def load_datasets(args):
if args.matchedTraing2ndDataDir == '': # Load the 1 dataset
# Load dataset API (Alread logged in the args log step)
train_dataset = dataset_cvppp.Fine_Tune_CVPPP_Dataset(blur_images=args.blurImages, num_imgs=args.numSamples)
train_dataset.load_cvppp(args.dataDir, 'train')
train_dataset.prepare()
crossVal_dataset = dataset_cvppp.Fine_Tune_CVPPP_Dataset(blur_images=args.blurImages) # Use all for crossVal
crossVal_dataset.load_cvppp(args.dataDir, 'crossVal')
crossVal_dataset.prepare()
else: # Load both
print("Using matched Training")
train_dataset1 = dataset_cvppp.CVPPP_Dataset(blur_images=args.blurImages)
train_dataset1.load_cvppp(args.dataDir, 'train')
train_dataset1.prepare()
crossVal_dataset1 = dataset_cvppp.CVPPP_Dataset(blur_images=args.blurImages)
crossVal_dataset1.load_cvppp(args.dataDir, 'crossVal')
crossVal_dataset1.prepare()
train_dataset2 = dataset_cvppp.CVPPP_Dataset(blur_images=args.blurImages)
train_dataset2.load_cvppp(args.matchedTraing2ndDataDir, 'train')
train_dataset2.prepare()
crossVal_dataset2 = dataset_cvppp.CVPPP_Dataset(blur_images=args.blurImages)
crossVal_dataset2.load_cvppp(args.matchedTraing2ndDataDir, 'crossVal')
crossVal_dataset2.prepare()
train_dataset = (train_dataset1, train_dataset2)
crossVal_dataset = (crossVal_dataset1, crossVal_dataset2)
return train_dataset, crossVal_dataset
def run_example_augmentations():
"""
Function which can be called to visualise some example augmentations and ensure that they are valid
This uses the image loading function inside maskRCNN so the user can be confident in the data being fed to the network
"""
parser = argparse.ArgumentParser(
description='Visualise example augmentations')
parser.add_argument(
'--dataDir',
type=str,
required=True,
help=
'Directory containing training data stored in the expected format. See dataset_cvppp.py'
)
parser.add_argument('--outputDir',
type=str,
required=True,
help='Directory to save example images to')
parser.add_argument('--numImages',
type=int,
default=30,
help='How many images to save')
parser.add_argument('--blurImages', dest='blurImages', action='store_true')
parser.add_argument('--dontBlurImages',
dest='blurImages',
action='store_false')
parser.set_defaults(blurImages=True)
args = parser.parse_args()
# Create output dir
assert not os.path.isdir(args.outputDir), "output dir already exists"
os.mkdir(args.outputDir)
# # Init dataset
train_dataset = dataset_cvppp.CVPPP_Dataset()
train_dataset.load_cvppp(args.dataDir, 'train')
train_dataset.prepare()
# Init config
configuration = config_cvppp.TrainConfig()
# Init augmentation
augmentation = get_augmentation_sequence()
# Generate images
for i in range(args.numImages):
image, meta, class_ids, bbox, mask = model.load_image_gt(
train_dataset, configuration, i, augmentation=augmentation)
rgb_mask = mask_to_rgb(mask)
im_path = os.path.join(args.outputDir, str(i) + '_image.png')
mask_path = os.path.join(args.outputDir, str(i) + '_mask.png')
io.imsave(im_path, image)
io.imsave(mask_path, rgb_mask)
print("Saved example", i)
def evaluate_model():
"""
The main evaluation procedure
"""
args = arguments()
# Create output dir
assert not os.path.isdir(args.outputDir), "output dir already exists"
os.mkdir(args.outputDir)
# Init config
configuration = config_cvppp.InferenceConfig()
# Init model
inference_model = model.MaskRCNN(mode="inference",
config=configuration,
model_dir=args.outputDir)
assert os.path.exists(
args.weightsPath), "Weights file does not exist at " + args.weightsPath
inference_model.load_weights(args.weightsPath, by_name=True)
# Load dataset API
test_dataset = dataset_cvppp.CVPPP_Dataset()
if os.path.isdir(os.path.join(args.dataDir, 'test')):
test_dataset.load_cvppp(args.dataDir, 'test')
else:
test_dataset.load_cvppp(args.dataDir,
'') # Assume it is just this directory
test_dataset.prepare()
# init metrics
dice = MetricTracker()
DiC = MetricTracker()
abs_DiC = MetricTracker()
mAP = MetricTracker()
# save predictions
with open(os.path.join(args.outputDir, RESULTS_FILENAME),
'a') as results_file:
results_file.write("Filename, Path, mAP, SBD \n")
for image_id in test_dataset.image_ids:
# Activate the choice of image
test_dataset.image_reference(image_id)
image, image_meta, gt_class_id, gt_bbox, gt_mask =\
model.load_image_gt(test_dataset, configuration, image_id, use_mini_mask=False)
image_path = test_dataset.image_reference(image_id)['path']
# Run inference
results = inference_model.detect([image], verbose=False)
r = results[0]
results_file.write(
os.path.basename(image_path) + ', ' + image_path + ', ')
# mAP @ IoU 0.5 metric
mAPResult, precisions, recalls, overlaps = \
utils.compute_ap(gt_bbox, gt_class_id, gt_mask,
r["rois"], r["class_ids"], r["scores"], r['masks'])
mAP.add(mAPResult, 0)
results_file.write(str(mAPResult)[:6] + ', ')
# Dice metric
dice_res, dice_std = CvpppSymmetricBestDice.calculate(
r['masks'], gt_mask)
dice.add(dice_res, dice_std)
results_file.write(str(dice_res)[:6] + '\n')
# Save visualisation of prediction
save_path = os.path.join(args.outputDir,
os.path.basename(image_path))
visualise_prediction(image, r['masks'], gt_mask, save_path)
# save overall results
with open(os.path.join(args.outputDir, OVERALL_RESULTS_FILENAME),
'a') as results_file:
results_file.write("Metric, Result, STD\n")
overall_mAP = mAP.calc_mean()
results_file.write("mAP, " + str(overall_mAP[0])[:6] + ', ' +
str(overall_mAP[1])[:6] + '\n')
overall_dice = dice.calc_mean()
results_file.write("Symmetric Best Dice, " + str(overall_dice[0])[:6] +
', ' + str(overall_dice[1])[:6] + '\n')