def map_eval(eval_file, token_length, tokenizer, device, model, label_list):
model.eval()
datasets, labels = get_datasets(eval_file)
total_batches = 0
total_avp = 0.0
total_mrr = 0.0
# scores, labels = [], []
for k, dataset in tqdm(datasets.items(), desc="Eval datasets"):
examples = []
for i, data in enumerate(dataset):
examples.append(InputExample(i, data[0], data[1], '0'))
eval_features = convert_examples_to_features(examples, label_list,
token_length, tokenizer)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long).to(device)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long).to(device)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long).to(device)
# all_label_ids = torch.tensor(
# [f.label_id for f in eval_features], dtype=torch.long).to(device)
x_input_ids = torch.tensor([f.input_ids_x for f in eval_features],
dtype=torch.long).to(device)
x_input_mask = torch.tensor([f.input_mask_x for f in eval_features],
dtype=torch.long).to(device)
x_segment_ids = torch.tensor([f.segment_ids_x for f in eval_features],
dtype=torch.long).to(device)
y_input_ids = torch.tensor([f.input_ids_y for f in eval_features],
dtype=torch.long).to(device)
y_input_mask = torch.tensor([f.input_mask_y for f in eval_features],
dtype=torch.long).to(device)
y_segment_ids = torch.tensor([f.segment_ids_y for f in eval_features],
dtype=torch.long).to(device)
with torch.no_grad():
logits = model(x_input_ids, x_input_mask, x_segment_ids,
y_input_ids, y_input_mask, y_segment_ids,
all_input_ids, all_segment_ids, all_input_mask)
score = F.softmax(logits, dim=1)[:, 1].cpu().numpy()
label = np.array(list(map(int, labels[k])))
# print(score, label)
# scores.append(score)
# labels.append(label)
total_avp += mean_average_precision(label, score)
total_mrr += mean_reciprocal_rank(label, score)
total_batches += 1
mAP = total_avp / total_batches
mRR = total_mrr / total_batches
logger.info("map is : {}, mrr is : {}".format(mAP, mRR))
data = {'map': mAP, 'mrr': mRR}
with open('./result.json', 'w', encoding='utf-8') as f:
json.dump(data, f)
def evaluate(args,
model,
tokenizer,
processor,
eval_dataset,
matched_questions_indexs,
prefix=""):
results = {}
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
os.makedirs(args.output_dir)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
# multi-gpu eval
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
eval_sampler = SequentialSampler(eval_dataset)
eval_dataloader = DataLoader(eval_dataset,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
# Eval!
logger.info(" Num examples = %d", len(eval_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
epoch_iterator = tqdm(eval_dataloader,
desc="Eval_Iteration",
disable=args.local_rank not in [-1, 0])
eval_original_dataset, _ = load_and_cache_examples(
args, args.task_name, tokenizer, processor, data_type='eval_original')
eval_original_sampler = SequentialSampler(eval_original_dataset)
eval_original_dataloader = DataLoader(eval_original_dataset,
sampler=eval_original_sampler,
batch_size=args.eval_batch_size)
original_iterator = tqdm(eval_original_dataloader,
desc="Original_Iteration",
disable=args.local_rank not in [-1, 0])
original_embeddings = []
eval_question_embeddings = []
for step, batch in enumerate(original_iterator):
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
original_inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"token_type_ids": batch[2]
}
original_outputs = model(**original_inputs)[0].mean(1)
original_embeddings.append(original_outputs)
original_embeddings = torch.cat(
[embed.cpu().data for embed in original_embeddings]).numpy()
for step, batch in enumerate(epoch_iterator):
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
eval_question_inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"token_type_ids": batch[2]
}
eval_questions_outputs = model(**eval_question_inputs)[0].mean(1)
eval_question_embeddings.append(eval_questions_outputs)
eval_question_embeddings = torch.cat(
[o.cpu().data for o in eval_question_embeddings]).numpy()
scores = cosine_similarity(eval_question_embeddings, original_embeddings)
sorted_indices = scores.argsort()[:, ::-1]
mmr = mean_reciprocal_rank(matched_questions_indexs == sorted_indices)
map = mean_average_precision(matched_questions_indexs == sorted_indices)
print("mean reciprocal rank: {}".format(mmr))
print("mean average precision: {}".format(map))
results['mmr'] = mmr
results['map'] = map
output_eval_file = os.path.join(args.output_dir, prefix,
"eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in results.items():
logger.info(" %s = %s", key, str(value))
writer.write("%s = %s\n" % (key, value))
return results
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--train_file",
default='../data/right_samples.csv',
type=str,
required=False,
help="training file")
parser.add_argument("--evaluate_file",
default='../data/eval_touzi.xlsx',
type=str,
required=False,
help="training file")
parser.add_argument("--do_evaluate",
action="store_true",
help="Whether to run evaluate.")
