def softmax_test(target):
x = np.array([[9, 2, 5, 0, 0], [7, 5, 0, 0, 0]])
expected_output = np.array([[
9.80897665e-01, 8.94462891e-04, 1.79657674e-02, 1.21052389e-04,
1.21052389e-04
],
[
8.78679856e-01, 1.18916387e-01,
8.01252314e-04, 8.01252314e-04,
8.01252314e-04
]])
test_cases = [{
"name": "datatype_check",
"input": [x],
"expected": np.ndarray,
"error": "The function should return a numpy array."
}, {
"name": "shape_check",
"input": [x],
"expected": expected_output,
"error": "Wrong shape"
}, {
"name": "equation_output_check",
"input": [x],
"expected": expected_output,
"error": "Wrong output"
}]
test(test_cases, target)
def image2vector_test(target):
image = np.array([[[0.67826139, 0.29380381], [0.90714982, 0.52835647],
[0.4215251, 0.45017551]],
[[0.92814219, 0.96677647], [0.85304703, 0.52351845],
[0.19981397, 0.27417313]],
[[0.60659855, 0.00533165], [0.10820313, 0.49978937],
[0.34144279, 0.94630077]]])
expected_output = np.array([[0.67826139], [0.29380381], [0.90714982],
[0.52835647], [0.4215251], [0.45017551],
[0.92814219], [0.96677647], [0.85304703],
[0.52351845], [0.19981397], [0.27417313],
[0.60659855], [0.00533165], [0.10820313],
[0.49978937], [0.34144279], [0.94630077]])
test_cases = [{
"name": "datatype_check",
"input": [image],
"expected": np.ndarray,
"error": "The function should return a numpy array."
}, {
"name": "shape_check",
"input": [image],
"expected": expected_output,
"error": "Wrong shape"
}, {
"name": "equation_output_check",
"input": [image],
"expected": expected_output,
"error": "Wrong output"
}]
test(test_cases, target)
def test():
"Please, try to update that everytime it is a change in the tests."
## Init tests
t0 = time.time()
print("")
print(message_init_tests)
print(ref_computer_stats)
print(message_ref_computer % time_ref_computer)
## Performing tests
with warnings.catch_warnings():
warnings.simplefilter("ignore")
#warnings.simplefilter("error")
test_pythonUtils.test()
test_io.test()
test_utils.test()
test_neighsinfo.test()
test_preprocess.test()
test_retriever.test()
test_spatial_discretizer.test()
test_descriptormodels.test()
test_features_objects.test()
test_features_retriever.test()
test_perturbation.test()
test_spatial_relations.test()
test_spdescriptormodels.test()
test_interpolation.test()
test_sampling.test()
test_api.test()
## Closing tests
time_own_computer = str(np.round(time.time()-t0, 2))
print(message_own_computer % time_own_computer)
def basic_sigmoid_test(target):
x = 1
expected_output = 0.7310585786300049
test_cases = [{
"name": "datatype_check",
"input": [x],
"expected": float,
"error": "Datatype mismatch."
}, {
"name": "equation_output_check",
"input": [x],
"expected": expected_output,
"error": "Wrong output."
}]
test(test_cases, target)
def L1_test(target):
yhat = np.array([.9, 0.2, 0.1, .4, .9])
y = np.array([1, 0, 0, 1, 1])
expected_output = 1.1
test_cases = [{
"name": "datatype_check",
"input": [yhat, y],
"expected": float,
"error": "The function should return a float."
