Finetuning Caffe Imagenet on CIFAR10 data

[Gil Levi]

Hi,

Following the flickr fine-tune and the pascal fine-tune examples, I trying to fine-tune the imagenet network on the CIFAR10 dataset (I resized the images to 256x256).

However, the loss doesn't seem to be reducing. 

Can someone please help me to detect the problem?

Thanks,

Gil

Implementation details:

Here is my train_val prototxt file:

name: "FlickrStyleCaffeNet"

layers {

  name: "data"

  type: DATA

  top: "data"

  top: "label"

  data_param {

    source: "/home/michael/CIFAR10/data_256_resized/resized_train_leveldb"

    mean_file: "/home/michael/CIFAR10/data_256_resized/imagenet_mean_cifar_resized.binaryproto"

    batch_size: 50

    crop_size: 227

    mirror: true

  }

  include: { phase: TRAIN }

}

layers {

  name: "data"

  type: DATA

  top: "data"

  top: "label"

  data_param {

    source: "/home/michael/CIFAR10/data_256_resized/resized_val_leveldb"

    mean_file: "/home/michael/CIFAR10/data_256_resized/imagenet_mean_cifar_resized.binaryproto"

    batch_size: 50

    crop_size: 227

    mirror: false

  }

  include: { phase: TEST }

}

layers {

  name: "conv1"

  type: CONVOLUTION

  bottom: "data"

  top: "conv1"

  blobs_lr: 1

  blobs_lr: 2

  weight_decay: 1

  weight_decay: 0

  convolution_param {

    num_output: 96

    kernel_size: 11

    stride: 4

    weight_filler {

      type: "gaussian"

      std: 0.01

    }

    bias_filler {

      type: "constant"

      value: 0

    }

  }

}

layers {

  name: "relu1"

  type: RELU

  bottom: "conv1"

  top: "conv1"

}

layers {

  name: "pool1"

  type: POOLING

  bottom: "conv1"

  top: "pool1"

  pooling_param {

    pool: MAX

    kernel_size: 3

    stride: 2

  }

}

layers {

  name: "norm1"

  type: LRN

  bottom: "pool1"

  top: "norm1"

  lrn_param {

    local_size: 5

    alpha: 0.0001

    beta: 0.75

  }

}

layers {

  name: "conv2"

  type: CONVOLUTION

  bottom: "norm1"

  top: "conv2"

  blobs_lr: 1

  blobs_lr: 2

  weight_decay: 1

  weight_decay: 0

  convolution_param {

    num_output: 256

    pad: 2

    kernel_size: 5

    group: 2

    weight_filler {

      type: "gaussian"

      std: 0.01

    }

    bias_filler {

      type: "constant"

      value: 1

    }

  }

}

layers {

  name: "relu2"

  type: RELU

  bottom: "conv2"

  top: "conv2"

}

layers {

  name: "pool2"

  type: POOLING

  bottom: "conv2"

  top: "pool2"

  pooling_param {

    pool: MAX

    kernel_size: 3

    stride: 2

  }

}

layers {

  name: "norm2"

  type: LRN

  bottom: "pool2"

  top: "norm2"

  lrn_param {

    local_size: 5

    alpha: 0.0001

    beta: 0.75

  }

}

layers {

  name: "conv3"

  type: CONVOLUTION

  bottom: "norm2"

  top: "conv3"

  blobs_lr: 1

  blobs_lr: 2

  weight_decay: 1

  weight_decay: 0

  convolution_param {

    num_output: 384

    pad: 1

    kernel_size: 3

    weight_filler {

      type: "gaussian"

      std: 0.01

    }

    bias_filler {

      type: "constant"

      value: 0

    }

  }

}

layers {

  name: "relu3"

  type: RELU

  bottom: "conv3"

  top: "conv3"

}

layers {

  name: "conv4"

  type: CONVOLUTION

  bottom: "conv3"

  top: "conv4"

  blobs_lr: 1

  blobs_lr: 2

  weight_decay: 1

  weight_decay: 0

  convolution_param {

    num_output: 384

    pad: 1

    kernel_size: 3

    group: 2

    weight_filler {

      type: "gaussian"

      std: 0.01

    }

    bias_filler {

      type: "constant"

      value: 1

    }

  }

}

layers {

  name: "relu4"

  type: RELU

  bottom: "conv4"

  top: "conv4"

}

layers {

  name: "conv5"

  type: CONVOLUTION

  bottom: "conv4"

  top: "conv5"

  blobs_lr: 1

  blobs_lr: 2

  weight_decay: 1

  weight_decay: 0

  convolution_param {

    num_output: 256

    pad: 1

    kernel_size: 3

    group: 2

    weight_filler {

      type: "gaussian"

      std: 0.01

    }

    bias_filler {

      type: "constant"

      value: 1

    }

  }

}

layers {

  name: "relu5"

  type: RELU

  bottom: "conv5"

  top: "conv5"

