def __init__(self, filename, internal=False):
with tf.name_scope("EXRIOTensor"):
data = tf.io.read_file(filename)
shapes, dtypes, channels = core_ops.io_decode_exr_info(data)
parts = []
index = 0
for (shape, dtypes, channels) in zip(shapes.numpy(),
dtypes.numpy(),
channels.numpy()):
# Remove trailing 0 from dtypes
while dtypes[-1] == 0:
dtypes.pop()
channels.pop()
spec = tuple([
tf.TensorSpec(tf.TensorShape(shape), dtype)
for dtype in dtypes
])
columns = [channel.decode() for channel in channels]
elements = [
io_tensor_ops.TensorIOTensor(core_ops.io_decode_exr(
data, index, channel, dtype=dtype),
internal=internal)
for (channel, dtype) in zip(columns, dtypes)
]
parts.append(
EXRPartIOTensor(spec, columns, elements,
internal=internal))
index += 1
spec = tuple([part.spec for part in parts])
columns = [i for i, _ in enumerate(parts)]
super(EXRIOTensor, self).__init__(spec,
columns,
parts,
internal=internal)
def __init__(self,
filename,
internal=False):
with tf.name_scope("TIFFIOTensor"):
# TIFF can fit into memory so load TIFF first
data = tf.io.read_file(filename)
info = core_ops.io_decode_tiff_info(data)
spec = tuple([
tf.TensorSpec(tf.TensorShape(
e.numpy().tolist() + [4]), tf.uint8) for e in info])
columns = [i for i, _ in enumerate(spec)]
elements = [
io_tensor_ops.TensorIOTensor(
core_ops.io_decode_tiff(data, i),
internal=internal) for i in columns]
super(TIFFIOTensor, self).__init__(
spec, columns, elements,
internal=internal)
def from_tensor(cls, tensor, **kwargs):
"""Converts a `tf.Tensor` into a `IOTensor`.
Examples:
```python
```
Args:
tensor: The `Tensor` to convert.
Returns:
A `IOTensor`.
Raises:
ValueError: If tensor is not a `Tensor`.
"""
with tf.name_scope(kwargs.get("name", "IOFromTensor")):
return io_tensor_ops.TensorIOTensor(tensor, internal=True)
def __init__(self, filename, internal=False):
with tf.name_scope("TIFFIOTensor"):
# TIFF can fit into memory so load TIFF first
data = tf.io.read_file(filename)
shapes, dtypes = core_ops.io_decode_tiff_info(data)
# NOTE: While shapes returned correctly handles 3 or 4 channels
# we can only handle RGBA so fix shape as 4 for now,
# until decode_tiff is updated.
spec = tuple([
tf.TensorSpec(tf.TensorShape(shape.tolist()[0:2] + [4]), dtype)
for (shape, dtype) in zip(shapes.numpy(), dtypes.numpy())
])
columns = [i for i, _ in enumerate(spec)]
elements = [
io_tensor_ops.TensorIOTensor(core_ops.io_decode_tiff(data, i),
internal=internal)
for i in columns
]
super().__init__(spec, columns, elements, internal=internal)