}, {
"name": "equation_output_check",
"input": [yhat, y],
"expected": expected_output,
"error": "Wrong output"
}]
test(test_cases, target)
def main(args):
train_loader, val_loader = custom_data_loader.customDataloader(args)
model = custom_model.buildModel(args)
optimizer, scheduler, records = solver_utils.configOptimizer(args, model)
criterion = solver_utils.Criterion(args)
recorder = recorders.Records(args.log_dir, records)
tf_train_writer, tf_test_writer = tfboard.tensorboard_init()
for epoch in range(args.start_epoch, args.epochs+1):
scheduler.step()
recorder.insertRecord('train', 'lr', epoch, scheduler.get_lr()[0])
train_utils.train(args, train_loader, model, criterion, optimizer, log, epoch, recorder, tf_train_writer)
if epoch % args.save_intv == 0:
model_utils.saveCheckpoint(args.cp_dir, epoch, model, optimizer, recorder.records, args)
if epoch % args.val_intv == 0:
test_utils.test(args, 'val', val_loader, model, log, epoch, recorder, tf_test_writer)
def sigmoid_derivative_test(target):
x = np.array([1, 2, 3])
expected_output = np.array([0.19661193, 0.10499359, 0.04517666])
test_cases = [{
"name": "datatype_check",
"input": [x],
"expected": np.ndarray,
"error": "The function should return a numpy array."
}, {
"name": "shape_check",
"input": [x],
"expected": expected_output,
"error": "Wrong shape."
}, {
"name": "equation_output_check",
"input": [x],
"expected": expected_output,
"error": "Wrong output."
}]
test(test_cases, target)
def sigmoid_test(target):
x = np.array([1, 2, 3])
expected_output = np.array([0.73105858, 0.88079708, 0.95257413])
test_cases = [{
"name": "datatype_check",
"input": [x],
"expected": np.ndarray,
"error": "Datatype mismatch."
}, {
"name": "shape_check",
"input": [x],
"expected": expected_output,
"error": "Wrong shape."
}, {
"name": "equation_output_check",
"input": [x],
"expected": expected_output,
"error": "Wrong output."
}]
test(test_cases, target)
def final_test(model, CFG, final_test_loaders):
test_loaders_src, test_loaders_tar = final_test_loaders # 测试集，源域与目标域的
# 开始测试
acc_test_srcs, acc_test_tars = [], [] # 正确率容器
for i in range(len(test_loaders_tar)): # 遍历所有测试集
#test_loader_src, test_loader_tar = test_loaders_src[i], test_loaders_tar[i] # 一个有s，一个没有
acc_collection_test, loss_collection_test, _ = test(
model, test_loaders_src[i], test_loaders_tar[i], 0,
CFG) # 依次对每个测试集进行测试
acc_test_srcs.append(acc_collection_test[0]) # 将第0个和第6个测试集的正确率作为最终投票结果
acc_test_tars.append(acc_collection_test[6])
return acc_test_srcs, acc_test_tars
def normalizeRows_test(target):
x = np.array([[0, 3, 4], [1, 6, 4]])
expected_output = np.array([[0., 0.6, 0.8],
[0.13736056, 0.82416338, 0.54944226]])
test_cases = [{
"name": "datatype_check",
"input": [x],
"expected": np.ndarray,
"error": "The function should return a numpy array."
}, {
"name": "shape_check",
"input": [x],
"expected": expected_output,
"error": "Wrong shape"
}, {
"name": "equation_output_check",
"input": [x],
"expected": expected_output,
"error": "Wrong output"
}]
test(test_cases, target)
### Test - stage 6 ### Part of: Blivet test collection ### Author: [email protected] ### This program is under GPL licence. import classes import test_utils loginst_test = test_utils.init_logging(0, None, True) test_utils.create_new_alloc_table("vdb") test_utils.create_new_partition("vdb", "extended", 1, -1) tsep = classes.SystemExtended_Scan('vdb', 1) tbep = classes.BlivetInitialization('vdb', 1).child loginst_test.debug("Comparing object attributes") ia = test_utils.test(tsep, tbep) test_utils.write_issues(ia, "Extended partition test", 6)
finish = 512 + start
## Test partitions
list_of_tests = []
list_of_blivet = []
list_of_ia = []
loginst_test.debug("Setting SystemPartitionFormatted_Scan")
for inc in range(4):
loginst_test.debug("Setting partition {}".format(inc + 1))
if inc == 3:
test_utils.create_new_partition("vdb", "primary", start, -1)
else:
test_utils.create_new_partition("vdb", "primary", start, finish)
start = finish + 1
finish = start + 512
loginst_test.debug("Setting partition {} - formatting to {}".format(inc + 1, "ext4"))
test_utils.format_new_partition("{}{}".format("vdb", inc + 1), "ext4")
## init objects
list_of_tests.append(classes.SystemPartitionFormatted_Scan('vdb', inc + 1))
list_of_blivet.append(classes.BlivetInitialization('vdb', inc + 1).child)
## Store in arrays
loginst_test.debug("Comparing SystemPartitionFormatted_Scan - partition {} with Blivet instance.".format(inc + 1))
list_of_ia.append(test_utils.test(list_of_tests[inc], list_of_blivet[inc]))
