search

IBBM / Cascaded-FCN

Source code for the MICCAI 2016 Paper "Automatic Liver and Lesion Segmentation in CT Using Cascaded Fully Convolutional NeuralNetworks and 3D Conditional Random Fields"
Other
304 stars 127 forks source link

Pretrained TensorFlow Models #25

Closed zakizhou closed 6 years ago

zakizhou commented 6 years ago

Can you provide the pre-trained tensorflow model(including graph and ckpt) for us? Training from scratch really takes a lot of time and need gpus with big memory. Thank you very much!

alizeshana commented 6 years ago

Do let me know if you get it.

zakizhou commented 6 years ago

@alizeshana I forked that tensorflow-implementation branch and successfully started the training but then I got memory error because I only have one gtx 1060 6G card. If you have more powerful gpus maybe you can train it yourself

alizeshana commented 6 years ago

@zakizhou I am facing the same issue on my gtx 740M 2GB. :p

Do let me know if you find the trained model.

mohamed-ezz commented 6 years ago

I think it's not very hard to convert the .caffeweights provided in this repo to a Keras .hdf5.

This is roughly how I would do it :

  1. Define the same network in Keras (this may be the most time consuming). Make sure there are no implementation differences for any of the layers between Caffe and Keras.
  2. A script Loads the .caffeweights and prototxt using caffe python library
  3. Compile the keras architecture
  4. Use keras's model.layers[i].set_weights(caffeweights), while figuring out any reshaping needed. Do this for every layer.
  5. Save Keras's model as .hdf5 model.save_weights(...)
mohamed-ezz commented 6 years ago

If you manage to do this, I would appreciate sharing your work with others by submitting a PR with the Tensorflow/Keras weights.