'LayerParameter' object has no attribute 'scale'
zht3344 opened this issue · 15 comments
I was having some problems when running the following line ,
Write the train prototxt in a file
f = open(MODEL_FOLDER + 'train_segnet.prototxt', 'w')
f.write(str(net_arch.to_proto()))
I got the error:
'LayerParameter' object has no attribute 'scale'
I don't know what cause the error,how can i resolve this problem?
the following error message is showed:
File "training.py", line 213, in
f.write(str(net_arch.to_proto()))
File "/home/tukrin1/caffe/python/caffe/net_spec.py", line 189, in to_proto
top._to_proto(layers, names, autonames)
File "/home/tukrin1/caffe/python/caffe/net_spec.py", line 97, in _to_proto
return self.fn._to_proto(layers, names, autonames)
File "/home/tukrin1/caffe/python/caffe/net_spec.py", line 158, in _to_proto
assign_proto(layer, k, v)
File "/home/tukrin1/caffe/python/caffe/net_spec.py", line 64, in assign_proto
is_repeated_field = hasattr(getattr(proto, name), 'extend')
AttributeError: 'LayerParameter' object has no attribute 'scale'
Could anyone help me with this?
I appreciate your help!
Are you using our Caffe fork ?
You should have done :
# Initialize and checkout our custom Caffe fork (upsample branch !)
git submodule init
git submodule updateThen, you can use Caffe in Python either by adding our fork to the PYTHON_PATH or manually in your script :
import sys
sys.path.insert(0, '/path/to/DeepNetsForEO/caffe/python')@nshaud Thank you, I change your Caffe for,now it work. but when i trained the caffemodel, i got the wrong result:
I0314 17:45:18.172921 6023 solver.cpp:238] Train net output #0: accuracy = 0.404126
I0314 17:45:18.172945 6023 solver.cpp:238] Train net output #1: loss = 1.32147 (* 1 = 1.32147 loss)
I0314 17:45:18.172952 6023 solver.cpp:238] Train net output #2: per_class_accuracy = 0.712315
I0314 17:45:18.172974 6023 solver.cpp:238] Train net output #3: per_class_accuracy = 0.291166
I0314 17:45:18.172979 6023 solver.cpp:238] Train net output #4: per_class_accuracy = 0.00170741
I0314 17:45:18.172986 6023 solver.cpp:238] Train net output #5: per_class_accuracy = 4.86287e-05
I0314 17:45:18.172992 6023 solver.cpp:238] Train net output #6: per_class_accuracy = 0
I0314 17:45:18.173008 6023 solver.cpp:238] Train net output #7: per_class_accuracy = 0
I0314 17:45:18.173024 6023 sgd_solver.cpp:105] Iteration 800, lr = 0.01
I also don't understand why set BGR=True, i use the Vaihingen irrg data , is it reverse the channel order to g r ir? can you give me some advices.
Yes, BGR=True reverses the order of the channels. E.g, if you use IRRG data, the LMDB will actually be GRIR. Our pre-trained weights use this format.
If you are training your own model, it should take around 40000 iterations (with a batch size of 10) before convergence.
@nshaud I increased the training times to 40000,but it still wrong. I found that at the beginning of the train net output the first four class accuracy are not zero,however,at the end of the training only the top of 2 are not zero :
I0315 14:30:36.392210 2502 solver.cpp:238] Train net output #0: accuracy = 0.16449
I0315 14:30:36.392230 2502 solver.cpp:238] Train net output #1: loss = 2.09245 (* 1 = 2.09245 loss)
I0315 14:30:36.392246 2502 solver.cpp:238] Train net output #2: per_class_accuracy = 0.270436
I0315 14:30:36.392251 2502 solver.cpp:238] Train net output #3: per_class_accuracy = 0.121069
I0315 14:30:36.392258 2502 solver.cpp:238] Train net output #4: per_class_accuracy = 0.0145857
I0315 14:30:36.392266 2502 solver.cpp:238] Train net output #5: per_class_accuracy = 0.175652
I0315 14:30:36.392271 2502 solver.cpp:238] Train net output #6: per_class_accuracy = 0.239834
I0315 14:30:36.392287 2502 solver.cpp:238] Train net output #7: per_class_accuracy = 0
I0315 14:30:36.392294 2502 sgd_solver.cpp:105] Iteration 0, lr = 0.01
/usr/lib/python2.7/dist-packages/numpy/core/_methods.py:59: RuntimeWarning: .
