def evaluate_acc(net, data_iter, ctx):
data_iter.reset()
box_metric = metric.MAE()
outs, labels = None, None
for i, batch in enumerate(data_iter):
data = batch.data[0].as_in_context(ctx)
label = batch.label[0].as_in_context(ctx)
# print('acc',label.shape)
anchors, box_preds, cls_preds = net(data)
#MultiBoxTraget 作用是将生成的anchors与哪些ground truth对应，提取出anchors的偏移和对应的类型
#预测的误差是每次网络输出的预测框g与anchors的差分别/anchor[xywh]，然后作为smoothL1（label-g）解算，g才是预测
# 正负样本比例1：3
box_offset, box_mask, cls_labels = MultiBoxTarget(
anchors,
label,
cls_preds.transpose((0, 2, 1)),
negative_mining_ratio=3.0)
box_metric.update([box_offset], [box_preds * box_mask])
cls_probs = nd.SoftmaxActivation(cls_preds.transpose((0, 2, 1)),
mode='channel')
#对输出的bbox通过NMS极大值抑制算法筛选检测框
out = MultiBoxDetection(cls_probs,
box_preds,
anchors,
force_suppress=True,
clip=False,
nms_threshold=0.45)
if outs is None:
outs = out
labels = label
else:
outs = nd.concat(outs, out, dim=0)
labels = nd.concat(labels, label, dim=0)
AP = evaluate_MAP(outs, labels)
return AP, box_metric
def detect_image(img_path):
if not os.path.exists(img_path):
print('can not find image: ', img_path)
# img = Image.open(img_file)
#print img_path
img = cv2.imread(img_path)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = cv2.resize(img, (cfg.img_size, cfg.img_size))
# img = ImageOps.fit(img, [data_shape, data_shape], Image.ANTIALIAS)
origin_img = img.copy()
img = (img / 255. - cfg.mean) / cfg.std
img = np.transpose(img, (2, 0, 1))
img = img[np.newaxis, :]
img = F.array(img)
print('input image shape: ', img.shape)
ctx = mx.gpu(0)
net = build_ssd("test", 300, ctx)
net.initialize(mx.init.Xavier(magnitude=2), ctx=ctx)
net.collect_params().reset_ctx(ctx)
params = 'model/ssd.params'
net.load_params(params, ctx=ctx)
anchors, cls_preds, box_preds = net(img.as_in_context(ctx))
print('anchors', anchors)
print('class predictions', cls_preds)
print('box delta predictions', box_preds)
# convert predictions to probabilities using softmax
cls_probs = F.SoftmaxActivation(F.transpose(cls_preds, (0, 2, 1)),
mode='channel')
# apply shifts to anchors boxes, non-maximum-suppression, etc...
output = MultiBoxDetection(*[cls_probs, box_preds, anchors],
force_suppress=True,
clip=True,
nms_threshold=0.01)
output = output.asnumpy()
pens = dict()
plt.imshow(origin_img)
thresh = 0.3
for det in output[0]:
cid = int(det[0])
if cid < 0:
continue
score = det[1]
if score < thresh:
continue
if cid not in pens:
pens[cid] = (random.random(), random.random(), random.random())
scales = [origin_img.shape[1], origin_img.shape[0]] * 2
xmin, ymin, xmax, ymax = [
int(p * s) for p, s in zip(det[2:6].tolist(), scales)
]
rect = plt.Rectangle((xmin, ymin),
xmax - xmin,
ymax - ymin,
fill=False,
edgecolor=pens[cid],
linewidth=3)
plt.gca().add_patch(rect)
voc_class_name = [
'person', 'bird', 'cat', 'cow', 'dog', 'horse', 'sheep',
'aeroplane', 'bicycle', 'boat', 'bus', 'car', 'motorbike', 'train',
'bottle', 'chair', 'diningtable', 'pottedplant', 'sofa',
'tvmonitor'
]
text = voc_class_name[cid]
plt.gca().text(xmin,
