model.load_weights(args.weights_path)
print(model.summary())
image_path = args.image_path
input_size = model.input_shape[1]
seguir = True
while seguir:
try:
# Load and process image
im_array = IOProcessor().process_image(image_path, input_size)
# Do the freaking prediction
pred = model.predict(im_array)[0]
# Filter by threshold
pred = np.where(pred >= 0.5, 1, 0) # the threshold can be in argument
# Convert binary values to classes
classes = []
for i, p in enumerate(pred):
if p == 1: classes.append(list(opts.CLASSES.keys())[i])
# Add classification information on the image and display it
im = Image.open(image_path).convert("RGBA")
draw_text(im, classes, show=True)
except FileNotFoundError:
logger.alert(
"File not found. The provided image path is not correct, please try a new one"
)
# Ask for input again!
logger.alert("Provide a new .jpg image path, or say 'quit' to... quit")
image_path = input()
if image_path.strip().lower() == "quit":
seguir = False
# Save everything in h5!!
h5_name = "voc_classification_trainval_224.h5"
h5_path = os.path.join(args.outputs_path, h5_name)
with h5py.File(h5_path, "w") as file:
file.create_dataset("features", data=features, dtype='float32')
file.create_dataset("labels", data=labels, dtype='float32')
logger.success("Created h5 file {} successfully".format(h5_name))
logger.log("Absolute path of h5 file:", os.path.abspath(h5_path))
#### Try to do that on next experiment ####
# Update on datastore if it is not a windows platform
if not sys.platform.startswith("win"):
logger.alert("This is not a windows platform, we try to update our features and labels on a datastore")
from azureml.core import Workspace, Dataset
from azureml.data.datapath import DataPath
# Get workspace
ws = Workspace(
subscription_id=args.subscription_id,
resource_group=args.resource_group,
workspace_name=args.workspace_name
)
files = [
h5_path
]
datastore = ws.get_default_datastore()
datastore.upload_files(
files=files,
