def detect(image_path, model_path, yolo_weights=None):
"""
Introduction
------------
加载模型，进行预测
Parameters
----------
model_path: 模型路径
image_path: 图片路径
"""
image = Image.open(image_path)
resize_image = letterbox_image(image, (416, 416))
image_data = np.array(resize_image, dtype=np.float32)
image_data /= 255.
image_data = np.expand_dims(image_data, axis=0)
pb_graph = tf.Graph()
with pb_graph.as_default():
input_image_shape = tf.placeholder(dtype=tf.int32,
shape=(2, ),
name="pred_im_shape")
input_image = tf.placeholder(shape=[None, 416, 416, 3],
dtype=tf.float32,
name='pred_input_img')
predictor = yolo_predictor(config.obj_threshold,
config.nms_iou_threshold,
config.classes_path, config.anchors_path)
boxes, scores, classes = predictor.predict(input_image,
input_image_shape)
print(input_image_shape)
print(input_image)
print(boxes)
print(scores)
print(classes)
with tf.Session(graph=pb_graph) as sess:
saver = tf.train.Saver()
saver.restore(sess, model_path)
out_boxes, out_scores, out_classes = sess.run(
[boxes, scores, classes],
feed_dict={
input_image: image_data,
input_image_shape: [image.size[1], image.size[0]]
})
graph_def = tf.get_default_graph().as_graph_def()
out_put_name_list = [
'predict/pred_boxes', 'predict/pred_scores', 'predict/pred_classes'
]
out_put_grah_def = tf.graph_util.convert_variables_to_constants(
sess, graph_def, out_put_name_list)
pb_file_path = 'F:\\github_working\\version_2_190114\\alsochen-tensorflow-yolo3-threeoutput\\tensorflow-yolo3\\pb_file\\model.pb'
with tf.gfile.GFile(pb_file_path, 'wb') as f:
f.write(out_put_grah_def.SerializeToString())
print("pb save done")
print('Found {} boxes for {}'.format(len(out_boxes), 'img'))
def eval_neg(model_path, neg_path, yolo_weights=None):
"""
Introduction
------------
计算模型在negtive datasets验证集上的MAP, 用于评价模型
"""
input_image_shape = tf.placeholder(dtype=tf.int32, shape=(2, ))
input_image = tf.placeholder(shape=[None, 416, 416, 3], dtype=tf.float32)
predictor = yolo_predictor(config.obj_threshold, config.nms_threshold,
config.classes_path, config.anchors_path)
boxes, scores, classes = predictor.predict(input_image, input_image_shape)
image_files = os.listdir(neg_path)
tp_nums = 0
all_nums = len(image_files)
with tf.Session() as sess:
if yolo_weights is not None:
with tf.variable_scope('predict'):
boxes, scores, classes = predictor.predict(
input_image, input_image_shape)
load_op = load_weights(tf.global_variables(scope='predict'),
weights_file=yolo_weights)
sess.run(load_op)
else:
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(model_path)
saver.restore(sess, ckpt.model_checkpoint_path)
# saver.restore(sess, model_path)
for image_file in image_files:
image = Image.open(neg_path + image_file)
resize_image = letterbox_image(image, (416, 416))
image_data = np.array(resize_image, dtype=np.float32)
image_data /= 255.
image_data = np.expand_dims(image_data, axis=0)
out_boxes, out_scores, out_classes = sess.run(
[boxes, scores, classes],
feed_dict={
input_image: image_data,
input_image_shape: [image.size[1], image.size[0]]
})
print(image_file)
print(out_classes)
print(out_scores)
if len(out_classes) == 0:
tp_nums += 1
print(tp_nums / all_nums)
def detect(image_path, model_path, yolo_weights=None):
"""
Introduction
------------
加载模型，进行预测
Parameters
----------
model_path: 模型路径
image_path: 图片路径
"""
image = Image.open(image_path)
resize_image = letterbox_image(image, (192, 192))
image_data = np.array(resize_image, dtype=np.float32)
image_data /= 255.
