How to read PNG encoded images from LMDB database with python ?

[Erogol]

I conerted all ImageNet into PNG encoded LMDB format. However, I suspect to have something wrong for the images and I like to check whether they are in correct format and they have correct labelling. For this purpose, how can I read images from LMDB by their labels. I followed the below code but datum returns 0 width height and channel. I guess this is because the images are encoded.

import caffe
import os
import lmdb
import numpy
import matplotlib.pyplot as plt
%matplotlib inline


LMDB_PATH = "/imagenet/ilsvrc12_train_lmdb/"


env = lmdb.open(LMDB_PATH, readonly=True, lock=False)


visualize = True


datum = caffe.proto.caffe_pb2.Datum()




with env.begin() as txn:
    cursor = txn.cursor()
    for key, value in cursor:
        # print(key, value)
        # convert to datum
        datum.ParseFromString(value)
        # Read the datum.data
        image = caffe.io.datum_to_array(datum)
        #img_data = img_data.reshape(datum.channels, datum.height, datum.width)
        if visualize:
            plt.imshow(img_data.transpose([1,2,0]))
            plt.show(block=False)


        #print key

[Mohamed Omran]

as

To get key and value, you want to call cursor.key() and cursor.value() respectively.

Here's some code for inspecting data inside lmdbs:

import os

import sys

import numpy as np

import lmdb

import matplotlib.pyplot as plt

caffe_root = #<CAFFE_ROOT>

sys.path.append(os.path.join(caffe_root, 'python'))

import caffe

def read_images_from_lmdb(db_path) :

    

    """

    Loops over image data in the lmdb, and displays information about each datum

    Assumes that data dimensions are as follows: (channels, height, width)

    """

    ax = plt.subplot(111)

    plt.hold(False)

    lmdb_env = lmdb.open(db_path, readonly=True)    

    with lmdb_env.begin() as lmdb_txn :

        lmdb_cursor = lmdb_txn.cursor() 

        #for it in lmdb_cursor.iternext() :

        while lmdb_cursor.next() :

            value = lmdb_cursor.value()

            key = lmdb_cursor.key()

            

            datum = caffe.proto.caffe_pb2.Datum()

            datum.ParseFromString(value)

            image = np.zeros((datum.channels, datum.height, datum.width))

            image = caffe.io.datum_to_array(datum)   

            image = np.transpose(image, (1, 2, 0))    # -> height, width, channels

            image = image[:,:,::-1]                   # BGR -> RGB

              

            print("key: ", key) 

            print("image shape: " + str(image.shape) + ", data type: " + str(image.dtype) + ", random pixel value: " +  str(image[150,150,0]))

                        

            ax.imshow(np.squeeze(image))

            plt.draw()

            plt.waitforbuttonpress()

            

    plt.show() 

    lmdb_txn.abort()

    lmdb_env.close()

    return

    

def main():

    """

    Set the db_path, and you're good to go.

    """

    

    db_path = #<DB_PATH>

    if not os.path.exists(db_path) :

        raise Exception('db not found')

    read_images_from_lmdb(db_path)

        

    return

 

   

if __name__ == "__main__" :

    main()

[Erogol]

I tried your code but the problem is same. Still cannot get any values for channel, width and height. As I pointed, I believe problem is reading the string representation of encoded PNG.