ymin - 2,
'{:s} {:.3f}'.format(text, score),
bbox=dict(facecolor=pens[cid], alpha=0.5),
fontsize=12,
color='white')
plt.axis('off')
# plt.savefig('result.png', dpi=100)
plt.show()
def predict(x):
""" 预测函数会输出所有边框，每个边框由[class_id, confidence, xmin, ymin, xmax, ymax] """
anchors, cls_preds, box_preds = net(x.as_in_context(ctx))
cls_probs = nd.SoftmaxActivation(cls_preds.transpose(0,2,1), mode='channel')
return MultiBoxDetection(cls_probs, box_preds, anchors, force_suppress=True, clip=False)
def mytrain(net,
train_data,
valid_data,
ctx,
start_epoch,
end_epoch,
cls_loss,
box_loss,
trainer=None):
if trainer is None:
# trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': 0.01,'momentum':0.9, 'wd':5e-1})
trainer = gluon.Trainer(net.collect_params(), 'sgd', {
'learning_rate': 0.1,
'wd': 1e-3
})
box_metric = metric.MAE()
for e in range(start_epoch, end_epoch):
# print(e)
train_data.reset()
valid_data.reset()
box_metric.reset()
tic = time.time()
_loss = [0, 0]
if e == 100 or e == 120 or e == 150 or e == 180 or e == 200:
trainer.set_learning_rate(trainer.learning_rate * 0.2)
outs, labels = None, None
for i, batch in enumerate(train_data):
data = batch.data[0].as_in_context(ctx)
label = batch.label[0].as_in_context(ctx)
# print(label.shape)
with autograd.record():
anchors, box_preds, cls_preds = net(data)
# print(anchors.shape,box_preds.shape,cls_preds.shape)
# negative_mining_ratio，在生成的mask中增加*3的反例参加loss的计算。
box_offset, box_mask, cls_labels = MultiBoxTarget(
anchors,
label,
cls_preds.transpose(axes=(0, 2, 1)),
negative_mining_ratio=3.0) # , overlap_threshold=0.75)
loss1 = cls_loss(cls_preds, cls_labels)
loss2 = box_loss(box_preds, box_offset, box_mask)
loss = loss1 + loss2
# print(loss1.shape,loss2.shape)
loss.backward()
trainer.step(data.shape[0])
_loss[0] += nd.mean(loss1).asscalar()
_loss[1] += nd.mean(loss2).asscalar()
cls_probs = nd.SoftmaxActivation(cls_preds.transpose((0, 2, 1)),
mode='channel')
out = MultiBoxDetection(cls_probs,
box_preds,
anchors,
force_suppress=True,
clip=False,
nms_threshold=0.45)
if outs is None:
outs = out
labels = label
else:
outs = nd.concat(outs, out, dim=0)
labels = nd.concat(labels, label, dim=0)
box_metric.update([box_offset], [box_preds * box_mask])
train_AP = evaluate_MAP(outs, labels)
valid_AP, val_box_metric = evaluate_acc(net, valid_data, ctx)
info["train_ap"].append(train_AP)
info["valid_ap"].append(valid_AP)
info["loss"].append(_loss)
if (e + 1) % 10 == 0:
print("epoch: %d time: %.2f loss: %.4f, %.4f lr: %.5f" %
(e, time.time() - tic, _loss[0], _loss[1],
trainer.learning_rate))
print("train mae: %.4f AP: %.4f" % (box_metric.get()[1], train_AP))
print("valid mae: %.4f AP: %.4f" %
(val_box_metric.get()[1], valid_AP))
if True:
info["loss"] = np.array(info["loss"])
info["cls_loss"] = info["loss"][:, 0]
info["box_loss"] = info["loss"][:, 1]
plt.figure(figsize=(12, 4))
plt.subplot(121)
plot("train_ap")
plot("valid_ap")
plt.legend(loc="upper right")
plt.subplot(122)
plot("cls_loss")
plot("box_loss")
plt.legend(loc="upper right")
plt.savefig('loss_curve.png')