image_data = np.expand_dims(image_data, axis=0)
input_image_shape = tf.placeholder(dtype=tf.int32, shape=(2, ))
input_image = tf.placeholder(shape=[None, 192, 192, 3], dtype=tf.float32)
with tf.variable_scope("model_gd"):
predictor = yolo_predictor(config.obj_threshold, config.nms_threshold,
config.classes_path, config.anchors_path2)
boxes, scores, classes = predictor.predict(input_image,
input_image_shape)
with tf.Session() as sess:
if yolo_weights is not None:
with tf.variable_scope('predict'):
boxes, scores, classes = predictor.predict(
input_image, input_image_shape)
load_op = load_weights(tf.global_variables(scope='predict'),
weights_file=yolo_weights)
sess.run(load_op)
else:
saver = tf.train.Saver()
model_file = tf.train.latest_checkpoint(model_path)
saver.restore(sess, model_file)
out_boxes, out_scores, out_classes = sess.run(
[boxes, scores, classes],
feed_dict={
input_image: image_data,
input_image_shape: [image.size[1], image.size[0]]
})
print('Found {} boxes for {}'.format(len(out_boxes), 'img'))
font = ImageFont.truetype(font='font/FiraMono-Medium.otf',
size=np.floor(3e-2 * image.size[1] +
0.5).astype('int32'))
thickness = (image.size[0] + image.size[1]) // 300
for i, c in reversed(list(enumerate(out_classes))):
predicted_class = predictor.class_names[c]
box = out_boxes[i]
score = out_scores[i]
label = '{} {:.2f}'.format(predicted_class, score)
draw = ImageDraw.Draw(image)
label_size = draw.textsize(label, font)
top, left, bottom, right = box
top = max(0, np.floor(top + 0.5).astype('int32'))
left = max(0, np.floor(left + 0.5).astype('int32'))
bottom = min(image.size[1], np.floor(bottom + 0.5).astype('int32'))
right = min(image.size[0], np.floor(right + 0.5).astype('int32'))
data = str(label) + "," + str(left) + "." + str(top) + "," + str(
right) + "," + str(bottom) + "\n"
with open('./res/data.txt', "a") as f:
f.write(data)
print(label, (left, top), (right, bottom))
if top - label_size[1] >= 0:
text_origin = np.array([left, top - label_size[1]])
else:
text_origin = np.array([left, top + 1])
# My kingdom for a good redistributable image drawing library.
for i in range(thickness):
draw.rectangle([left + i, top + i, right - i, bottom - i],
outline=predictor.colors[c])
draw.rectangle(
[tuple(text_origin),
tuple(text_origin + label_size)],
fill=predictor.colors[c])
draw.text(text_origin, label, fill=(0, 0, 0), font=font)
draw.text(text_origin, label, fill=(0, 0, 0), font=font)
del draw
image.show()
image.save('./res/1.jpg')
def detect(image_path, model_path, yolo_weights=None):
"""
Introduction
------------
加载模型，进行预测
Parameters
----------
model_path: 模型路径
image_path: 图片路径
"""
image = Image.open(image_path)
resize_image = letterbox_image(image,
(416, 416)) #resize 和padding图像 与训练的时候保持一致
image_data = np.array(resize_image, dtype=np.float32)
image_data /= 255.
