File size: 7,600 Bytes
6fc683c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
import logging
import torch

from collections import OrderedDict

from transformers.modeling_bert import (BertConfig, BertEncoder,
                                        BertIntermediate, BertLayer,
                                        BertModel, BertOutput,
                                        BertSelfAttention,
                                        BertSelfOutput)
from transformers.modeling_roberta import (RobertaEmbeddings,
                                           RobertaForMaskedLM,
                                           RobertaForSequenceClassification,
                                           RobertaModel)

logger = logging.getLogger(__name__)


# NOTE transformers should be 2.5.1

def convert_cxlm_to_transformers(ckpt_path):
    ckpt = torch.load(ckpt_path, map_location="cpu")
    args = ckpt["args"]

    config = BertConfig(
        # vocab_size_or_config_json_file=250002,
        vocab_size=250002,
        hidden_size=args.encoder_embed_dim,
        num_hidden_layers=args.encoder_layers,
        num_attention_heads=args.encoder_attention_heads,
        intermediate_size=args.encoder_ffn_embed_dim,
        max_position_embeddings=args.max_positions + 2,
        type_vocab_size=1,
        layer_norm_eps=1e-5,  # PyTorch default used in fairseq
    )

    print("Our BERT config:", config)

    stat_dict = ckpt["model"]
    new_stat_dict = {}

    model = RobertaForMaskedLM(config)
    model.eval()

    sent_enc = "decoder.sentence_encoder"
    new_stat_dict["roberta.embeddings.word_embeddings.weight"] = stat_dict[sent_enc + ".embed_tokens.weight"]
    new_stat_dict["roberta.embeddings.position_embeddings.weight"] = stat_dict[sent_enc + ".embed_positions.weight"]

    new_stat_dict["roberta.embeddings.token_type_embeddings.weight"] = torch.zeros_like(
        model.roberta.embeddings.token_type_embeddings.weight)

    new_stat_dict["roberta.embeddings.LayerNorm.weight"] = stat_dict[sent_enc + ".emb_layer_norm.weight"]
    new_stat_dict["roberta.embeddings.LayerNorm.bias"] = stat_dict[sent_enc + ".emb_layer_norm.bias"]

    for i in range(config.num_hidden_layers):
        # Encoder: start of layer
        # layer: BertLayer = model.roberta.encoder.layer[i]
        layer = "roberta.encoder.layer.%d" % i
        roberta_layer = sent_enc + (".layers.%d" % i)

        ### self attention
        # self_attn: BertSelfAttention = layer.attention.self
        self_attn = layer + ".attention.self"
        assert (
                stat_dict[roberta_layer + ".self_attn.k_proj.weight"].data.shape == \
                stat_dict[roberta_layer + ".self_attn.q_proj.weight"].data.shape == \
                stat_dict[roberta_layer + ".self_attn.v_proj.weight"].data.shape == \
                torch.Size((config.hidden_size, config.hidden_size))
        )

        new_stat_dict[self_attn + ".query.weight"] = stat_dict[roberta_layer + ".self_attn.q_proj.weight"]
        new_stat_dict[self_attn + ".query.bias"] = stat_dict[roberta_layer + ".self_attn.q_proj.bias"]
        new_stat_dict[self_attn + ".key.weight"] = stat_dict[roberta_layer + ".self_attn.k_proj.weight"]
        new_stat_dict[self_attn + ".key.bias"] = stat_dict[roberta_layer + ".self_attn.k_proj.bias"]
        new_stat_dict[self_attn + ".value.weight"] = stat_dict[roberta_layer + ".self_attn.v_proj.weight"]
        new_stat_dict[self_attn + ".value.bias"] = stat_dict[roberta_layer + ".self_attn.v_proj.bias"]

        ### self-attention output
        # self_output: BertSelfOutput = layer.attention.output
        self_output = layer + ".attention.output"
        assert (
                model.roberta.encoder.layer[i].attention.output.dense.weight.shape == stat_dict[
            roberta_layer + ".self_attn.out_proj.weight"].shape
        )
        new_stat_dict[self_output + ".dense.weight"] = stat_dict[roberta_layer + ".self_attn.out_proj.weight"]
        new_stat_dict[self_output + ".dense.bias"] = stat_dict[roberta_layer + ".self_attn.out_proj.bias"]
        new_stat_dict[self_output + ".LayerNorm.weight"] = stat_dict[roberta_layer + ".self_attn_layer_norm.weight"]
        new_stat_dict[self_output + ".LayerNorm.bias"] = stat_dict[roberta_layer + ".self_attn_layer_norm.bias"]

