Convert own dataset mean to numpy succeed but error when running the prediction.

[teknosains]

Using this script, i successfully converty the binaryproto mean into numpy
import sys
from caffe.proto import caffe_pb2
from convert import blobproto_to_array
import numpy as np

blob = caffe_pb2.BlobProto()
data = open("../../examples/test/my_mean.binaryproto", "rb").read()
blob.ParseFromString(data)
nparray = blobproto_to_array(blob)
f = file("my_mean.npy","wb")
np.save(f,nparray)
f.close()

but when I run the prediction i got this error

Traceback (most recent call last):
File "tester.py", line 23, in
scores = net.predict(caffe_root + 'examples/test/'+filename)
File "/home/ubuntu/caffe/caffe-dev/python/caffe/imagenet/wrapper.py", line 85, in predict
input_blob = [prepare_image(filename, self._center_only)]
File "/home/ubuntu/caffe/caffe-dev/python/caffe/imagenet/wrapper.py", line 65, in prepare_image
img_reshape -= IMAGENET_MEAN
ValueError: operands could not be broadcast together with shapes (256,256,3) (1,3,256,256) (256,256,3)

The prediction is just fine when using ilsvrc12 mean, but its error when i run my own mean from my dataset

Thx

[teknosains]

I change wrapper.py , on prepare_image function

the_mean = IMAGENET_MEAN.reshape(256,256,3)

im not sure it is correct or not...but it works

[shelhamer]

The image mean doesn't actually vary much pixel-to-pixel and channel-wise means work almost equally well so it doesn't matter too much.

At any rate, try it with the new python interface in the latest master. It's a bit more batteries-included.

[sunbaigui]

Hi @shelhamer ,
when I run @codetrash 's solution, I've occurred the same error, @codetrash just change the code in wrapper.py to "the_mean = IMAGENET_MEAN.reshape(256,256,3)". Does it right or not? I am new to caffe, it's a little bit hard for me to understand your last comment.
Is there any new features in master to support this kind convert? Looking forward to your early reply.

[teknosains]

@sunbaigui are u using your own generated dataset from your own images ?

I had those error until i fix it by reshaping the mean IMAGENET_MEAN.reshape(256,256,3) .

but when i used the latest branch master, i did not get those errors anymore

[sunbaigui]

@codetrash which latest branch master do you really use, I am using the version download at 24 April, can you kindly give me your email or facebook? I'd like to ask you a few questions.

[teknosains]

@sunbaigui im using master committed this month. here's my facebook : budyks

[sunbaigui]

@codetrash when I run it under the latest branch master, another error occurred below:
AttributeError: 'module' object has no attribute 'BufferedIOBase'

Have you met this error before?

[teknosains]

@sunbaigui well, i never had it...

maybe you should check all installation dependencies

[shelhamer]

Don't import/run Caffe inside the wrapper directory
CAFFE_ROOT/python/caffe. The caffe.io submodule is conflicting with the
python standard library io module because caffe/io.py is in the same dir
and had precedence.

Note your PYTHONPATH should point to the parent dir of caffe (python/caffe
that is) and not the module dir itself.

Le jeudi 29 mai 2014, sunbaigui [email protected] a écrit :

@codetrash https://github.com/codetrash when I run it under the latest
branch master, another error occurred below:
AttributeError: 'module' object has no attribute 'BufferedIOBase'

Have you met this error before?

—
Reply to this email directly or view it on GitHub
#420 (comment).

[Yangqing]

If you would like to inspect the image mean, don't do reshape (this will mess up with the order). What you will need to do is

mean_channel_last = IMAGENET_MEAN[0].swapaxes(0,1).swapaxes(1,2)

so that you get a correct 3-channel mean file. You can then use e.g. pyplot to show the mean (It's a pretty boring, gray-ish image)

[sunbaigui]

@Yangqing actually, I just wanna generate corresponding npy format mean file from binaryproto.
I've generated the npy format mean file by using function below:
def convert(binaryproto_file_path, npy_file_path):
blob = caffe_pb2.BlobProto()
data = open(binaryproto_file_path, "rb").read()
blob.ParseFromString(data)
nparray = blobproto_to_array(blob)
try:
f = file(npy_file_path,"wb")
np.save(f,nparray)
f.close()
except:
print "save nparray to file %s error!" % npy_file_path

But when I run apdated wrapper.py to extract feature, there's an error occurred in code line:
img_reshape -= IMAGENET_MEAN

Error: img_reshape -= IMAGENET_MEAN
ValueError: operands could not be broadcast together with shapes (256,256,3) (1,3,256,256) (256,256,3)

So how should I modify this code line to make it right?

