Initial commit
Browse files- adjacency_matrix/graph_extended_comments.pkl +3 -0
- config.json +3 -2
- modeling_vcgn.py +95 -12
- pytorch_model.bin +2 -2
adjacency_matrix/graph_extended_comments.pkl
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:b0508863549ac3faea223be7d93bef5ec24b70af65124223fec485e1021b0f3e
|
3 |
+
size 829003020
|
config.json
CHANGED
@@ -1,9 +1,10 @@
|
|
1 |
{
|
2 |
"attention_probs_dropout_prob": 0.1,
|
|
|
3 |
"classifier_dropout": null,
|
4 |
"do_lower_case": 1,
|
5 |
"do_remove_accents": 0,
|
6 |
-
"gcn_adj_matrix": "adjacency_matrix/
|
7 |
"gcn_embedding_dim": 32,
|
8 |
"gradient_checkpointing": false,
|
9 |
"hidden_act": "gelu",
|
@@ -33,7 +34,7 @@
|
|
33 |
"pad_token_id": 0,
|
34 |
"position_embedding_type": "absolute",
|
35 |
"tf_threshold": 0.0,
|
36 |
-
"transformers_version": "4.
|
37 |
"type_vocab_size": 2,
|
38 |
"use_cache": true,
|
39 |
"vocab_size": 37788,
|
|
|
1 |
{
|
2 |
"attention_probs_dropout_prob": 0.1,
|
3 |
+
"bert_model": "readerbench/RoBERT-base",
|
4 |
"classifier_dropout": null,
|
5 |
"do_lower_case": 1,
|
6 |
"do_remove_accents": 0,
|
7 |
+
"gcn_adj_matrix": "adjacency_matrix/graph_extended_comments.pkl",
|
8 |
"gcn_embedding_dim": 32,
|
9 |
"gradient_checkpointing": false,
|
10 |
"hidden_act": "gelu",
|
|
|
34 |
"pad_token_id": 0,
|
35 |
"position_embedding_type": "absolute",
|
36 |
"tf_threshold": 0.0,
|
37 |
+
"transformers_version": "4.31.0",
|
38 |
"type_vocab_size": 2,
|
39 |
"use_cache": true,
|
40 |
"vocab_size": 37788,
|
modeling_vcgn.py
CHANGED
@@ -1,4 +1,6 @@
|
|
|
|
1 |
import torch
|
|
|
2 |
from transformers import PreTrainedModel, BertTokenizer
|
3 |
from transformers.utils import is_remote_url, download_url
|
4 |
from pathlib import Path
|
@@ -49,6 +51,9 @@ def get_torch_gcn(gcn_vocab_adj_tf, gcn_vocab_adj,gcn_config:VGCNConfig):
|
|
49 |
adj = gcn_vocab_adj_list[i]
|
50 |
adj = normalize_adj(adj)
|
51 |
norm_gcn_vocab_adj_list.append(sparse_scipy2torch(adj.tocoo()))
|
|
|
|
|
|
|
52 |
|
53 |
del gcn_vocab_adj_list
|
54 |
|
@@ -66,7 +71,8 @@ class VCGNModelForTextClassification(PreTrainedModel):
|
|
66 |
self.remove_stop_words = False
|
67 |
self.tokenizer = None
|
68 |
self.norm_gcn_vocab_adj_list = None
|
69 |
-
|
|
|
70 |
|
71 |
self.load_adj_matrix(config.gcn_adj_matrix)
|
72 |
|
@@ -80,26 +86,97 @@ class VCGNModelForTextClassification(PreTrainedModel):
|
|
80 |
)
|
81 |
|
82 |
def load_adj_matrix(self, adj_matrix):
|
|
|
83 |
if Path(adj_matrix).is_file():
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
88 |
|
89 |
self.pre_trained_model_name = adj_config['bert_model']
|
90 |
self.remove_stop_words = adj_config['remove_stop_words']
|
91 |
self.tokenizer = BertTokenizer.from_pretrained(self.pre_trained_model_name)
|
92 |
-
self.norm_gcn_vocab_adj_list =
|
93 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
94 |
|
95 |
-
|
96 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
97 |
if labels is not None:
|
98 |
loss = torch.nn.cross_entropy(logits, labels)
|
99 |
return {"loss": loss, "logits": logits}
|
100 |
return {"logits": logits}
|
101 |
|
102 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
103 |
|
104 |
import torch
|
105 |
import torch.nn as nn
|
