def do_inference(graph: mvnc.Graph,
image_filename: str,
number_results: int = 5) -> (List[str], List[numpy.float16]):
""" executes one inference which will determine the top classifications for an image file.
Parameters
----------
graph : Graph
The graph to use for the inference. This should be initialize prior to calling
image_filename : string
The filename (full or relative path) to use as the input for the inference.
number_results : int
The number of results to return, defaults to 5
Returns
-------
labels : List[str]
The top labels for the inference. labels[i] corresponds to probabilities[i]
probabilities: List[numpy.float16]
The top probabilities for the inference. probabilities[i] corresponds to labels[i]
"""
# text labels for each of the possible classfications
labels = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
# Load image from disk and preprocess it to prepare it for the network
# assuming we are reading a .jpg or .png color image so need to convert it
# single channel gray scale image for mnist network.
# Then resize the image to the size of image the network was trained with.
# Next convert image to floating point format and normalize
# so each pixel is a value between 0.0 and 1.0
image_for_inference = cv2.imread(image_filename)
image_for_inference = cv2.cvtColor(image_for_inference, cv2.COLOR_BGR2GRAY)
image_for_inference = cv2.resize(image_for_inference,
NETWORK_IMAGE_DIMENSIONS)
image_for_inference = image_for_inference.astype(numpy.float32)
image_for_inference[:] = ((image_for_inference[:]) * (1.0 / 255.0))
# Start the inference by sending to the device/graph
graph.LoadTensor(image_for_inference.astype(numpy.float16), None)
# Get the result from the device/graph. userobj should be the
# same value that was passed in LoadTensor above.
output, userobj = graph.GetResult()
# sort indices in order of highest probabilities
five_highest_indices = (-output).argsort()[:number_results]
# get the labels and probabilities for the top results from the inference
inference_labels = []
inference_probabilities = []
for index in range(0, number_results):
inference_probabilities.append(str(
output[five_highest_indices[index]]))
inference_labels.append(labels[five_highest_indices[index]])
return inference_labels, inference_probabilities
def do_inference(
self,
graph: mvnc.Graph,
img,
number_results: int = 5) -> (List[str], List[numpy.float16]):
labels = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
image_for_inference = img
image_for_inference = cv2.cvtColor(image_for_inference,
cv2.COLOR_BGR2GRAY)
image_for_inference = cv2.resize(image_for_inference,
NETWORK_IMAGE_DIMENSIONS)
ret, image_for_inference = cv2.threshold(image_for_inference, 120, 255,
cv2.THRESH_BINARY_INV)
cv2.imshow("image_for_inference:", image_for_inference)
image_for_inference = image_for_inference.astype(numpy.float32)
image_for_inference[:] = ((image_for_inference[:]) * (1.0 / 255.0))
# Start the inference by sending to the device/graph
graph.LoadTensor(image_for_inference.astype(numpy.float16), None)
# Get the result from the device/graph. userobj should be the
# same value that was passed in LoadTensor above.
output, userobj = graph.GetResult()
# sort indices in order of highest probabilities
five_highest_indices = (-output).argsort()[:number_results]
# get the labels and probabilities for the top results from the inference
inference_labels = []
inference_probabilities = []
for index in range(0, number_results):
inference_probabilities.append(
str(output[five_highest_indices[index]]))
inference_labels.append(labels[five_highest_indices[index]])
return inference_labels, inference_probabilities