image_data = np.expand_dims(image_data, axis=0) #扩展为[1 416 416 3]用于输入网络
input_image_shape = tf.placeholder(dtype=tf.int32, shape=(2, ))
input_image = tf.placeholder(shape=[None, 416, 416, 3], dtype=tf.float32)
predictor = yolo_predictor(config.obj_threshold, config.nms_threshold,
config.classes_path, config.anchors_path)
boxes, scores, classes = predictor.predict(input_image, input_image_shape)
with tf.Session() as sess:
if yolo_weights is not None:
with tf.variable_scope('predict'):
boxes, scores, classes = predictor.predict(
input_image, input_image_shape)
load_op = load_weights(tf.global_variables(scope='predict'),
weights_file=yolo_weights)
sess.run(load_op)
else:
saver = tf.train.Saver()
saver.restore(sess, model_path)
#运行检测
out_boxes, out_scores, out_classes = sess.run(
[boxes, scores, classes],
feed_dict={
input_image: image_data,
input_image_shape: [image.size[1], image.size[0]]
})
print('Found {} boxes for {}'.format(len(out_boxes), 'img'))
font = ImageFont.truetype(font='font/FiraMono-Medium.otf',
size=np.floor(3e-2 * image.size[1] +
0.5).astype('int32'))
thickness = (image.size[0] + image.size[1]) // 300
for i, c in reversed(list(enumerate(
out_classes))): #使用了reversed但是 枚举的索引与元素的对应关系没变，reversed意义何在？
predicted_class = predictor.class_names[c]
box = out_boxes[i]
score = out_scores[i]
label = '{} {:.2f}'.format(predicted_class, score)
draw = ImageDraw.Draw(image) #使用PIL里面的函数
label_size = draw.textsize(label, font)
top, left, bottom, right = box #这里解析 box的形式是 y1 x1 y2 x2
top = max(
0,
np.floor(top + 0.5).astype(
'int32')) #因为是预测给出的框的大小，有可能越界，判断一下边界还是必要的，加0.5是为了四舍五入，向上取整
left = max(0, np.floor(left + 0.5).astype('int32'))
bottom = min(image.size[1], np.floor(bottom + 0.5).astype('int32'))
right = min(image.size[0], np.floor(right + 0.5).astype('int32'))
print(label, (left, top), (right, bottom))
if top - label_size[1] >= 0:
text_origin = np.array([
left, top - label_size[1]
]) #标签的文本一般放在框的上面，如果框的上面的高度大于label_size[1]，则可以在上面画框，否则在下面画框
else:
text_origin = np.array([left, top + 1])
# My kingdom for a good redistributable image drawing library.
for i in range(thickness):
draw.rectangle([left + i, top + i, right - i, bottom - i],
outline=predictor.colors[c])
draw.rectangle(
[tuple(text_origin),
tuple(text_origin + label_size)],
fill=predictor.colors[c])
draw.text(text_origin, label, fill=(0, 0, 0), font=font)
del draw
image.show()
image.save('./test/result1.jpg')
def eval(model_path, min_Iou=0.5, yolo_weights=None):
"""
Introduction
------------
计算模型在coco验证集上的MAP, 用于评价模型
"""
ground_truth = {}
class_pred = defaultdict(list)
gt_counter_per_class = defaultdict(int)
input_image_shape = tf.placeholder(dtype=tf.int32, shape=(2, ))
input_image = tf.placeholder(shape=[None, 416, 416, 3], dtype=tf.float32)
predictor = yolo_predictor(config.obj_threshold, config.nms_threshold,
config.classes_path, config.anchors_path)
boxes, scores, classes = predictor.predict(input_image, input_image_shape)
val_Reader = Reader("val",
config.data_dir,
config.anchors_path,
config.num_classes,
input_shape=config.input_shape,
max_boxes=config.max_boxes)
image_files, bboxes_data = val_Reader.read_annotations()
allBBox = 0
with tf.Session() as sess:
if yolo_weights is not None:
with tf.variable_scope('predict'):
boxes, scores, classes = predictor.predict(
input_image, input_image_shape)
load_op = load_weights(tf.global_variables(scope='predict'),
weights_file=yolo_weights)
sess.run(load_op)
else:
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(model_path)
#saver.restore(sess, model_path)
saver.restore(sess, ckpt.model_checkpoint_path)
for index in range(len(image_files)):
val_bboxes = []
image_file = image_files[index]
file_id = os.path.split(image_file)[-1].split('.')[0]
for bbox in bboxes_data[index]:
left, top, right, bottom, class_id = bbox[0], bbox[1], bbox[
2], bbox[3], bbox[4]
class_name = val_Reader.class_names[int(class_id)]
bbox = [float(left), float(top), float(right), float(bottom)]
val_bboxes.append({
"class_name": class_name,
"bbox": bbox,
"used": False
})
gt_counter_per_class[class_name] += 1
ground_truth[file_id] = val_bboxes
image = Image.open(image_file)
resize_image = letterbox_image(image, (416, 416))
image_data = np.array(resize_image, dtype=np.float32)
image_data /= 255.
