def __tensor__(self, shape, dtype):
channels = None
h = None
w = None
if shape is not None:
shape_len = len(shape)
# NOTE(adamb) Technically we can load images, convert them to
# single channel (if they aren't already) and reshape them to
# [h, w]. Revisit the code below if we ever want this.
if shape_len < 3:
return None
# NOTE(adamb) Technically we can handle .gif files, which decode to
# a shape of length 4. For this to work properly, we'll need to
# know which one we're dealing with and return None if the
# format doesn't match the requested shape.
if shape_len > 3:
return TensorShape(shape[-3:])
channels = shape[-1]
h = shape[-3]
w = shape[-2]
contents = as_tensor(self.bytes, None, None)
image = tf.image.decode_image(contents, channels=channels)
if dtype is not None and dtype != image.dtype:
image = tf.image.convert_image_dtype(image, dtype=dtype)
if w is not None or h is not None:
image = tf.image.resize_image_with_crop_or_pad(image,
target_height=h,
target_width=w)
return image
def terminal_write(o):
tensor = as_tensor(o, None, None)
accept_mimetypes = {"image/png": imgcat, "text/plain": lambda x: x}
encoded = _encode(tensor, accept_mimetypes)
if encoded is None:
encoded = tensor
print(pprint.pformat(encoded), file=sys.stdout, flush=True)
def session_handle_for(value, signature_def_input):
if isinstance(value, float):
return None
# TODO(adamb) This might append to the default graph. These appends
# should be cached and idempotent. The added nodes should be
# reused if possible.
return as_tensor(
value,
tf.TensorShape(signature_def_input.tensor_shape).as_list(),
tf.as_dtype(signature_def_input.dtype))
def __tensor__(self, shape, dtype):
channels = None
if shape is not None:
shape_len = len(shape)
if shape_len < 3 or shape_len > 4:
raise Exception("Unsupported shape: %s" % shape)
channels = shape[-1]
contents = as_tensor(self.bytes, None, None)
image = tf.image.decode_image(contents, channels=channels)
if dtype is not None and dtype != image.dtype:
image = tf.image.convert_image_dtype(image, dtype=dtype)
h = shape[-3]
w = shape[-2]
if w is not None or h is not None:
image = tf.image.resize_image_with_crop_or_pad(image,
target_height=h,
target_width=w)
return image
def session_handle_for(value, signature_def_input):
return as_tensor(
value,
tf.TensorShape(signature_def_input.tensor_shape).as_list(),
tf.as_dtype(signature_def_input.dtype))
def do_test(self, expect, shape, dtype, data):
result = as_tensor(data, shape, dtype)
self.assertEqual(expect, result)