pudae
commented
6 years ago Hi ilkarman~
Seems to complain about finding biases:
I turned off bias with _slim.argscope. Try like this, plz.
with slim.arg_scope(densenet.densenet_arg_scope):
logits, endpoints = densenet.densenet121(
X,
num_classes=1000,
[is_training=True,
reuse=None)Also was curious if there was a reason that shape is channels-last, since I thought channels-first is faster for cuDNN training?
You're right. But I just used slim's default data format, NHWC. If you want to use NCHW, you can do it simply adding data_format argument to slim.arg_scope.
Thanks~
ilkarman
Hi Pudae, thanks a lot for this super-neat implementation. I am trying to load the model (without the pre-processing functions, etc) so that I can chop off the last fc layer and stick my own (to re-train on my data-set). However, I was having a bit of an issue loading the model without any of the functions attached:
Seems to complain about finding biases:
However, the end-points seem ok:
('densenet121/dense_block4', <tf.Tensor 'densenet121/dense_block4/conv_block16/concat:0' shape=(?, 7, 7, 1024) dtype=float32>), ('densenet121/logits', <tf.Tensor 'densenet121/logits/Relu:0' shape=(?, 1, 1, 1000) dtype=float32>), ('predictions', <tf.Tensor 'densenet121/predictions/Reshape_1:0' shape=(?, 1, 1, 1000) dtype=float32>)])So that I can just extract: 'densenet121/dense_block4/conv_block16/concat:0' and add my own: ('densenet121/logits', <tf.Tensor 'densenet121/logits/Relu:0' shape=(?, 1, 1, 16) dtype=float32>), ('predictions', <tf.Tensor 'densenet121/predictions/Reshape_1:0' shape=(?, 1, 1, 16) dtype=float32>)])
Also was curious if there was a reason that shape is channels-last, since I thought channels-first is faster for cuDNN training?
Thanks