def _save_image(self, dpoint: str, img_filepath: str) -> None:
"""Save image to filesystem."""
if not os.path.exists(img_filepath):
# Downloading image from url as numpy array
image_np = url_to_image(dpoint)
cv2.imwrite( # pylint: disable=no-member
img_filepath, np.flip(image_np, axis=2))
def show_data(self, start: int = 0, end: int = 20) -> None:
"""Show the video data."""
ims = []
fig, ax = plt.subplots()
num_frames = len(self.items_list)
for i in tqdm(range(start, min(end, num_frames))):
frame = url_to_image(url=self.items_list[i])
height, width, _ = frame.shape
im = plt.imshow(frame, animated=True)
ims_ = self.labels.show(
ax=ax,
width=width,
height=height,
frameindex=i,
num_frames=len(self.items_list),
)
ims_ = [im] + ims_ # the frame data im needs to be first
ims.append(ims_)
animation.ArtistAnimation(
fig,
ims,
interval=50,
blit=True,
repeat_delay=1000,
)
plt.show()
def test_url_to_image_invalid() -> None:
"""Test the utility function with invalid urls"""
url = "https://not-a-real-url.jpg"
url2 = "http://127.0.0.1:8080/not-real.jpg"
# first image
try:
url_to_image(url)
assert False
except:
assert True
# second image
try:
url_to_image(url2)
assert False
except:
assert True
def test_url_to_image() -> None:
"""Test the conversion of a url to image data."""
url = "https://upload.wikimedia.org/wikipedia/en/a/a9/Example.jpg"
local = "tests/Example.jpg"
# get local image
# pylint: disable=no-member
img_local = cv2.imread(local)
img_local = np.flip(img_local, axis=2)
# get remote image
img_remote = url_to_image(url)
comparison = img_local == img_remote
assert comparison.all()
def get_data(presigned_url: str) -> np.ndarray:
"""Get the frame/image data."""
return url_to_image(presigned_url)
def get_datapoint(
self,
org_id: str,
label_set_name: str,
dp_id: str,
task_type: str,
taxonomy: dict,
) -> Union[Image, Video]:
"""Get all relevant information related to a datapoint."""
query_string = """
query ($orgId: UUID!, $dpId: UUID!, $name:String!) {
labelData(orgId: $orgId, dpId: $dpId, customGroupName: $name) {
blob
dataPoint {
items:items(presigned:true)
items_not_signed:items(presigned:false)
name
},
createdBy,
taskType,
dataType,
labels {
category,
attributes {
... on LabelAttributeInt {
name
valint: value
}
... on LabelAttributeBool {
name
valbool: value
}
... on LabelAttributeFloat {
name
valfloat: value
}
... on LabelAttributeString {
name
valstr: value
}
}
labelid,
frameclassify,
frameindex,
trackid,
keyframe,
end,
bbox2d {
xnorm,
ynorm,
wnorm,
hnorm,
},
pixel {
imagesize,
regions,
holes
}
}
}
}
"""
# EXECUTE THE QUERY
query_variables = {
"orgId": org_id,
"name": label_set_name,
"dpId": dp_id
}
query = dict(query=query_string, variables=query_variables)
result = self._execute_query(query)
# IMAGE DATA
if result["labelData"]["dataType"] == "IMAGE":
# Parse result
# print("RESULT", result)
with open("temp.json", "w+") as file:
json.dump(result, file, indent=2)
signed_image_url = result["labelData"]["dataPoint"]["items"][0]
unsigned_image_url = result["labelData"]["dataPoint"][
"items_not_signed"][0]
labels = result["labelData"]["labels"]
created_by = result["labelData"]["createdBy"]
image_data = url_to_image(signed_image_url) # Get image array
dpoint = Image(
created_by=created_by,
org_id=org_id,
label_set_name=label_set_name,
taxonomy=taxonomy,
remote_labels=labels,
dp_id=dp_id,
image_url=signed_image_url,
image_url_not_signed=unsigned_image_url,
image_data=image_data,
task_type=task_type,
)
return dpoint
# VIDEO DATA
else:
# Parse the result
items = result["labelData"]["dataPoint"]["items"]
items_not_signed = result["labelData"]["dataPoint"][
"items_not_signed"]
labels = result["labelData"]["labels"]
with open("result.json", "w+") as file:
json.dump(result, file, indent=2)
name = result["labelData"]["dataPoint"]["name"]
dpoint_vid = Video(
org_id=org_id,
label_set_name=label_set_name,
taxonomy=taxonomy,
task_type=task_type,
remote_labels=labels,
dp_id=dp_id,
video_name=name,
items_list=items,
items_list_not_signed=items_not_signed,
)
return dpoint_vid
def cache_bbox(self) -> None:
"""Cache video bounding boxes."""