image_data = np.expand_dims(image_data, axis=0)
out_boxes, out_scores, out_classes = sess.run(
[boxes, scores, classes],
feed_dict={
input_image: image_data,
input_image_shape: [image.size[1], image.size[0]]
})
allBBox += len(out_boxes)
print("detect {}/{} found boxes: {},allBBox:{}".format(
index, len(image_files), len(out_boxes), allBBox))
for o, c in enumerate(out_classes):
predicted_class = val_Reader.class_names[c]
box = out_boxes[o]
score = out_scores[o]
top, left, bottom, right = box
top = max(0, np.floor(top + 0.5).astype('int32'))
left = max(0, np.floor(left + 0.5).astype('int32'))
bottom = min(image.size[1],
np.floor(bottom + 0.5).astype('int32'))
right = min(image.size[0],
np.floor(right + 0.5).astype('int32'))
bbox = [left, top, right, bottom]
class_pred[predicted_class].append({
"confidence": str(score),
"file_id": file_id,
"bbox": bbox
})
# 计算每个类别的AP
sum_AP = 0.0
sum_rec = 0.0
sum_prec = 0.0
count_true_positives = {}
for class_index, class_name in enumerate(
sorted(gt_counter_per_class.keys())):
count_true_positives[class_name] = 0
predictions_data = class_pred[class_name]
# 该类别总共有多少个box
nd = len(predictions_data)
tp = [0] * nd # true positive
fp = [0] * nd # false positive
for idx, prediction in enumerate(predictions_data):
file_id = prediction['file_id']
ground_truth_data = ground_truth[file_id]
bbox_pred = prediction['bbox']
Iou_max = -1
gt_match = None
for obj in ground_truth_data:
if obj['class_name'] == class_name:
bbox_gt = obj['bbox']
bbox_intersect = [
max(bbox_pred[0], bbox_gt[0]),
max(bbox_gt[1], bbox_pred[1]),
min(bbox_gt[2], bbox_pred[2]),
min(bbox_gt[3], bbox_pred[3])
]
intersect_weight = bbox_intersect[2] - bbox_intersect[0] + 1
intersect_high = bbox_intersect[3] - bbox_intersect[1] + 1
if intersect_high > 0 and intersect_weight > 0:
union_area = (bbox_pred[2] - bbox_pred[0] + 1) * (
bbox_pred[3] - bbox_pred[1] +
1) + (bbox_gt[2] - bbox_gt[0] +
1) * (bbox_gt[3] - bbox_gt[1] +
1) - intersect_weight * intersect_high
Iou = intersect_high * intersect_weight / union_area
if Iou > Iou_max:
Iou_max = Iou
gt_match = obj
if Iou_max > min_Iou:
if not gt_match['used'] and gt_match is not None:
tp[idx] = 1
gt_match['used'] = True
else:
fp[idx] = 1
else:
fp[idx] = 1
# 计算精度和召回率
sum_class = 0
for idx, val in enumerate(fp):
fp[idx] += sum_class
sum_class += val
sum_class = 0
for idx, val in enumerate(tp):
tp[idx] += sum_class
sum_class += val
rec = tp[:]
for idx, val in enumerate(tp):
rec[idx] = tp[idx] / gt_counter_per_class[class_name]
prec = tp[:]
for idx, val in enumerate(tp):
prec[idx] = tp[idx] / (fp[idx] + tp[idx])
ap, mrec, mprec = voc_ap(rec, prec)
sum_AP += ap
sum_rec += (mrec[-2])
sum_prec += sum(mprec) / (allBBox + 2)
f1 = 2 * sum_rec * sum_prec / (sum_rec + sum_prec)
MAP = sum_AP / len(gt_counter_per_class) * 100
#rec = sum_rec / len(gt_counter_per_class) * 100
#prec = sum_prec / len(gt_counter_per_class) * 100
print("The Model Eval MAP: {},prec:{},rec:{},f1:{}".format(
MAP, sum_prec, sum_rec, f1))
def detect(image, yolo_weights = config.yolo3_weights_path,image_size=(416,416)):
"""
Introduction
------------
加载模型，进行预测
Parameters
----------
model_path: 模型路径
image_path: 图片路径
"""
image = Image.open(image)
if image_size != (None, None):
assert image_size[0] % 32 == 0, 'Multiples of 32 required'
assert image_size[1] % 32 == 0, 'Multiples of 32 required'
resize_image = letterbox_image(image, tuple(reversed(image_size)))
else:
new_image_size = (image.width - (image.width % 32),
image.height - (image.height % 32))
resize_image = letterbox_image(image, new_image_size)
image_data = np.array(resize_image, dtype = 'float32')
image_data /= 255.
