def evaluate_accuracy(data_iter, net, ctx):
acc = nd.array([0], ctx=ctx)
for X, y in data_iter:
# 如果ctx是GPU，则将数据集复制到GPU上
X, y = X.as_in_context(ctx), y.as_in_context(ctx)
acc += gb.accuracy(net(X), y)
return acc.asscalar() / len(data_iter)
def train_ch5(net, train_iter, test_iter, batch_size, trainer, ctx,
num_epochs):
"""Train and evaluate a model on CPU or GPU."""
print('training on', ctx)
loss = gloss.SoftmaxCrossEntropyLoss()
for epoch in range(1, num_epochs + 1):
train_l_sum = 0
train_acc_sum = 0
num_add = 0
start = time.time()
for X, y in train_iter:
X, y = X.as_in_context(ctx), y.as_in_context(ctx)
with autograd.record():
y_hat = net(X)
l = loss(y_hat, y)
l.backward()
trainer.step(batch_size)
train_l_sum += l.mean().asscalar()
train_acc_sum += accuracy(y_hat, y)
num_add += 1
# test_acc = evaluate_accuracy(test_iter, net, ctx)
test_acc = 0
print('epoch %d, loss %.4f, train acc %.3f, test acc %.3f, '
'time %.1f sec'
% (epoch, train_l_sum / num_add,
train_acc_sum / num_add, test_acc, time.time() - start))
def train(net,train_iterator,test_iterator,num_epochs,lr,wd,ctx,lr_period,lr_decay):
# 只训练我们定义的输出网络
trainer=gluon.Trainer(net.output_new.collect_params(),'sgd',
{'learning_rate':lr,'momentum':0.9,'wd':wd})
prev_time=datetime.datetime.now()
for epoch in range(num_epochs):
# train_iter=data_iter(prefix,foldername,batch_size,True)
train_iter=train_iterator(batch_size)
test_iter=test_iterator(batch_size,False)
train_l=0.0
train_acc=0.0
train_n=0
# if epoch > 0 and epoch % lr_period == 0:
# trainer.set_learning_rate(trainer.learning_rate*lr_decay)
num=0.
for X,y,n in train_iter:
if n==0:
continue
y=y.astype('float32').as_in_context(ctx)
output_features=net.features(X.as_in_context(ctx))
with autograd.record():
y_hat=net.output_new(output_features)
l=loss(y_hat,y)
l.backward()
trainer.step(n)
train_l+=l.mean().asscalar()
train_acc+=gb.accuracy(y_hat,y)
num+=1
train_n+=n
# print ("train samples:%d"%(train_n))
cur_time=datetime.datetime.now()
h,remainder=divmod((cur_time-prev_time).seconds,3600)
m,s=divmod(remainder,60)
time_s="time %02d:%02d:%02d"%(h,m,s)
test_acc=evaluate_accuracy(test_iter,net,ctx)
if num==0.:
epoch_s=("epoch %d, loss %f, train acc %f, test acc %f, "
% (epoch, train_l ,
train_acc, test_acc))
else:
epoch_s=("epoch %d, loss %f, train acc %f, test acc %f, "
% (epoch, train_l / num,
train_acc / num, test_acc))
prev_time=cur_time
print(epoch_s+time_s+',lr '+str(trainer.learning_rate))
def evaluate_accuracy(data_iter,net,ctx):
acc=0.
num=0.
