def test_create_detection_mapping():
"""Test create_detection_mapping()."""
# credit to voxel51 crew for a helpful test suite from which this test borrows
# https://github.com/voxel51/fiftyone/blob/a7c2b36a4f101330fa8edec35a9bdee841886f96/tests/unittests/view_tests.py#L59
dataset = fo.Dataset(name="test_detection_mapping")
dataset.add_sample(
fo.Sample(
filepath="filepath1.jpg",
tags=["train"],
test_dets=fo.Detections(detections=[
fo.Detection(
label="friend",
confidence=0.9,
bounding_box=[0, 0, 0.5, 0.5],
),
fo.Detection(
label="stopper",
confidence=0.1,
bounding_box=[0, 0, 0.5, 0.5],
),
]),
another_field=51,
))
test_output = create_detection_mapping("test_detection_mapping",
label_field="test_dets",
training_tag="train")
assert isinstance(test_output, str)
assert (
test_output ==
'item {\n name: "friend"\n id: 1\n}\nitem {\n name: "stopper"\n id: 2\n}\n'
)
def test__create_list_of_class_names():
"""Test _create_list_of_class_names."""
# credit to voxel51 crew for a helpful test suite from which this test borrows
# https://github.com/voxel51/fiftyone/blob/a7c2b36a4f101330fa8edec35a9bdee841886f96/tests/unittests/view_tests.py#L59
dataset = fo.Dataset()
dataset.add_sample(
fo.Sample(
filepath="filepath1.jpg",
tags=["test"],
test_dets=fo.Detections(detections=[
fo.Detection(
label="friend",
confidence=0.9,
bounding_box=[0, 0, 0.5, 0.5],
),
fo.Detection(
label="stopper",
confidence=0.1,
bounding_box=[0, 0, 0.5, 0.5],
),
fo.Detection(
label="big bro",
confidence=0.6,
bounding_box=[0, 0, 0.1, 0.5],
),
]),
another_field=51,
))
test_list = _create_list_of_class_names(dataset, label_field="test_dets")
assert set(test_list) == set(["friend", "stopper", "big bro"])
def _make_detection_dataset(
img, images_dir, num_samples=4, num_objects_per_sample=3
):
exts = [".jpg", ".png"]
samples = []
for idx in range(num_samples):
filepath = os.path.join(
images_dir, "%06d%s" % (idx, exts[idx % len(exts)])
)
etai.write(img, filepath)
detections = []
for _ in range(num_objects_per_sample):
label = random.choice(["cat", "dog", "bird", "rabbit"])
bounding_box = [
0.8 * random.random(),
0.8 * random.random(),
0.2,
0.2,
]
detections.append(
fo.Detection(label=label, bounding_box=bounding_box)
)
samples.append(
fo.Sample(
filepath=filepath,
ground_truth=fo.Detections(detections=detections),
)
)
dataset = fo.Dataset()
dataset.add_samples(samples)
return dataset
def _create_synth_fiftyone_dataset(tmpdir):
img_dir = Path(tmpdir / "fo_imgs")
img_dir.mkdir()
Image.new('RGB', (1920, 1080)).save(img_dir / "sample_one.png")
Image.new('RGB', (1920, 1080)).save(img_dir / "sample_two.png")
dataset = fo.Dataset.from_dir(
img_dir,
dataset_type=fo.types.ImageDirectory,
)
sample1 = dataset[str(img_dir / "sample_one.png")]
sample2 = dataset[str(img_dir / "sample_two.png")]
d1 = fo.Detection(
label="person",
bounding_box=[0.3, 0.4, 0.2, 0.2],
)
d2 = fo.Detection(
label="person",
bounding_box=[0.05, 0.10, 0.28, 0.15],
)
d3 = fo.Detection(
label="person",
bounding_box=[0.23, 0.14, 0.09, 0.18],
)
d1["iscrowd"] = 1
d2["iscrowd"] = 0
d3["iscrowd"] = 0
sample1["ground_truth"] = fo.Detections(detections=[d1])
sample2["ground_truth"] = fo.Detections(detections=[d2, d3])
sample1.save()
sample2.save()
return dataset
def create_fo_sample(image: Image, labels: str, boxes):
"""
Args
-----------
image: PIL image
labels: label name
boxes: xyxy format
"""
assert len(labels) == len(boxes)
detections = []
for i in range(len(labels)):
detections.append(fo.Detection(label=labels[i], bounding_box=boxes[i]))
sample = fo.Sample(filepath=image.filename)
sample["ground_truth"] = fo.Detections(detections=detections)
return sample