parser.add_argument("--do_predict",
default=True,
action="store_true",
help="Whether to run predict.")
parser.add_argument("--batch_size",
default=16,
type=int,
required=False,
help="batch size for train and eval")
args = parser.parse_args()
if not os.path.exists("embeddings.pkl"):
train_df = pd.read_csv(args.train_file, sep='\t')
candidate_title = train_df['best_title'].tolist()
candidate_reply = train_df["reply"].tolist()
titels = []
for title in tqdm(candidate_title, desc='对原问题进行分词ing'):
titels.append(seg.cut(title))
embeddings = []
for i in trange(0, len(titels), 16, desc='获取ELMo的句子表示'):
mini_embeddings = e.sents2elmo(titels[i:min(len(titels), i + 16)])
for mini_embedding in mini_embeddings:
# 获取句子向量，对词取平均
embeddings.append(np.mean(mini_embedding, axis=0))
if i == 0:
print(len(embeddings))
print(embeddings[0].shape)
print("原始问题句子向量表示获取完毕，保存ing")
with open("embeddings.pkl", 'wb') as fout:
pickle.dump([candidate_title, candidate_reply, embeddings], fout)
else:
with open("embeddings.pkl", 'rb') as fint:
candidate_title, candidate_reply, embeddings = pickle.load(fint)
if args.do_evaluate:
evulate_df = pd.read_excel(args.evaluate_file, '投资知道')
# code.interact(local=locals())
evulate_df = evulate_df[['问题', '匹配问题']]
evulate_df = evulate_df[evulate_df['问题'].notna()]
evulate_df = evulate_df[evulate_df['匹配问题'].notna()]
questions = evulate_df['问题'].tolist()
matched_questions = evulate_df['匹配问题'].tolist()
matched_questions_indexs = []
for k, q in enumerate(matched_questions):
flag = False
for i, _q in enumerate(candidate_title):
if q == _q:
matched_questions_indexs.append([i])
flag = True
break
if not flag:
matched_questions_indexs.append([-1])
matched_questions_indexs = np.asarray(matched_questions_indexs)
# print("size of matched_questions_index:{}".format(matched_questions_indexs.shape))
questions = [seg.cut(question.strip()) for question in questions]
question_embedding = [
np.mean(emb, 0) for emb in e.sents2elmo(questions)
]
scores = cosine_similarity(question_embedding, embeddings)
# print("scores shape : {}".format(scores.shape))
sorted_indices = scores.argsort()[:, ::-1]
# code.interact(local=locals())
mmr = mean_reciprocal_rank(sorted_indices == matched_questions_indexs)
map = mean_average_precision(
sorted_indices == matched_questions_indexs)
logger.info("mean reciprocal rank: {}".format(mmr))
logger.info("mean average precision: {}".format(map))
if args.do_predict:
while True:
title = input("你的问题是？\n")
if len(str(title).strip()) == 0:
continue
title = [seg.cut(str(title).strip())]
title_embedding = np.mean(e.sents2elmo(title)[0], 0).reshape(1, -1)
scores = cosine_similarity(title_embedding, embeddings)[0]
top5_indices = scores.argsort()[-5:][::-1]
for index in top5_indices:
print("可能的答案，参考问题：" + candidate_title[index] + "\t答案：" +
candidate_reply[index] + "\t得分：" + str(scores[index]))
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--data_dir",
default="../data/right_samples.csv",
type=str,
required=False,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task.",
)
parser.add_argument(
"--evaluate_dir",
default="../data/eval_touzi.xlsx",
type=str,
required=False,
help=
"The evaluate data dir. Should contain the .tsv files (or other data files) for the task.",
)
parser.add_argument(
"--model_type",
default="bert",
type=str,
required=False,
help="Model type selected in the list:",
)
parser.add_argument(
"--model_name_or_path",