test_n=0
for X,y,n in data_iter:
if n==0:
continue
y=y.astype('float32').as_in_context(ctx)
output_features=net.features(X.as_in_context(ctx))
y_hat=net.output_new(output_features)
acc+=gb.accuracy(y_hat,y)
num+=1
test_n+=n
# print ("test samples:%d"%(test_n))
if num==0.:
return acc
else:
return acc/num
def train_ch5(net, train_iter, test_iter, loss, batch_size, trainer,
num_epochs):
for epoch in range(1, num_epochs + 1):
train_l_sum = 0
train_acc_sum = 0
start = time()
for X, y in train_iter:
with autograd.record():
y_hat = net(X)
l = loss(y_hat, y)
l.backward()
trainer.step(batch_size)
train_l_sum += l.mean().asscalar()
train_acc_sum += gb.accuracy(y_hat, y)
test_acc = evaluate_accuracy(test_iter, net)
print(
"epoch %d, loss %.4f, train acc %.3f, test acc %.3f, time %1f sec"
% (epoch, train_l_sum / len(train_iter),
train_acc_sum / len(train_iter), test_acc, time() - start))
def train_ch5(net, train_iter, test_iter, batch_size, trainer, ctx,
num_epochs):
print('training on', ctx)
loss = gloss.SoftmaxCrossEntropyLoss()
for epoch in range(num_epochs):
train_l_sum, train_acc_sum, start = 0, 0, time.time()
for X, y in train_iter:
X, y = X.as_in_context(ctx), y.as_in_context(ctx)
with autograd.record():
y_hat = net(X)
l = loss(y_hat, y)
l.backward()
trainer.step(batch_size)
train_l_sum += l.mean().asscalar()
train_acc_sum += gb.accuracy(y_hat, y)
test_acc = evaluate_accuracy(test_iter, net, ctx)
print(
'epoch %d, loss %.4f, train acc %.3f, test acc %.3f, time %.1f sec'
% (epoch + 1, train_l_sum / len(train_iter),
train_acc_sum / len(train_iter), test_acc, time.time() - start))
def train(net, train_data, valid_data, num_epochs, lr, wd, ctx, lr_period,
lr_decay):
trainer = gluon.Trainer(net.collect_params(), 'sgd', {
'learning_rate': lr,
'momentum': 0.9,
'wd': wd
})
prev_time = datetime.datetime.now()
for epoch in range(num_epochs):
train_loss = 0.0
train_acc = 0.0
if epoch > 0 and epoch % lr_period == 0:
trainer.set_learning_rate(trainer.learning_rate * lr_decay)
for data, label in train_data:
label = label.astype('float32').as_in_context(ctx)
with autograd.record():
output = net(data.as_in_context(ctx))
loss = softmax_cross_entropy(output, label)
loss.backward()
trainer.step(batch_size)
train_loss += nd.mean(loss).asscalar()
train_acc += gb.accuracy(output, label)
cur_time = datetime.datetime.now()
h, remainder = divmod((cur_time - prev_time).seconds, 3600)
m, s = divmod(remainder, 60)
time_str = "Time %02d:%02d:%02d" % (h, m, s)
if valid_data is not None:
valid_acc = gb.evaluate_accuracy(valid_data, net, ctx)
epoch_str = ("Epoch %d. Loss: %f, Train acc %f, Valid acc %f, " %
(epoch, train_loss / len(train_data),
train_acc / len(train_data), valid_acc))
else:
epoch_str = ("Epoch %d. Loss: %f, Train acc %f, " %
(epoch, train_loss / len(train_data),
train_acc / len(train_data)))
prev_time = cur_time
print(epoch_str + time_str + ', lr ' + str(trainer.learning_rate))
def vgg():
print 'programe begin'
train_pic_list, train_label_list = get_pic_dogandcat(TRAIN_DATA_SIZE)
test_pic_list, test_label_list = get_pic_dogandcat(TEST_DATA_SIZE)
train_dataset = gluon.data.ArrayDataset(train_pic_list, train_label_list)
train_data_iter = gluon.data.DataLoader(train_dataset,
batch_size=BATCH_SIZE,
shuffle=True)
test_dataset = gluon.data.ArrayDataset(test_pic_list, test_label_list)
test_data_iter = gluon.data.DataLoader(test_dataset,
batch_size=TEST_DATA_SIZE,
shuffle=True)
# train_data_iter = mx.image.ImageIter(batch_size=BATCH_SIZE, data_shape=(3, 224, 224),
# path_imglist='train.lst')
# train_data_iter.reset()
print 'dataset created'
with mx.Context(mx.gpu()):
net = nn.Sequential()
net.add(
nn.Conv2D(channels=64,
kernel_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
activation='relu',
in_channels=3))
net.add(
nn.Conv2D(channels=64,
kernel_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
activation='relu',
in_channels=64))
net.add(nn.MaxPool2D(pool_size=2, strides=2))
net.add(
nn.Conv2D(channels=128,
kernel_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
