def main():
# if args.use_cuda is True:
# use deterministic algorithm for cnn
# torch.backends.cudnn.deterministic = True
args.kernel_sizes = [int(k) for k in args.kernel_sizes.split(',')]
# save file
mulu = datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S')
args.mulu = mulu
args.save_dir = os.path.join(args.save_dir, mulu)
if not os.path.isdir(args.save_dir):
os.makedirs(args.save_dir)
Temp_Test_Result = "./Temp_Test_Result"
if os.path.exists(Temp_Test_Result):
shutil.rmtree(Temp_Test_Result)
if not os.path.isdir(Temp_Test_Result):
os.makedirs(Temp_Test_Result)
pretrained_text_field = data.Field(lower=True)
pretrained_label_field = data.Field(sequential=False)
build_Pretrained_vocab(pretrained_text_field,
pretrained_label_field,
path_file=args.word_Embedding_Path)
# print(text_field.vocab.stoi)
args.pretrained_text_field = pretrained_text_field
args.pretrained_label_field = pretrained_label_field
args.embed_num = len(pretrained_text_field.vocab)
embed, pretrained_embed_dim = model_SumPooling_Pretrained.load_pretrain(
file=args.word_Embedding_Path, args=args)
args.embed = embed
args.pretrained_embed_dim = pretrained_embed_dim
cv_result = []
# CV loop start
for id in range(args.nfold):
print("\nthe {} CV file".format(id))
# build vocab and iterator
text_field = data.Field(lower=True)
label_field = data.Field(sequential=False)
cv_spilit_file(args.train_path, args.nfold, test_id=id)
train_iter, test_iter = load_data(text_field,
label_field,
path_file=args.train_path,
device=args.gpu_device,
repeat=False,
shuffle=args.epochs_shuffle,
sort=False)
args.text_field = text_field
args.class_num = len(label_field.vocab) - 1
args.PaddingID = pretrained_text_field.vocab.stoi[
pretrained_text_field.pad_token]
print("embed_num : {}, class_num : {}".format(args.embed_num,
args.class_num))
print("PaddingID {}".format(args.PaddingID))
# pretrained word embedding
# if args.word_Embedding:
# pretrain_embed = load_pretrained_emb_zeros(path=args.word_Embedding_Path,
# text_field_words_dict=text_field.vocab.itos,
# pad=text_field.pad_token)
# calculate_oov(path=args.word_Embedding_Path, text_field_words_dict=text_field.vocab.itos,
# pad=text_field.pad_token)
# args.pretrained_weight = pretrain_embed
# print params
show_params()
# load model and start train
if args.CNN is True:
print("loading SumPooling model.....")
# model = model_CNN.CNN_Text(args)
model = model_SumPooling_Pretrained.SumPooling(args)
# for param in model.parameters():
# param.requires_grad = False
shutil.copy("./models/model_SumPooling_Pretrained.py",
args.save_dir)
print(model)
if args.use_cuda is True:
print("using cuda......")
model = model.cuda()
print("CNN training start......")
if os.path.exists("./Test_Result.txt"):
os.remove("./Test_Result.txt")
model_count = train_CV.train(train_iter, test_iter, model, args)
cv_result.append(calculate_result(id=id))
print(cv_result)
cv_mean = cal_mean(cv_result)
print("The best result is {:.6f} ".format(cv_mean))
file = open("./Temp_Test_Result/Final_Result.txt", "a")
for index, value in enumerate(cv_result):
stra = str(index + 1) + " " + str(value)
file.write(stra)
file.write("\n")
file.write("mean_value " + str(cv_mean))
file.close()
shutil.copytree("./Temp_Test_Result/",
"./snapshot/" + mulu + "/Temp_Test_Result")
def main():
# if args.use_cuda is True:
# use deterministic algorithm for cnn
# torch.backends.cudnn.deterministic = True
args.kernel_sizes = [int(k) for k in args.kernel_sizes.split(',')]
# save file
mulu = datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S')
args.mulu = mulu
args.save_dir = os.path.join(args.save_dir, mulu)
if not os.path.isdir(args.save_dir):
os.makedirs(args.save_dir)
# build vocab and iterator
text_field = data.Field(lower=True)
label_field = data.Field(sequential=False)
if args.SST_1 is True:
print("loading sst-1 dataset......")