@@ -130,7 +207,13 @@ class VocabGraphConvolution(nn.Module):
|
|
130 |
"""
|
131 |
def __init__(self,adj_matrix,voc_dim, num_adj, hid_dim, out_dim, dropout_rate=0.2):
|
132 |
super(VocabGraphConvolution, self).__init__()
|
133 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
134 |
self.voc_dim=voc_dim
|
135 |
self.num_adj=num_adj
|
136 |
self.hid_dim=hid_dim
|
@@ -147,7 +230,7 @@ class VocabGraphConvolution(nn.Module):
|
|
147 |
|
148 |
def reset_parameters(self):
|
149 |
for n,p in self.named_parameters():
|
150 |
-
if n.startswith('W')
|
151 |
init.kaiming_uniform_(p, a=math.sqrt(5))
|
152 |
|
153 |
def forward(self, X_dv, add_linear_mapping_term=False):
|
|
|
1 |
+
from typing import List, Union
|
2 |
import torch
|
3 |
+
import torch.nn.functional as F
|
4 |
from transformers import PreTrainedModel, BertTokenizer
|
5 |
from transformers.utils import is_remote_url, download_url
|
6 |
from pathlib import Path
|
|
|
51 |
adj = gcn_vocab_adj_list[i]
|
52 |
adj = normalize_adj(adj)
|
53 |
norm_gcn_vocab_adj_list.append(sparse_scipy2torch(adj.tocoo()))
|
54 |
+
|
55 |
+
for t in norm_gcn_vocab_adj_list:
|
56 |
+
t.requires_grad = False
|
57 |
|
58 |
del gcn_vocab_adj_list
|
59 |
|
|
|
71 |
self.remove_stop_words = False
|
72 |
self.tokenizer = None
|
73 |
self.norm_gcn_vocab_adj_list = None
|
74 |
+
self.gcn_vocab_size = config.vocab_size
|
75 |
+
|
76 |
|
77 |
self.load_adj_matrix(config.gcn_adj_matrix)
|
78 |
|
|
|
86 |
)
|
87 |
|
88 |
def load_adj_matrix(self, adj_matrix):
|
89 |
+
filename = None
|
90 |
if Path(adj_matrix).is_file():
|
91 |
+
filename = Path(adj_matrix)
|
92 |
+
#load file
|
93 |
+
elif (Path(__file__).parent / Path(adj_matrix)).is_file():
|
94 |
+
filename = Path(__file__).parent / Path(adj_matrix)
|
95 |
+
elif is_remote_url(adj_matrix):
|
96 |
+
filename = download_url(adj_matrix)
|
97 |
+
|
98 |
+
|
99 |
+
gcn_vocab_adj_tf, gcn_vocab_adj, adj_config = pkl.load(open(filename, 'rb'))
|
100 |
+
|
101 |
|
102 |
self.pre_trained_model_name = adj_config['bert_model']
|
103 |
self.remove_stop_words = adj_config['remove_stop_words']
|
104 |
self.tokenizer = BertTokenizer.from_pretrained(self.pre_trained_model_name)
|
105 |
+
self.norm_gcn_vocab_adj_list = get_torch_gcn(gcn_vocab_adj_tf, gcn_vocab_adj, self.config)
|
106 |
+
|
107 |
+
def _prep_batch(self, batch: torch.Tensor):
|
108 |
+
|
109 |
+
vocab_size = self.tokenizer.vocab_size
|
110 |
+
|
111 |
+
batch_gcn_swop_eye = F.one_hot(batch, vocab_size).float().to(self.device) # shape (batch_size, seq_len, vocab_size)
|
112 |
+
batch_gcn_swop_eye = batch_gcn_swop_eye.transpose(1,2) # shape (batch_size, vocab_size, seq_len)
|
113 |
+
# set all [PAD] tokens to 0
|
114 |
+
batch_gcn_swop_eye[:, self.tokenizer.pad_token_id, :] = 0
|
115 |
+
batch_gcn_swop_eye[:, self.tokenizer.cls_token_id, :] = 0
|
116 |
+
batch_gcn_swop_eye[:, self.tokenizer.sep_token_id, :] = 0
|
117 |
+
|
118 |
+
batch_gcn_swop_eye = F.pad(batch_gcn_swop_eye,(0,self.config.gcn_embedding_dim,0,0,0,0),value=0)
|
119 |
+
|
120 |
+
batch = F.pad(batch, (0, self.config.gcn_embedding_dim), 'constant', 0)
|
121 |
+
|
122 |
+
#fill gcn tokens with [SEP]
|
123 |
+
mask = torch.zeros(batch.shape[0], batch.shape[1] + 1, dtype=batch.dtype, device=self.device)
|
124 |
+
