This is a 'canonical' TensorFlow implementation of VGG, using TF's own library mechanisms to simplify model creation, saving, and loading. The result is a very compact model. Features:
- Both VGG16 and VGG19 are implemented in the same code (can be selected with
vgg_typeparameter) - Writes network graph definition as a MetaGraphDef file.
- Loads/saves network weights using the
tf.Savermechanism.
The way to use this source code is as follows. First, the MetaGraphDef file has to be generated:
python vgg.py <patch_size> <16|19> <num_classes> <output_file.meta>patch_size is the size of the input patch in pixels. It is best to pick a multiple of 32 smaller than or equal to 224, for example 192. Choose either '16' or '19' to select the vgg type. Select the dropout_fraction to be 0.0 if testing and, for instance, 0.5 for training. num_classes is the number of output classes. VOC is 1000; for many segmentation purposes it can be 2. Finally, output_file is the output MetaGraphDef file to write.
The second script does the 'net surgery' to re-use trained variables from the stock VGG model. If variable shapes match, they are used, otherwise the variables are initialized. The following command does this:
python src/load_weights.py models/vgg16.meta models/vgg16.h5 models/vgg16_customIt loads the MetaGraphDef file we just generated, along with the stock vgg16.h5 weights, and outputs a vgg16_custom file using the tf.train.Saver() mechanism. Make sure that cache/vgg16.h5 or cache/vgg19.h5 exist and contain the proper initializions for the VGG model.
Now, we are ready to run the training and/or testing operations:
python src/train.py cache/colorectal.h5 models/vgg16_custom-0Make sure that the cache/colorectal.h5 file exists and contains the training data.
TensorFlow is not Python.