image_data = np.expand_dims(image_data, axis = 0)
print(image_data.shape)
input_image_shape = tf.placeholder(dtype = tf.int32, shape = (2,))
input_image = tf.placeholder(shape = [None, 416, 416, 3], dtype = tf.float32)
predictor = yolo_predictor(config.obj_threshold, config.nms_threshold, config.classes_path, config.anchors_path)
boxes, scores, classes = predictor.predict(input_image, input_image_shape)
with tf.Session() as sess:
if yolo_weights is not None:
print("yes")
with tf.variable_scope('predict'):
boxes, scores, classes = predictor.predict(input_image, input_image_shape)
load_op = load_weights(tf.global_variables(scope = 'predict'), weights_file = yolo_weights)
sess.run(load_op)
else:
saver = tf.train.Saver()
saver.restore(sess, config.yolo3_weights_path)
out_boxes, out_scores, out_classes = sess.run(
[boxes, scores, classes],
feed_dict={
input_image: image_data,
input_image_shape: [image.size[1], image.size[0]]
})
print('Found {} boxes for {}'.format(len(out_boxes), 'img'))
font = ImageFont.truetype(font = 'font/FiraMono-Medium.otf', size = np.floor(3e-2 * image.size[1] + 0.5).astype('int32'))
thickness = (image.size[0] + image.size[1]) // 300
for i, c in reversed(list(enumerate(out_classes))):
predicted_class = predictor.class_names[c]
box = out_boxes[i]
score = out_scores[i]
label = '{} {:.2f}'.format(predicted_class, score)
draw = ImageDraw.Draw(image)
label_size = draw.textsize(label, font)
top, left, bottom, right = box
top = max(0, np.floor(top + 0.5).astype('int32'))
left = max(0, np.floor(left + 0.5).astype('int32'))
bottom = min(image.size[1], np.floor(bottom + 0.5).astype('int32'))
right = min(image.size[0], np.floor(right + 0.5).astype('int32'))
print(label, (left, top), (right, bottom))
if top - label_size[1] >= 0:
text_origin = np.array([left, top - label_size[1]])
else:
text_origin = np.array([left, top + 1])
# My kingdom for a good redistributable image drawing library.
for i in range(thickness):
draw.rectangle(
[left + i, top + i, right - i, bottom - i],
outline = predictor.colors[c])
draw.rectangle(
[tuple(text_origin), tuple(text_origin + label_size)],
fill = predictor.colors[c])
draw.text(text_origin, label, fill=(0, 0, 0), font=font)
del draw
result = cv2.cvtColor(np.asarray(image), cv2.COLOR_RGB2BGR)
result = np.asarray(result)
cv2.imwrite("./output.png", result)
def detect(model_path, yolo_weights=None, image_path=None):
"""
Introduction
------------
加载模型，进行预测
Parameters
----------
model_path: 模型路径
image_path: 图片路径
"""
cap = None
if image_path == None:
cap = cv2.VideoCapture(0)
input_image_shape = tf.placeholder(dtype=tf.int32, shape=(2, ))
input_image = tf.placeholder(shape=[None, 416, 416, 3], dtype=tf.float32)
predictor = yolo_predictor(config.obj_threshold, config.nms_threshold,
config.classes_path, config.anchors_path)
boxes, scores, classes = predictor.predict(input_image, input_image_shape)
with tf.Session() as sess:
if yolo_weights is not None:
with tf.variable_scope('predict'):
boxes, scores, classes = predictor.predict(
input_image, input_image_shape)
load_op = load_weights(tf.global_variables(scope='predict'),
weights_file=yolo_weights)
sess.run(load_op)
else:
saver = tf.train.Saver()
saver.restore(
sess,
"./test_model/model.ckpt-192192/model.ckpt-44865") # emotion
# saver.restore(sess, "./test_model/model.ckpt-192192/model.ckpt-19940") # detection
while True:
start_time = time.time()
if image_path == None:
ret, image = cap.read()
if ret == 0:
break
[h, w] = image.shape[:2]
print(h, w)
image = cv2.flip(image, 1)
image_np = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = Image.fromarray(image_np)
else:
image = Image.open(image_path)
resize_image = letterbox_image(image, (416, 416))
image_data = np.array(resize_image, dtype=np.float32)
image_data /= 255.
