chenxi116 / tf-deeplab Goto Github PK

In the solver.prototxt of the original implementation, iter_size: 10 specifies that gradient updates only happen every 10 batches. However, the training script here doesnt do such a thing

In the model, you initialize factor as:

factor = tf.get_variable( 'factor', 1, tf.float32, initializer=tf.constant_initializer(0.0, tf.float32), trainable=False)

which makes everything nan on first pass

the operation already include a softmax function. tf.nn.sparse_softmax_cross_entropy_with_logits.

and before sparse_softmax_cross_entropy_with_logits you also have a softmax, which might slight influence the result.

When I ran python caffemodel2npy.py deploy.prototxt ../deeplab/ResNet101/init.caffemodel/model/ResNet101_init.npy, an error occurred as the following.

WARNING: Logging before InitGoogleLogging() is written to STDERR
W0725 18:40:56.169318 9626 _caffe.cpp:122] DEPRECATION WARNING - deprecated use of Python interface
W0725 18:40:56.169384 9626 _caffe.cpp:123] Use this instead (with the named "weights" parameter):
W0725 18:40:56.169395 9626 _caffe.cpp:125] Net('deploy.prototxt', 1, weights='../deeplab/ResNet101/init.caffemodel')
[libprotobuf ERROR google/protobuf/text_format.cc:245] Error parsing text-format caffe.NetParameter: 15:16: Message type "caffe.LayerParameter" has no field named "interp_param".
F0725 18:40:56.173895 9626 upgrade_proto.cpp:79] Check failed: ReadProtoFromTextFile(param_file, param) Failed to parse NetParameter file: deploy.prototxt
*** Check failure stack trace: ***
Abort (core dumped)

What can the problem possibly be?

The default end_learning_rate is 0.0001 (https://www.tensorflow.org/api_docs/python/tf/train/polynomial_decay). However, according to https://stackoverflow.com/questions/30033096/what-is-lr-policy-in-caffe/30045244, the default caffe behavior is end_learning_rate=0.

I am using the windows caffe python3.5 vs2015 release version. But it reports

Thanks!