def train():
batch_size = 32
train_iter, _ = d2l.load_data_pikachu(batch_size)
ctx, net = d2l.try_gpu(), TinySSD(num_classes=1)
net.initialize(init=init.Xavier(), ctx=ctx)
trainer = gluon.Trainer(net.collect_params(), 'sgd', {
'learning_rate': 0.2,
'wd': 5e-4
})
cls_loss = gloss.SoftmaxCrossEntropyLoss()
bbox_loss = gloss.L1Loss()
def calc_loss(cls_preds, cls_labels, bbox_preds, bbox_labels, bbox_masks):
cls = cls_loss(cls_preds, cls_labels)
bbox = bbox_loss(bbox_preds * bbox_masks, bbox_labels * bbox_masks)
return cls + bbox
def cls_eval(cls_preds, cls_labels):
# 由于类别预测结果放在最后一维，argmax需要指定最后一维
return (cls_preds.argmax(axis=-1) == cls_labels).sum().asscalar()
def bbox_eval(bbox_preds, bbox_labels, bbox_masks):
return ((bbox_labels - bbox_preds) * bbox_masks).abs().sum().asscalar()
for epoch in range(20):
acc_sum, mae_sum, n, m = 0.0, 0.0, 0, 0
train_iter.reset() # 从头读取数据
start = time.time()
for batch in train_iter:
X = batch.data[0].as_in_context(ctx)
Y = batch.label[0].as_in_context(ctx)
with autograd.record():
# 生成多尺度的锚框，为每个锚框预测类别和偏移量
anchors, cls_preds, bbox_preds = net(X)
# 为每个锚框标注类别和偏移量
bbox_labels, bbox_masks, cls_labels = \
contrib.nd.MultiBoxTarget(anchors, Y, cls_preds.transpose((0, 2, 1)))
# 根据类别和偏移量的预测和标注值计算损失函数
l = calc_loss(cls_preds, cls_labels, bbox_preds, bbox_labels,
bbox_masks)
l.backward()
trainer.step(batch_size)
acc_sum += cls_eval(cls_preds, cls_labels)
n += cls_labels.size
mae_sum += bbox_eval(bbox_preds, bbox_labels, bbox_masks)
m += bbox_labels.size
if (epoch + 1) % 5 == 0:
print(
'epoch %2d, class err %.2e, bbox mae %.2e, time %.1f sec' %
(epoch + 1, 1 - acc_sum / n, mae_sum / m, time.time() - start))
net = TinySSD(num_classes=1)
net.initialize()
X = nd.zeros((32, 3, 256, 256))
anchors, cls_preds, bbox_preds = net(X)
print('output anchors:', anchors.shape)
print('output class preds:', cls_preds.shape)
print('output bbox preds:', bbox_preds.shape)
#9.7.2-训练
#读取数据和初始化
batch_size = 32
print('try_gpu')
train_iter, _ = d2l.load_data_pikachu(batch_size)
ctx, net = d2l.try_gpu(), TinySSD(num_classes=1)
net.initialize(init=init.Xavier(), ctx=ctx)
print("start train...\n")
trainer = gluon.Trainer(net.collect_params(), 'sgd', {
'learning_rate': 0.2,
'wd': 5e-4
})
print("end train...\n")
#定义损失函数和评价函数
cls_loss = gloss.SoftmaxCrossEntropyLoss()
bbox_loss = gloss.L1Loss()
def calc_loss(cls_preds, cls_labels, bbox_preds, bbox_labels, bbox_masks):