## Store in file.
test_utils.write_issues(list_of_ia[inc], "Multi partition - part {}".format(inc + 1), 5)
presort=False)
# 训练该轮并记录数据
clf_tree = clf_tree.fit(carSet.data,
carSet.target,
sample_weight=weightArrays)
trs.append(clf_tree)
preds = clf_tree.predict(carSet.data)
#print(preds)
pbs = clf_tree.predict_proba(carSet.data)
# 调整数据, 准备下一轮循环 确定r和alpha权重
r = 0.0
r = 1 - clf_tree.score(
carSet.data, carSet.target, sample_weight=weightArrays)
alpha = 0.5 * math.log((1 - r) / r)
alphas.append(alpha)
for i in range(len(carSet.data)):
for l0 in range(labels):
for l1 in range(labels):
D[i, l0,
l1] = D[i, l0, l1] * math.exp(0.5 * alpha *
(pbs[i][l0] - pbs[i][l1]))
Z = sum(sum(sum(D)))
D /= Z
print("决策树训练结束!\n")
print("进行预测\n")
# 测试数据与期望目标
test_utils.test(test_data, test_target, trs, alphas, numRound, labels)
### Test - stage 2 ### Part of: Blivet test collection ### Author: [email protected] ### This program is under GPL licence. import classes import test_utils loginst_test = test_utils.init_logging(0, None, True) test_utils.create_new_alloc_table("vdb") loginst_test.debug("Setting up classes.SystemDiskFormatted_Scan") test_system_formatted = classes.SystemDiskFormatted_Scan('vdb') test_blivet_formatted = classes.BlivetInitialization('vdb').disk loginst_test.debug("Comparing object attributes") ia = test_utils.test(test_system_formatted, test_blivet_formatted) test_utils.write_issues(ia, "Formatted disk", 2)
def main(args):
test_loader = custom_data_loader.benchmarkLoader(args)
model = custom_model.buildModel(args)
recorder = recorders.Records(args.log_dir)
test_utils.test(args, 'test', test_loader, model, log, 1, recorder)
print('\ncreating top-5 nodule dataset...')
data, label = create_test_dataset(args.sorted_slices_dir, 1)
testds = prep_dataset(data, label)
test_loader = data_utils.DataLoader(testds, batch_size=args.batch_size, shuffle=False)
print(' dataset ready!')
model = Predictor()
if args.cuda:
model.cuda()
if args.cuda:
model.load_state_dict(torch.load(args.resume))
else:
model.load_state_dict(torch.load(args.resume, map_location=torch.device('cpu')))
output1 = np.load('src/preds3D_val.npy') #subtraction of year 2000 and 1999 masks + sum + sigmoid scores
output, label = test(model, test_loader, args)
output = output.data.cpu().numpy()
PP = 0.68
outputx = output[:,1]*PP + output1[:,0]*(1-PP)
AUC = roc_auc_score(label, outputx)
print('\n%s\nAUC: %.4f\n%s'%(28*'*',AUC,28*'*'))
print('\ntotal elapsed time: %0.2f min\n' % ((time.time() - ini_t_time)/60.0))
### Test - stage 3 ### Part of: Blivet test collection ### Author: [email protected] ### This program is under GPL licence. import classes import test_utils loginst_test = test_utils.init_logging(0, None, True) test_utils.create_new_alloc_table("vdb") test_utils.create_new_partition("vdb", "primary", 1, -1) loginst_test.debug("Setting up classes.SystemPartition_Scan") test_blivet_partition = classes.BlivetInitialization('vdb', 1).child test_system_partition = classes.SystemPartition_Scan('vdb', 1) loginst_test.debug("Comparing object attributes") ia = test_utils.test(test_system_partition, test_blivet_partition) test_utils.write_issues(ia, "Partitioned disk", 3)
### Test - stage 1 ### Part of: Blivet test collection ### Author: [email protected] ### This program is under GPL licence. import classes import test_utils loginst_test = test_utils.init_logging(0, None, True) loginst_test.debug("Setting up classes.SystemDisk_Scan") test_system = classes.SystemDisk_Scan('vdb') test_blivet = classes.BlivetInitialization('vdb').disk loginst_test.debug("Comparing object attributes") ia = test_utils.test(test_system, test_blivet) test_utils.write_issues(ia, "Basic disk", 1)
# Basic automated test_utils for the search API
#
from test_utils import test
def check_photo(animals):
for animal in animals:
try:
if (animal['MainPhoto']['default'].find('jpg/320-') > 0):
return True
except:
"ignore"
return False
test1 = test('/search?AnimalType=ALL&Location=ALL&StatusCategory=available',
lambda check: check_photo(check['response']))
details = [
test('/animal?Id=%d' % animal['AnimalId'], lambda ok: True)
for animal in test1['response'][0:9]
]
sortable = [a['response']['Id'] for a in details]
if (sortable != sorted(sortable)):
print("the sortable items are not in order: " + str(sortable))
raise
test(
'/search?AnimalType=Cat&Location=ALL&StatusCategory=rescue',
lambda check: set(['Cat']) == set(
[animal['AnimalType'] for animal in check['response']]))
def train(loaders, model, optimizer, scheduler, CFG):
acc_test_tar_max = 0 # 最大的目标域测试正确率，用于在最大测试正确率的时候保存pt文件
train_loader_src, test_loader_src, train_loader_tar, test_loader_tar = loaders # 样本加载器
len_src_loader, len_tar_loader = len(train_loader_src), len(
train_loader_tar) # 样本中的minibatch的数目
train_loss_clf_src, train_loss_clf_tar, train_loss_tran, train_loss_total = utils.AverageMeter(
), utils.AverageMeter(), utils.AverageMeter(), utils.AverageMeter(
) # 用于存储误差
acc_collection, loss_collection,acc_collection_test, loss_collection_test, assistant_collection = [],[],[],[],[] # 用于收集整个过程中的正确率与误差
loss_24, loss_24_epoch, loss_24_collection,loss_24_collection_test = [],[],[],[] # 24个loss保存容器
#for ep in range(CFG['epoch']): # 遍历每一个epoch
for ep in range(50): # 遍历每一个epoch
# labda随着epoch调整
labda1 = 2 / (1 +
math.exp(-10 *
(ep) / CFG['epoch'])) - 1 # 随着ep的变化，labda会变化
CFG['lambda'] = 1 * labda1 # 调整分类损失和迁移损失之间的动态平衡
# 学习速率衰减
if ep > 50: # 设置学习速率从epoch=0开始衰减
if ('RevGrad' in CFG['tranmodel']): # RevGrad模型单独考虑
scheduler[0].step() # FE+TC的学习速率
scheduler[1].step() # DomClf的学习速率
else:
scheduler.step() # FE+TC的学习速率
model.train() # 开始训练，向模型申明目前是训练阶段，
# 训练监测指标的存储器
corr_src, corr_tar = 0, 0 # 正确样本数统计，对于没有投票功能的网络模型
corr_multi_src, corr_multi_tar = [0, 0, 0, 0,
0], [0, 0, 0, 0,
0] # 正确样本数统计，对于有投票多分类器的网络
corr_voted_src, corr_voted_tar = 0, 0 # 投票结果的正确数统计
loss_tran_multi, loss_clf_multi_src, loss_sum = [0, 0, 0, 0, 0], [
0, 0, 0, 0, 0
], [0, 0, 0, 0, 0] # loss分支记录
