def load_images(img_path, flip, rotate, zoom, shear, batch_size, img_size,
seed):
logger.info(
'Loading training data from {}. {}, rotate={}, zoom={}, shear={}. Batch size={}. Image size={}.'.
format(img_path, 'flip' if flip else 'no flip', rotate, zoom, shear,
batch_size, img_size))
training = image.ImageDataGenerator(
horizontal_flip=flip,
vertical_flip=flip,
rotation_range=rotate,
zoom_range=zoom,
shear_range=shear)
train_gen = image.DirectoryIterator(
os.path.join(img_path, 'training'),
training,
batch_size=batch_size,
target_size=(img_size, img_size),
seed=seed)
validation = image.ImageDataGenerator()
valid_gen = image.DirectoryIterator(
os.path.join(img_path, 'validation'),
validation,
batch_size=batch_size,
target_size=(img_size, img_size),
seed=seed)
return train_gen, valid_gen
def evaluate(model_root,
model,
images,
image_size,
num_batches,
seed,
top_n=None):
imagegen = image.ImageDataGenerator()
test_gen = image.DirectoryIterator(os.path.join(images, 'testing'),
imagegen,
target_size=(image_size, image_size),
seed=seed,
shuffle=True)
# Get classes sorted by their value
classes = [
x[0]
for x in sorted(test_gen.class_indices.items(), key=lambda x: x[1])
]
metrics_path = model_root + "_metrics.csv"
cm_path = model_root + "_cm.png"
metrics, _, _ = score_keras.evaluate_model(model,
test_gen,
classes,
num_batches,
metrics_path,
cm_path,
top_n=top_n)
return classes, cm_path, metrics
def evaluate(model_root, model, images, image_size, seed, top_n=None):
imagegen = image.ImageDataGenerator()
test_gen = image.DirectoryIterator(os.path.join(images, 'testing'),
imagegen,
target_size=(image_size, image_size),
seed=seed,
shuffle=True)
# Get classes sorted by their value
classes = [
x[0]
for x in sorted(test_gen.class_indices.items(), key=lambda x: x[1])
]
metrics_path = model_root.with_name(model_root.name + "_metrics.csv")
cm_path = model_root.with_name(model_root.name + "_cm.png")
metrics, _, _ = score_keras.evaluate_model(model,
test_gen,
classes,
metrics_path,
cm_path,
top_n=top_n)
class_map = model_root.with_name(model_root.name + "_classes.csv")
try:
with class_map.open('w', encoding='utf-8') as cmfp:
cmfp.write('Class,ID\n')
for i, c in enumerate(classes):
cmfp.write('"{}",{}\n'.format(c, i))
except:
logger.warn('Failed to write class map file.')
return classes, cm_path, metrics
def load_model_and_images(model_file, image_dir, seed):
testpath = image_dir
imagegen = image.ImageDataGenerator()
test_gen = image.DirectoryIterator(testpath,
imagegen,
target_size=(224, 224),
seed=seed,
shuffle=True)
model = load_model(model_file)
# Get classes sorted by their value
classes = [
x[0]
for x in sorted(test_gen.class_indices.items(), key=lambda x: x[1])
]
return model, test_gen, classes
def extract_neg_hard_pairs():
path = os.path.join(os.path.join(os.path.join(base_path, "img"), "BOTH"),
"train")
datagen_anchor = image.ImageDataGenerator(rescale=1 / 255.)
datagen_negative = image.ImageDataGenerator(rescale=1 / 255.)
iterator_anchor = image.DirectoryIterator(
directory=path,
image_data_generator=datagen_anchor,
batch_size=1,
target_size=(64, 64))
iterator_negative = image.DirectoryIterator(
directory=path,
image_data_generator=datagen_negative,
batch_size=1,
target_size=(64, 64))
def euclidean_dist(a, n):
return np.sum(np.square((a - n))) / 64
for i in tqdm(range(len(iterator_anchor)),
ncols=100,
desc="Training",
bar_format="{l_bar}%s{bar}%s{r_bar}" %
(Fore.GREEN, Fore.RESET)):
info_anchor = iterator_anchor.filenames[i].split("/")
cls_idx_anchor = info_anchor[0]
item_idx_anchor = info_anchor[1]
item_name_anchor = info_anchor[2]
img_anchor, y_anchor = next(iterator_anchor)
closest_dist = 1e9
closest_cls_idx = -1
closest_item_idx = -1
closest_item_name = ""
dist = 0
for j in range(len(iterator_negative)):
info_negative = iterator_negative.filenames[j].split("/")
cls_idx_negative = info_negative[0]
item_idx_negative = info_negative[1]
item_name_negative = info_negative[2]
img_negative, y_negative = next(iterator_negative)
# dist = euclidean_dist(img_anchor, img_negative)
if cls_idx_anchor != cls_idx_negative and closest_dist > dist:
closest_dist = dist
closest_cls_idx = cls_idx_negative
closest_item_idx = item_idx_negative
closest_item_name = item_name_negative
pair_dict = {
"c_idx_anc": cls_idx_anchor,
"c_idx_neg": closest_cls_idx,
"i_idx_anc": item_idx_anchor,
"i_idx_neg": closest_item_idx,
"i_name_anc": item_name_anchor,
"i_name_neg": closest_item_name,
"distance": closest_dist
}
print(pair_dict)
with open('./data/Anno/hard_neg_pairs.csv', 'w') as output_file:
writer = csv.writer(output_file)
for key, value in pair_dict.items():
writer.writerow([key, value])