net.add(gluon.nn.Dense(10))
# net.initialize()
return net
if __name__ == '__main__':
train_data, test_data = utils1.load_data_fashion_mnist(batch_size)
entrory_loss = gluon.loss.SoftmaxCrossEntropyLoss()
lr = 0.5
net = mynet()
net.initialize()
trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': 0.5})
epoches = 5
for e in range(epoches):
train_loss = 0
test_acc = 0
train_acc = 0
for data, label in train_data:
with autograd.record():
out = net(data)
loss = entrory_loss(out, label)
loss.backward()
trainer.step(batch_size)
train_loss += nd.mean(loss).asscalar()
train_acc += utils1.accuracy(out, label)
test_acc = utils1.evaluate_accuracy(test_data, net)
print("Epoch %d. Loss: %f, Train acc %f, Test acc %f" % (
e, train_loss / len(train_data),
train_acc / len(train_data), test_acc))
def myTrain(net,
batch_size,
train_data,
valid_data,
epoches,
lr,
wd,
ctx,
lr_period,
lr_decay,
verbose=False):
trainer = mx.gluon.Trainer(net.collect_params(), 'sgd', {
'learning_rate': lr,
'momentum': 0.9,
'wd': wd
})
prev_time = datetime.datetime.now()
train_loss_record = []
valid_loss_record = [] # epoches recycle record loss
train_acc_record = []
valid_acc_record = []
focalloss = netlib.FocalLoss()
for e in range(epoches):
train_loss = 0.0
train_acc = 0.0
# if e > 99 and e < 251 and e % 10 == 0:
# trainer.set_learning_rate(trainer.learning_rate * lr_decay) # decrease lr
# if e == 60 or e == 120 or e == 160:
# trainer.set_learning_rate(trainer.learning_rate * lr_decay) # decrease lr
if e > 150 and e % 20 == 0:
trainer.set_learning_rate(trainer.learning_rate *
lr_decay) # decrease
# print('train len:',len(train_data))
for data, label in train_data:
label = label.reshape(
shape=(label.shape[0], )) # be careful:it turns to vector
label = label.astype('float32').as_in_context(ctx)
with autograd.record():
output = net(data.as_in_context(ctx))
loss = softmax_cross_entrory(output, label) #
# loss = focalloss(output, label)# focal loss
loss.backward()
trainer.step(batch_size)
train_loss += nd.mean(loss).asscalar()
train_acc += utils1.accuracy(output, label)
train_loss_record.append(train_loss / len(train_data))
train_acc_record.append(train_acc / len(train_data))
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 = evaluate_accuracy(valid_data, net, ctx)
valid_acc_record.append(valid_acc)
if verbose:
###valid data loss
valid_loss = 0
for data, valid_label in valid_data:
valid_label = valid_label.reshape(
shape=(valid_label.shape[0],
)) # be careful:it turns to vector
valid_label = valid_label.astype('float32').as_in_context(
ctx)
# with autograd.predict_mode():
out = net(data.as_in_context(ctx))
loss = softmax_cross_entrory(out, valid_label)
# loss = focalloss(out, valid_label) # focal loss
valid_loss += nd.mean(loss).asscalar()
# valid_loss = nd.mean(loss).asscalar( # only used valid loss of every batch(vaild_data)
valid_loss_record.append(
valid_loss /
len(valid_data)) # record every batch loss of valid data
epoch_str = (
"Epoch %d. Train Loss: %f,Valid Loss: %f, Train acc %f, Valid acc %f, "
%
(e, train_loss / len(train_data), valid_loss /
len(valid_data), train_acc / len(train_data), valid_acc))
else:
epoch_str = (
"Epoch %d. Train Loss: %f, Train acc %f, Valid acc %f, " %
(e, train_loss / len(train_data),
train_acc / len(train_data), valid_acc))
else:
epoch_str = (
"Epoch %d. Loss: %f, Train acc %f, " %
(e, train_loss / len(train_data), train_acc / len(train_data)))
prev_time = cur_time
print(epoch_str + 'lr=' + str(trainer.learning_rate) + ',' + time_str)
# plot loss and acc
fig, (fig1, fig2) = plt.subplots(1, 2)
if verbose:
fig1.plot(train_loss_record, 'b')
fig1.legend(['train'])
fig2.plot(train_acc_record, 'b')
fig2.legend(['train_acc'])
if valid_data is not None:
fig1.plot(valid_loss_record, 'r')
fig1.legend(['train', 'test'])
fig2.plot(valid_acc_record, 'r')
fig2.legend(['train_acc', 'valid_acc'])
else:
fig1.plot(train_loss_record, 'b')
fig1.legend(['train'])
fig2.plot(train_acc_record, 'b')
fig2.plot(valid_acc_record, 'r')
fig2.legend(['train_acc', 'valid_acc'])
fig.show()
fig.savefig('./CIFAR10_result.png')