mask2 = torch.zeros(batch.shape[0], batch.shape[1] + 1, dtype=batch.dtype, device=self.device)
|
125 |
+
|
126 |
+
pos_start = (batch==self.tokenizer.pad_token_id).int().argmax(1)
|
127 |
+
|
128 |
+
mask[(torch.arange(batch.shape[0]), pos_start)] = 1
|
129 |
+
mask2[(torch.arange(batch.shape[0]), pos_start+self.config.gcn_embedding_dim)] = 1
|
130 |
+
|
131 |
+
mask = mask.cumsum(1)[:, :-1].bool()
|
132 |
+
mask2 = mask2.cumsum(1)[:, :-1].bool()
|
133 |
|
134 |
+
mask = mask & ~mask2
|
135 |
+
|
136 |
+
batch.masked_fill_(mask, self.tokenizer.sep_token_id)
|
137 |
+
|
138 |
+
return batch, batch_gcn_swop_eye
|
139 |
+
|
140 |
+
def text_to_batch(self, text: Union[List[str], str]):
|
141 |
+
if isinstance(text, str):
|
142 |
+
text = [text]
|
143 |
+
encoded = self.tokenizer.batch_encode_plus(text, padding=True, truncation=True, return_tensors='pt', max_length=self.config.max_seq_len-self.config.gcn_embedding_dim)
|
144 |
+
return encoded['input_ids'].to(self.device)
|
145 |
+
|
146 |
+
def forward(self, input:Union[torch.Tensor, List[str], str], labels=None):
|
147 |
+
|
148 |
+
if not isinstance(input, torch.Tensor):
|
149 |
+
input = self.text_to_batch(input)
|
150 |
+
|
151 |
+
input, batch_gcn_swop_eye = self._prep_batch(input)
|
152 |
+
|
153 |
+
segment_ids = torch.zeros_like(input).int().to(self.device)
|
154 |
+
input_mask = (input>0).int().to(self.device)
|
155 |
+
|
156 |
+
|
157 |
+
logits = self.model(batch_gcn_swop_eye, input, segment_ids, input_mask )
|
158 |
if labels is not None:
|
159 |
loss = torch.nn.cross_entropy(logits, labels)
|
160 |
return {"loss": loss, "logits": logits}
|
161 |
return {"logits": logits}
|
162 |
|
163 |
+
def predict(self, text: Union[List[str], str], as_dict=True):
|
164 |
+
with torch.no_grad():
|
165 |
+
logits = self.forward(text)['logits']
|
166 |
+
if as_dict:
|
167 |
+
label_id = torch.argmax(logits, dim=1).cpu().numpy()
|
168 |
+
label = [self.config.id2label[l] for l in label_id]
|
169 |
+
return {
|
170 |
+
"logits": logits,
|
171 |
+
"label_id": label_id,
|
172 |
+
"label": label,
|
173 |
+
}
|
174 |
+
else:
|
175 |
+
return torch.argmax(logits, dim=1).cpu().numpy()
|
176 |
+
|
177 |
+
@property
|
178 |
+
def device(self):
|
179 |
+
return next(self.parameters()).device
|
180 |
|
181 |
import torch
|
182 |
import torch.nn as nn
|
|
|
207 |
"""
|
208 |
def __init__(self,adj_matrix,voc_dim, num_adj, hid_dim, out_dim, dropout_rate=0.2):
|
209 |
super(VocabGraphConvolution, self).__init__()
|
210 |
+
if type(adj_matrix) is not list:
|
211 |
+
self.adj_matrix=adj_matrix
|
212 |
+
else:
|
213 |
+
self.adj_matrix=torch.nn.ParameterList([torch.nn.Parameter(x) for x in adj_matrix])
|
214 |
+
for p in self.adj_matrix:
|
215 |
+
p.requires_grad=False
|
216 |
+
|
217 |
self.voc_dim=voc_dim
|
218 |
self.num_adj=num_adj
|
219 |
self.hid_dim=hid_dim
|
|
|
230 |
|
231 |
def reset_parameters(self):
|
232 |
for n,p in self.named_parameters():
|
233 |
+
if n.startswith('W') :
|
234 |
init.kaiming_uniform_(p, a=math.sqrt(5))
|
235 |
|
236 |
def forward(self, X_dv, add_linear_mapping_term=False):
|
pytorch_model.bin
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:2dd4760540bf1667e77b45ab271e0a87376a97ecb0ea7ab669391e45a5606820
|
3 |
+
size 481615461
|