wyli
commented
6 years ago Hi @johnnychhsu an example of transfer learning is presented here -- https://github.com/NifTK/NiftyNet/pull/140. It's still a work in progress, mainly we need a generic user interface (via config file) to handle these changes so that the user is able to specify which variables to be initialised from checkpoint files and the others from random initialisation. If you are also interested in this direction, we could collaborate on it.
Alternatively if you want to have full control of the graph, the code example attached could be a good starting point. A caveat is that the preprocessing layers (not included here) should exactly follow the ones used for training e.g. https://github.com/NifTK/NiftyNet/blob/v0.3.0/niftynet/application/segmentation_application.py#L178.
import tensorflow as tf
import os
from niftynet.io.image_reader import ImageReader
from niftynet.engine.sampler_resize import ResizeSampler
import numpy as np
os.environ["CUDA_VISIBLE_DEVICES"]='1'
##### Address of the model to be restored
check_point_location='/home/niftynet/models/dense_vnet_abdominal_ct/models/model.ckpt-3000'
#####
##### Create a sampler
data_param = {'image': {'path_to_search': '~/niftynet/data/dense_vnet_abdominal_ct',
'filename_contains': 'CT', 'spatial_window_size': (144, 144, 144)}}
reader = ImageReader().initialise(data_param)
sampler = ResizeSampler(
reader=reader,
data_param=data_param,
batch_size=1,
shuffle_buffer=True,
queue_length=35)
#####
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False)) as sess:
sampler.run_threads(sess, tf.train.Coordinator(), num_threads=1)
from niftynet.network.dense_vnet import DenseVNet
data_dict = sampler.pop_batch_op()
net_logits = DenseVNet(num_classes=9)(data_dict['image'])
# restore the variables
saver = tf.train.Saver()
saver.restore(sess, check_point_location)
net_logits = sess.run(net_logits)
print(net_logits.shape)
johnnychhsu
AkhilaPerumalla123
alekswithakayy
koriavinash1
jillianlee
I am wondering if we can use NiftyNet pre-trained model in a way like TensorflowHub, such as the example from TensorflowHub :
I want to do transfer learning using the pre-trained model on my dataset, I think I can modify the config file to do what I want, but I am used to the tensorflowhub way. Thus asking this question. So the question are :
Thank you!