FCN-xs EXAMPLE

This folder contains an example implementation for Fully Convolutional Networks (FCN) in MXNet.
The example is based on the FCN paper by long et al. of UC Berkeley.

Sample results

We have trained a simple fcn-xs model, the hyper-parameters are below:

(when using the newest mxnet, you'd better using larger learning rate, such as 1e-4, 1e-5, 1e-6 instead, because the newest mxnet will do gradient normalization in SoftmaxOutput)

The training dataset size is only 2027, and the validation dataset size is 462.

Training the model

Step 1: setup pre-requisites

• Install python package Pillow (required by image_segment.py).

pip install --upgrade Pillow

• Setup your working directory. Assume your working directory is ~/train_fcn_xs, and MXNet is built as ~/mxnet. Copy example scripts into the working directory.

cp ~/mxnet/example/fcn-xs/* .

Step 2: Download the vgg16fc model and training data

• vgg16fc model: you can download the VGG_FC_ILSVRC_16_layers-symbol.json and VGG_FC_ILSVRC_16_layers-0074.params from baidu yun, dropbox.
  this is the fully convolution style of the origin VGG_ILSVRC_16_layers.caffemodel, and the corresponding VGG_ILSVRC_16_layers_deploy.prototxt, the vgg16 model has license for non-commercial use only.
• Training data: download the VOC2012.rar robots.ox.ac.uk, and extract it into .\VOC2012
• Mapping files: download train.lst, val.lst from baidu yun into the .\VOC2012 directory

Once you completed all these steps, your working directory should contain a .\VOC2012 directory, which contains the following: JPEGImages folder, SegmentationClass folder, train.lst, val.lst

Step 3: Train the fcn-xs model

• Based on your hardware, configure GPU or CPU for training in fcn_xs.py. It is recommended to use GPU due to the computational complexity and data load.

# ctx = mx.cpu(0)
ctx = mx.gpu(0)

• It is recommended to train fcn-32s and fcn-16s before training the fcn-8s model

To train the fcn-32s model, run the following:

python -u fcn_xs.py --model=fcn32s --prefix=VGG_FC_ILSVRC_16_layers --epoch=74 --init-type=vgg16

• In the fcn_xs.py, you may need to change the directory root_dir, flist_name, ``fcnxs_model_prefix``` for your own data.
• When you train fcn-16s or fcn-8s model, you should change the code in run_fcnxs.sh corresponding, such as when train fcn-16s, comment out the fcn32s script, then it will like this:

 python -u fcn_xs.py --model=fcn16s --prefix=FCN32s_VGG16 --epoch=31 --init-type=fcnxs

• The output log may look like this(when training fcn-8s):

INFO:root:Start training with gpu(3)
INFO:root:Epoch[0] Batch [50]   Speed: 1.16 samples/sec Train-accuracy=0.894318
INFO:root:Epoch[0] Batch [100]  Speed: 1.11 samples/sec Train-accuracy=0.904681
INFO:root:Epoch[0] Batch [150]  Speed: 1.13 samples/sec Train-accuracy=0.908053
INFO:root:Epoch[0] Batch [200]  Speed: 1.12 samples/sec Train-accuracy=0.912219
INFO:root:Epoch[0] Batch [250]  Speed: 1.13 samples/sec Train-accuracy=0.914238
INFO:root:Epoch[0] Batch [300]  Speed: 1.13 samples/sec Train-accuracy=0.912170
INFO:root:Epoch[0] Batch [350]  Speed: 1.12 samples/sec Train-accuracy=0.912080

Using the pre-trained model for image segmentation

To try out the pre-trained model, follow these steps:

• Download the pre-trained symbol and weights from yun.baidu. You should download these files: FCN8s_VGG16-symbol.json and FCN8s_VGG16-0019.params
• Run the segmentation script, providing it your input image path: python image_segmentaion.py --input <your JPG image path>
• The segmented output .png file will be generated in the working directory

Tips

• This example runs full image size training, so there is no need to resize or crop input images to the same size. Accordingly, batch_size during training is set to 1.
• The fcn-xs model is based on vgg16 model, with some crop, deconv, element-sum layer added, so the model is quite big, moreover, the example is using whole image size training, if the input image is large(such as 700*500), then memory consumption may be high. Due to that, I suggest you use GPU with at least 12GB memory for training.
• If you don't have access to GPU with 12GB memory for training, I suggest you change the cut_off_size to a small value when constructing the FileIter, example below:

train_dataiter = FileIter(
      root_dir             = "./VOC2012",
      flist_name           = "train.lst",
      cut_off_size         = 400,
      rgb_mean             = (123.68, 116.779, 103.939),
      )