def test_filter_detections(self):
self.sample1["test_dets"] = fo.Detections(detections=[
fo.Detection(
label="friend",
confidence=0.9,
bounding_box=[0, 0, 0.5, 0.5],
),
fo.Detection(
label="friend",
confidence=0.3,
bounding_box=[0.25, 0, 0.5, 0.1],
),
fo.Detection(
label="stopper",
confidence=0.1,
bounding_box=[0, 0, 0.5, 0.5],
),
fo.Detection(
label="big bro",
confidence=0.6,
bounding_box=[0, 0, 0.1, 0.5],
),
])
self.sample1.save()
self.sample2["test_dets"] = fo.Detections(detections=[
fo.Detection(
label="friend",
confidence=0.99,
bounding_box=[0, 0, 1, 1],
),
fo.Detection(
label="tricam",
confidence=0.2,
bounding_box=[0, 0, 0.5, 0.5],
),
fo.Detection(
label="hex",
confidence=0.8,
bounding_box=[0.35, 0, 0.2, 0.25],
),
])
self.sample2.save()
view = self.dataset.filter_detections(
"test_dets", (F("confidence") > 0.5) & (F("label") == "friend"))
for sv in view:
for det in sv.test_dets.detections:
self.assertGreater(det.confidence, 0.5)
self.assertEqual(det.label, "friend")
def add_predictions(dataset, sample_label, device, model):
classes = dataset.default_classes
with fo.ProgressBar() as pb:
for sample in pb(dataset):
# Load image
image = Image.open(sample.filepath)
image = F.to_tensor(image).to(device)
c, h, w = image.shape
preds = model([image])[0]
# Non-Max suppression.
# Indices to keep.
idx = nms(boxes=preds["boxes"], scores=preds["scores"], iou_threshold=0.1)
boxes = preds["boxes"][idx]
labels = preds["labels"][idx]
scores = preds["scores"][idx]
labels = labels.cpu().detach().numpy()
scores = scores.cpu().detach().numpy()
boxes = boxes.cpu().detach().numpy()
detections = []
for label, score, box in zip(labels, scores, boxes):
x1, y1, x2, y2 = box
rel_box = [
x1 / w,
y1 / h,
(x2 - x1) / w,
(y2 - y1) / h,
] # fiftyone format
detections.append(
fo.Detection(
label=classes[label], bounding_box=rel_box, confidence=score
)
)
sample[sample_label] = fo.Detections(detections=detections)
sample.save()
def train_with_hydra(cfg: DictConfig):
cfg.inference.base_path= cfg.inference.model_path_to_load.split("train/",1)[0] +"inference"
print("INFERENCE RESULTS WILL BE SAVED {}".format(cfg.inference.base_path))
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# For inferece set always batch_size = 1
cfg.inference.batch_size = 1
createFolderForExplanation(cfg)
# ----------
# INSTANTIATE TEST TRANSFORM DATA
# -----------
# Dataclass for custom Image transform See dataset configuration in .yaml
@dataclass
class ObjectDetectionInputTransform(InputTransform):
def predict_per_sample_transform(self):
return instantiate(cfg.dataset.test_transform, _convert_="all")
# ----------
# INSTANTIATE COCO DATA
# -----------
# Load COCO formatted dataset in fiftyone
fiftyone_predict_dataset = fo.Dataset.from_dir(
dataset_type=fo.types.COCODetectionDataset,
data_path=cfg.inference.path_folder_images,
labels_path=cfg.inference.path_folder_annotations,
include_id=True,
label_field="ground_truth"
)
class_labels = fiftyone_predict_dataset.default_classes
# create datamodule to load data in lightning
datamodule = ObjectDetectionData.from_fiftyone(
predict_dataset=fiftyone_predict_dataset,
predict_transform=ObjectDetectionInputTransform,
label_field="ground_truth_detections",
batch_size=cfg.inference.batch_size
)
# ----------
# INSTANTIATE MODEL AND TRAINER
# -----------
model = instantiate(cfg.model.model)
# if((cfg.inference.model_path_to_load != None) and (cfg.inference.model_path_to_load != "")):
# print("LOAD PRETRAINED WEIGHT")
# model = model.load_from_checkpoint(
# cfg.inference.model_path_to_load)
# instantiate trainer
trainer = instantiate(cfg.trainer.default)
