.65 78.90/94.47 79.63/94.97
resnext101_32x4d 78.70/94.21 80.53/95.11 80.09/95.03 80.81/95.41
resnext101_64x4d 79.40/94.59 81.12/95.41 80.74/95.37 81.52/95.69
wrn50_2(resnet50_1x128d) 77.87/93.87 79.91/94.94 79.32/94.72 80.17/95.13
  • The pre-train models are tested on original caffe by evaluation_cls.py, but ceil_mode:false(pooling_layer) is used for the models converted from torch, the detail in https://github.com/BVLC/caffe/pull/3057/files. If you remove ceil_mode:false, the performance will decline about 1% top1.
  • 224x224(base_size=256) and 320x320(base_size=320) crop size for resnet-v2/resnext/wrn, 299x299(base_size=320) and 395x395(base_size=395) crop size for inception.
  1. Top-1/5 accuracy with different crop sizes. teaser
  • Figure: Accuracy curves of inception_v3(left) and resnet101_v2(right) with different crop sizes.
  1. Download url and forward/backward time cost for each model.

Forward/Backward time cost is evaluated with one image/mini-batch using cuDNN 5.1 on a Pascal Titan X GPU.

We use

  ~/caffe/build/tools/caffe -model deploy.prototxt time -gpu -iterations 1000

to test the forward/backward time cost, the result is really different with time cost of evaluation_cls.py

Network F/B(224/299) F/B(320/395) Download Source
resnet101-v2 22.31/22.75ms 26.02/29.50ms 170.3MB craftGBD
resnet152-v2 32.11/32.54ms 37.46/41.84ms 230.2MB craftGBD
resnet269-v2 58.20/59.15ms 69.43/77.26ms 390.4MB craftGBD
inception-v3 21.79/19.82ms 22.14/24.88ms 91.1MB mxnet
inception-v4 32.96/32.19ms 36.04/41.91ms 163.1MB tensorflow_slim
inception-resnet-v2 49.06/54.83ms 54.06/66.38ms 213.4MB tensorflow_slim
resnext50_32x4d 17.29/20.08ms 19.02/23.81ms 95.8MB facebookresearch
resnext101_32x4d 30.73/35.75ms 34.33/41.02ms 169.1MB facebookresearch
resnext101_64x4d 42.07/64.58ms 51.99/77.71ms 319.2MB facebookresearch
wrn50_2(resnet50_1x128d) 16.48/25.28ms 20.99/35.04ms 263.1MB szagoruyko

Check the performance

  1. Download the ILSVRC 2012 classification val set 6.3GB, and put the extracted images into the directory:

    ~/Database/ILSVRC2012
    
  2. Check the resnet-v2 (101, 152 and 269) performance, the settings of evaluation_cls.py:

    val_file = 'ILSVRC2012_val.txt' # download from this folder, label range 0~999
    ... ...
    model_weights = 'resnet-v2/resnet101_v2.caffemodel' # download as below
    model_deploy = 'resnet-v2/deploy_resnet101_v2.prototxt' # check the parameters of input_shape
    ... ...
    mean_value = np.array([102.9801, 115.9465, 122.7717])  # BGR
    std = np.array([1.0, 1.0, 1.0])  # BGR
    crop_num = 1    # perform center(single)-crop
    

    Check the inception-v3 performance, the settings of evaluation_cls.py:

    val_file = 'ILSVRC2015_val.txt' # download from this folder, label range 0~999
    ... ...
    model_weights = 'inception_v3/inception_v3.caffemodel' # download as below
    model_deploy = 'inception_v3/deploy_inception_v3.prototxt' # check the parameters of input_shape
    ... ...
    mean_value = np.array([128.0, 128.0, 128.0])  # BGR
    std = np.array([128.0, 128.0, 128.0])  # BGR
    crop_num = 1    # perform center(single)-crop
    

    Check the inception-resnet-v2 (inception-v4) performance, the settings of evaluation_cls.py:

    val_file = 'ILSVRC2012_val.txt' # download from this folder, label range 0~999
    ... ...
    model_weights = 'inception_resnet_v2/inception_resnet_v2.caffemodel' # download as below
    model_deploy = 'inception_resnet_v2/deploy_inception_resnet_v2.prototxt' # check the parameters of input_shape
    ... ...
    mean_value = np.array([128.0, 128.0, 128.0])  # BGR
    std = np.array([128.0, 128.0, 128.0])  # BGR
    crop_num = 1    # perform center(single)-crop
    

    Check the resnext (50_32x4d, 101_32x4d and 101_64x4d) or wrn50_2 performance, the settings of evaluation_cls.py:

    val_file = 'ILSVRC2012_val.txt' # download from this folder, label range 0~999
    ... ...
    model_weights = 'inception_resnet_v2/inception_resnet_v2.caffemodel' # download as below
    model_deploy = 'inception_resnet_v2/deploy_inception_resnet_v2.prototxt' # check the parameters of input_shape
    ... ...
    mean_value = np.array([103.52, 116.28, 123.675])  # BGR
    std = np.array([57.375, 57.12, 58.395])  # BGR
    crop_num = 1    # perform center(single)-crop
    
  3. then

    python evaluation_cls.py
    

Acknowlegement

I greatly thank Yangqing Jia and BVLC group for developing Caffe

And I would like to thank all the authors of every cnn model