train/scripts/convert_gpt2_from_huggingface_to_tencentpretrain.py

import argparse
import collections
import torch


parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument("--input_model_path", type=str, default="models/input_model.bin",
                        help=".")
parser.add_argument("--output_model_path", type=str, default="models/output_model.bin",
                        help=".")
parser.add_argument("--layers_num", type=int, default=12)

args = parser.parse_args()

input_model = torch.load(args.input_model_path, map_location="cpu")

output_model = collections.OrderedDict()
emb_size = input_model["transformer.h." + str(0) + ".attn.c_attn.weight"].shape[0]

output_model["embedding.word.embedding.weight"] = input_model["transformer.wte.weight"]
output_model["embedding.pos.embedding.weight"] = input_model["transformer.wpe.weight"]

for i in range(args.layers_num):
    for j in range(3):
        output_model["encoder.transformer." + str(i) + ".self_attn.linear_layers." + str(j) + ".weight"] = \
            input_model["transformer.h." + str(i) + ".attn.c_attn.weight"].t()[j*emb_size:(j+1)*emb_size, :]
        output_model["encoder.transformer." + str(i) + ".self_attn.linear_layers." + str(j) + ".bias"] = \
            input_model["transformer.h." + str(i) + ".attn.c_attn.bias"][j*emb_size:(j+1)*emb_size]

    output_model["encoder.transformer." + str(i) + ".self_attn.final_linear.weight"] = \
        input_model["transformer.h." + str(i) + ".attn.c_proj.weight"].t()
    output_model["encoder.transformer." + str(i) + ".self_attn.final_linear.bias"] = \
        input_model["transformer.h." + str(i) + ".attn.c_proj.bias"]

    output_model["encoder.transformer." + str(i) + ".layer_norm_1.gamma"] = \
        input_model["transformer.h." + str(i) + ".ln_1.weight"]
    output_model["encoder.transformer." + str(i) + ".layer_norm_1.beta"] = \
        input_model["transformer.h." + str(i) + ".ln_1.bias"]

    output_model["encoder.transformer." + str(i) + ".feed_forward.linear_1.weight"] = \
        input_model["transformer.h." + str(i) + ".mlp.c_fc.weight"].t()
    output_model["encoder.transformer." + str(i) + ".feed_forward.linear_1.bias"] = \
        input_model["transformer.h." + str(i) + ".mlp.c_fc.bias"]
    output_model["encoder.transformer." + str(i) + ".feed_forward.linear_2.weight"] = \
        input_model["transformer.h." + str(i) + ".mlp.c_proj.weight"].t()
    output_model["encoder.transformer." + str(i) + ".feed_forward.linear_2.bias"] = \
        input_model["transformer.h." + str(i) + ".mlp.c_proj.bias"]

    output_model["encoder.transformer." + str(i) + ".layer_norm_2.gamma"] = \
        input_model["transformer.h." + str(i) + ".ln_2.weight"]
    output_model["encoder.transformer." + str(i) + ".layer_norm_2.beta"] = \
        input_model["transformer.h." + str(i) + ".ln_2.bias"]


output_model["encoder.layer_norm.gamma"] = input_model["transformer.ln_f.weight"]
output_model["encoder.layer_norm.beta"] = input_model["transformer.ln_f.bias"]
output_model["target.lm.output_layer.weight"] = input_model["embedding.word.embedding.weight"]

torch.save(output_model, args.output_model_path)