# ----------
# RUN PREDICTION
# -----------
# All prediction in your dataset
predictions = trainer.predict(model, datamodule=datamodule)
# model needs to put on gpu after train.predict in order to run explanation on gpu
if(torch.cuda.is_available()):
model = model.to(device)
# model in eval model for explainability
model.eval()
# over all predictions
for i, sample in enumerate(fiftyone_predict_dataset):
detections = []
# 0 is because batch_size = 1
#over all bboxes in one image
for bbox,score,label_index in zip(predictions[i][0][DataKeys.PREDS]["bboxes"],predictions[i][0][DataKeys.PREDS]["scores"],predictions[i][0][DataKeys.PREDS]["labels"]):
# Original Image Dimension
img_width_orig = sample.metadata["width"]
img_height_orig = sample.metadata["height"]
# model input dimension
model_input_height, model_input_width = predictions[i][0][DataKeys.METADATA]["size"]
# Coordinate rescale format is (x1,y1,x2,y2)
rescale_coordinate = scale_coords(
(model_input_height, model_input_width), bbox, (img_height_orig, img_width_orig))
# Need to change from (x1,y1,x2,y2)-> (x1,y1,width,heigth). Coco annotations are in this format
rescale_coordinate[2] = rescale_coordinate[2] - \
rescale_coordinate[0]
rescale_coordinate[3] = rescale_coordinate[3] - \
rescale_coordinate[1]
# Coordinates in Fiftyone are normalized for Coco
rescale_coordinate[[0, 2]] /= img_width_orig
rescale_coordinate[[1, 3]] /= img_height_orig
# Detection threshold
if(score> cfg.inference.threshold):
detections.append(
fo.Detection(
label=class_labels[label_index],
bounding_box=rescale_coordinate.tolist(),
confidence=score
))
# save new detection to fifyone dataset
sample["inference"] = fo.Detections(detections=detections)
sample.save()
results = fiftyone_predict_dataset.evaluate_detections(
"inference",
gt_field="ground_truth_detections",
eval_key="objectDetection_eval",
compute_mAP=True,
)
# Save Confusion Matrix
plot = results.plot_confusion_matrix(classes=class_labels,
backend="matplotlib", figsize=(6, 6))
plot.savefig(cfg.inference.confusion_matrix.path_to_confusion_matrix_image)
print(results.mAP())
print(results.metrics())
def run_detection_model_tiled(
dataset_name,
training_name,
prediction_field,
sample_tag,
tile_string,
tile_overlap:int,
iou_threshold:float,
):
"""Runs the detection model over the entire dataset using a tiling approach
Args:
interpreter: The ``tf.lite.Interpreter`` to update.
size (tuple): The original image size as (width, height) tuple.
resize: A function that takes a (width, height) tuple, and returns an
image resized to those dimensions.
Returns:
The resized tensor with zero-padding as tuple
(resized_tensor, resize_ratio).
"""
model_path = (
"/tf/model-export/" + training_name + "/image_tensor_saved_model/saved_model"
)
min_score = 0.50 # This is the minimum score for adding a prediction. This helps keep out bad predictions but it may need to be adjusted if your model is not that good yet.
input_tensor_size = 512
logging.info("Loading model...")
start_time = time.time()
tf.keras.backend.clear_session()
detect_fn = tf.saved_model.load(model_path)
infer = detect_fn.signatures["serving_default"]
print(infer.structured_outputs)
print(infer)
end_time = time.time()
elapsed_time = end_time - start_time
logging.info("Loading model took: " + str(elapsed_time) + "s")
category_index = _load_label_map(training_name)
dataset = fo.load_dataset(dataset_name)
# Go through all of the samples in the dataset
for sample in dataset.match_tags(sample_tag).select_fields("filepath"):
start_time = time.time()
img = load_img(
sample.filepath,
)
img_size = img.size
img_width, img_height = img_size
objects_by_label = dict()
