aapot
commited on
Commit
•
3323445
1
Parent(s):
eb43aa0
Add convbert generator model
Browse files- config.json +29 -0
- convert_original_convbert_tf_checkpoint_to_generator_pytorch.py +150 -0
- pytorch_model.bin +3 -0
- special_tokens_map.json +1 -0
- tokenizer.json +0 -0
- tokenizer_config.json +1 -0
- vocab.txt +0 -0
config.json
ADDED
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"architectures": [
|
3 |
+
"ConvBertForMaskedLM"
|
4 |
+
],
|
5 |
+
"attention_probs_dropout_prob": 0.1,
|
6 |
+
"bos_token_id": 0,
|
7 |
+
"classifier_dropout": null,
|
8 |
+
"conv_kernel_size": 9,
|
9 |
+
"embedding_size": 768,
|
10 |
+
"eos_token_id": 2,
|
11 |
+
"head_ratio": 2,
|
12 |
+
"hidden_act": "gelu",
|
13 |
+
"hidden_dropout_prob": 0.1,
|
14 |
+
"hidden_size": 256,
|
15 |
+
"initializer_range": 0.02,
|
16 |
+
"intermediate_size": 1024,
|
17 |
+
"layer_norm_eps": 1e-12,
|
18 |
+
"max_position_embeddings": 512,
|
19 |
+
"model_type": "convbert",
|
20 |
+
"num_attention_heads": 4,
|
21 |
+
"num_groups": 1,
|
22 |
+
"num_hidden_layers": 12,
|
23 |
+
"pad_token_id": 0,
|
24 |
+
"torch_dtype": "float32",
|
25 |
+
"transformers_version": "4.17.0.dev0",
|
26 |
+
"type_vocab_size": 2,
|
27 |
+
"use_cache": true,
|
28 |
+
"vocab_size": 50265
|
29 |
+
}
|
convert_original_convbert_tf_checkpoint_to_generator_pytorch.py
ADDED
@@ -0,0 +1,150 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Adapted from https://github.com/huggingface/transformers/issues/9920#issuecomment-770970712
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import os
|
5 |
+
|
6 |
+
import tensorflow as tf
|
7 |
+
|
8 |
+
from transformers import ConvBertConfig, ConvBertForMaskedLM, ConvBertPreTrainedModel
|
9 |
+
from transformers.utils import logging
|
10 |
+
from operator import attrgetter
|
11 |
+
|
12 |
+
logger = logging.get_logger(__name__)
|
13 |
+
|
14 |
+
config_file = "/researchdisk/convbert-base-generator-finnish/config.json"
|
15 |
+
tf_path = "/researchdisk/convbert-base-finnish/renamed-model.ckpt"
|
16 |
+
pytorch_dump_path = "/researchdisk/convbert-base-generator-finnish"
|
17 |
+
config = ConvBertConfig.from_json_file(config_file)
|
18 |
+
|
19 |
+
model = ConvBertForMaskedLM(config)
|
20 |
+
|
21 |
+
def load_tf_weights_in_convbert(model, config, tf_checkpoint_path):
|
22 |
+
"""Load tf checkpoints in a pytorch model."""
|
23 |
+
try:
|
24 |
+
import tensorflow as tf
|
25 |
+
except ImportError:
|
26 |
+
logger.error(
|
27 |
+
"Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
|
28 |
+
"https://www.tensorflow.org/install/ for installation instructions."
