Problem: huge accuracy degradation after conversion of ONNX model to RKNN and inference on Radxa Zero 3W board.
Full description: I develop a deep learning model using yolov5n architecture (applying optimization mentioned in rknn-model-zoo, SiLU -> ReLU etc.). I did conversion with default code example from https://docs.radxa.com/en/rock5/rock5c/app-development/rknn_toolkit_lite2_yolov5 . The test inference after the conversion on my Linux x86_64 machine predicted acceptable results (photo below). Also, I tried to add MMSE quantization method, pruning, channel quantization, add more images for quantization set (~30+ images). All of them resulted in good performance on my host machine.
However, all my attempts of inference on my Radxa Zero 3W with the same code (changing only rknn to rknn_lite) resulted in massive accuracy degradation (photo below):
Here is the inference code I’ve used:
from rknnlite.api import RKNNLite
import cv2
import numpy as np
import time
RKNN_MODEL = '/home/radxa/VideoObjectDetection/yolov5/models/default_yolov5/yolov5n_AerialObjectDetection_cars_simple_400epochs_normal.rknn'
IMG_PATH = '/home/radxa/VideoObjectDetection/yolov5/images/cars_bridge.jpg'
RESULT_PATH = 'result/cars_bridge_pred_400epochs_normal.png'
IMG_SIZE = 800
OBJ_THRESH = 0.05
NMS_THRESH = 0.45
CLASSES = ("car", "motorbike")
def xywh2xyxy(x):
# Convert [x, y, w, h] to [x1, y1, x2, y2]
y = np.copy(x)
y[:, 0] = x[:, 0] - x[:, 2] / 2 # top left x
y[:, 1] = x[:, 1] - x[:, 3] / 2 # top left y
y[:, 2] = x[:, 0] + x[:, 2] / 2 # bottom right x
y[:, 3] = x[:, 1] + x[:, 3] / 2 # bottom right y
return y
def process(input, mask, anchors):
anchors = [anchors[i] for i in mask]
grid_h, grid_w = map(int, input.shape[0:2])
box_confidence = input[..., 4]
box_confidence = np.expand_dims(box_confidence, axis=-1)
box_class_probs = input[..., 5:]
box_xy = input[..., :2]*2 - 0.5
col = np.tile(np.arange(0, grid_w), grid_w).reshape(-1, grid_w)
row = np.tile(np.arange(0, grid_h).reshape(-1, 1), grid_h)
col = col.reshape(grid_h, grid_w, 1, 1).repeat(3, axis=-2)
row = row.reshape(grid_h, grid_w, 1, 1).repeat(3, axis=-2)
grid = np.concatenate((col, row), axis=-1)
box_xy += grid
box_xy *= int(IMG_SIZE/grid_h)
box_wh = pow(input[..., 2:4]*2, 2)
box_wh = box_wh * anchors
box = np.concatenate((box_xy, box_wh), axis=-1)
return box, box_confidence, box_class_probs
def filter_boxes(boxes, box_confidences, box_class_probs):
"""Filter boxes with box threshold. It's a bit different with origin yolov5 post process!
# Arguments
boxes: ndarray, boxes of objects.
box_confidences: ndarray, confidences of objects.
box_class_probs: ndarray, class_probs of objects.
# Returns
boxes: ndarray, filtered boxes.
classes: ndarray, classes for boxes.
scores: ndarray, scores for boxes.
"""
boxes = boxes.reshape(-1, 4)
box_confidences = box_confidences.reshape(-1)
box_class_probs = box_class_probs.reshape(-1, box_class_probs.shape[-1])
_box_pos = np.where(box_confidences >= OBJ_THRESH)
boxes = boxes[_box_pos]
box_confidences = box_confidences[_box_pos]
box_class_probs = box_class_probs[_box_pos]
class_max_score = np.max(box_class_probs, axis=-1)
classes = np.argmax(box_class_probs, axis=-1)
_class_pos = np.where(class_max_score >= OBJ_THRESH)
boxes = boxes[_class_pos]
classes = classes[_class_pos]
scores = (class_max_score* box_confidences)[_class_pos]
return boxes, classes, scores
def nms_boxes(boxes, scores):
"""Suppress non-maximal boxes.
# Arguments
boxes: ndarray, boxes of objects.
scores: ndarray, scores of objects.
# Returns
keep: ndarray, index of effective boxes.