exportDetections = []
predicted_objects = []
tile_sizes = []
for tile_size in tile_string.split(","):
tile_size = tile_size.split("x")
tile_sizes.append([int(tile_size[0]), int(tile_size[1])])
# Collect all of the detections for each tile size:
for tile_size in tile_sizes:
tile_width, tile_height = tile_size
# For tiles that are smaller that the image size, calculated all of the different
# Sub images that are needed
for tile_location in _tiles_location_gen(img_size, tile_size, tile_overlap):
tile = img.crop(tile_location)
old_size = tile.size # old_size[0] is in (width, height) format
ratio = float(input_tensor_size) / max(old_size)
if ratio > 1:
continue
new_size = tuple([int(x * ratio) for x in old_size])
im = tile.resize(new_size, Image.ANTIALIAS)
# create a new image and paste the resized on it
new_im = Image.new("RGB", (input_tensor_size, input_tensor_size))
new_im.paste(
im, (0, 0)
) # ((input_tensor_size-new_size[0])//2, (input_tensor_size-new_size[1])//2))
img_array = img_to_array(new_im, dtype="uint8")
img_batch = np.array([img_array])
detections = detect_fn(img_batch)
for i, detectScore in enumerate(detections["detection_scores"][0]):
if detectScore > min_score:
x1 = (
detections["detection_boxes"][0][i][1].numpy()
* input_tensor_size
) # tile_width
y1 = (
detections["detection_boxes"][0][i][0].numpy()
* input_tensor_size
) # tile_height
x2 = (
detections["detection_boxes"][0][i][3].numpy()
* input_tensor_size
) # tile_width
y2 = (
detections["detection_boxes"][0][i][2].numpy()
* input_tensor_size
) # tile_height
bbox = [x1, y1, x2, y2]
scaled_bbox = []
for number in bbox:
scaled_bbox.append(number / ratio)
repositioned_bbox = _reposition_bounding_box(
scaled_bbox, tile_location
)
confidence = detections["detection_scores"][0][i]
label = _find_class_name(
category_index, int(detections["detection_classes"][0][i])
)
objects_by_label.setdefault(label, []).append(
Object(label, confidence, repositioned_bbox)
)
predicted_objects.append(Object(label, confidence, repositioned_bbox))
# for label, objects in objects_by_label.items():
# idxs = _non_max_suppression(objects, iou_threshold)
# for idx in idxs:
# x1 = objects[idx].bbox[0] / img_width
# y1 = objects[idx].bbox[1] / img_height
# x2 = objects[idx].bbox[2] / img_width
# y2 = objects[idx].bbox[3] / img_height
# w = x2 - x1
# h = y2 - y1
# bbox = [x1, y1, w, h]
# exportDetections.append(
# fo.Detection(
# label=objects[idx].label,
# bounding_box=bbox,
# confidence=objects[idx].score,
# )
# )
objects = predicted_objects
idxs = _non_max_suppression(objects, iou_threshold)
for idx in idxs:
x1 = objects[idx].bbox[0] / img_width
y1 = objects[idx].bbox[1] / img_height
x2 = objects[idx].bbox[2] / img_width
y2 = objects[idx].bbox[3] / img_height
w = x2 - x1
h = y2 - y1
bbox = [x1, y1, w, h]
exportDetections.append(
fo.Detection(
label=objects[idx].label,
bounding_box=bbox,
confidence=objects[idx].score,
)
)
# Store detections in a field name of your choice
sample[prediction_field] = fo.Detections(detections=exportDetections)
sample.save()
end_time = time.time()
print("{} - Processing {} took: {}s".format(len(exportDetections),sample.filepath, end_time - start_time))
for detect in exportDetections:
print("\t - {} {}%".format(detect.label,detect.confidence))
def run_detection_model(dataset_name, training_name, prediction_field):
model_path = (
"/tf/model-export/" + training_name + "/image_tensor_saved_model/saved_model"
)
min_score = 0.5 # This is the minimum score for adding a prediction. This helps keep out bad predictions but it may need to be adjusted if your model is not that good yet.
logging.info("Loading model...")
