keras ERNIE

Pre-trained ERNIE models could be loaded for feature extraction and prediction.

Install

pip install keras-ernie

Usage

• Download pre-trained ERNIE models
• Load the pre-trained ERNIE models
• Convert pre-trained ERNIE model to Tensor model

Download Pre-trained ERNIE Models

Notes: Currently, only the following models are supported.

Model	Description
ERNIE 1.0 Base for Chinese	with params, config and vocabs
ERNIE 1.0 Base for Chinese(max-len-512)	with params, config and vocabs
ERNIE 2.0 Base for English	with params, config and vocabs

Load Pre-trained ERNIE Models

import os
from keras_ernie import load_from_checkpoint

ernie_path = "/root/ERNIE_stable-1.0.1"
init_checkpoint = os.path.join(ernie_path, 'params')
ernie_config_path = os.path.join(ernie_path, 'ernie_config.json')
ernie_vocab_path = os.path.join(ernie_path, 'vocab.txt')
ernie_version = "stable-1.0.1"

model = load_from_checkpoint(init_checkpoint, ernie_config_path, ernie_vocab_path, ernie_version,
            max_seq_len=128, num_labels=2, use_fp16=False, training=False, seq_len=None, name='ernie')
model.summary()

Convert Pre-trained ERNIE Model To Tensor Model

import os
from keras_ernie import ErnieArgs
from keras_ernie import convert_paddle_to_tensor

ernie_path = "/root/ERNIE_stable-1.0.1"
init_checkpoint = os.path.join(ernie_path, 'params')
ernie_config_path = os.path.join(ernie_path, 'ernie_config.json')
ernie_vocab_path = os.path.join(ernie_path, 'vocab.txt')
tensor_checkpoints_dir = "/root/checkpoints"

args = ErnieArgs(init_checkpoint, ernie_config_path, ernie_vocab_path,
        max_seq_len=128, num_labels=2, use_fp16=False)

convert_paddle_to_tensor(args, tensor_checkpoints_dir)