[drubiano]

Hi, I am working with this file as well. To fix the error "ValueError: operands could not be broadcast together with shapes" I simply changed the blobproto_to_array by removing the blob.num argument from the reshape. I made it into a new function as well so it does not mess with anything anywhere else.

[sunbaigui]

thanks to @drubiano
suddenly, I got @Yangqing 's idea, I am agree with @drubiano ,I've put the nparray tranfer code from @Yangqing in convert_binaryproto2npy.py file to make sure it does not mess with anything anywhere else.

[teknosains]

I convert the mean binaryproto to npy and test it but error raised :

    Traceback (most recent call last):
   File "classify.py", line 143, in <module>
    main(sys.argv)
  File "classify.py", line 99, in main
   input_scale=args.input_scale, channel_swap=channel_swap)
  File "/home/ubuntu/Documents/caffe-master2/python/caffe/classifier.py", line 34, in __init__
  self.set_mean(self.inputs[0], mean_file)
    File "/home/ubuntu/Documents/caffe-master2/python/caffe/pycaffe.py", line 200, in _Net_set_mean
mean = caffe.io.resize_image(normal_mean.transpose((1,2,0)),
 ValueError: axes don't match array

until i tried @drubiano suggestion to remove blob.num :

   '''return np.array(blob.data).reshape(
            blob.num, blob.channels, blob.height, blob.width)'''
   #blob.num removeed
    return np.array(blob.data).reshape(blob.channels, blob.height, blob.width)

and it works, any idea what happened ? @sunbaigui @shelhamer @Yangqing

NOTE : the error occurred when im using the latest caffe (branch master) , with caffe-dev (branch dev) it works just fine

[shelhamer]

The mean is for a single input, not an entire batch, so it should only have
dimensions K x H x W. Technically, this could be a blob of dimensions 1 x K
x H x W but as a coding convenience it is three dimensional.

On Tue, Jun 24, 2014 at 5:30 PM, Budy [email protected] wrote:

I convert the binaryproto to npy and test it but error raised :

Traceback (most recent call last):

File "classify.py", line 143, in
main(sys.argv)
File "classify.py", line 99, in main
input_scale=args.input_scale, channel_swap=channel_swap)
File "/home/ubuntu/Documents/caffe-master2/python/caffe/classifier.py", line 34, in init
self.set_mean(self.inputs[0], mean_file)
File "/home/ubuntu/Documents/caffe-master2/python/caffe/pycaffe.py", line 200, in _Net_set_mean
mean = caffe.io.resize_image(normal_mean.transpose((1,2,0)),
ValueError: axes don't match array

until i tried @drubiano https://github.com/drubiano suggestion to
remove blob.num :

'''return np.array(blob.data).reshape(
blob.num, blob.channels, blob.height, blob.width)'''
#blob.num removeed
return np.array(blob.data).reshape(blob.channels, blob.height, blob.width)

and it works, any idea what happened ? @sunbaigui
https://github.com/sunbaigui @shelhamer https://github.com/shelhamer
@Yangqing https://github.com/Yangqing

—
Reply to this email directly or view it on GitHub
#420 (comment).

[teknosains]

@shelhamer u mean the Converter code (proto mean to npy) for single input only ?

[Cheng-Wang]

hello, I have similar error, how to fix it ? I tried those methods, but doesn't work, thanks!

811 08:28:01.393484 20592 net.cpp:74] Creating Layer prob
I0811 08:28:01.393498 20592 net.cpp:84] prob <- fc8
I0811 08:28:01.393509 20592 net.cpp:110] prob -> prob
I0811 08:28:01.393525 20592 net.cpp:125] Top shape: 10 10 1 1 (100)
I0811 08:28:01.393537 20592 net.cpp:151] prob needs backward computation.
I0811 08:28:01.393548 20592 net.cpp:162] This network produces output prob
I0811 08:28:01.393561 20592 net.cpp:173] Collecting Learning Rate and Weight Decay.
I0811 08:28:01.393580 20592 net.cpp:166] Network initialization done.
I0811 08:28:01.393594 20592 net.cpp:167] Memory required for Data 41943640
Traceback (most recent call last):
File "cheng_feature_visualization.py", line 26, in
scores = net.predict(caffe_root + 'examples/cheng_test/n00453396_157.jpg')
File "../caffe1/python/caffe/imagenet/wrapper.py", line 84, in predict
input_blob = [prepare_image(filename, self._center_only)]
File "../caffe1/python/caffe/imagenet/wrapper.py", line 64, in prepare_image
img_reshape -= IMAGENET_MEAN
ValueError: operands could not be broadcast together with shapes (256,256,3) (3,256,256) (256,256,3)