# 样本生成器
iter_src, iter_tar = iter(train_loader_src), iter(
train_loader_tar) # 样本batch的迭代器
n_batch = min(len_src_loader,
len_tar_loader) # batch数，当源域与目标域的batch数目不一样的时候使用
# 各类损失函数
loss_clf_fun = torch.nn.CrossEntropyLoss() # 交叉熵损失函数
#loss_clf_fun = F.nll_loss # CrossEntropyLoss()=log_softmax() + NLLLoss()
loss_mse_fun = torch.nn.MSELoss() # MSE损失函数
time_cost = 0 # 训练耗时
for ib in range(n_batch): # 遍历所有n_batch
data_src, label_src = iter_src.next() # 训练与测试样本的迭代器
data_tar, label_tar = iter_tar.next()
data_src, label_src = data_src.to(DEVICE), label_src.to(
DEVICE) # 将当前batch拷贝到指定的运算设备上
data_tar, label_tar = data_tar.to(DEVICE), label_tar.to(
DEVICE) # 每次只拷贝1个batch的样本进GPU，避免占用GPU内存过大
if ('RevGrad' in CFG['tranmodel']): #
optimizer[0].zero_grad() # 每个batch结束后都对梯度进行一次清理
optimizer[1].zero_grad() # 每个batch结束后都对梯度进行一次清理
else:
optimizer.zero_grad() # 每个batch结束后都对梯度进行一次清理
clfout_src, feamap_src = model(data_src) # 源域与目标域样本模型前馈
clfout_tar, feamap_tar = model(data_tar)
model.train()
###############################################
start_time = time.clock() # 训练开始时间
###############################################
''' MSTVM 投票网络'''
###############################################
if CFG['tranmodel'] == 'DDC_MSTVM':
loss_tran = 0
loss_multi_tran = []
for ic in range(len(feamap_src)): # 遍历每个灯笼分支
loss_tran_temp = loss_adapt_fun(feamap_src[ic],
feamap_tar[ic], 'mmd')
loss_tran += loss_tran_temp
loss_multi_tran.append(loss_tran_temp)
return_packt = voter_multi_classifier(
clfout_src[1:], clfout_tar[1:], label_src, label_tar,
corr_multi_src, corr_multi_tar, corr_voted_src,
corr_voted_tar)
corr_multi_src, corr_multi_tar, corr_voted_src, corr_voted_tar, loss_multi_clsclf_src, loss_multi_clsclf_tar, loss_clsclf_src, loss_clsclf_tar = return_packt
loss = loss_clsclf_src + CFG['lambda'] * loss_tran
loss_24 = [loss, 0, 0, 0, 0, 0,
loss_clsclf_src] + loss_multi_clsclf_src + [
loss_clsclf_tar
] + loss_multi_clsclf_tar + [loss_tran
] + loss_multi_tran
###############################################
if CFG['tranmodel'] == 'RevGrad_MSTVM':
loss_domclf_src, loss_multi_domclf_src = RevGrad_assistant(
model.DomClf_net, feamap_src, label_src, 0) # 梯度翻转的辅助函数
loss_domclf_tar, loss_multi_domclf_tar = RevGrad_assistant(
model.DomClf_net, feamap_tar, label_tar, 1)
loss_tran = loss_domclf_src + loss_domclf_tar
return_packt = voter_multi_classifier(
clfout_src[1:], clfout_tar[1:], label_src, label_tar,
corr_multi_src, corr_multi_tar, corr_voted_src,
corr_voted_tar)
corr_multi_src, corr_multi_tar, corr_voted_src, corr_voted_tar, loss_multi_clsclf_src, loss_multi_clsclf_tar, loss_clsclf_src, loss_clsclf_tar = return_packt
loss_domclf = loss_tran # 用于domclf的单独参数更新
loss = loss_clsclf_src + CFG['lambda'] * loss_tran
loss_multi_tran = []
for il in range(len(loss_multi_domclf_src)):
loss_multi_tran_temp = loss_multi_domclf_src[
il] + loss_multi_domclf_tar[il]
loss_multi_tran.append(loss_multi_tran_temp)
loss_24 = [loss, 0, 0, 0, 0, 0,
loss_clsclf_src] + loss_multi_clsclf_src + [