default='D:\\NLP\\my-wholes-models\\chinese_wwm_pytorch\\',
type=str,
required=False,
help="Path to pre-trained model or shortcut name selected in the list",
)
parser.add_argument(
"--output_type",
default="avg",
type=str,
required=False,
choices=["pooled", "avg"],
help="the type of choice output vector",
)
parser.add_argument(
"--task_name",
default="faq",
type=str,
required=False,
help="The name of the task to train selected in the list: " +
", ".join(processors.keys()),
)
# Other parameters
parser.add_argument(
"--config_name",
default="",
type=str,
help="Pretrained config name or path if not the same as model_name")
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help="Pretrained tokenizer name or path if not the same as model_name",
)
parser.add_argument(
"--cache_dir",
default="",
type=str,
help=
"Where do you want to store the pre-trained models downloaded from s3",
)
parser.add_argument(
"--max_seq_length",
default=256,
type=int,
help=
"The maximum total input sequence length after tokenization. Sequences longer "
"than this will be truncated, sequences shorter will be padded.",
)
parser.add_argument("--do_predict",
default=True,
action="store_true",
help="Whether to run training.")
parser.add_argument("--do_eval",
action="store_true",
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
default=True,
action="store_true",
help="Set this flag if you are using an uncased model.")
parser.add_argument("--per_gpu_train_batch_size",
default=8,
type=int,
help="Batch size per GPU/CPU for training.")
parser.add_argument("--per_gpu_eval_batch_size",
default=8,
type=int,
help="Batch size per GPU/CPU for evaluation.")
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay",
default=0.0,
type=float,
help="Weight decay if we apply some.")
parser.add_argument("--adam_epsilon",
default=1e-8,
type=float,
help="Epsilon for Adam optimizer.")
parser.add_argument("--max_grad_norm",
default=1.0,
type=float,
help="Max gradient norm.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--max_steps",
default=-1,
type=int,
help=
"If > 0: set total number of training steps to perform. Override num_train_epochs.",
)
parser.add_argument("--warmup_steps",
default=0,
type=int,
help="Linear warmup over warmup_steps.")
parser.add_argument("--logging_steps",
type=int,
default=1000,
help="Log every X updates steps.")
parser.add_argument("--save_steps",
type=int,
default=5000,
help="Save checkpoint every X updates steps.")
parser.add_argument(
"--eval_all_checkpoints",
action="store_true",
help=
"Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
)
parser.add_argument("--no_cuda",
action="store_true",
help="Avoid using CUDA when available")
parser.add_argument(
"--overwrite_cache",
action="store_true",
help="Overwrite the cached training and evaluation sets")
parser.add_argument("--seed",
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
"--fp16",
action="store_true",
help=
"Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
)
parser.add_argument(
"--fp16_opt_level",
type=str,
default="O1",
help=
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html",
)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="For distributed training: local_rank")
parser.add_argument("--server_ip",
type=str,
default="",
help="For distant debugging.")
parser.add_argument("--server_port",
type=str,
default="",
help="For distant debugging.")