start_time = time.time()
tf.keras.backend.clear_session()
detect_fn = tf.saved_model.load(model_path)
infer = detect_fn.signatures["serving_default"]
end_time = time.time()
elapsed_time = end_time - start_time
logging.info("Loading model took: " + str(elapsed_time) + "s")
category_index = _load_label_map(training_name)
dataset = fo.load_dataset(dataset_name)
for sample in dataset.select_fields("filepath"):
start_time = time.time()
img = load_img(sample.filepath)
img_array = img_to_array(img)
input_tensor = np.expand_dims(img_array, 0)
detections = detect_fn(input_tensor)
exportDetections = []
for i, detectScore in enumerate(detections["detection_scores"][0]):
if detectScore > min_score:
print(
"\t- {}: {}".format(
_find_class_name(category_index, int(detections["detection_classes"][0][i])),
detections["detection_scores"][0][i],
)
)
label = _find_class_name(category_index, int(detections["detection_classes"][0][i]))
confidence = detections["detection_scores"][0][i]
# TF Obj Detect bounding boxes are: [ymin, xmin, ymax, xmax]
# For Voxel 51 - Bounding box coordinates should be relative values
# in [0, 1] in the following format:
# [top-left-x, top-left-y, width, height]
x1 = detections["detection_boxes"][0][i][1]
y1 = detections["detection_boxes"][0][i][0]
x2 = detections["detection_boxes"][0][i][3]
y2 = detections["detection_boxes"][0][i][2]
w = x2 - x1
h = y2 - y1
bbox = [x1, y1, w, h]
exportDetections.append(
fo.Detection(label=label, bounding_box=bbox, confidence=confidence)
)
# Store detections in a field name of your choice
sample[prediction_field] = fo.Detections(detections=exportDetections)
sample.save()
end_time = time.time()
print("Processing {} took: {}s".format(sample.filepath, end_time - start_time))
for mask in instances.pred_masks
]
for rle in rles:
rle["counts"] = rle["counts"].decode("utf-8")
# Convert detections to FiftyOne format
detections = []
for label, score, box in zip(labels, scores, boxes):
# Convert to [top-left-x, top-left-y, width, height]
# in relative coordinates in [0, 1] x [0, 1]
x1, y1, x2, y2 = box
rel_box = [x1 / w, y1 / h, (x2 - x1) / w, (y2 - y1) / h]
detections.append(
fo.Detection(label=classes[label],
bounding_box=rel_box,
confidence=score,
masks=rles))
# Save predictions to dataset
sample["mask_rcnn"] = fo.Detections(detections=detections)
sample.save()
print("Finished adding predictions")
# In[ ]:
session.view = predictions_view
# In[ ]:
# Only contains detections with confidence >= 0.15
def test_accuracy_yolov4(capsys):
data_path = "/home/Develop/Dataset/Coco2017/validation/val2017"
labels_path = "/home/Develop/Dataset/Coco2017/validation/valAnnotation/instances_val2017.json"
# The type of the dataset being imported
dataset_type = fo.types.COCODetectionDataset
# Coco By default has 90 class
# Detection use only 80 output
# This is needed to map the 80 output from model prediction directly
CocoMap=[1,2,3,4,5,6,7,8,
9,10,11,13,14,15,16,17,
18,19,20,21,22,23,24,25,
27,28,31,32,33,34,35,36,
37,38,39,40,41,42,43,44,
46,47,48,49,50,51,52,53,
54,55,56,57,58,59,60,61,
62,63,64,65,67,70,72,73,
74,75,76,77,78,79,80,81,
82,84,85,86,87,88,89,90]
dataset = foz.load_zoo_dataset(
"coco-2017",
split="validation",
dataset_name="evaluate-detections-tutorial",
)
dataset.persistent = True
detectionDir="/home/Develop/Dataset/Coco2017/validation/val2017Pred/"
#dataset.delete_sample_field("faster_rcnn")
# KEY
# Need to map Output classes from Ai4prod Yolo evaluation to Coco Class.
# This happen beacuse on Coco annotation we have 90 classes while from Yolov4 prediction we have only 80 Class
# So if Yolov4 output class is 0 on Coco annoation is class 1 CocoMap[0]=1
with capsys.disabled():
classes = dataset.default_classes
with fo.ProgressBar() as pb:
for sample in pb(dataset):
image = Image.open(sample.filepath)
w, h = image.size
head, tail = os.path.split(sample.filepath)
filename, file_extension = os.path.splitext(tail)
# one image can have multiple detections
detections=[]
cvsPath= detectionDir + filename +".txt"
if (os.path.isfile(cvsPath)):
with open(cvsPath,"r") as file:
reader = csv.reader(file)
for row in reader:
if(row[0]=="x1"):
pass
else:
# Detection must be int value respect to original image size
# left_x ,left_y , width , height
x1= float(row[0])
y1= float(row[1])
width= float(row[2])
height= float(row[3])
# Coordinate Normalization
rel_box = [x1/w, y1/h, width/w, height/h]
score= row[4]
cls = CocoMap[int(row[5])]
# fiftyone require the correct Coco class
label = classes[CocoMap[int(row[5])]]
detections.append(
fo.Detection(
label=label,
bounding_box=rel_box,
confidence=score
)
)
else:
print("FILE NOT FOUND")
#Save predictions data in Dataset
sample["yolov4"] = fo.Detections(detections=detections)
sample.save()
print("EVALUATION")
# eval_key= use always the same key to evaluate yolov4 predictions
results = dataset.evaluate_detections(
"yolov4",
gt_field="ground_truth",
eval_key="yolov4_eval",
compute_mAP=True,
)
print(results.mAP())
print(results.metrics())
assert results.mAp()> 0.47