loss_clsclf_tar
] + loss_multi_clsclf_tar + [loss_tran
] + loss_multi_tran
###############################################
''' 汇总 '''
###############################################
# 损失函数反向传播
# 损失函数反向传播
if ('RevGrad' in CFG['tranmodel']):
loss_domclf.backward(retain_graph=True) # 域判别器 误差反向传播，求梯度
optimizer[1].step() # 域判别器 误将delta更新在权值与偏置值上
loss.backward() # 主干部分 误差反向传播，求梯度
optimizer[0].step() # 主干部分 将delta更新在权值与偏置值上
else:
loss.backward() # 误差反向传播，求梯度
optimizer.step() # 将delta更新在权值与偏置值上
# 时间结算
end_time = time.clock() # 训练结束时间
time_cost += end_time - start_time # 累计每一个batch的训练耗时
train_loss_clf_src.update(loss_clsclf_src.item()) # 在loss的和上，加一个值
train_loss_clf_tar.update(loss_clsclf_tar.item())
train_loss_tran.update(loss_tran.item())
train_loss_total.update(loss.item())
if ib % CFG['log_interval'] == 0: # 每隔多少个batch停顿一下
print(
'Train Epoch: [{}/{} ({:02d}%)], cls_Loss: {:.4f}, loss_tran: {:.4f}, total_Loss: {:.4f}'
.format(ep + 1, CFG['epoch'], int(100. * ib / n_batch),
train_loss_clf_src.avg, train_loss_tran.avg,
train_loss_total.avg))
# loss统计
for i24 in range(24):
loss_24[i24] = loss_24[i24].item(
) if loss_24[i24] != 0 else loss_24[i24] # 三目运算符
## 如果loss_24[i24]!=0，则表示loss_24[i24]是touch.tensor，因此使用.item()来将其转换为普通float； 如果loss_24[i24]==0，则表明其只是一个填充占位符，因此直接用
loss_24_epoch.append(
loss_24
) # # loss_24_epoch中记录着本格epoch中所有batch计算所得的24个loss，每一行就是一个batch所计算出的loss
loss_24_epoch_avg = [] # 求取平均损失的容器
num_batch = len(loss_24_epoch) # batch数
for i1 in range(24): # 遍历24个loss值，
loss_temp = 0
for i2 in range(num_batch): # 遍历每个batch
loss_temp += loss_24_epoch[i2][i1] # 把每个loss值沿着batch方向求和
loss_temp_avg = loss_temp / num_batch # 求每个batch的平均值
loss_24_epoch_avg.append(loss_temp_avg) # 本次epoch下的平均损失
loss_24_collection.append(loss_24_epoch_avg) # 将每个epoch下的24个平均损失收集起来
multi_acc_src, multi_acc_tar = [], [] # 多分类器的正确率容器
for ic in range(5): # 遍历5个分支分类器
acc_src_temp = 100. * corr_multi_src[ic].cpu().numpy() / len(
train_loader_src.dataset) # 利用分类正确的样本数计算正确率
acc_tar_temp = 100. * corr_multi_tar[ic].cpu().numpy() / len(
train_loader_tar.dataset)
multi_acc_src.append(
acc_src_temp) # multi_acc_src是个list，里面存放五个分支的正确率
multi_acc_tar.append(acc_tar_temp)
acc_voted_src = 100. * corr_voted_src.cpu().numpy() / len(
train_loader_src.dataset) # 投票后的正确率
acc_voted_tar = 100. * corr_voted_tar.cpu().numpy() / len(
train_loader_tar.dataset)
# 要return的结果
acc_all = [acc_voted_src] + multi_acc_src + [
acc_voted_tar
] + multi_acc_tar # 将训练所得的所有正确率汇集到一起
acc_collection.append(
acc_all) # # 从左到右依次是：1个源域投票正确率、5个源域分支正确率、1个目标域投票正确率、5个目标域分支正确率
loss_all = [
train_loss_clf_src.avg, train_loss_clf_tar.avg,
train_loss_tran.avg, train_loss_total.avg
]
loss_collection.append(
loss_all) # 从左到右依次是：源域测试集分类损失，目标域测试集分类损失，测试集迁移损失，测试集总损失
# 屏幕打印
print(multi_acc_src) # 打印当前epoch下的5个分支源域正确率
print(multi_acc_tar) # 打印当前epoch下的5个分支目标域正确率
print([acc_voted_src, acc_voted_tar]) # 打印源域和目标域投票后的正确率
print('Train: source_acc:{:.2f}%({}/{}), target_acc:{:.2f}%({}/{})\n'.