args = parser.parse_args()
# Setup distant debugging if needed
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
# Setup CUDA, GPU & distributed training
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend="nccl")
args.n_gpu = 1
args.device = device
# Setup logging
# set_log(logger)
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
filename="BERT.log",
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
args.local_rank,
device,
args.n_gpu,
bool(args.local_rank != -1),
args.fp16,
)
set_seed(args)
args.task_name = args.task_name.lower()
args.model_type = args.model_type.lower()
config_class, model_class, tokenizer_class = MODEL_CLASS[args.model_type]
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
finetuning_task=args.task_name,
cache_dir=args.cache_dir if args.cache_dir else "./cache")
tokenizer = tokenizer_class.from_pretrained(
args.tokenizer_name
if args.tokenizer_name else args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else "./cache")
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool('.ckpt' in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else "./cache")
model = model.to(device)
if not os.path.exists("embeddings.pkl"):
eval_dataset, candidate_title, candidate_reply = load_examples(
args, tokenizer)
outputs = evaluate(args, model, eval_dataset)
candidate_embeddings = torch.cat([o.cpu().data
for o in outputs]).numpy()
torch.save([candidate_title, candidate_reply, candidate_embeddings],
"embeddings.pkl")
else:
candidate_title, candidate_reply, candidate_embeddings = torch.load(
"embeddings.pkl")
if args.do_eval:
evulate_df = pd.read_excel(args.evaluate_dir, '投资知道')
evulate_df = evulate_df[['问题', '匹配问题']]
evulate_df = evulate_df[evulate_df['问题'].notna()]
evulate_df = evulate_df[evulate_df['匹配问题'].notna()]
questions = evulate_df['问题'].tolist()
matched_questions = evulate_df['匹配问题'].tolist()
matched_questions_indexs = []
# 找出匹配问题对应的index
for k, q in enumerate(matched_questions):
flag = False
for i, _q in enumerate(candidate_title):
if q == _q:
matched_questions_indexs.append([i])
flag = True
break
if not flag:
matched_questions_indexs.append([-1])
matched_questions_indexs = np.asarray(matched_questions_indexs)
examples = [
InputExample(guid='eva_' + str(idx),
text_a=title,
text_b=None,
label=1) for idx, title in enumerate(questions)
]
dataset = from_examples2dataset(args, examples, tokenizer)
outputs = evaluate(args, model, dataset)
question_embedding = torch.cat([o.cpu().data for o in outputs]).numpy()
scores = cosine_similarity(question_embedding, candidate_embeddings)
sorted_indices = scores.argsort()[:, ::-1]
mmr = mean_reciprocal_rank(matched_questions_indexs == sorted_indices)
map = mean_average_precision(
matched_questions_indexs == sorted_indices)
print("mean reciprocal rank: {}".format(mmr))
print("mean average precision: {}".format(map))
logger.info("====" * 100)
logger.info("mean reciprocal rank: {}".format(mmr))
logger.info("mean average precision: {}".format(map))
logger.info("====" * 100)
if args.do_predict:
while True:
question = input("你的问题是?\n")
if len(str(question)) == 0:
continue
examples = [
InputExample(guid=0, text_a=question, text_b=None, label=1)
]
dataset = from_examples2dataset(args, examples, tokenizer)
outputs = evaluate(args, model, dataset)
question_embedding = torch.cat([o.cpu().data
for o in outputs]).numpy()
scores = cosine_similarity(question_embedding,
candidate_embeddings)[0]
top5 = scores.argsort()[-5:][::-1]
for index in top5:
print("可能得答案，参考问题为:{},答案:{},得分:{}".format(
candidate_title[index], candidate_reply[index],
str(scores[index])))
print()
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--train_file",
default=None,
type=str,
required=True,
help="training file")
parser.add_argument("--batch_size",
default=64,
type=int,
required=False,
help="batch size for train and eval")
args = parser.parse_args()
# load dataset
if not os.path.exists(args.train_file + "_embeddings.pkl"):
train_df = pd.read_csv(args.train_file)
candidates = train_df[train_df["is_best"] == 1][["title", "reply"]]
candidate_title = candidates["title"].tolist()
candidate_reply = candidates["reply"].tolist()
titles = [seg.cut(title) for title in candidates["title"]]
embeddings = e.sents2elmo(titles)
# list of numpy arrays, each array with shape [seq_len * 1024]
# code.interact(local=locals())
candidate_embeddings = [
np.mean(embedding, 0) for embedding in embeddings
] # a list of 1024 dimensional vectors
with open(args.train_file + "_embeddings.pkl", "wb") as fout:
pickle.dump(
[candidate_title, candidate_reply, candidate_embeddings], fout)
else:
with open(args.train_file + "_embeddings.pkl", "rb") as fin:
candidate_title, candidate_reply, candidate_embeddings = pickle.load(
fin)
df = pd.read_excel("../../dataset/faq/验证数据.xlsx", "投资知道")
df = df[["问题", "匹配问题"]]
df = df[df["匹配问题"].notna()]
df = df[df["问题"].notna()]
questions = df["问题"].tolist()
matched_questions = df["匹配问题"].tolist()
matched_questions_index = []
for q in matched_questions:
flg = False
for i, _q in enumerate(candidate_title):
if q == _q:
matched_questions_index.append([i])
flg = True
break
if flg == False:
matched_questions_index.append([-1])
matched_questions_index = np.asarray(matched_questions_index)
questions = [seg.cut(q.strip()) for q in questions]
question_embedding = [np.mean(emb, 0) for emb in e.sents2elmo(questions)
] # 得到了新问题的ELMo embedding
scores = cosine_similarity(question_embedding,
candidate_embeddings) # 计算新问题和所有候选问题的分数
sorted_indices = scores.argsort()[:, ::-1] #[-5:][::-1] # 对分数做排名
# code.interact(local=locals())
mmr = mean_reciprocal_rank(
sorted_indices ==
matched_questions_index) # 利用分数排名来计算mean receiprocal rank
print("mean reciprocal rank: {}".format(mmr))
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task.",
)
parser.add_argument(
"--model_type",
default=None,
type=str,
required=True,
help="Model type selected in the list: " +
", ".join(MODEL_CLASSES.keys()),
)
parser.add_argument(
"--model_name_or_path",
default=None,
type=str,
required=True,
help="Path to pre-trained model or shortcut name selected in the list",
)
parser.add_argument(
"--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train selected in the list: " +
", ".join(processors.keys()),
)
# Other parameters
parser.add_argument(
"--config_name",
default="",
type=str,
help="Pretrained config name or path if not the same as model_name")
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help="Pretrained tokenizer name or path if not the same as model_name",
)
parser.add_argument(
"--cache_dir",
default="",
type=str,
help=
"Where do you want to store the pre-trained models downloaded from s3",
)
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after tokenization. Sequences longer "
"than this will be truncated, sequences shorter will be padded.",
)
parser.add_argument("--do_train",
action="store_true",
help="Whether to run training.")
parser.add_argument("--do_eval",
action="store_true",
help="Whether to run eval on the dev set.")
parser.add_argument(
"--evaluate_during_training",
action="store_true",
help="Rul evaluation during training at each logging step.")
parser.add_argument(
"--do_lower_case",
action="store_true",
help="Set this flag if you are using an uncased model.")
parser.add_argument("--per_gpu_train_batch_size",
default=8,
type=int,
help="Batch size per GPU/CPU for training.")
parser.add_argument("--per_gpu_eval_batch_size",
default=8,
type=int,
help="Batch size per GPU/CPU for evaluation.")
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay",
default=0.0,
type=float,
help="Weight decay if we apply some.")
parser.add_argument("--adam_epsilon",
default=1e-8,
type=float,
help="Epsilon for Adam optimizer.")
parser.add_argument("--max_grad_norm",
default=1.0,
type=float,
help="Max gradient norm.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--max_steps",
default=-1,
type=int,
help=
"If > 0: set total number of training steps to perform. Override num_train_epochs.",
)
parser.add_argument("--warmup_steps",
default=0,
type=int,
help="Linear warmup over warmup_steps.")
parser.add_argument("--logging_steps",
type=int,
default=50,
help="Log every X updates steps.")
parser.add_argument("--save_steps",
type=int,
default=50,
help="Save checkpoint every X updates steps.")
parser.add_argument(
"--eval_all_checkpoints",
action="store_true",
help=
"Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
)
parser.add_argument("--no_cuda",
action="store_true",
help="Avoid using CUDA when available")
parser.add_argument(
"--overwrite_cache",
action="store_true",
help="Overwrite the cached training and evaluation sets")
parser.add_argument("--seed",
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
"--fp16",
action="store_true",
help=
"Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
)
parser.add_argument(
"--fp16_opt_level",
type=str,
default="O1",
help=
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html",
)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="For distributed training: local_rank")
parser.add_argument("--server_ip",
type=str,
default="",
help="For distant debugging.")
parser.add_argument("--server_port",
type=str,
default="",
help="For distant debugging.")