image_data = np.expand_dims(image_data, axis=0)
out_boxes, out_scores, out_classes = sess.run(
[boxes, scores, classes],
feed_dict={
input_image: image_data,
input_image_shape: [image.size[1], image.size[0]]
})
print('Found {} boxes for {}'.format(len(out_boxes), 'img'))
font = ImageFont.truetype(font='font/FiraMono-Medium.otf',
size=np.floor(3e-2 * image.size[1] +
0.5).astype('int32'))
thickness = (image.size[0] + image.size[1]) // 300
for i, c in reversed(list(enumerate(out_classes))):
c = int(c[0])
print("i:{}, c:{}, type:{}".format(i, c, type(c)))
if c > 2:
continue
predicted_class = predictor.class_names[c]
box = out_boxes[i]
score = out_scores[i]
label = '{} {:.2f}'.format(predicted_class, score)
draw = ImageDraw.Draw(image)
label_size = draw.textsize(label, font)
top, left, bottom, right = box
top = max(0, np.floor(top + 0.5).astype('int32'))
left = max(0, np.floor(left + 0.5).astype('int32'))
bottom = min(image.size[1],
np.floor(bottom + 0.5).astype('int32'))
right = min(image.size[0],
np.floor(right + 0.5).astype('int32'))
print(label, (left, top), (right, bottom))
if top - label_size[1] >= 0:
text_origin = np.array([left, top - label_size[1]])
else:
text_origin = np.array([left, top + 1])
duration = time.time() - start_time
# My kingdom for a good redistributable image drawing library.
for i in range(thickness):
draw.rectangle([left + i, top + i, right - i, bottom - i],
outline=predictor.colors[c])
draw.rectangle(
[tuple(text_origin),
tuple(text_origin + label_size)],
fill=predictor.colors[c])
frame_rate = '{:.2f}'.format(1.0 / duration)
draw.text(text_origin, label, fill=(0, 0, 0), font=font)
draw.text(np.array([0, 0]),
frame_rate,
fill=(0, 0, 0),
font=font)
del draw
# image.show()
# image.save('./result1.jpg')
# cv_img = cv2.CreateImageHeader(image.size, cv2.IPL_DEPTH_8U, 3) # RGB image
# cv2.SetData(cv_img, image.tostring(), image.size[0]*3)
if image_path != None:
print('just one image')
image.show()
image.save('./result1.jpg')
break
else:
open_cv_image = np.array(image)[:, :, ::-1].copy()
cv2.imshow('cimage', open_cv_image)
k = cv2.waitKey(1) & 0xff
if k == ord('q') or k == 27:
break
def detect(image_path, model_path, yolo_weights = None):
image = Image.open(image_path)
resize_image = letterbox_image(image,(416,416))
image_data = np.array(resize_image, dtype=np.float32)
#皈一
image_data/=255
image_data = np.expand_dims(image_data,axis=0)
input_image_shape = tf.placeholder(dtype=tf.int32, shape=(2, ))
#圖像
input_image = tf.placeholder(shape = [None,416,416,3], dtype=tf.float32)
predictor = yolo_predictor(config.obj_threshold,config.nms_threshold, config.classes_path, config.anchors_path)
with tf.Session() as sess:
if yolo_weights is not None:
with tf.variable_scope('predict'):
boxes, scores, classes = predictor.predict(input_image, input_image_shape)
#載入模型
load_op = load_weights(tf.global_variables(scores="predict"), weights_file=yolo_weights)
sess.run(load_op)
#進行預測
out_boxes, out_scores, out_classes = sess.run(
[boxes, scores,classes],
feed_dict={
input_image:image_data,
input_image_shape:[image.size[1],image.size[0]]
}
)
else:
boxes, out_scores, out_classes = predictor.predict(input_image, input_image_shape)
saver = tf.train.Saver()
saver.restore(sess,model_path)
out_boxes, out_scores, out_classes = sess.run(
[boxes, scores, classes],
feed_dict={
input_image:image_data,
input_image_shape:[image.size[1], image.size[0]]
}
)
print('Found {} boxes for {}'.format(len(out_boxes),'img'))