"""
x = boxes[:, 0]
y = boxes[:, 1]
w = boxes[:, 2] - boxes[:, 0]
h = boxes[:, 3] - boxes[:, 1]
areas = w * h
order = scores.argsort()[::-1]
keep = []
while order.size > 0:
i = order[0]
keep.append(i)
xx1 = np.maximum(x[i], x[order[1:]])
yy1 = np.maximum(y[i], y[order[1:]])
xx2 = np.minimum(x[i] + w[i], x[order[1:]] + w[order[1:]])
yy2 = np.minimum(y[i] + h[i], y[order[1:]] + h[order[1:]])
w1 = np.maximum(0.0, xx2 - xx1 + 0.00001)
h1 = np.maximum(0.0, yy2 - yy1 + 0.00001)
inter = w1 * h1
ovr = inter / (areas[i] + areas[order[1:]] - inter)
inds = np.where(ovr <= NMS_THRESH)[0]
order = order[inds + 1]
keep = np.array(keep)
return keep
def yolov5_post_process(input_data):
masks = [[0, 1, 2], [3, 4, 5], [6, 7, 8]]
anchors = [
[4.5273438, 6.5234375], [13.1953125, 6.9804688], [17.390625, 12.125],
[27.40625, 14.640625], [26.84375, 20.375], [45.9375, 23.203125],
[49.46875, 37.5], [88.5, 61.125], [115.1875, 91.375]
]
boxes, classes, scores = [], [], []
for input, mask in zip(input_data, masks):
b, c, s = process(input, mask, anchors)
b, c, s = filter_boxes(b, c, s)
boxes.append(b)
classes.append(c)
scores.append(s)
boxes = np.concatenate(boxes)
boxes = xywh2xyxy(boxes)
classes = np.concatenate(classes)
scores = np.concatenate(scores)
nboxes, nclasses, nscores = [], [], []
for c in set(classes):
inds = np.where(classes == c)
b = boxes[inds]
c = classes[inds]
s = scores[inds]
keep = nms_boxes(b, s)
nboxes.append(b[keep])
nclasses.append(c[keep])
nscores.append(s[keep])
if not nclasses and not nscores:
return None, None, None
boxes = np.concatenate(nboxes)
classes = np.concatenate(nclasses)
scores = np.concatenate(nscores)
return boxes, classes, scores
def draw(image, boxes, scores, classes):
"""Draw the boxes on the image.
# Argument:
image: original image.
boxes: ndarray, boxes of objects.
classes: ndarray, classes of objects.
scores: ndarray, scores of objects.
all_classes: all classes name.
"""
print("{:^12} {:^12} {}".format('class', 'score', 'xmin, ymin, xmax, ymax'))
print('-' * 50)
for box, score, cl in zip(boxes, scores, classes):
top, left, right, bottom = box
top = int(top)
left = int(left)
right = int(right)
bottom = int(bottom)
cv2.rectangle(image, (top, left), (right, bottom), (255, 0, 0), 2)
cv2.putText(image, '{0} {1:.2f}'.format(CLASSES[cl], score),
(top, left - 6),
cv2.FONT_HERSHEY_SIMPLEX,
0.6, (0, 0, 255), 2)
print("{:^12} {:^12.3f} [{:>4}, {:>4}, {:>4}, {:>4}]".format(CLASSES[cl], score, top, left, right, bottom))
# Load the YOLOv5 RKNN model
rknn_lite = RKNNLite()
# Load the RKNN model
ret = rknn_lite.load_rknn(RKNN_MODEL)
if ret != 0:
print('Load RKNN model failed!')
exit(ret)
# Initialize the RKNN runtime environment
ret = rknn_lite.init_runtime()
if ret != 0:
print('Init runtime environment failed!')
exit(ret)
# Load and preprocess the image
img = cv2.imread(IMG_PATH)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# Do not normalize input image, use int8 format
img_resized = cv2.resize(img, (IMG_SIZE, IMG_SIZE))
img_input = img_resized.astype(np.int8)
# Add a batch dimension and convert to NHWC format (necessary for RKNN input)
img_input = np.expand_dims(img_input, 0)
outputs = rknn_lite.inference(inputs=[img_input])
input0_data = outputs[0]
input1_data = outputs[1]
input2_data = outputs[2]
input0_data = input0_data.reshape([3, -1]+list(input0_data.shape[-2:]))
input1_data = input1_data.reshape([3, -1]+list(input1_data.shape[-2:]))
input2_data = input2_data.reshape([3, -1]+list(input2_data.shape[-2:]))
input_data = list()
input_data.append(np.transpose(input0_data, (2, 3, 0, 1)))
input_data.append(np.transpose(input1_data, (2, 3, 0, 1)))
input_data.append(np.transpose(input2_data, (2, 3, 0, 1)))
boxes, classes, scores = yolov5_post_process(input_data)
img_1 = cv2.cvtColor(img_resized, cv2.COLOR_RGB2BGR)
if boxes is not None:
draw(img_1, boxes, scores, classes)
cv2.imwrite(RESULT_PATH, img_1)
# De-initialize and release resources
rknn_lite.release()