args = parser.parse_args()
# Setup distant debugging if needed
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
# Setup CUDA, GPU & distributed training
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend="nccl")
args.n_gpu = 1
args.device = device
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
args.local_rank,
device,
args.n_gpu,
bool(args.local_rank != -1),
args.fp16,
)
# Set seed
set_seed(args)
# Prepare GLUE task
args.task_name = args.task_name.lower()
# Load pretrained model and tokenizer
if args.local_rank not in [-1, 0]:
torch.distributed.barrier(
) # Make sure only the first process in distributed training will download model & vocab
args.model_type = args.model_type.lower()
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
finetuning_task=args.task_name,
cache_dir=args.cache_dir if args.cache_dir else None,
)
tokenizer = tokenizer_class.from_pretrained(
args.tokenizer_name
if args.tokenizer_name else args.model_name_or_path,
do_lower_case=args.do_lower_case,
cache_dir=args.cache_dir if args.cache_dir else None,
)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None,
)
if args.local_rank == 0:
torch.distributed.barrier(
) # Make sure only the first process in distributed training will download model & vocab
model.to(args.device)
if not os.path.exists(args.data_dir + "_bert_embeddings.pkl"):
logger.info("Training/evaluation parameters %s", args)
eval_dataset, candidate_title, candidate_reply = load_examples(
args, args.task_name, tokenizer)
sequence_outputs = evaluate(args, model, eval_dataset)
candidate_embeddings = torch.cat([o for o in sequence_outputs]).numpy()
with open(args.data_dir + "_bert_embeddings.pkl", "wb") as fout:
pickle.dump(
[candidate_title, candidate_reply, candidate_embeddings], fout)
# 18677 * 768
else:
with open(args.data_dir + "_bert_embeddings.pkl", "rb") as fin:
candidate_title, candidate_reply, candidate_embeddings = pickle.load(
fin)
# code.interact(local=locals())
# load dataset
# if not os.path.exists("embeddings.pkl"):
# # code.interact(local=locals())
# candidate_embeddings = [np.mean(embedding, 0) for embedding in embeddings]
# with open("embeddings.pkl", "wb") as fout:
# pickle.dump([candidate_title, candidate_reply, embeddings], fout)
# else:
# with open("embeddings.pkl", "rb") as fin:
# candidate_title, candidate_reply, embeddings = pickle.load(fin)
df = pd.read_excel("../../dataset/faq/验证数据.xlsx", "投资知道")
df = df[["问题", "匹配问题"]]
df = df[df["匹配问题"].notna()]
df = df[df["问题"].notna()]
questions = df["问题"].tolist()
matched_questions = df["匹配问题"].tolist()
matched_questions_index = []
for q in matched_questions:
flg = False
for i, _q in enumerate(candidate_title):
if q == _q:
matched_questions_index.append([i])
flg = True
break
if flg == False:
matched_questions_index.append([-1])
matched_questions_index = np.asarray(matched_questions_index)
examples = [
InputExample(guid=0, text_a=title, text_b=None, label=1)
for title in questions
]
features = convert_examples_to_features(
examples,
tokenizer,
label_list=[1],
output_mode="classification",
max_length=args.max_seq_length,
pad_on_left=bool(
args.model_type in ["xlnet"]), # pad on the left for xlnet
pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
pad_token_segment_id=4 if args.model_type in ["xlnet"] else 0,
)
# Convert to Tensors and build dataset
all_input_ids = torch.tensor([f.input_ids for f in features],
dtype=torch.long)
all_attention_mask = torch.tensor([f.attention_mask for f in features],
dtype=torch.long)
all_token_type_ids = torch.tensor([f.token_type_ids for f in features],
dtype=torch.long)
dataset = TensorDataset(all_input_ids, all_attention_mask,
all_token_type_ids)
sequence_outputs = evaluate(args, model, dataset)
# code.interact(local=locals())
question_embedding = torch.cat([o for o in sequence_outputs]).numpy()
scores = cosine_similarity(question_embedding, candidate_embeddings)
sorted_indices = scores.argsort()[:, ::-1] #[-5:][::-1]
# code.interact(local=locals())
mmr = mean_reciprocal_rank(sorted_indices == matched_questions_index)
print("mean reciprocal rank: {}".format(mmr))