Caffe image formats

[Patrick Grossmann]

Dear all,

I am trying to learn a deep convolution network with Caffe, and for some reasons my image input data are required to be .tiff files. I tried the IMAGE_DATA layer which failed, so I converted all the images to lmdb using the convert_imageset tool. My training runs without errors on these lmdb folders, but I am unsure whether the convert_imageset file actually understands .tiff files. Could I get an confirmation that Caffe understands my lmdb files correctly, please?

Thank you very much!

[Luke Yeager]

You should be able to tell whether it's working or not based on the accuracy that you get during the test phase (look for lines like "Test net output #0: accuracy = 0.0362").

It looks like caffe can handle any image format that OpenCV can handle, since that is the library they use to read images.

[Yoann]

As mentioned by Luke you can check the accuracy or check directly in the lmdb file if the data has been written correctly.

Here is a python script that can extract an image (actually the image associated with the key 1 in the lmdb file in my example) and save it as png to be able to visualize one image stored in the lmdb file.

PS: the last lines to reshape the data may be different depending on how you have converted your images in the datum format

import lmdb

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

db_path = '../data/liris-accede/train_data_lmdb/'

lmdb_env = lmdb.open(db_path)  # equivalent to mdb_env_open()

lmdb_txn = lmdb_env.begin()  # equivalent to mdb_txn_begin()

lmdb_cursor = lmdb_txn.cursor()  # equivalent to mdb_cursor_open()

lmdb_cursor.get('{:0>10d}'.format(1)) #  get the data associated with the 'key' 1, change the value to get other images

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))

image = image[:, :, (2, 1, 0)]

image = image.astype(np.uint8)

mpimg.imsave('out.png', image)

[Patrick Grossmann]

Dear Luke and Yoann, thank you very much for pointing out how to confirm everything was read correctly into Caffe! Thank you, Yoann, also for the script, which worked almost out of the box for me!

[Matthew Roos]

I tried to do this, using Yoann's code, but either I'm not actually opening the lmdb or something similar is occuring.  My only change to Yoann's code is to define the path of the mnist lmdb:
db_path = '/home/mroos/caffe-master/examples/mnist/mnist_train_lmdb/'

But after trying to open and read from the lmdb, the returned value and key are empty strings:

>>> lmdb_cursor = lmdb_txn.cursor()  # equivalent to mdb_cursor_open()
>>> lmdb_cursor.get('{:0>10d}'.format(1)) #  get the data associated with the 'key' 1, change the value to get other images
>>> value = lmdb_cursor.value()
>>> key = lmdb_cursor.key()

>>> value
''
>>> key
''

I'm not getting any errors. I'm a novice to python and know almost nothing about DBs, so any suggestions would be appreciated!

[Yoann]

Maybe the key "{:0>10d}'.format(1)" doesn't exist in MNIST lmdb.

You can try to pick the first key from the lmdb file using something like "lmdb_cursor.first()", at least to understand how the keys have been created for your lmdb file (you can iterate using next()).

Please look at http://lmdb.readthedocs.org/en/release/ for the doc

[niey...@gmail.com]

can i extract the image from lmdb file as a string of bytes directly,rather than the image?~

在 2015年2月10日星期二 UTC+8下午11:00:18，Yoann写道：

[Jeremy Pinto]

Hello Yoann, is there a reason why you transpose the images twice in a row (context) : 

image = np.transpose(image, (1, 2, 0))

image = image[:, :, (2, 1, 0)]

seems to me like that just undoes the initial transpose?