I want to read the data stored in a TFRecord file that I've used as a train record in TF Object Detection API.
However, I get an InvalidArgumentError: Input to reshape is a tensor with 91090 values, but the requested shape has 921600. I don't understand what the root of the error is, even though the discrepancy seems to be a factor of 10.
Question: How can I read the file without this error?
- I can't rule out that the error is from creating the record, or if the error is in how I read it. Therefore, I've included my code for both.
- I'm able to run object_detection/train.py with the data, AND generate a frozen graph from the trained model.
- From this answer (and its mentioned GitHub issue), I figured out that I had to convert my PNG images into JPG, hence the
as_jpg-part (see my code below). - I used the code from this answer as a starting point to read the file.
- I use Tensorflow 1.7.0, Python 3.5
There is only one class: "Human". The record has 1000 images; each image can have a single bounding box, or multiple. (One for each human in the respective image.)
How I read the TFRecord: As mentioned above: I used the code from this answer as a starting point to read the file:
train_record = 'train.record'
def read_and_decode(filename_queue):
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(
serialized_example,
# Defaults are not specified since both keys are required.
features={
'image/height': tf.FixedLenFeature([], tf.int64),
'image/width': tf.FixedLenFeature([], tf.int64),
'image/source_id': tf.FixedLenFeature([], tf.string),
'image/encoded': tf.FixedLenFeature([], tf.string),
'image/format': tf.FixedLenFeature([], tf.string),
'image/object/bbox/xmin': tf.VarLenFeature(tf.float32),
'image/object/bbox/xmax': tf.VarLenFeature(tf.float32),
'image/object/bbox/ymin': tf.VarLenFeature(tf.float32),
'image/object/bbox/ymax': tf.VarLenFeature(tf.float32),
'image/object/class/text': tf.VarLenFeature(tf.string),
'image/object/class/label': tf.VarLenFeature(tf.int64)
})
image = tf.decode_raw(features['image/encoded'], tf.uint8)
# label = tf.cast(features['image/object/class/label'], tf.int32)
height = tf.cast(features['image/height'], tf.int32)
width = tf.cast(features['image/width'], tf.int32)
return image, height, width
def get_all_records(FILE):
with tf.Session() as sess:
filename_queue = tf.train.string_input_producer([ FILE ])
image, height, width = read_and_decode(filename_queue)
image = tf.reshape(image, tf.stack([height, width, 3]))
image.set_shape([640,480,3])
init_op = tf.initialize_all_variables()
sess.run(init_op)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
for i in range(1):
example, l = sess.run([image])
img = Image.fromarray(example, 'RGB')
img.save( "output/" + str(i) + '-train.png')
print (example,l)
coord.request_stop()
coord.join(threads)
get_all_records(train_record)
Creation:
I've made a class Image to logically model the image, and a class Rect to represent the bounding boxes and labels. This is not very relevant, but the code below is makes use of them when variable img or rect is seen.
A relevant part might be the get_bytes()-method, which is more a wrapper for using PIL's Image.open(file_path):
class Image:
# ... rest of class
def open_img(self):
if self.file_path is not None:
return Image.open(self.file_path)
def get_bytes(self, as_jpg=False):
if self.file_path is None:
return None
if as_jpg:
# Convert to jpg:
with BytesIO() as f:
self.open_img().convert('RGB').save(f, format='JPEG', quality=95)
return f.getvalue()
else: # Assume png
return np.array(self.open_img().convert('RGB')).tobytes()
How I created the Examples:
use_jpg = True
def create_tf_example(img):
image_format= b'jpg' if use_jpg else b'png'
encoded_image_data = img.get_bytes(as_jpg=use_jpg) # Encoded image bytes
relative_path = img.get_file_path()
if relative_path is None or not img.has_person():
return None # Ignore images without humans or image data
else:
filename = str(Path(relative_path).resolve()) # Absolute filename of the image. Empty if image is not from file
xmins = [] # List of normalized left x coordinates in bounding box (1 per box)
xmaxs = [] # List of normalized right x coordinates in bounding box (1 per box)
ymins = [] # List of normalized top y coordinates in bounding box (1 per box)
ymaxs = [] # List of normalized bottom y coordinates in bounding box (1 per box)
classes_text = [] # List of string class name of bounding box (1 per box)
classes = [] # List of integer class id of bounding box (1 per box)
for rect in img.rects:
if not rect.is_person:
continue # For now, ignore negative samples as TF does this by default
else:
xmin, xmax, ymin, ymax = rect.get_normalized_xy_min_max()
xmins.append(xmin)
xmaxs.append(xmax)
ymins.append(ymin)
ymaxs.append(ymax)
# Human class:
classes.append(1)
classes_text.append('Human'.encode())
return tf.train.Example(features=tf.train.Features(feature={
'image/height': dataset_util.int64_feature(height),
'image/width': dataset_util.int64_feature(width),
# 'image/filename': dataset_util.bytes_feature(filename.encode()),
'image/source_id': dataset_util.bytes_feature(filename.encode()),
'image/encoded': dataset_util.bytes_feature(encoded_image_data),
'image/format': dataset_util.bytes_feature(image_format),
'image/object/bbox/xmin': dataset_util.float_list_feature(xmins),
'image/object/bbox/xmax': dataset_util.float_list_feature(xmaxs),
'image/object/bbox/ymin': dataset_util.float_list_feature(ymins),
'image/object/bbox/ymax': dataset_util.float_list_feature(ymaxs),
'image/object/class/text': dataset_util.bytes_list_feature(classes_text),
'image/object/class/label': dataset_util.int64_list_feature(classes),
}))
How I created the TFRecord:
def convert_to_tfrecord(imgs, output_file_path):
with tf.python_io.TFRecordWriter(output_file_path) as writer:
for img in imgs:
tf_example = create_tf_example(img)
if tf_example is not None:
writer.write(tf_example.SerializeToString())
convert_to_tfrecord(train_imgs, 'train.record')
convert_to_tfrecord(validation_imgs, 'validate.record')
convert_to_tfrecord(test_imgs, 'test.record')
From the dataset_util module:
def int64_feature(value):
return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))
def int64_list_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]))
def bytes_list_feature(value):
return tf.train.Feature(bytes_list=tf.train.BytesList(value=value))
def float_list_feature(value):
return tf.train.Feature(float_list=tf.train.FloatList(value=value))