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

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

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.

0. Top-1/5 accuracy of pre-train models in this repository.

• 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.

0. Top-1/5 accuracy with different crop sizes.

• Figure: Accuracy curves of inception_v3(left) and resnet101_v2(right) with different crop sizes.

0. 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

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

   ~/Database/ILSVRC2012
   

1. 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
   

2. then

   python evaluation_cls.py
   

I greatly thank Yangqing Jia and BVLC group for developing Caffe

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