fully_convolutional_networks running error

[xray1111]

Hi,

I clone the notebook files from your git site: git clone https://github.com/de-code/tensorflow_notes.git

But when I run fully_convolutional_networks.ipynb, there's still an error:

ValueError Traceback (most recent call last) in () 47 pred, fcn_16s_variables_mapping = FCN_8s(image_batch_tensor=image_batch_tensor, 48 number_of_classes=number_of_classes, ---> 49 is_training=False) 50 51 # The op for initializing the variables.

/home1/data/xiaoren/tf-image-segmentation/tf_image_segmentation/utils/inference.pyc in new_network_definition(*args, **kwargs) 51 kwargs['image_batch_tensor'] = resized_images_batch 52 ---> 53 all_outputs = network_definition(*args, **kwargs) 54 55 all_outputs = list(all_outputs)

/home1/data/xiaoren/tf-image-segmentation/tf_image_segmentation/models/fcn_8s.py in FCN_8s(image_batch_tensor, number_of_classes, is_training) 77 is_training=is_training, 78 spatial_squeeze=False, ---> 79 fc_conv_padding='SAME') 80 81

/home1/data/xiaoren/tf-image-segmentation/models/slim/nets/vgg.py in vgg_16(inputs, num_classes, is_training, dropout_keep_prob, spatial_squeeze, scope, fc_conv_padding) 164 with slim.arg_scope([slim.conv2d, slim.fully_connected, slim.max_pool2d], 165 outputs_collections=end_points_collection): --> 166 net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1') 167 net = slim.max_pool2d(net, [2, 2], scope='pool1') 168 net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')

...... I'm not sure if it because the code below, but I don't know where to download that check point file(model_fcn8s_final.ckpt):

fcn_16s_checkpoint_path = '/home1/data/xiaoren/tf-image-segmentation/ckpt/model_fcn8s_final.ckpt'

I really need your help. Thank you!

BTW, I've clone the tensorflow/models instead of from the author's site, because of version 1.0 problem.

[de-code]

My 'new-tf-api-python3-and-relative-paths' branch / PR hasn't fixed anything for the fully_convolutional_networks notebook and won't help you in this case.

In the GitHub README of tf-image-segmentation there is a link to the what appears to be the checkpoints (haven't tried it myself).

[araitatsuya]

Same here. I could not find model_fcn8s_final.ckpt.

I tested the dropbox link. url = "https://www.dropbox.com/s/7r6lnilgt78ljia/fcn_8s.tar.gz"

It returned with an error... ReadError: not a gzip file

Edit It was my bonehead mistake. Dropbox did not allow to directly download and uncompress fcn_8s checkpoint files from the python codes. I manually downloaded the tarball and opened it inside the checkpoint directory.

import tarfile tarfile.open('fcn_8s.tar.gz', 'r:gz').extractall('.')