bojone / cdial-gpt-tf Goto Github PK
View Code? Open in Web Editor NEW用bert4keras加载CDial-GPT
用bert4keras加载CDial-GPT
设置了环境变量 TF_KERAS=1
代码如下:
#! -*- coding: utf-8 -*-
# bert4keras加载CDial-GPT
# https://github.com/bojone/CDial-GPT-tf
import numpy as np
from bert4keras.models import build_transformer_model
from bert4keras.tokenizers import Tokenizer
from bert4keras.snippets import AutoRegressiveDecoder
from bert4keras.snippets import uniout
from tensorflow import keras
import json
import requests
config_path = './GPT_LCCC-large-tf/config.json'
checkpoint_path = './GPT_LCCC-large-tf/gpt_model.ckpt'
dict_path = './GPT_LCCC-large-tf/vocab.txt'
tokenizer = Tokenizer(dict_path, do_lower_case=True) # 建立分词器
speakers = [
tokenizer.token_to_id('[speaker1]'),
tokenizer.token_to_id('[speaker2]')
]
model = build_transformer_model(
config_path=config_path,
checkpoint_path=checkpoint_path,
model='gpt_openai'
) # 建立模型,加载权重
print(type(model))
model.summary(line_length=150)
# model.save('./b4k_model')
# k_model = keras.models.load_model('./b4k_model')
mls_uri = 'http://cdial-gpt.t9k-sample.n.ksvc.tensorstack.net/v1/models/cdial-gpt:predict'
class ChatBot(AutoRegressiveDecoder):
"""基于随机采样对话机器人
"""
@AutoRegressiveDecoder.wraps(default_rtype='probas')
def predict(self, inputs, output_ids, states):
token_ids, segment_ids = inputs
curr_segment_ids = np.zeros_like(output_ids) + token_ids[0, -1]
token_ids = np.concatenate([token_ids, output_ids], 1)
segment_ids = np.concatenate([segment_ids, curr_segment_ids], 1)
serving_url = mls_uri
instances = [{'Input-Segment': segment_ids.tolist(), 'Input-Token': token_ids.tolist()}]
data = {'signature_name': 'serving_default', 'instances': instances}
data = json.dumps(data)
headers = {"content-type": "application/json"}
json_response = requests.post(serving_url, data=data, headers=headers)
try:
pred_res = json.loads(json_response.text)
except json.decoder.JSONDecodeError:
raise Exception("Invalid response: {}".format(json_response))
print(type(pred_res))
print(pred_res)
pred = pred_res['predictions']
return pred[:, -1]
def response(self, texts, topk=5):
token_ids = [tokenizer._token_start_id, speakers[0]]
segment_ids = [tokenizer._token_start_id, speakers[0]]
for i, text in enumerate(texts):
ids = tokenizer.encode(text)[0][1:-1] + [speakers[(i + 1) % 2]]
token_ids.extend(ids)
segment_ids.extend([speakers[i % 2]] * len(ids))
segment_ids[-1] = speakers[(i + 1) % 2]
results = self.random_sample([token_ids, segment_ids], 1, topk)
return tokenizer.decode(results[0])
chatbot = ChatBot(start_id=None, end_id=tokenizer._token_end_id, maxlen=32)
while True:
raw_text = input(">>> ")
print(chatbot.response([raw_text]))
"""
回复是随机的,例如:你还有我 | 那就不要爱我 | 你是不是傻 | 等等。
"""
报错:
{'error': 'Input to reshape is a tensor with 3840 values, but the requested shape has 5\n\t [[{{node model/Transformer-0-MultiHeadSelfAttention/dense/Tensordot/Reshape}}]]'}
在tf2.0.0 和2.3.0 下均报该错误
本TF版predict 输入为 https://github.com/bojone/CDial-GPT-tf/blob/master/example.py#L37
return model.predict([token_ids, segment_ids])[:, -1]
原版模型输入为 https://github.com/thu-coai/CDial-GPT/blob/master/interact.py#L81
logits, *_ = model(input_ids, token_type_ids=token_type_ids)
我使用 pytorch 版的输入 input_ids, token_type_ids
作为 TF版的输入 token_ids, segment_ids
,但两个模型的输出结果虽然shape相同,但数值差别很大。
请问什么原因呢?感谢!
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.