train_iter, dev_iter, test_iter = load_SST_1(
text_field,
label_field,
train_path=args.train_path,
dev_path=args.dev_path,
test_path=args.test_path,
device=args.gpu_device,
repeat=False,
shuffle=args.epochs_shuffle,
sort=False)
if args.SST_2 is True:
print("loading sst-2 dataset......")
train_iter, dev_iter, test_iter = load_SST_2(
text_field,
label_field,
train_path=args.train_path,
dev_path=args.dev_path,
test_path=args.test_path,
device=args.gpu_device,
repeat=False,
shuffle=args.epochs_shuffle,
sort=False)
if args.TREC is True:
print("loading TREC dataset......")
train_iter, test_iter = load_TREC(text_field,
label_field,
train_path=args.train_path,
test_path=args.test_path,
device=args.gpu_device,
repeat=False,
shuffle=args.epochs_shuffle,
sort=False)
args.embed_num = len(text_field.vocab)
args.class_num = len(label_field.vocab) - 1
args.PaddingID = text_field.vocab.stoi[text_field.pad_token]
print("embed_num : {}, class_num : {}".format(args.embed_num,
args.class_num))
print("PaddingID {}".format(args.PaddingID))
# pretrained word embedding
if args.word_Embedding:
pretrain_embed = load_pretrained_emb_zeros(
path=args.word_Embedding_Path,
text_field_words_dict=text_field.vocab.itos,
pad=text_field.pad_token)
calculate_oov(path=args.word_Embedding_Path,
text_field_words_dict=text_field.vocab.itos,
pad=text_field.pad_token)
args.pretrained_weight = pretrain_embed
# print params
show_params()
# load model and start train
if args.CNN is True:
print("loading CNN model.....")
# model = model_CNN.CNN_Text(args)
model = model_SumPooling.SumPooling(args)
# for param in model.parameters():
# param.requires_grad = False
shutil.copy("./models/model_CNN.py", args.save_dir)
print(model)
if args.use_cuda is True:
print("using cuda......")
model = model.cuda()
print("CNN training start......")
if os.path.exists("./Test_Result.txt"):
os.remove("./Test_Result.txt")
if args.SST_1 is True or args.SST_2 is True:
model_count = train_ALL_CNN.train(train_iter, dev_iter, test_iter,
model, args)
if args.TREC is True:
model_count = train_CV.train(train_iter, test_iter, model, args)
# calculate the best result
cal_result()
def main():
assert args.test_interval == args.dev_interval
assert args.CV is False, "CrossFold Dataset, Please Run main_hyperparams_CV.py"
# if args.use_cuda is True:
# use deterministic algorithm for cnn
# torch.backends.cudnn.deterministic = True
# save file
mulu = datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S')
args.mulu = mulu
args.save_dir = os.path.join(args.save_dir, mulu)
if not os.path.isdir(args.save_dir):
os.makedirs(args.save_dir)
# build vocab and iterator
text_field = data.Field(lower=True)
label_field = data.Field(sequential=False)
if args.IMDB is True:
print("loading IMDB dataset......")
train_iter, test_iter = load_IMDB(
text_field,
label_field,
train_path=args.train_path,
test_path=args.test_path,
device=args.gpu_device,
repeat=False,
shuffle=args.epochs_shuffle,
sort=False,
)
args.embed_num = len(text_field.vocab)
args.class_num = len(label_field.vocab) - 1
args.PaddingID = text_field.vocab.stoi[text_field.pad_token]
print("embed_num : {}, class_num : {}".format(args.embed_num,
args.class_num))
print("PaddingID {}".format(args.PaddingID))
# pretrained word embedding
if args.word_Embedding:
pretrain_embed = load_pretrained_emb_zeros(
path=args.word_Embedding_Path,
text_field_words_dict=text_field.vocab.itos,
pad=text_field.pad_token)
calculate_oov(path=args.word_Embedding_Path,
text_field_words_dict=text_field.vocab.itos,
pad=text_field.pad_token)
args.pretrained_weight = pretrain_embed
# print params
show_params()
# load model and start train
if args.SumPooling is True:
print("loading SumPooling model.....")
model = model_SumPooling.SumPooling(args)
shutil.copy("./models/model_SumPooling.py", args.save_dir)
print(model)
if args.use_cuda is True:
print("using cuda......")
model = model.cuda()
print("CNN training start......")
if os.path.exists("./Test_Result.txt"):
os.remove("./Test_Result.txt")
if args.IMDB is True:
model_count = train_CV.train(train_iter, test_iter, model, args)
# calculate the best result
cal_result()