activation='relu',
in_channels=64))
net.add(
nn.Conv2D(channels=128,
kernel_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
activation='relu',
in_channels=128))
net.add(nn.MaxPool2D(pool_size=2, strides=2))
net.add(
nn.Conv2D(channels=256,
kernel_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
activation='relu',
in_channels=128))
net.add(
nn.Conv2D(channels=256,
kernel_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
activation='relu',
in_channels=256))
net.add(
nn.Conv2D(channels=256,
kernel_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
activation='relu',
in_channels=256))
net.add(nn.MaxPool2D(pool_size=2, strides=2))
net.add(
nn.Conv2D(channels=512,
kernel_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
activation='relu',
in_channels=256))
net.add(
nn.Conv2D(channels=512,
kernel_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
activation='relu',
in_channels=512))
net.add(
nn.Conv2D(channels=512,
kernel_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
activation='relu',
in_channels=512))
net.add(nn.MaxPool2D(pool_size=2, strides=2))
net.add(
nn.Conv2D(channels=512,
kernel_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
activation='relu',
in_channels=512))
net.add(
nn.Conv2D(channels=512,
kernel_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
activation='relu',
in_channels=512))
net.add(
nn.Conv2D(channels=512,
kernel_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
activation='relu',
in_channels=512))
net.add(nn.MaxPool2D(pool_size=2, strides=2))
net.add(nn.Dense(4096, activation='relu', in_units=7 * 7 * 512))
net.add(nn.Dropout(DROPOUT_RATE))
net.add(nn.Dense(4096, activation='relu', in_units=4096))
net.add(nn.Dropout(DROPOUT_RATE))
net.add(nn.Dense(2, in_units=4096))
print 'net created'
net.initialize()
####################
X = nd.random.uniform(shape=(BATCH_SIZE, 3, 224, 224))
for blk in net:
X = blk(X)
print(blk.name, 'output shape:\t', X.shape)
# exit()
# net(train_pic_list[0])
####################
loss = gluon.loss.SoftmaxCrossEntropyLoss()
learning_rate = 0.01
trainer = gluon.Trainer(net.collect_params(), 'adam',
{'learning_rate': learning_rate})
num_epoch = 1000
print 'train begin'
for epoch in range(num_epoch):
start_time = time.time()
train_l_sum = 0
train_acc_sum = 0
num_add = 1
for X, y in train_data_iter:
X = X.copyto(mx.gpu())
y = y.copyto(mx.gpu())
with autograd.record():
y_predict = net(X)
l = loss(net(X), y)
l.backward()
trainer.step(batch_size=BATCH_SIZE)
train_l_sum += l.mean().asscalar()
train_acc_sum += gb.accuracy(y_predict, y)
num_add += 1
if (epoch != 0 and epoch % 3 == 0):
learning_rate = learning_rate * 0.9
trainer.set_learning_rate(learning_rate)
test_acc = evaluate_accuracy(test_data_iter, net)
# test_acc = 0
time1 = time.time() - start_time
print(
'epoch %3d, loss %.8f, train acc %.8f, test acc %.8f, learning_rate: %.8f, Time: %.3fs, predict_time: %dmin%ds'
% (epoch + 1, train_l_sum / num_add, train_acc_sum / num_add,
test_acc, trainer.learning_rate, time1,
int(time1 * (num_epoch - epoch - 1) / 60), time1 *
(num_epoch - epoch - 1) % 60))
def evaluate_accuracy(data_iter, net):
acc = nd.array([0])
for X, y in data_iter:
acc += gb.accuracy(net(X), y)
return acc.asscalar() / len(data_iter)
# print small_conv_arch
ctx = mx.gpu()
net = vgg()
net.initialize(init.Xavier(), ctx=ctx)
train_iter, test_iter = get_iter()
loss = gloss.SoftmaxCrossEntropyLoss()
num_epochs = 100
batch_size = BATCH_SIZE
trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': 0.001})
for epoch in range(1, num_epochs + 1):
train_l_sum = 0
train_acc_sum = 0
num_add = 0
start = time.time()
for X, y in train_iter:
X, y = X.as_in_context(ctx), y.as_in_context(ctx)
with autograd.record():
y_hat = net(X)
l = loss(y_hat, y)
l.backward()
trainer.step(batch_size)
train_l_sum += l.mean().asscalar()
train_acc_sum += accuracy(y_hat, y)
num_add += 1
# test_acc = evaluate_accuracy(test_iter, net, ctx)
test_acc = 0
print('epoch %d, loss %.4f, train acc %.3f, test acc %.3f, '