# Create YOLO style meta data files
num_classes = len(list(self.labelset.taxonomy.keys()))
train_file = "train.txt"
names_file = "names.txt"
data_file = "obj.data"
backup_dir = "backup/"
# obj.data
with open(self.cache_dir + "/" + data_file, "w+") as file:
file.write("classes = %s" % num_classes)
file.write("\n")
file.write("train = %s" % ("data" + "/" + train_file))
file.write("\n")
file.write("names = %s" % ("data" + "/" + names_file))
file.write("\n")
file.write("backup = %s" % backup_dir)
# names.txt
with open(self.cache_dir + "/" + names_file, "w+") as file:
class_names = list(self.labelset.taxonomy.keys())
for name in class_names:
file.write(name + "\n")
# obj_train_data/
os.mkdir(os.path.join(self.cache_dir, "obj_train_data"))
taxonomy_mapper = {
name: idx
for idx, name in enumerate(list(self.labelset.taxonomy.keys()))
}
image_filepaths = []
for i in range(len(self.labelset.dp_ids)):
print(
colored("[INFO]:", "blue"),
"Exporting %s/%s video(s)" %
(i + 1, len(self.labelset.dp_ids)),
)
# Create tqdm loader
dp = self.labelset.__getitem__(i)
pbar = tqdm(total=len(dp.items_list))
# Interpolate the video labels, and get per frame labels
framelabels = dp.labels.interpolate_labels(
num_frames=len(dp.items_list))
# Iterate through video frames
for idx, item in enumerate(dp.items_list):
pbar.update(1)
dp_entry = {}
dp_entry["url"] = dp.items_list_not_signed[idx]
dp_entry["labels"] = []
# write image data to file
image_filepath = os.path.join(
self.cache_dir,
"obj_train_data",
str(dp.items_list_not_signed[idx].replace("/",
"_").replace(
":", "_")),
)
image_filepaths.append(
os.path.join(
"obj_train_data",
str(dp.items_list_not_signed[idx].replace(
"/", "_").replace(":", "_")),
))
frame_data = np.flip(url_to_image(item), axis=2)
cv2.imwrite(image_filepath, frame_data)
# create the label file name
filext_idx = image_filepath.rfind(".")
if not filext_idx:
filext_idx = -1
label_filepath = image_filepath[0:filext_idx] + ".txt"
# write labels to the txt file
frame_labels = framelabels[idx]
# write labels to the txt file
with open(label_filepath, "w+") as file:
for label in frame_labels:
class_idx = taxonomy_mapper[label.category]
file.write("%d %.6f %.6f %.6f %.6f \n" % (
class_idx,
label.xnorm,
label.ynorm,
label.wnorm,
label.hnorm,
))
pbar.close()
# create train.txt file
with open(os.path.join(self.cache_dir, train_file), "w+") as file:
for filename in image_filepaths:
file.write(filename + "\n")
def cache_classify(self) -> None:
"""Cache video classification."""
output = []
for i in range(len(self.labelset.dp_ids)):
print(
colored("[INFO]:", "blue"),
"Exporting %s/%s video(s)" %
(i + 1, len(self.labelset.dp_ids)),
)
dp = self.labelset.__getitem__(i)
video_name = dp.video_name
labels = dp.labels
# Loop through the key frame labels, and export by interpolating
# the non key frames.
start_idx = labels.labels[0].frameindex
end_idx = len(dp.items_list)
curr_idx = start_idx # the current real frame index
frame_label = labels.labels[curr_idx].category[0][-1]
key_index = 0 # the current key frame index
pbar = tqdm(total=end_idx)
while curr_idx != end_idx:
# check if frame is key frame, if yes get the new labels, else keep old
if (key_index < len(labels.labels)
and labels.labels[key_index].frameindex == curr_idx):
frame_label = labels.labels[key_index].category[0][-1]
key_index += 1
# check export format type
if self.format == "redbrick-json":
output.append({
"url": dp.items_list_not_signed[curr_idx],
"class": frame_label,
"name": video_name,
})
else:
# export the frame to the relvant folder
if not os.path.isdir(
os.path.join(self.cache_dir, video_name,
frame_label)):
os.makedirs(
os.path.join(self.cache_dir, video_name,
frame_label))
frame_data = np.flip(url_to_image(dp.items_list[curr_idx]),
axis=2)
frame_filename = (
dp.items_list_not_signed[curr_idx].replace(
"/", "_").replace(":", "_"))
cv2.imwrite(
os.path.join(self.cache_dir, video_name, frame_label,
frame_filename),
frame_data,
)
# update the index
curr_idx += 1
pbar.update(1)
pbar.close()
if self.format == "redbrick-json":
with open(os.path.join(self.cache_dir, "output.json"),
"w+") as file:
json.dump(output, file, indent=2)