keras ERNIE
Pre-trained ERNIE models could be loaded for feature extraction and prediction.
Install
pip install keras-ernieUsage
- 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)