'time %.1f sec'
% (epoch, train_l_sum / num_add,
train_acc_sum / num_add, test_acc, time.time() - start))
def train(net, train_data, valid_data, num_epochs, lr, wd, ctx, lr_period,
lr_decay):
trainer = gluon.Trainer(net.collect_params(), 'sgd', {
'learning_rate': lr,
'momentum': 0.9,
'wd': wd
})
prev_time = datetime.datetime.now()
for epoch in range(num_epochs):
train_l, train_acc = 0.0, 0.0
if epoch > 0 and epoch % lr_period == 0:
trainer.set_learning_rate(trainer.learning_rate * lr_decay)
print('e=' + str(epoch) + ' lr=' +
str(trainer.learning_rate * lr_decay))
iii = 0
for X, y in train_data:
iii += 1
print('... {:.1f}% ...'.format(iii * 100 / len(train_data)),
end='\r')
# X=nd.swapaxes(X,1,3)
# X=nd.swapaxes(X,2,3)
y = y.as_in_context(ctx)
# import pdb; pdb.set_trace() ###
with autograd.record():
y_hat = net(X.astype('float32').as_in_context(ctx))
l = loss(y_hat, y)
l.backward()
trainer.step(batch_size)
train_l += l.mean().asscalar()
train_acc += gb.accuracy(y_hat, y)
cur_time = datetime.datetime.now()
h, remainder = divmod((cur_time - prev_time).seconds, 3600)
m, s = divmod(remainder, 60)
time_s = "time %02d:%02d:%02d" % (h, m, s)
if valid_data is not None:
valid_acc = 0
iii = 0
for X, y in valid_data:
iii += 1
print('... {:.1f}% ...'.format(iii * 100 / len(valid_data)),
end='\r')
# X=nd.swapaxes(X,1,3)
# X=nd.swapaxes(X,2,3)
y = y.as_in_context(ctx)
val_y_hat = net(X.astype('float32').as_in_context(ctx))
valid_acc += gb.accuracy(val_y_hat, y)
epoch_s = ("epoch %d, loss %f, train acc %f, valid acc %f, " %
(epoch + 1, train_l / len(train_data), train_acc /
len(train_data), valid_acc / len(valid_data)))
else:
epoch_s = ("epoch %d, loss %f, train acc %f, " %
(epoch + 1, train_l / len(train_data),
train_acc / len(train_data)))
prev_time = cur_time
print(epoch_s + time_s) # + ', lr ' + str(trainer.learning_rate))
net.save_parameters('./params_ffn/dbc_' + str(epoch + 16) + '.param')
print('model saved')
# 然后通过lr_scheduler.learning_rate -= 0.01 让学习率随着epoch线性递减。
learning_rate = 0.001
trainer = gluon.Trainer(net.collect_params(), 'sgd',
{'learning_rate': learning_rate})
num_eproch = 100
for epoch in range(num_eproch):
start_time = time.time()
train_l_sum = 0
train_acc_sum = 0
num_add = 0
for X, y in train_data_iter:
with autograd.record():
y_predict = net(X)
l = loss(net(X), y)
l.backward()
trainer.step(batch_size=BATCH_SIZE)
train_l_sum += l.mean().asscalar()
train_acc_sum += gb.accuracy(y_predict, y)
num_add += 1
if (epoch != 0 and epoch % 3 == 0):
learning_rate = learning_rate * 0.9
trainer.set_learning_rate(learning_rate)
test_acc = evaluate_accuracy(test_data_iter, net)
time1 = time.time() - start_time
print(
'epoch %3d, loss %.8f, train acc %.8f, test acc %.8f, learning_rate: %.8f, Time: %.3fs, predict_time: %dmin%ds'
%
(epoch + 1, train_l_sum / len(train_data_iter), train_acc_sum /
len(train_data_iter), test_acc, trainer.learning_rate, time1,
int(time1 * (num_eproch - epoch - 1) / 60), time1 *
(num_eproch - epoch - 1) % 60))
train_iter.reset()
test_iter.reset()
num_batch = 1
train_l_sum = 0
train_acc_sum = 0
train_time_sum = 0
read_time_sum = 0
test_time_sum = 0
for data_cpu in train_iter:
start1 = time.time()
X = data_cpu.data[0].copyto(mx.gpu()) / 255
Y_true = data_cpu.label[0].copyto(mx.gpu())
read_time_sum += (time.time() - start1)
start2 = time.time()
with autograd.record():
y_predict = net(X)
l = loss(y_predict, Y_true)
l.backward()
trainer.step(batch_size=BATCH_SIZE)
train_time_sum += (time.time() - start2)
start3 = time.time()
train_l_sum += l.mean().asscalar()
train_acc_sum += gb.accuracy(y_predict, Y_true)
num_batch += 1
test_time_sum += (time.time() - start3)
print epoch, num_batch, train_l_sum / num_batch, train_acc_sum / num_batch
print 'read_time_sum:', read_time_sum, ' arg:', read_time_sum / num_batch
print 'train_time_sum:', train_time_sum, ' arg:', train_time_sum / num_batch
print 'test_time_sum:', test_time_sum, ' arg:', test_time_sum / num_batch
print 'total_time:', total_time_start - time.time()