N = 20
figsize = (10, 8)
axis_off = True
# ***************************************************************************
model_name_norm = 'FT' if model_name != 'SIMPLE' else model_name
path_man = PathManager(dataset=dataset_name,
d_type=dataset_subtype,
model_name=model_name_norm,
version=version)
################################################
#----------------------读取模型参数------------------
################################################
if model_name == 'SIMPLE':
model_cfg = TrainingConfigManager(path_man.Doc() + 'config.json')
else:
model_cfg = TrainingConfigManager('../run/runConfig.json')
modelParams = model_cfg.modelParams()
dataset = SeqFileDataset(path_man.FileData(), path_man.FileSeqLen(), N=N)
dataloader = DataLoader(dataset,
batch_size=N,
collate_fn=batchSequenceWithoutPad)
if model_name != 'Random':
state_dict = t.load(path_man.Model() + '_v%s.0' % version)
word_matrix = state_dict['Embedding.weight']
else:
word_matrix = t.Tensor(
cuda=True, expand=expand,
parallel=modelParams['data_parallel_devices'])
val_task = ImpEpisodeTask(k, qk, n, N, val_dataset,
cuda=True, expand=expand,
parallel=modelParams['data_parallel_devices'])
else:
train_task = ProtoEpisodeTask(k, qk, n, N, train_dataset,
cuda=True, expand=expand,
parallel=modelParams['data_parallel_devices'])
val_task = ProtoEpisodeTask(k, qk, n, N, val_dataset,
cuda=True, expand=expand,
parallel=modelParams['data_parallel_devices'])
stat = TrainStatManager(model_save_path=train_path_manager.Model(),
stat_save_path=train_path_manager.Doc(),
train_report_iter=ValCycle,
criteria=criteria)
if RecordGradient:
types.append('line')
titles.append('gradient')
xlabels.append('iterations')
ylabels.append('gradient norm')
legends.append(['Encoder Gradient'])
if UseVisdom:
plot = VisdomPlot(env_title='train monitoring',
types=types,
titles=titles,
xlabels=xlabels,
################################################
printState('init model...')
word_matrix = t.Tensor(np.load(test_path_manager.WordEmbedMatrix(), allow_pickle=True))
loss = t.nn.NLLLoss().cuda() if loss_func_name == 'nll' else t.nn.MSELoss().cuda()
model = FT(n=n, loss_fn=loss,
pretrained_matrix=word_matrix, **modelParams)
# model.load_state_dict(state_dict)
model = model.cuda()
statParamNumber(model)
if os.path.exists(test_path_manager.Doc()):
deleteDir(test_path_manager.Doc())
os.mkdir(test_path_manager.Doc())
shutil.copy('../models/FT.py', test_path_manager.Doc()+"FT.py")
shutil.copy('./ftConfig.json', test_path_manager.Doc()+"config.json")
print('doc path:', test_path_manager.Doc())
stat.startTimer()
################################################
#--------------------开始测试------------------
################################################
metrics = np.zeros(4,)
with t.autograd.set_detect_anomaly(False):
for epoch in range(TestingEpoch):
################################################
#----------------------定义数据------------------
################################################
printState('init managers...')
test_path_manager = PathManager(dataset=data_folder,
d_type=USED_SUB_DATASET,
model_name=model_name,
version=version)
################################################
#----------------------读取模型参数------------------
################################################
param_config_path = 'runConfig.json' if MODEL_RANDOM_STATE else test_path_manager.Doc(
) + 'config.json'
model_cfg = TrainingConfigManager(param_config_path)
modelParams = model_cfg.modelParams()
LRDecayIters, LRDecayGamma, optimizer_type,\
weight_decay, loss_func, default_lr, lrs, \
taskBatchSize, criteria = model_cfg.trainingParams()
test_dataset = SeqFileDataset(test_path_manager.FileData(),
test_path_manager.FileSeqLen(), N)
expand = True if loss_func == 'mse' else False
if model_type in ADAPTED_MODELS:
test_task = AdaptEpisodeTask(k,