mean.binaryproto visualization for uint8 data

[Yoann]

Hello all,

I've successfully saved my dataset in LMDB format as float data, created the mean binaryproto file and visualized this file to make sure that everything was ok. However, it is not possible to crop float data in train_val.prototxt and thus I decided to save my data as uint8.

To save my data I just replaced "float" by "np.uint8" in my python script (see below). Then I called the same tool ./build/tools/compute_image_mean (...). And finally I used the same python script to visualize the mean file (see below). I just added the line 'out = out.astype(np.uint8)'.

But, the mean file saved as png appears completely black. 

It seems normal since the 'arr' variable looks like:

[[[ 0.00309061  0.00339213  0.00437208 ...,  0.00248756  0.0027137

    0.00241218]

  [ 0.00361827  0.00346751  0.00505051 ...,  0.0027137   0.00309061

    0.00278908]

  [ 0.0042967   0.00414594  0.00482436 ...,  0.00407056  0.00407056

    0.00361827] etc....

Did I miss something?

Here is the script to save the binaryproto as a png file (written by another caffe user on another thread):

import numpy as np
import matplotlib.pyplot as plt
import matplotlib.image as mpimg

 

# Make sure that caffe is on the python path:
caffe_root = '../'  # this file is expected to be in {caffe_root}/examples
import sys
sys.path.insert(0, caffe_root + 'python')
import caffe

 

blob = caffe.proto.caffe_pb2.BlobProto()
data = open( sys.argv[1] , 'rb' ).read()
blob.ParseFromString(data)
arr = np.array( caffe.io.blobproto_to_array(blob) )
out = arr[0]
out = np.ascontiguousarray(out.transpose(1,2,0))
out = out.astype(np.uint8)
mpimg.imsave(sys.argv[2], out)

Here is the loow where I save data as uint8 (in_ is the path of one image):

for idx in range(int(math.ceil(len(Inputs)/1000.0))):
in_db_data = lmdb.open(lmdb_data_name, map_size=int(1e12))
with in_db_data.begin(write=True) as in_txn:
for in_idx, in_ in enumerate(Inputs[(1000*idx):(1000*(idx+1))]):
im = caffe.io.load_image(in_)
#plt.imshow(im)
#plt.show()
im_dat = caffe.io.array_to_datum(im.astype(float).transpose((2, 0, 1)))
in_txn.put('{:0>10d}'.format(1000*idx + in_idx), im_dat.SerializeToString())
string_ = str(1000*idx+in_idx+1) + ' / ' + str(len(Inputs))
sys.stdout.write("\r%s" % string_)
sys.stdout.flush()
in_db_data.close()
print('')

[Yoann]

To get a point of comparison, the 'arr' variable for the mean file generated on the float data looks like:

[[[ 0.29685655  0.29646629  0.29816759 ...,  0.29304227  0.29310983

    0.29361176]

  [ 0.3050448   0.30444458  0.3056635  ...,  0.30154204  0.30190951

    0.30188578]

  [ 0.3077313   0.30739287  0.30698517 ...,  0.30331087  0.30332285

    0.30343989]

Which makes more sense.

Thus, do you know what is the problem?

[Yoann]

My bad, never thought to read the caffe.io.load_image which loads an image as float.

Creating a new function reading an image as uint8 fixed my problem.

[Yasser Souri]

Here is a trick I have used to visualize mean images:

https://gist.github.com/yassersouri/f617bf7eff9172290b4f

[Anurag Gupta]

    import os, sys, glob, caffe

    import numpy as np

    mean_file= "path/to/file/mean.binaryproto"

    #convert mean file to image

    blob= caffe.proto.caffe_pb2.BlobProto()

    try:

        data = open( mean_file, 'rb' ).read()

    except:

        data = open( mean_file, 'r' ).read()

    blob.ParseFromString(data)

    arr = np.uint8(np.array( caffe.io.blobproto_to_array(blob) )[0])

    #a= arr[0];    b= arr[1];    c= arr[2]

    img= np.zeros([128,200,3])

    img[:,:,0]= arr[0];    img[:,:,1]= arr[1];    img[:,:,2]= arr[2]

    import cv2

    cv2.imwrite(mean_file.replace(".binaryproto", ".bmp"), img)