format(acc_voted_src, corr_voted_src,
len(train_loader_src.dataset), acc_voted_tar,
corr_voted_tar, len(train_loader_tar.dataset)))
print('########\n')
# 测试集返回
acc_collection_test_temp, loss_collection_test_temp, loss_24_epoch_avg_temp = test(
model, test_loader_src, test_loader_tar, ep, CFG)
acc_collection_test.append(acc_collection_test_temp) # 将测试正确率qppend进去
loss_collection_test.append(loss_collection_test_temp) # 将测试损失append进去
loss_24_collection_test.append(loss_24_epoch_avg_temp)
if ('RevGrad' in CFG['tranmodel']) or ('DCTLN' in CFG['tranmodel']):
assistant_collection.append(
[scheduler[0].get_lr()[0], CFG['lambda'],
time_cost]) # scheduler[0]yu scheduler[1]的lr衰减速率相同
else:
assistant_collection.append(
[scheduler.get_lr()[0], CFG['lambda'],
time_cost]) # 将辅助记录append进去
# 判断当前是否保存pt文件
result_trace = [
acc_collection, loss_collection, acc_collection_test,
loss_collection_test, assistant_collection
] # 所有的记录合并在一起
acc_test_tar_current = acc_collection_test_temp[
6] # 当前的测试正确率是第7个，也就是前面6个是训练正确率
if acc_test_tar_current > acc_test_tar_max: # 如果当前正确率大于最大正确率
acc_test_tar_max = acc_test_tar_current # 将最大正确率交换成当前正确率
if ('RevGrad' in CFG['tranmodel']):
pt = {
'model': model.state_dict(),
'optimizer0': optimizer[0].state_dict(),
'optimizer1': optimizer[1].state_dict(),
'scheduler0': scheduler[0].state_dict(),
'scheduler1': scheduler[1].state_dict(),
'CFG': CFG,
'result_trace': result_trace
}
else:
pt = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'CFG': CFG,
'result_trace': result_trace
}
torch.save(pt, CFG['dir_now'] + '/' + CFG['tranmodel'] +
'_maxacc.pt') # 并且，保存当前正确率下的模型
trace = [
acc_collection, loss_collection, loss_24_collection,
acc_collection_test, loss_collection_test, loss_24_collection_test,
assistant_collection
] # 记录器
return trace
load_model = False
if args.resume != '':
model_dict = model.state_dict()
pretrained_dict = torch.load(args.resume)
pretrained_dict = {
k: v
for k, v in pretrained_dict.items()
if k in model_dict and v.size() == model_dict[k].size()
}
model_dict.update(pretrained_dict)
model.load_state_dict(model_dict)
res_flag = 1
if args.cuda:
model.cuda()
if __name__ == '__main__':
best_loss = None
if load_model:
best_loss = test(0)
epoch = 0
output, label = tu.test(epoch, model, test_loader, args)
output = output.data.cpu().numpy()
print('score(s): \n')
print(output[:, 1])
print('total elapsed time: %0.2f min\n' % ((time.time() - ini_t_time) / 60.0))
### Test - stage 4 ### Part of: Blivet test collection ### Author: [email protected] ### This program is under GPL licence. import classes import test_utils loginst_test = test_utils.init_logging(0, None, True) test_utils.create_new_alloc_table("vdb") test_utils.create_new_partition("vdb", "primary", 1, -1) test_utils.format_new_partition("{}{}".format("vdb", 1), "ext4") loginst_test.debug("Setting up SystemPartitionFormatted_Scan") tspf = classes.SystemPartitionFormatted_Scan("vdb", 1) tbpf = classes.BlivetInitialization("vdb", 1).child loginst_test.debug("Comparing object attributes") ia = test_utils.test(tspf, tbpf) test_utils.write_issues(ia, "Single partition formatted", 4)
