Convert TensorFlow XLNet Model

Pre-trained models for XLNet (Bidirectional Encoder Representations from Transformers) are publicly available.

Download the Pre-Trained Base XLNet Model

Download and unzip an archive with the XLNet-Base, Cased.

After the archive is unzipped, the directory cased_L-12_H-768_A-12 is created and contains the following files:

  • TensorFlow checkpoint (xlnet_model.ckpt) containing the pre-trained weights (which is actually 3 files)

  • sentence piece model (spiece.model) used for (de)tokenization

  • config file (xlnet_config.json) which specifies the hyperparameters of the model

To get pb-file from the archive contents, you need to do the following.

  1. Run commands

cd ~
mkdir XLNet-Base
cd XLNet-Base
git clone https://github.com/zihangdai/xlnet
wget https://storage.googleapis.com/xlnet/released_models/cased_L-12_H-768_A-12.zip
unzip cased_L-12_H-768_A-12.zip
mkdir try_save
  1. Save and run the following Python script in ~/XLNet-Base/xlnet :

Note The original model repository has been tested with TensorFlow 1.13.1 under Python2.

from collections import namedtuple

import tensorflow as tf
from tensorflow.python.framework import graph_io

import model_utils
import xlnet

LENGTHS = 50
BATCH = 1
OUTPUT_DIR = '~/XLNet-Base/try_save/'
INIT_CKPT_PATH = '~/XLNet-Base/xlnet_cased_L-12_H-768_A-12/xlnet_model.ckpt'
XLNET_CONFIG_PATH = '~/XLNet-Base/xlnet_cased_L-12_H-768_A-12/xlnet_config.json'

FLags = namedtuple('FLags', 'use_tpu init_checkpoint')
FLAGS = FLags(use_tpu=False, init_checkpoint=INIT_CKPT_PATH)

xlnet_config = xlnet.XLNetConfig(json_path=XLNET_CONFIG_PATH)
run_config = xlnet.RunConfig(is_training=False, use_tpu=False, use_bfloat16=False, dropout=0.1, dropatt=0.1,)


sentence_features_input_idx = tf.compat.v1.placeholder(tf.int32, shape=[LENGTHS, BATCH], name='input_ids')
sentence_features_segment_ids = tf.compat.v1.placeholder(tf.int32, shape=[LENGTHS, BATCH], name='seg_ids')
sentence_features_input_mask = tf.compat.v1.placeholder(tf.float32, shape=[LENGTHS, BATCH], name='input_mask')

with tf.compat.v1.Session() as sess:
    xlnet_model = xlnet.XLNetModel(xlnet_config=xlnet_config, run_config=run_config,
                                   input_ids=sentence_features_input_idx,
                                   seg_ids=sentence_features_segment_ids,
                                   input_mask=sentence_features_input_mask)

    sess.run(tf.compat.v1.global_variables_initializer())
    model_utils.init_from_checkpoint(FLAGS, True)

    # Save the variables to disk.
    saver = tf.compat.v1.train.Saver()

    # Saving checkpoint
    save_path = saver.save(sess, OUTPUT_DIR + "model.ckpt")

    # Freezing model
    outputs = ['model/transformer/dropout_2/Identity']
    graph_def_freezed = tf.compat.v1.graph_util.convert_variables_to_constants(sess, sess.graph.as_graph_def(), outputs)

    # Saving non-frozen and frozen  model to pb
    graph_io.write_graph(sess.graph.as_graph_def(), OUTPUT_DIR, 'model.pb', as_text=False)
    graph_io.write_graph(graph_def_freezed,OUTPUT_DIR, 'model_frozen.pb',
                         as_text=False)

    # Write to tensorboard
    with tf.compat.v1.summary.FileWriter(logdir=OUTPUT_DIR, graph_def=graph_def_freezed) as writer:
        writer.flush()

Download the Pre-Trained Large XLNet Model

Download and unzip an archive with the XLNet-Large, Cased.

