Unable to create a tensorflow tfrecord for an image label

Question

Hi I want to create a tfrecord for images and their one hot array labels.Im able to acheive it for the images,but not for the labels.I referred to this SOF link,but getting the same error.Below is my code.

def _int64_feature(value):
return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))


def _bytes_feature(value):
return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))


for i in range(len(train_addrs)):

  print('reading image no {0} : and image address {1}'.format(i,train_addrs[i]))

  img = load_image(train_addrs[i])#loading the preprocessed image

  label = train_labels[i]#loading associated one-hot array

  print('label is ',label) #array([0, 1]) of type uint8 ,I tried with int64,int32 also;but no use

  feature = {'train/label':_int64_feature(label),
             'train/image':_bytes_feature(tf.compat.as_bytes(img.tostring())) #this part works 
           }    


  example = tf.train.Example(features=tf.train.Features(feature=feature))

  serToString = example.SerializeToString()

  writer.write(serToString)

When I execute this code,Im getting the following error.

 TypeError: array([0, 1]) has type <type 'numpy.ndarray'>, but expected one of: (<type 'int'>, <type 'long'>)

Im not sure where am I going wrong?Any help would be really helpful.

Ishant Mrinal · Answer 1 · 2017-08-31T12:31:36.050

1

Since you defined label as _int64_feature, you have to use a int for label not a numpy array

  label = train_labels[i]#loading associated one-hot array
  label = np.argmax(label)

you can convert them to one_hot format while reading the data.

If you want to pass it as a list; modify your function definition

def _int64_feature(value):
    return tf.train.Feature(int64_list=tf.train.Int64List(value=value))

edited Aug 31 '17 at 12:31

answered Aug 31 '17 at 11:59

Ishant Mrinal

4,898
3
29
47

Thanks for the reply Ishant .But in the above link,Steven has passed a list,it seems.Is that possible? – george Aug 31 '17 at 12:13
yes ! you need to modify your function for that, see my updated answer. – Ishant Mrinal Aug 31 '17 at 12:32
Thank u very much ishant..could you suggest some tutorials on feeding the data to tensorflow pipeline for training?; – george Aug 31 '17 at 16:31
you can refer https://indico.io/blog/tensorflow-data-inputs-part1-placeholders-protobufs-queues/ – Ishant Mrinal Aug 31 '17 at 16:54

Unable to create a tensorflow tfrecord for an image label

1 Answers1