tensorflow模型转pytorch

具体的应用场景是将BERT的CKPT转化为pytorch的model。
把下面的代码保存为transfer.py
比如下述训练好的check point。

执行：transfer.py model.ckpt-2543 model/rf-bert 前面是训练第2543次的check point 后面是保存的路径，就会得到：

transfer.py:

import tensorflow as tf
import deepdish as dd
import argparse
import os
import numpy as np

def tr(v):
    # tensorflow weights to pytorch weights
    if v.ndim == 4:
        return np.ascontiguousarray(v.transpose(3,2,0,1))
    elif v.ndim == 2:
        return np.ascontiguousarray(v.transpose())
    return v

def read_ckpt(ckpt):
    # https://github.com/tensorflow/tensorflow/issues/1823
    reader = tf.train.NewCheckpointReader(ckpt)
    weights = {n: reader.get_tensor(n) for (n, _) in reader.get_variable_to_shape_map().items()}
    pyweights = {k: tr(v) for (k, v) in weights.items()}
    return pyweights
if __name__ == '__main__':
    parser = argparse.ArgumentParser(description="Converts ckpt weights to deepdish hdf5")
    parser.add_argument("infile", type=str,
                        help="Path to the ckpt.")  # ***model.ckpt-22177***
    parser.add_argument("outfile", type=str, nargs='?', default='',
                        help="Output file (inferred if missing).")
    args = parser.parse_args()
    if args.outfile == '':
        args.outfile = os.path.splitext(args.infile)[0] + '.h5'
    outdir = os.path.dirname(args.outfile)
    if not os.path.exists(outdir):
        os.makedirs(outdir)
    weights = read_ckpt(args.infile)
    dd.io.save(args.outfile, weights)
————————————————
版权声明：本文为CSDN博主「 杨杨」的原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接及本声明。
原文链接：https://blog.csdn.net/weixin_42699651/article/details/88932670