font = ImageFont.truetype(font='font/FiraMono-Medium.otf',size=np.floor(3e-2 * image.size[1] + 0.5).astype('int32'))
thickness = (image.size[0] + image.size[1]) // 300
for i, c in reversed(list(enumerate(out_classes))):
predicted_class = predictor.class_names[c]
box = out_boxes[i]
score = out_scores[i]
#打印
label = '{} {:.2f}'.format(predicted_class, score)
draw = ImageDraw.Draw(image)
label_size = draw.textsize(label,font)
top, left, bottom, right = box
top = max(0,np.floor(top+0.5).astype('int32'))
left = max(0,np.floor(left+0.5).astype('int32'))
bottom = min(image.size[1]-1,np.floor(bottom+0.5).astype('int32'))
right = min(image.size[0]-1,np.floor(right+0.5).astype('int32'))
print(label, (left,top),(right,bottom))
print(label_size)
if top - label_size[1] >= 0:
text_origin = np.array(left,top - label_size[1])
else:
text_origin = np.array([left, top+1])
for i in range(thickness):
draw.rectangle(
[left+i,top+i,right-i,bottom-i],
outline=predictor.colors[c]
)
draw.rectangle(
[tuple(text_origin), tuple(text_origin + label_size)],
fill=predictor.colors[c]
)
draw.text(text_origin, label,fill=(0,0,0), font=font)
del draw
image.show()
image.save('./img/result1111.jpg')
def detect(image_path, model_path, yolo_weights=None):
"""
Introduction
------------
加载模型，进行预测
Parameters
----------
model_path: 模型路径
image_path: 图片路径
"""
image = Image.open(image_path)
resize_image = letterbox_image(image, (416, 416))
image_data = np.array(resize_image, dtype=np.float32)
image_data /= 255.
image_data = np.expand_dims(image_data, axis=0)
input_image_shape = tf.placeholder(dtype=tf.int32, shape=(2, ))
input_image = tf.placeholder(shape=[None, 416, 416, 3], dtype=tf.float32)
predictor = yolo_predictor(config.obj_threshold, config.nms_iou_threshold,
config.classes_path, config.anchors_path)
boxes, scores, classes = predictor.predict(input_image, input_image_shape)
with tf.Session() as sess:
if yolo_weights is not None:
with tf.variable_scope('predict'):
boxes, scores, classes = predictor.predict(
input_image, input_image_shape)
load_op = load_weights(tf.global_variables(scope='predict'),
weights_file=yolo_weights)
sess.run(load_op)
else:
saver = tf.train.Saver()
saver.restore(sess, model_path)
out_boxes, out_scores, out_classes = sess.run(
[boxes, scores, classes],
feed_dict={
input_image: image_data,
input_image_shape: [image.size[1], image.size[0]]
})
print([out_boxes, out_scores, out_classes])
print('Found {} boxes for {}'.format(len(out_boxes), 'img'))
font = ImageFont.truetype(
font=
'F:\\github_working\\version_2_190114\\aloyschen-tensorflow-yolo3\\tensorflow-yolo3\\font\\FiraMono-Medium.otf',
size=np.floor(3e-2 * image.size[1] + 0.5).astype('int32'))
thickness = (image.size[0] + image.size[1]) // 300
for i, c in reversed(list(enumerate(out_classes))):
predicted_class = predictor.class_names[c]
box = out_boxes[i]
score = out_scores[i]
label = '{} {:.2f}'.format(predicted_class, score)
draw = ImageDraw.Draw(image)
label_size = draw.textsize(label, font)
top, left, bottom, right = box
top = max(0, np.floor(top + 0.5).astype('int32'))
left = max(0, np.floor(left + 0.5).astype('int32'))
bottom = min(image.size[1], np.floor(bottom + 0.5).astype('int32'))
right = min(image.size[0], np.floor(right + 0.5).astype('int32'))
print(label, (left, top), (right, bottom))
if top - label_size[1] >= 0:
text_origin = np.array([left, top - label_size[1]])
else:
text_origin = np.array([left, top + 1])