warnings.warn("Mean of empty slice.", RuntimeWarning)
/usr/lib/python2.7/dist-packages/numpy/core/_methods.py:70: RuntimeWarning: invalid value encountered in double_scalars
ret = ret.dtype.type(ret / rcount)
iter 0, train_loss=2.092453
0%| | 100/40000 [00:20<2:11:38, 5.05it/s]I0315 14:30:56.584589 2502 solver.cpp:219] Iteration 100 (4.95231 iter/s, 20.1926s/100 iters), loss = 1.52628
I0315 14:30:56.584635 2502 solver.cpp:238] Train net output #0: accuracy = 0.303955
I0315 14:30:56.584646 2502 solver.cpp:238] Train net output #1: loss = 1.52628 (* 1 = 1.52628 loss)
I0315 14:30:56.584655 2502 solver.cpp:238] Train net output #2: per_class_accuracy = 0.772755
I0315 14:30:56.584677 2502 solver.cpp:238] Train net output #3: per_class_accuracy = 0.23393
I0315 14:30:56.584699 2502 solver.cpp:238] Train net output #4: per_class_accuracy = 0.00012333
I0315 14:30:56.584704 2502 solver.cpp:238] Train net output #5: per_class_accuracy = 0.00130501
I0315 14:30:56.584710 2502 solver.cpp:238] Train net output #6: per_class_accuracy = 0
I0315 14:30:56.584717 2502 solver.cpp:238] Train net output #7: per_class_accuracy = 0
I0315 14:30:56.584724 2502 sgd_solver.cpp:105] Iteration 100, lr = 0.01
iter 100, train_loss=1.526283
0%|▏ | 200/40000 [00:40<2:14:46, 4.92it/s]I0315 14:31:16.698926 2502 solver.cpp:219] Iteration 200 (4.97154 iter/s, 20.1145s/100 iters), loss = 1.3434
100%|██████████████████ | 39900/40000 [3:56:21<00:19, 5.20it/s]I0315 18:26:57.442383 2502 solver.cpp:219] Iteration 39900 (1.57393 iter/s, 63.5351s/100 iters), loss = 1.32824
I0315 18:26:57.442427 2502 solver.cpp:238] Train net output #0: accuracy = 0.353693
I0315 18:26:57.442440 2502 solver.cpp:238] Train net output #1: loss = 1.32824 (* 1 = 1.32824 loss)
I0315 18:26:57.442446 2502 solver.cpp:238] Train net output #2: per_class_accuracy = 0.00656828
I0315 18:26:57.442471 2502 solver.cpp:238] Train net output #3: per_class_accuracy = 0.999583
I0315 18:26:57.442476 2502 solver.cpp:238] Train net output #4: per_class_accuracy = 0
I0315 18:26:57.442482 2502 solver.cpp:238] Train net output #5: per_class_accuracy = 0
I0315 18:26:57.442488 2502 solver.cpp:238] Train net output #6: per_class_accuracy = 0
I0315 18:26:57.442503 2502 solver.cpp:238] Train net output #7: per_class_accuracy = 0
I0315 18:26:57.442510 2502 sgd_solver.cpp:105] Iteration 39900, lr = 0.001
Besides,there are several warning messages:
I0315 14:30:35.096953 2502 layer_factory.cpp:91] cuDNN does not support multiple tops. Using Caffe's own pooling layer.
upsample_layer.cpp:31] Params 'pad_out_{}_' are deprecated. Please declare upsample height and width useing the upsample_h, upsample_w parameters.
need help thanks:)
Those warnings are not critical, you can safely ignore them.
On what data are you training the network ? The Vaihingen dataset ?
yes, i use The Vaihingen dataset
It seems that you are only learning the 'building' class. Can you check the classification achieves good results using the provided pre-trained weights ?
@nshaud
yes,i run the reference.py with the segnet_vaihingen_128x128_fold1_iter_60000.caffemodel ,and have got a good result
Did you change anything in the training script file ? Or in the training prototxt ?
@nshaud i just change the path of dataset in the prototxt:
data_param {
source: "path/to/your/train/lmdb" # E
Could you paste the full prototxt that you used ? How did you generate the LMDB ? Can you check (using Python) that the Caffe Blobs inside the network are OK ? i.e. by using the Python Caffe API checking that net.blobs['data'].data and net.blobs['label'].data have not been corrupted somehow ?
@nshaud thanks for the solution,I re-download the code and regenerate the database,now the traning results are correct. I have modified the code before , it maybe have made something wrong
Hello
I have also same problem ('LayerParameter' object has no attribute 'scale' ). but I could not solve it. would you please guide me step by step. I used caffe version 1.0.0 on windows7.