def alexnet_layers(n, MODEL_NAME):
branchyNet = AlexNet().build(n)
branchyNet.to_gpu()
branchyNet.training()
branchyNet.verbose = False
# 在主网络上进行训练
main_loss, main_acc, main_time = utils.train(branchyNet,
X_train,
Y_train,
main=True,
batchsize=TRAIN_BATCHSIZE,
num_epoch=TRAIN_NUM_EPOCHES)
print("main_time:", main_time)
# 在带有分支的网络上进行训练
branch_loss, branch_acc, branch_time = utils.train(
branchyNet,
X_train,
Y_train,
batchsize=TRAIN_BATCHSIZE,
num_epoch=TRAIN_NUM_EPOCHES)
print("branch_time:", branch_time)
# 保存已经训练好的网络模型
with open(MODEL_PATH + MODEL_NAME, "wb") as f:
dill.dump(branchyNet, f)
# 加载已经保存好的模型
with open(MODEL_PATH + MODEL_NAME, "rb") as f:
branchyNet = dill.load(f)
# 网络模型测试，获取分支网络的准确率
branchyNet.testing()
branchyNet.verbose = False
branchyNet.to_gpu()
# 退出点阈值设置
thresholdExits = [0.05, 0.01]
branchyNet.thresholdExits = thresholdExits
# 根据退出点阈值来获取分支网络的总体准确率
overacc, _, _, accbreakdowns, _ = utils.test(branchyNet,
X_test,
Y_test,
batchsize=TEST_BATCHSIZE)
return accbreakdowns
# 保存训练好的网络模型
with open(MODEL_NAME, "wb") as f:
dill.dump(branchyNet, f)
# 加载已保存好的模型
with open(MODEL_NAME, "rb") as f:
branchyNet = dill.load(f)
# 网络模型测试，获取主网络测试时间以及准确率
branchyNet.testing()
branchyNet.verbose = False
branchyNet.to_gpu()
g_baseacc, g_basediff, _, _, _ = utils.test(branchyNet,
X_test,
Y_test,
main=True,
batchsize=TEST_BATCHSIZE)
g_basediff = (g_basediff / float(len(Y_test))) * 1000.
print("g_baseacc:", g_baseacc)
print("g_basediff:", g_basediff)
# 退出点阈值设置
thresholds = [
0.0001, 0.0005, 0.001, 0.005, 0.01, 0.025, 0.05, 0.075, 0.1, 0.25, 0.5,
0.75, 1., 2., 3., 5., 10.
]
# 根据退出点阈值来获取代入单个退出点阈值的网络测试准确率、测试时间以及退出点样本数（带分支网络）,最大信息熵
g_ts, g_accs, g_diffs, g_exits, g_entropies = utils.screen_branchy(
branchyNet,
# Settings
TEST_BATCHSIZE = 1
print '2. set network to inference mode'
branchyNet.to_gpu()
branchyNet.testing()
branchyNet.verbose = False
branchyNet.gpu = True # ren +
# Test main
g_baseacc, g_basediff, g_num_exits, g_accbreakdowns = utils.test(branchyNet, x_test, y_test, main=True,
batchsize=TEST_BATCHSIZE)
print 'main accuracy: ', g_baseacc
g_basediff = (g_basediff / float(len(y_test))) * 1000.
branchyNet.to_cpu()
with open("_models/test_resnet56_cifar10_gpu_(g_baseacc).pkl", "w") as f:
dill.dump({'g_baseacc': g_baseacc}, f)
with open("_models/test_resnet56_cifar10_gpu_(g_basediff).pkl", "w") as f:
dill.dump({'g_basediff': g_basediff}, f)
with open("_models/test_resnet56_cifar10_gpu_(g_num_exits).pkl", "w") as f:
dill.dump({'g_num_exits': g_num_exits}, f)
with open("_models/test_resnet56_cifar10_gpu_(g_accbreakdowns).pkl", "w") as f:
dill.dump({'g_accbreakdowns': g_accbreakdowns}, f)