        ### intermediate
        # intermediate: BertIntermediate = layer.intermediate
        intermediate = layer + ".intermediate"
        assert (
                model.roberta.encoder.layer[i].intermediate.dense.weight.shape == stat_dict[
            roberta_layer + ".fc1.weight"].shape
        )
        # TODO
        new_stat_dict[intermediate + ".dense.weight"] = stat_dict[roberta_layer + ".fc1.weight"]
        new_stat_dict[intermediate + ".dense.bias"] = stat_dict[roberta_layer + ".fc1.bias"]

        ### output
        # bert_output: BertOutput = layer.output
        bert_output = layer + ".output"
        assert (
                model.roberta.encoder.layer[i].output.dense.weight.shape == stat_dict[
            roberta_layer + ".fc2.weight"].shape
        )
        new_stat_dict[bert_output + ".dense.weight"] = stat_dict[roberta_layer + ".fc2.weight"]
        new_stat_dict[bert_output + ".dense.bias"] = stat_dict[roberta_layer + ".fc2.bias"]
        new_stat_dict[bert_output + ".LayerNorm.weight"] = stat_dict[roberta_layer + ".final_layer_norm.weight"]
        new_stat_dict[bert_output + ".LayerNorm.bias"] = stat_dict[roberta_layer + ".final_layer_norm.bias"]
        #### end of layer

    new_stat_dict["lm_head.dense.weight"] = stat_dict["decoder.lm_head.dense.weight"]
    new_stat_dict["lm_head.dense.bias"] = stat_dict["decoder.lm_head.dense.bias"]
    new_stat_dict["lm_head.layer_norm.weight"] = stat_dict["decoder.lm_head.layer_norm.weight"]
    new_stat_dict["lm_head.layer_norm.bias"] = stat_dict["decoder.lm_head.layer_norm.bias"]
    new_stat_dict["lm_head.decoder.weight"] = stat_dict["decoder.lm_head.weight"]
    new_stat_dict["lm_head.bias"] = stat_dict["decoder.lm_head.bias"]
    new_stat_dict["lm_head.decoder.bias"] = stat_dict["decoder.lm_head.bias"]

    new_stat_dict["roberta.pooler.dense.weight"] = model.roberta.pooler.dense.weight
    new_stat_dict["roberta.pooler.dense.bias"] = model.roberta.pooler.dense.bias

    return new_stat_dict


def update_hf_sd(old_sd, xlmr_path):
    x = torch.load(xlmr_path, map_location="cpu")
    m = old_sd
    d = OrderedDict()
    for k, v in m.items():
        if k == 'roberta.pooler.dense.weight':
            d[k] = x[k].half().clone()
        elif k not in ('proj_matrix_fast', 'lm_head.decoder.bias', 'roberta.pooler.dense.weight'):
            d[k] = v.data.half().clone()

    assert set(d.keys()) == set(x.keys())

    for k in d.keys():
        assert d[k].size() == x[k].size()

    for k in d.keys():
        if k != 'roberta.pooler.dense.weight':
            assert (d[k].float() - m[k].float()).abs().max().item() <= 1e-4

    return d


def convert_pt_to_hf(xlmr_path, inf, logger=None):
    if logger:
        logger.info("converting pt file at {} to hf file.".format(inf))
    sd = convert_cxlm_to_transformers(inf)
    return update_hf_sd(sd, xlmr_path)


if __name__ == "__main__":
    import os

    xlmr_path = "/home/v-zechi/data/unilm/zechi/exp/res/huggingface/hf-ckpt/xlmr-large/pytorch_model.bin"
    inf = "/home/v-zechi/data/unilm/zechi/exp/cxlm_exp/dump-ifx94-large/checkpoint_2_200000.pt"
    outf = "/home/v-zechi/data/unilm/zechi/usw/res/infoxlm-models/huggingface/infoxlm-large-without-meta/pytorch_model.bin"

    assert not os.path.exists(outf)
    sd = convert_cxlm_to_transformers(inf)
    sd2 = update_hf_sd(sd, xlmr_path)
    torch.save(sd2, outf)