def test_identical(self, x, score, rep):
""" Ensure that non_max_suppression works correctly for identical boxes and that ordering doesn't matter. """
x = x.cumsum(axis=1)
x = x.repeat(rep).reshape(x.shape[1], rep).T
score = np.array([score] * rep)
idx = np.random.randint(len(x))
score[idx] = 1000
nms = non_max_suppression(x, score)
msg = "non_max_suppression failed to produce the expected output when all detections are identical"
assert_array_equal(nms, np.array([idx]), msg)
nms = non_max_suppression(x, score, threshold=1)
msg = "non_max_suppression failed to produce the expected output for identical detections with threshold 1"
assert_array_equal(nms, np.array(range(len(x))), msg)
def scan_fridge(fridge_image, model, classifying_model):
training_mean = 200.62128
training_std = 42.149155
device = 'cuda:1' if torch.cuda.is_available() else 'cpu'
img = np.copy(fridge_image).astype(np.float32)
img-=training_mean
img/= training_std
img = img[np.newaxis]
img = torch.tensor(img.transpose(0, 3, 1, 2)).to(device)
#print(img.shape)
assert img.shape == torch.Size([1, 3, 800, 497]), "The fridge image is not in the right shape"
out_cls, out_reg = model(img)
box_preds, class_preds, scores = compute_detections(out_cls.squeeze(),
out_reg.squeeze(),
feature_map_width=13,
anchor_box_step=40,
anchor_box_size=80)
#Check this
threshold = 0
keep = scores > threshold
box_preds = box_preds[keep]
class_preds = class_preds[keep]
scores = scores[keep]
keep_idxs = non_max_suppression(box_preds, scores, threshold=0.1)
box_preds = box_preds[keep_idxs]
class_preds = class_preds[keep_idxs]
#print(box_preds)
food_images = []
food_coords = []
for class_pred, box_pred in zip(class_preds, box_preds):
if class_pred > 0:
x1, y1, x2, y2 = box_pred
if x1 < 0 or y1 < 0 or x2 >= 497 or y2 >= 800:
continue
cropped_image = crop_image(fridge_image.swapaxes(0,2).swapaxes(1,2),x1,y1)
#print("cropped image shape", cropped_image.shape)
food_images.append(cropped_image)
food_coords.append((x1,y1))
#print("one item in food_images", food_images[2].shape)
#print("coords", food_coords)
food_images = np.array(food_images)
#print("all images shape", food_images.shape)
if len(food_images)==0:
return []
#print(food_images.shape, food_images)
classes = classifying_model(food_images)
#print(classes.shape)
#print(food_coords[0][0])
#print(food_coords[0][1])
#print(labels[np.argmax(classes[0].flatten())])
#return food_images, [(food_coords[i][0], food_coords[i][1], labels[np.argmax(classes[i].flatten())], "food") for i in range(len(classes))]
#Check this: CHANGE back to this LATER
return [Item(food_coords[i][0], food_coords[i][1], labels[np.argmax(classes[i].flatten())], "food",None)for i in range(len(classes))]
def test_known_results(self, threshold, desired_nms):
""" Ensures that non_max_suppression works correctly for known values. """
boxes = np.array([[0, 0, 1, 1], [0.5, 0.5, 0.9, 0.9]])
scores = np.array([0, 1])
actual_nms = non_max_suppression(boxes, scores, threshold=threshold)
assert_array_equal(actual_nms, desired_nms)
def test_empty(self):
""" Ensure that non_max_suppression works correctly with zero detections. """
x = np.empty((0, 4))
scores = np.empty((0, ))
nms = non_max_suppression(x, scores)
msg = "non_max_suppression failed to produce the expected output for zero detections"
assert nms.shape == (0, ), msg
def test_shapes(self, boxes: ndarray, data: st.SearchStrategy):
""" Ensure that non_max_suppression produces the correct shape output, even for empty inputs. """
scores = boxes[:, 4]
boxes = boxes[:, :4].cumsum(
axis=1) # ensures no 0-width or -height boxes
N = scores.shape[0]
nms = non_max_suppression(boxes, scores)
assert (
nms.shape[0] <= N
) # we're suppressing, so we can never end up with more things than we started with
assert nms.ndim == 1
def scan_fridge(fridge_image, training_mean, training_std):
device = 'cuda:1' if torch.cuda.is_available() else 'cpu'
f = open("Actual_Model.p","rb")
model = pickle.load(f)
f.close()
img = np.copy(fridge_image).astype(np.float32)
img-=training_mean
img/= training_std
img = img[np.newaxis]
img = torch.tensor(img.transpose(0, 3, 1, 2)).to(device)
assert img.shape == torch.Size([1, 3, 800, 497]), "The fridge image is not in the right shape"
out_cls, out_reg = model(img)
box_preds, class_preds, scores = compute_detections(out_cls.squeeze(),
out_reg.squeeze(),
feature_map_width=13,
anchor_box_step=40,
anchor_box_size=80)
keep = scores > threshold
box_preds = box_preds[keep]
class_preds = class_preds[keep]
scores = scores[keep]
keep_idxs = non_max_suppression(box_preds, scores, threshold=0.1)
box_preds = box_preds[keep_idxs]
class_preds = class_preds[keep_idxs]
food_images = []
food_coords = []
for class_pred, box_pred in zip(class_preds, box_preds):
if class_pred > 0:
#print("Here")
x1, y1, x2, y2 = box_pred
food_images.append(crop_image(fridge_image,x1,y1))
food_cords.append((x1,y1))
food_images = np.array(food_images)
f = open("classifying_model1", "rb")
classifying_model = pickle.load(f)
f.close()
classes = classifying_model(food_images)
return [Item(food_coords[i][0], food_coords[i][1], labels[classes[i]], "food")]
def test_single_detections(self, x: ndarray, score):
""" Ensure that a single detection is not suppressed. """
nms = non_max_suppression(x, np.array([score]))
msg = "non_max_suppression failed to produce the expected output for a single detection"
assert_array_equal(nms, np.array([0]), msg)