After the archive is unzipped, the directory cased_L-12_H-1024_A-16 is created and contains the following files:

  • TensorFlow checkpoint (xlnet_model.ckpt) containing the pre-trained weights (which is actually 3 files)

  • sentence piece model (spiece.model) used for (de)tokenization

  • config file (xlnet_config.json) which specifies the hyperparameters of the model

To get pb-file from the archive contents, you need to do the following.

  1. Run commands

cd ~
mkdir XLNet-Large
cd XLNet-Large
git clone https://github.com/zihangdai/xlnet
wget https://storage.googleapis.com/xlnet/released_models/cased_L-24_H-1024_A-16.zip
unzip cased_L-24_H-1024_A-16.zip
mkdir try_save
  1. Save and run the following Python script in ~/XLNet-Large/xlnet :

from collections import namedtuple

import tensorflow as tf
from tensorflow.python.framework import graph_io

import model_utils
import xlnet

LENGTHS = 50
BATCH = 1
OUTPUT_DIR = '~/XLNet-Large/try_save'
INIT_CKPT_PATH = '~/XLNet-Large/cased_L-24_H-1024_A-16/xlnet_model.ckpt'
XLNET_CONFIG_PATH = '~/XLNet-Large/cased_L-24_H-1024_A-16/xlnet_config.json'

FLags = namedtuple('FLags', 'use_tpu init_checkpoint')
FLAGS = FLags(use_tpu=False, init_checkpoint=INIT_CKPT_PATH)

xlnet_config = xlnet.XLNetConfig(json_path=XLNET_CONFIG_PATH)
run_config = xlnet.RunConfig(is_training=False, use_tpu=False, use_bfloat16=False, dropout=0.1, dropatt=0.1,)


sentence_features_input_idx = tf.compat.v1.placeholder(tf.int32, shape=[LENGTHS, BATCH], name='input_ids')
sentence_features_segment_ids = tf.compat.v1.placeholder(tf.int32, shape=[LENGTHS, BATCH], name='seg_ids')
sentence_features_input_mask = tf.compat.v1.placeholder(tf.float32, shape=[LENGTHS, BATCH], name='input_mask')

with tf.compat.v1.Session() as sess:
    xlnet_model = xlnet.XLNetModel(xlnet_config=xlnet_config, run_config=run_config,
                                   input_ids=sentence_features_input_idx,
                                   seg_ids=sentence_features_segment_ids,
                                   input_mask=sentence_features_input_mask)

    sess.run(tf.compat.v1.global_variables_initializer())
    model_utils.init_from_checkpoint(FLAGS, True)

    # Save the variables to disk.
    saver = tf.compat.v1.train.Saver()

    # Saving checkpoint
    save_path = saver.save(sess, OUTPUT_DIR + "model.ckpt")

    # Freezing model
    outputs = ['model/transformer/dropout_2/Identity']
    graph_def_freezed = tf.compat.v1.graph_util.convert_variables_to_constants(sess, sess.graph.as_graph_def(), outputs)

    # Saving non-frozen and frozen  model to pb
    graph_io.write_graph(sess.graph.as_graph_def(), OUTPUT_DIR, 'model.pb', as_text=False)
    graph_io.write_graph(graph_def_freezed,OUTPUT_DIR, 'model_frozen.pb',
                         as_text=False)

    # Write to tensorboard
    with tf.compat.v1.summary.FileWriter(logdir=OUTPUT_DIR, graph_def=graph_def_freezed) as writer:
        writer.flush()

The script should save into ~/XLNet-Large/xlnet.

Convert frozen TensorFlow XLNet Model to IR

To generate the XLNet Intermediate Representation (IR) of the model, run the Model Optimizer with the following parameters:

mo --input_model path-to-model/model_frozen.pb \
   --input "input_mask[50 1],input_ids[50 1],seg_ids[50 1]"