}

layers {

  name: "pool5"

  type: POOLING

  bottom: "conv5"

  top: "pool5"

  pooling_param {

    pool: MAX

    kernel_size: 3

    stride: 2

  }

}

layers {

  name: "fc6"

  type: INNER_PRODUCT

  bottom: "pool5"

  top: "fc6"

  blobs_lr: 1

  blobs_lr: 2

  weight_decay: 1

  weight_decay: 0

  inner_product_param {

    num_output: 4096

    weight_filler {

      type: "gaussian"

      std: 0.005

    }

    bias_filler {

      type: "constant"

      value: 1

    }

  }

}

layers {

  name: "relu6"

  type: RELU

  bottom: "fc6"

  top: "fc6"

}

layers {

  name: "drop6"

  type: DROPOUT

  bottom: "fc6"

  top: "fc6"

  dropout_param {

    dropout_ratio: 0.5

  }

}

layers {

  name: "fc7"

  type: INNER_PRODUCT

  bottom: "fc6"

  top: "fc7"

  # Note that blobs_lr can be set to 0 to disable any fine-tuning of this, and any other, layer

  blobs_lr: 1

  blobs_lr: 2

  weight_decay: 1

  weight_decay: 0

  inner_product_param {

    num_output: 4096

    weight_filler {

      type: "gaussian"

      std: 0.005

    }

    bias_filler {

      type: "constant"

      value: 1

    }

  }

}

layers {

  name: "relu7"

  type: RELU

  bottom: "fc7"

  top: "fc7"

}

layers {

  name: "drop7"

  type: DROPOUT

  bottom: "fc7"

  top: "fc7"

  dropout_param {

    dropout_ratio: 0.5

  }

}

layers {

  name: "fc8_cifar"

  type: INNER_PRODUCT

  bottom: "fc7"

  top: "fc8_cifar"

  # blobs_lr is set to higher than for other layers, because this layer is starting from random while the others are already trained

  blobs_lr: 10

  blobs_lr: 20

  weight_decay: 1

  weight_decay: 0

  inner_product_param {

    num_output: 10

    weight_filler {

      type: "gaussian"

      std: 0.01

    }

    bias_filler {

      type: "constant"

      value: 0

    }

  }

}

layers {

  name: "loss"

  type: SOFTMAX_LOSS

  bottom: "fc8_cifar"

  bottom: "label"

}

layers {

  name: "accuracy"

  type: ACCURACY

  bottom: "fc8_flickr"

  bottom: "label"

  top: "accuracy"

  include: { phase: TEST }

}

And here is the solver file:

net: "finetune_train_val_prototxt.txt"

test_iter: 100

test_interval: 1000

# lr for fine-tuning should be lower than when starting from scratch

base_lr: 0.001

lr_policy: "step"

gamma: 0.1

# stepsize should also be lower, as we're closer to being done

stepsize: 20000

display: 20

max_iter: 100000

momentum: 0.9

weight_decay: 0.0005

snapshot: 10000

snapshot_prefix: "finetune_cifar"

# uncomment the following to default to CPU mode solving

solver_mode: CPU

[Yangqing Jia]

It most likely won't work. Imagine what the input would look like: they are 32x32 images resized to 256x256. The Imagenet model takes 7x7 patches, which only correspond to 0.875 pixels in the original image. As a result, they won't see anything interesting but an almost constant-colored patch.

Yangqing

--
You received this message because you are subscribed to the Google Groups "Caffe Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to caffe-users...@googlegroups.com.
To post to this group, send email to caffe...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/caffe-users/daf919fb-ba00-43c9-9c8e-e5b767e3ebac%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Gil Levi]

Thanks, I actually didn't feed the "train" binary with the initial weights, so I was training from scratch. 

Still, I'm taking your comment into consideration, but I will give it a shot and see what happens. 

Thanks again,

Gil 

[deepcnn]

Does caffe provide a Imagenet model takes 7x7 patches ?  Because I remember the Alex network takes 11*11 patches in the first convolutional layer. If it does, where could I get the model? Thanks evry much!

Yangqing Jia於 2014年10月8日星期三UTC+8上午2時25分38秒寫道：

[deepcnn]

A little mofication I want to suggest you: the name of accuracy layer should be cifar10, I am not sure whether this is a really possible lead to problem, but from the view of naming consistence, I think need change. 

layers {

  name: "loss"

  type: SOFTMAX_LOSS

  bottom: "fc8_cifar"

  bottom: "label"

}

layers {

  name: "accuracy"

  type: ACCURACY

  bottom: "fc8_flickr"  #---> fc8_cifar

  bottom: "label"

  top: "accuracy"

  include: { phase: TEST }

}

Gil Levi於 2014年10月8日星期三UTC+8上午2時32分19秒寫道：

[Gil Levi]

Thanks for your comment. 

