def training_data_generator(coco, start_index, num_images, bounding_box_count,
cell_width_px, cell_height_px, batch_size):
img_ids = get_downloaded_ids()
img_ids = list(filter(is_not_greyscale, img_ids))[:num_images]
while 1:
for img_id in img_ids:
image_batch = np.empty((batch_size, 640, 640, 3),
np.float) #hard code
y_true_batch = np.empty((batch_size, 10, 10, 5),
np.float) #hard code
for i in range(batch_size):
image = file_management.get_image(img_id)
image = np.divide(image, 256, dtype=np.float32)
try:
image = pad_image(image, PADDED_SIZE)
except Exception as e:
print(id)
raise e
ground_truth = get_y_true(coco, bounding_box_count,
cell_width_px, cell_height_px,
img_id)
image_batch[i, :, :, :] = image
y_true_batch[i, :, :, :] = ground_truth
yield (image_batch, y_true_batch)
def image_generator(image_ids, buffer_size=0):
for id in image_ids:
image = file_management.get_image(id)
image = np.divide(image, 256, dtype=np.float32)
try:
image = pad_image(image, PADDED_SIZE)
except Exception as e:
print(id)
raise e
yield image
def plot_annotations(img_id, annotations, color='r'):
img = get_image(img_id)
fig, ax = plt.subplots()
ax.imshow(img)
for ann in annotations:
rect_patch = patches.Rectangle((ann['bbox'][0], ann['bbox'][1]),
ann['bbox'][2],
ann['bbox'][3],
linewidth=2,
edgecolor=ann.get('color', color),
facecolor='none')
ax.add_patch(rect_patch)
plt.show()
)
args = vars(ap.parse_args())
# load the trained convolutional neural network
print("[INFO] loading network...")
# Load in the custom loss function
get_custom_objects().update({"QueueTime_loss": QueueTime_loss})
model = load_model(args["model"])
img_ids = args["image_ids"]
if args['all']:
img_ids = filter(is_not_greyscale, get_downloaded_ids())
for img_id in img_ids:
image = pad_image(get_image(img_id), PADDED_SIZE)
image = np.expand_dims(image, axis=0)
# classify the input image
print("[INFO] classifying image...")
y_pred = model.predict(image)[0]
# post_pred = QueueTime_post_process(y_pred)
post_pred = cnn_y_to_absolute(CELL_WIDTH, CELL_HEIGHT, y_pred)
# filter out all scores below threshold:
post_pred_filtered = list(
filter(lambda ann: ann['score'] > 0.001, post_pred))
# Add color key:
scores = [ann['score'] for ann in post_pred_filtered]
normalize_score = lambda score: (score - min(scores)) / (max(scores) - min(
def is_not_greyscale(img_id):
img = get_image(img_id)
return img.ndim == 3