def get_single_image(self, train_id, tag, step):
"""
Returns the actual image bytes for a given image.
Args:
train_id (str): The ID of the events data the image belongs to.
tag (str): The name of the tag the images belongs to.
step (int): The step of the image in the current reservoir. If step = -1, return image of final step.
Returns:
bytes, a byte string of the raw image bytes.
"""
Validation.check_param_empty(train_id=train_id, tag=tag, step=step)
step = to_int(step, "step")
try:
tensors = self._data_manager.list_tensors(train_id, tag)
except ParamValueError as ex:
raise ImageNotExistError(ex.message)
image = _find_image(tensors, step)
if image is None:
raise ImageNotExistError(
"Can not find the step with given train job id and tag.")
return image
def get_single_image(self, train_id, tag, step):
"""
Returns the actual image bytes for a given image.
Args:
train_id (str): The ID of the events data the image belongs to.
tag (str): The name of the tag the images belongs to.
step (int): The step of the image in the current reservoir.
Returns:
bytes, a byte string of the raw image bytes.
"""
Validation.check_param_empty(train_id=train_id, tag=tag, step=step)
step = to_int(step, "step")
try:
tensors = self._data_manager.list_tensors(train_id, tag)
except ParamValueError as ex:
raise ImageNotExistError(ex.message)
image = None
for tensor in tensors:
if tensor.step == step:
# Default value for bytes field is empty byte string normally,
# see also "Optional Fields And Default Values" in protobuf
# documentation.
image = tensor.value.encoded_image
break
if image is None:
raise ImageNotExistError("Can not find the step with given train job id and tag.")
return image
def _get_hoc_image(self, image_path, train_id):
"""Get hoc image for image data demonstration in UI."""
sample_id, label, layer = image_path.strip(".jpg").split("_")
layer = int(layer)
job = self.job_manager.get_job(train_id)
samples = job.samples
label_idx = job.labels.index(label)
chosen_sample = samples[int(sample_id)]
original_path_image = chosen_sample['image']
abs_image_path = os.path.join(self.job_manager.summary_base_dir,
_clean_train_id_b4_join(train_id),
original_path_image)
if self._is_forbidden(abs_image_path):
raise FileSystemPermissionError("Forbidden.")
image_type = ImageQueryTypes.OUTCOME.value
try:
image = Image.open(abs_image_path)
except FileNotFoundError:
raise ImageNotExistError(
f"train_id:{train_id} path:{image_path} type:{image_type}")
except PermissionError:
raise FileSystemPermissionError(
f"train_id:{train_id} path:{image_path} type:{image_type}")
except OSError:
raise UnknownError(
f"Invalid image file: train_id:{train_id} path:{image_path} type:{image_type}"
)
edit_steps = []
boxes = chosen_sample["hierarchical_occlusion"][label_idx][
"hoc_layers"][layer]["boxes"]
mask = chosen_sample["hierarchical_occlusion"][label_idx]["mask"]
for box in boxes:
edit_steps.append(EditStep(layer, *box))
image_cp = pil_apply_edit_steps(image, mask, edit_steps)
buffer = io.BytesIO()
image_cp.save(buffer, format=_PNG_FORMAT)
return buffer.getvalue()
def get_metadata_list(self, train_id, tag):
"""
Builds a JSON-serializable object with information about images.
Args:
train_id (str): The ID of the events data.
tag (str): The name of the tag the images all belong to.
Returns:
list[dict], a list of dictionaries containing the `wall_time`, `step`, `width`,
and `height` for each image.
[
{
"wall_time": ****,
"step": ****,
"width": ****,
"height": ****,
},
{...}
]
"""
Validation.check_param_empty(train_id=train_id, tag=tag)
result = []
try:
tensors = self._data_manager.list_tensors(train_id, tag)
except ParamValueError as ex:
raise ImageNotExistError(ex.message)
for tensor in tensors:
# no tensor_proto in TensorEvent
(width, height) = (tensor.value.width, tensor.value.height)
result.append({
'wall_time': tensor.wall_time,
'step': tensor.step,
'width': int(width),
'height': int(height),
})
return dict(metadatas=result)
def query_image_binary(self, train_id, image_path, image_type):
"""
Query image binary content.
Args:
train_id (str): Job ID.
image_path (str): Image path relative to explain job's summary directory.
image_type (str): Image type, 'original' or 'overlay'.
Returns:
bytes, image binary.
"""
abs_image_path = os.path.join(self.job_manager.summary_base_dir,
_clean_train_id_b4_join(train_id),
image_path)
if self._is_forbidden(abs_image_path):
raise FileSystemPermissionError("Forbidden.")
try:
if image_type != "overlay":
# no need to convert
with open(abs_image_path, "rb") as fp:
return fp.read()
image = Image.open(abs_image_path)
if image.mode == _RGBA_MODE:
# It is RGBA already, do not convert.
with open(abs_image_path, "rb") as fp:
return fp.read()
except FileNotFoundError:
raise ImageNotExistError(
f"train_id:{train_id} path:{image_path} type:{image_type}")
except PermissionError:
raise FileSystemPermissionError(
f"train_id:{train_id} path:{image_path} type:{image_type}")
except OSError:
raise UnknownError(
f"Invalid image file: train_id:{train_id} path:{image_path} type:{image_type}"
)
if image.mode == _SINGLE_CHANNEL_MODE:
saliency = np.asarray(image) / _UINT8_MAX
elif image.mode == _RGB_MODE:
saliency = np.asarray(image)
saliency = saliency[:, :, 0] / _UINT8_MAX
else:
raise UnknownError(f"Invalid overlay image mode:{image.mode}.")
saliency_stack = np.empty((saliency.shape[0], saliency.shape[1], 4))
for c in range(3):
saliency_stack[:, :, c] = saliency
rgba = saliency_stack * _SALIENCY_CMAP_HI
rgba += (1 - saliency_stack) * _SALIENCY_CMAP_LOW
rgba[:, :, 3] = saliency * _UINT8_MAX
overlay = Image.fromarray(np.uint8(rgba), mode=_RGBA_MODE)
buffer = io.BytesIO()
overlay.save(buffer, format=_PNG_FORMAT)
return buffer.getvalue()