Actually I got it working and currently the accuracy on the validation set is 87% after 3K iteration. Hope I don't have a bug somewhere. 

Regarding your question, the size of the filters in the first later of AlexNet and Caffe Imagenet are indeed 11x11, but there are new deeper models on the "Model Zoo" where the size of the filters is 3x3. 

Thanks again!

Gil. 

[Venkatesh N]

Hi,

 

How did you manage to resolve this issue? did you just resize 32x32 to 256x256 and used the fine-tuning on Caffe imagenet network? Were you able to find out the reason about why didn't the loss function decrease with your initial settings?

Thanks,

Venky

[Gil Levi]

Hi,

I resized the images to 256x256 when creating the leveldb. 

I have a bug somewhere, I don't really remember where, but I finally got it working.

Gil

[Venkatesh N]

Thanks, Gil. 

[Martin Yu]

Hi, Gil

I'm trying to resize cifar10 dataset and then wanna feed them into imagenet network. However, the images after resizing are weird. For example:

Top: 32x32 cifar10 image

Bottom: ( resized ) 256x256 image

I use caffe.io.resize_image to do resizing. How do you resize your images? Do you oversample them to obtain ten 227x227 crops? Thank you very much.

[Gil Levi]

Hi,

Sorry for the late response.

I'm not resizing my images. Where I created the leveldb, Caffe resized them automatically (I passed it the resize flag).

When testing the net, I just fed the 32x32x3 images to Caffe. 

I can send you the prediction scripts (in python). Keep in mind though that I'm using a somewhat old version of Caffe.

Gil. 

[H KK]

how to feed the train binary with the initial weights? The finetune-net.cpp seems be abolished...Thank U!

在 2014年10月8日星期三 UTC+8上午2:32:19，Gil Levi写道：

[Gil Levi]

Here:

TOOLS=/home/caffe-master/build/tools

TXT_DIR=/home//prototxt_files

$TOOLS/caffe train --solver=$TXT_DIR/finetune_solver_prototxt.txt -weights /home/reference_net/caffe_reference_imagenet_model

Gil

[李組賢]

HI~

How you resized the cifar10 data when creating lmdb. What can I do for it, Thx~ 

[Gil Levi]

I resized it to 256x256. 

[Nikiforos Pittaras]

Googlenet feature extraction works ok with me, after finetuning on 38 concepts.

I renamed and changed the output sizes if the fc  and loss layers.

[李組賢]

Thx for your response. But I mean what command should i add to convert 32X32 to 256X256  Thx~~ QAQ I am a new learner

Gil Levi於 2015年4月6日星期一 UTC+8下午11時18分24秒寫道：

[Gil Levi]

Here's an example to the shell script that I'm using for creating the leveldb:

TOOLS=/home/ubuntu/repositories/caffe/build/tools

DATA=/

DEF_FILES=/mnt/CASIA/CasiaWebFace/train_val_txt_files

OUT=/mnt/CASIA/CasiaWebFace/leveldb

# Set RESIZE=true to resize the images to 256x256. Leave as false if images have

# already been resized using another tool.

RESIZE=true

if $RESIZE; then

  RESIZE_HEIGHT=256

  RESIZE_WIDTH=256

else

  RESIZE_HEIGHT=0

  RESIZE_WIDTH=0

fi

if [ ! -d "$DATA" ]; then

  echo "Error: DATA is not a path to a directory: $DATA"

  echo "Set the DATA variable in create_imagenet.sh to the path" \

       "where the ImageNet training data is stored."

  exit 1

fi

echo "Creating  train leveldb..."

GLOG_logtostderr=1 $TOOLS/convert_imageset.bin --resize_height=$RESIZE_HEIGHT --resize_width=$RESIZE_WIDTH --shuffle  $DATA $DEF_FILES/train.txt $OUT/train_leveldb

[李組賢]

Can this do with .bin Data ? Datasets download from website is name .bin(binary)              THank you ~~~~~~~: ))))  

Gil Levi於 2015年4月7日星期二 UTC+8下午9時01分18秒寫道：

[Gil Levi]

I'm not sure. 

[李組賢]

Thank you. Can you share the URL where you downloaded that datasets?  : )

[Nazil Perveen]

Hi Gil,

I am new to finetuning in caffe and after going through different documentation, I just want to ask you one question what Yangqing suggest, that how the image size will be if you resize it from 32 x 32 to 256 x 256. Is that the correct way to fintune the model on small image-size dataset ?.

[Gil Levi]

Hi,

Yes, it actually worked. I got nice results (above 90%).

Best,

Gil