|
29 |
+
)
|
30 |
+
raise
|
31 |
+
tf_path = os.path.abspath(tf_checkpoint_path)
|
32 |
+
logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
|
33 |
+
# Load weights from TF model
|
34 |
+
init_vars = tf.train.list_variables(tf_path)
|
35 |
+
tf_data = {}
|
36 |
+
for name, shape in init_vars:
|
37 |
+
logger.info("Loading TF weight {} with shape {}".format(name, shape))
|
38 |
+
array = tf.train.load_variable(tf_path, name)
|
39 |
+
tf_data[name] = array
|
40 |
+
|
41 |
+
param_mapping = {
|
42 |
+
"convbert.embeddings.word_embeddings.weight": "electra/embeddings/word_embeddings",
|
43 |
+
"convbert.embeddings.position_embeddings.weight": "electra/embeddings/position_embeddings",
|
44 |
+
"convbert.embeddings.token_type_embeddings.weight": "electra/embeddings/token_type_embeddings",
|
45 |
+
"convbert.embeddings.LayerNorm.weight": "electra/embeddings/LayerNorm/gamma",
|
46 |
+
"convbert.embeddings.LayerNorm.bias": "electra/embeddings/LayerNorm/beta",
|
47 |
+
"convbert.embeddings_project.weight": "generator/embeddings_project/kernel",
|
48 |
+
"convbert.embeddings_project.bias": "generator/embeddings_project/bias",
|
49 |
+
"generator_predictions.LayerNorm.weight": "generator_predictions/LayerNorm/gamma",
|
50 |
+
"generator_predictions.LayerNorm.bias": "generator_predictions/LayerNorm/beta",
|
51 |
+
"generator_predictions.dense.weight": "generator_predictions/dense/kernel",
|
52 |
+
"generator_predictions.dense.bias": "generator_predictions/dense/bias",
|
53 |
+
"generator_lm_head.bias": "generator_predictions/output_bias"
|
54 |
+
}
|
55 |
+
if config.num_groups > 1:
|
56 |
+
group_dense_name = "g_dense"
|
57 |
+
else:
|
58 |
+
group_dense_name = "dense"
|
59 |
+
|
60 |
+
for j in range(config.num_hidden_layers):
|
61 |
+
param_mapping[
|
62 |
+
f"convbert.encoder.layer.{j}.attention.self.query.weight"
|
63 |
+
] = f"generator/encoder/layer_{j}/attention/self/query/kernel"
|
64 |
+
param_mapping[
|
65 |
+
f"convbert.encoder.layer.{j}.attention.self.query.bias"
|
66 |
+
] = f"generator/encoder/layer_{j}/attention/self/query/bias"
|
67 |
+
param_mapping[
|
68 |
+
f"convbert.encoder.layer.{j}.attention.self.key.weight"
|
69 |
+
] = f"generator/encoder/layer_{j}/attention/self/key/kernel"
|
70 |
+
param_mapping[
|
71 |
+
f"convbert.encoder.layer.{j}.attention.self.key.bias"
|
72 |
+
] = f"generator/encoder/layer_{j}/attention/self/key/bias"
|
73 |
+
param_mapping[
|
74 |
+
f"convbert.encoder.layer.{j}.attention.self.value.weight"
|
75 |
+
] = f"generator/encoder/layer_{j}/attention/self/value/kernel"
|
76 |
+
param_mapping[
|
77 |
+
f"convbert.encoder.layer.{j}.attention.self.value.bias"
|
78 |
+
] = f"generator/encoder/layer_{j}/attention/self/value/bias"
|
79 |
+
param_mapping[
|
80 |
+
f"convbert.encoder.layer.{j}.attention.self.key_conv_attn_layer.depthwise.weight"
|
81 |
+
] = f"generator/encoder/layer_{j}/attention/self/conv_attn_key/depthwise_kernel"
|
82 |
+
param_mapping[
|
83 |
+
f"convbert.encoder.layer.{j}.attention.self.key_conv_attn_layer.pointwise.weight"
|
84 |
+
] = f"generator/encoder/layer_{j}/attention/self/conv_attn_key/pointwise_kernel"
|
85 |
+
param_mapping[
|
86 |
+
f"convbert.encoder.layer.{j}.attention.self.key_conv_attn_layer.bias"
|
87 |
+
] = f"generator/encoder/layer_{j}/attention/self/conv_attn_key/bias"
|
88 |
+
param_mapping[
|
89 |
+
f"convbert.encoder.layer.{j}.attention.self.conv_kernel_layer.weight"
|
90 |
+
] = f"generator/encoder/layer_{j}/attention/self/conv_attn_kernel/kernel"
|
91 |
+
param_mapping[
|
92 |
+
f"convbert.encoder.layer.{j}.attention.self.conv_kernel_layer.bias"
|
93 |
+
] = f"generator/encoder/layer_{j}/attention/self/conv_attn_kernel/bias"
|
94 |
+
param_mapping[
|