[AlesZurek]

I got the same problem
I convert the mean binaryproto to npy.
And when I want to use it in caffe.Classifier() I got error:
"axes don't match array"
Is the solution to remove blob.num still actual or is there any other better solution?
I have got actual (2014-09-20) version of master branch.

[sleepsophia]

I convert the mean binaryproto to npy.
And when I want to use it in caffe.Classifier() I got error:
"mean shape incompatible with input shape"
and when i remove blob.num ,i got the same error,is there any solution?

[laotao]

@sunbaigui Where did you put convert_binaryproto2npy.py? Can you share it?

[adnan15110]

you can use the following code to convert between .binaryproto to .npy:

print "Usage: python convert_protomean.py proto.mean out.npy"
blob = caffe.proto.caffe_pb2.BlobProto()
BINARY_PROTO_FILE_NAME  ='image_mean.binaryproto'
BINARY_PROTO_FILE_PATH  = os.path.join(os.getcwd(),BINARY_PROTO_FILE_NAME)
NPY_FILE_NAME  ='image_mean.npy'
NPY_FILE_PATH  = os.path.join(os.getcwd(),NPY_FILE_NAME)

data = open( BINARY_PROTO_FILE_PATH, 'rb' ).read()
blob.ParseFromString(data)
arr = np.array( caffe.io.blobproto_to_array(blob) )
out = arr[0]
np.save( NPY_FILE_PATH , out )

[SongEsther]

@sleepsophia Hi, I get the same problem with you ,have you already solved it ?

[SongEsther]

@Yangqing Could you help me? I gengrate my own mean.npy and follow http://nbviewer.jupyter.org/github/BVLC/caffe/blob/master/examples/00-classification.ipynb . But when I run the below:
load the mean ImageNet image (as distributed with Caffe) for subtraction
mu = np.load(caffe_root + 'python/caffe/imagenet/mean.npy')
mu = mu.mean(1).mean(1) # average over pixels to obtain the mean (BGR) pixel values
print 'mean-subtracted values:', zip('BGR', mu)
transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
transformer.set_transpose('data', (2,0,1)) # move image channels to outermost dimension
transformer.set_mean('data', mu) # subtract the dataset-mean value in each channel
transformer.set_raw_scale('data', 255) # rescale from [0, 1] to [0, 255]
transformer.set_channel_swap('data', (2,1,0)) # swap channels from RGB to BGR

my problem:
mean-subtracted values: [('B', array([ 211.32064002, 220.19243949, 226.30641338, 228.79393605,
229.45166288, 228.21699033, 225.68690327, 222.26552255,
218.41735295, 214.56371852, 210.7109702 , 207.73977334,
205.35947091, 203.82138116, 203.28422383, 203.61476408,
205.16159385, 207.72180993, 211.22096089, 215.12253734,
219.32449286, 223.22309222, 226.59998049, 228.92342867,
230.11787251, 229.7456883 , 224.77160536, 218.04600089]))]
ValueError Traceback (most recent call last)
in ()
9
10 transformer.set_transpose('data', (2,0,1)) # move image channels to outermost dimension
---> 11 transformer.set_mean('data', mu) # subtract the dataset-mean value in each channel
12 transformer.set_raw_scale('data', 255) # rescale from [0, 1] to [0, 255]
13 transformer.set_channel_swap('data', (2,1,0)) # swap channels from RGB to BGR
/home/siat/caffe-master/python/caffe/io.py in set_mean(self, in_, mean)
257 raise ValueError('Mean shape invalid')
258 if ms != self.inputs[in_][1:]:
--> 259 raise ValueError('Mean shape incompatible with input shape.')
260 self.mean[in_] = mean
261
ValueError: Mean shape incompatible with input shape.