# My kingdom for a good redistributable image drawing library.
for i in range(thickness):
draw.rectangle([left + i, top + i, right - i, bottom - i],
outline=predictor.colors[c])
draw.rectangle(
[tuple(text_origin),
tuple(text_origin + label_size)],
fill=predictor.colors[c])
draw.text(text_origin, label, fill=(0, 0, 0), font=font)
del draw
image.show()
print("done")
img_save_path = "F:\\github_working\\version_2_190114\\aloyschen-tensorflow-yolo3\\single_img_out"
img_name = "valid_" + image_path.split("\\")[-1]
final_save_path = os.path.join(img_save_path, img_name)
image.save(final_save_path)
print("image save done")
def detect(image_path, model_path, yolo_weights=None):
start_time = time.time()
for image_id in name_space[0:5]:
reset_time = time.time()
tf.reset_default_graph()
reset_end = time.time()
st_image_prep = time.time()
image = Image.open(image_path + image_id) #variable was: image
resize_image = letterbox_image(image,
(416, 416)) # should be (320, 240)
image_data = np.array(resize_image, dtype=np.float32)
image_data /= 255.
image_data = np.expand_dims(image_data, axis=0)
input_image_shape = tf.placeholder(dtype=tf.int32, shape=(2, ))
input_image = tf.placeholder(shape=[None, 416, 416, 3],
dtype=tf.float32)
predictor = yolo_predictor(config.obj_threshold, config.nms_threshold,
config.classes_path, config.anchors_path)
# boxes, scores, classes = predictor.predict(input_image, input_image_shape)
en_image_prep = time.time()
with tf.Session() as sess:
st1_sess = time.time()
if yolo_weights is not None:
with tf.variable_scope('predict'):
st1_end = time.time()
st2_sess = time.time()
boxes, scores, classes = predictor.predict(
input_image, input_image_shape)
st2_end = time.time()
st3_sess = time.time()
load_op = load_weights(
tf.global_variables(scope='predict'),
weights_file=yolo_weights)
st3_end = time.time()
st4_sess = time.time()
sess.run(load_op)
st4_end = time.time()
end = time.time()
print(
'Image preprocessing takes: {}'.format(en_image_prep -
st_image_prep))
print(
'tf.reset_default_graph takes: {}'.format(reset_end -
reset_time))
print(
'if not None statement an with... time is: {}'.format(
st1_end - st1_sess))
print(
'boxes, scores, classes time is: {}'.format(st2_end -
st2_sess))
print(
'load_op = load_weights(tf.global time is: {}'.format(
st3_end - st3_sess))
print('sess.run(load_op) time is: {}'.format(st4_end -
st4_sess))
else:
print('session not running')
saver = tf.train.Saver()
saver.restore(sess, model_path)
out_boxes, out_scores, out_classes = sess.run(
[boxes, scores, classes],
feed_dict={
input_image: image_data,
input_image_shape: [image.size[1], image.size[0]]
})
print('Found {} boxes for {}'.format(len(out_boxes), 'image: ',
image_id))
font = ImageFont.truetype(font='font/FiraMono-Medium.otf',
size=np.floor(3e-2 * image.size[1] +
0.5).astype('int32'))
thickness = (image.size[0] + image.size[1]) // 300
for i, c in reversed(list(enumerate(out_classes))):
predicted_class = predictor.class_names[c]
box = out_boxes[i]
score = out_scores[i]
label = '{} {:.2f}'.format(predicted_class, score)
print(label)
print("--- %s seconds ---" % (time.time() - start_time))
draw = ImageDraw.Draw(image)
label_size = draw.textsize(label, font)
top, left, bottom, right = box
top = max(0, np.floor(top + 0.5).astype('int32'))
left = max(0, np.floor(left + 0.5).astype('int32'))
bottom = min(image.size[1],
np.floor(bottom + 0.5).astype('int32'))
right = min(image.size[0],
np.floor(right + 0.5).astype('int32'))
print(label, (left, top), (right, bottom))
if top - label_size[1] >= 0:
text_origin = np.array([left, top - label_size[1]])
else:
text_origin = np.array([left, top + 1])
for i in range(thickness):
draw.rectangle([left + i, top + i, right - i, bottom - i],
outline=predictor.colors[c])
draw.rectangle(
[tuple(text_origin),
tuple(text_origin + label_size)],
fill=predictor.colors[c])
draw.text(text_origin, label, fill=(0, 0, 0), font=font)
del draw
image.show()
image.save('./result1.jpg')