95 |
+
f"convbert.encoder.layer.{j}.attention.self.conv_out_layer.weight"
|
96 |
+
] = f"generator/encoder/layer_{j}/attention/self/conv_attn_point/kernel"
|
97 |
+
param_mapping[
|
98 |
+
f"convbert.encoder.layer.{j}.attention.self.conv_out_layer.bias"
|
99 |
+
] = f"generator/encoder/layer_{j}/attention/self/conv_attn_point/bias"
|
100 |
+
param_mapping[
|
101 |
+
f"convbert.encoder.layer.{j}.attention.output.dense.weight"
|
102 |
+
] = f"generator/encoder/layer_{j}/attention/output/dense/kernel"
|
103 |
+
param_mapping[
|
104 |
+
f"convbert.encoder.layer.{j}.attention.output.LayerNorm.weight"
|
105 |
+
] = f"generator/encoder/layer_{j}/attention/output/LayerNorm/gamma"
|
106 |
+
param_mapping[
|
107 |
+
f"convbert.encoder.layer.{j}.attention.output.dense.bias"
|
108 |
+
] = f"generator/encoder/layer_{j}/attention/output/dense/bias"
|
109 |
+
param_mapping[
|
110 |
+
f"convbert.encoder.layer.{j}.attention.output.LayerNorm.bias"
|
111 |
+
] = f"generator/encoder/layer_{j}/attention/output/LayerNorm/beta"
|
112 |
+
param_mapping[
|
113 |
+
f"convbert.encoder.layer.{j}.intermediate.dense.weight"
|
114 |
+
] = f"generator/encoder/layer_{j}/intermediate/{group_dense_name}/kernel"
|
115 |
+
param_mapping[
|
116 |
+
f"convbert.encoder.layer.{j}.intermediate.dense.bias"
|
117 |
+
] = f"generator/encoder/layer_{j}/intermediate/{group_dense_name}/bias"
|
118 |
+
param_mapping[
|
119 |
+
f"convbert.encoder.layer.{j}.output.dense.weight"
|
120 |
+
] = f"generator/encoder/layer_{j}/output/{group_dense_name}/kernel"
|
121 |
+
param_mapping[
|
122 |
+
f"convbert.encoder.layer.{j}.output.dense.bias"
|
123 |
+
] = f"generator/encoder/layer_{j}/output/{group_dense_name}/bias"
|
124 |
+
param_mapping[
|
125 |
+
f"convbert.encoder.layer.{j}.output.LayerNorm.weight"
|
126 |
+
] = f"generator/encoder/layer_{j}/output/LayerNorm/gamma"
|
127 |
+
param_mapping[f"convbert.encoder.layer.{j}.output.LayerNorm.bias"] = f"generator/encoder/layer_{j}/output/LayerNorm/beta"
|
128 |
+
|
129 |
+
for param in model.named_parameters():
|
130 |
+
param_name = param[0]
|
131 |
+
retriever = attrgetter(param_name)
|
132 |
+
result = retriever(model)
|
133 |
+
tf_name = param_mapping[param_name]
|
134 |
+
value = torch.from_numpy(tf_data[tf_name])
|
135 |
+
logger.info(f"TF: {tf_name}, PT: {param_name} ")
|
136 |
+
if tf_name.endswith("/kernel"):
|
137 |
+
if not tf_name.endswith("/intermediate/g_dense/kernel"):
|
138 |
+
if not tf_name.endswith("/output/g_dense/kernel"):
|
139 |
+
value = value.T
|
140 |
+
if tf_name.endswith("/depthwise_kernel"):
|
141 |
+
value = value.permute(1, 2, 0) # 2, 0, 1
|
142 |
+
if tf_name.endswith("/pointwise_kernel"):
|
143 |
+
value = value.permute(2, 1, 0) # 2, 1, 0
|
144 |
+
if tf_name.endswith("/conv_attn_key/bias"):
|
145 |
+
value = value.unsqueeze(-1)
|
146 |
+
result.data = value
|
147 |
+
return model
|
148 |
+
|
149 |
+
model = load_tf_weights_in_convbert(model, config, tf_path)
|
150 |
+
model.save_pretrained(pytorch_dump_path)
|
pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:403378cbb9fb6cf606824f4d46ccf64fa7564787ed8054dee65570e792656df7
|
3 |
+
size 194453503
|
special_tokens_map.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]"}
|
tokenizer.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
tokenizer_config.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"do_lower_case": false, "unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]", "tokenize_chinese_chars": true, "strip_accents": null, "model_max_length": 512, "special_tokens_map_file": null, "name_or_path": "bert-base-cased", "tokenizer_class": "BertTokenizer"}
|
vocab.txt
ADDED
The diff for this file is too large to render.
See raw diff
|
|