Upload model
Browse files- modeling_vnsabsa.py +130 -2
modeling_vnsabsa.py
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
from transformers import PreTrainedModel
|
2 |
-
from modules import SmartphoneBERT
|
3 |
import torch
|
|
|
4 |
|
5 |
from .configuration_vnsabsa import VnSmartphoneAbsaConfig
|
6 |
|
@@ -72,4 +72,132 @@ class VnSmartphoneAbsaModel(PreTrainedModel):
|
|
72 |
if a_i[-1] >= aspect_thresholds[-1]:
|
73 |
res_i["OTHERS"] = ""
|
74 |
|
75 |
-
return results
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
from transformers import PreTrainedModel
|
|
|
2 |
import torch
|
3 |
+
import torch.nn as nn
|
4 |
|
5 |
from .configuration_vnsabsa import VnSmartphoneAbsaConfig
|
6 |
|
|
|
72 |
if a_i[-1] >= aspect_thresholds[-1]:
|
73 |
res_i["OTHERS"] = ""
|
74 |
|
75 |
+
return results
|
76 |
+
|
77 |
+
|
78 |
+
class AspectClassifier(nn.Module):
|
79 |
+
def __init__(
|
80 |
+
self,
|
81 |
+
input_size: int,
|
82 |
+
dropout: float = 0.3,
|
83 |
+
hidden_size: int = 64,
|
84 |
+
*args, **kwargs
|
85 |
+
) -> None:
|
86 |
+
super().__init__(*args, **kwargs)
|
87 |
+
|
88 |
+
self.input_size = input_size
|
89 |
+
|
90 |
+
self.fc = nn.Sequential(
|
91 |
+
nn.Dropout(dropout),
|
92 |
+
nn.Linear(
|
93 |
+
in_features=input_size,
|
94 |
+
out_features=hidden_size
|
95 |
+
),
|
96 |
+
nn.ReLU(),
|
97 |
+
nn.Dropout(dropout),
|
98 |
+
nn.Linear(
|
99 |
+
in_features=hidden_size,
|
100 |
+
out_features=10+1
|
101 |
+
)
|
102 |
+
)
|
103 |
+
|
104 |
+
def forward(self, input: torch.Tensor):
|
105 |
+
x = self.fc(input)
|
106 |
+
return x
|
107 |
+
|
108 |
+
|
109 |
+
class PolarityClassifier(nn.Module):
|
110 |
+
def __init__(
|
111 |
+
self,
|
112 |
+
input_size: int,
|
113 |
+
dropout: float = 0.5,
|
114 |
+
hidden_size: int = 64,
|
115 |
+
*args, **kwargs
|
116 |
+
) -> None:
|
117 |
+
super().__init__(*args, **kwargs)
|
118 |
+
self.polarity_fcs = nn.ModuleList([
|
119 |
+
nn.Sequential(
|
120 |
+
nn.Dropout(dropout),
|
121 |
+
nn.Linear(
|
122 |
+
in_features=input_size,
|
123 |
+
out_features=hidden_size
|
124 |
+
),
|
125 |
+
nn.ReLU(),
|
126 |
+
nn.Dropout(dropout),
|
127 |
+
nn.Linear(
|
128 |
+
in_features=hidden_size,
|
129 |
+
out_features=3
|
130 |
+
)
|
131 |
+
)
|
132 |
+
for _ in torch.arange(10)
|
133 |
+
])
|
134 |
+
|
135 |
+
def forward(self, input: torch.Tensor):
|
136 |
+
polarities = torch.stack([
|
137 |
+
fc(input)
|
138 |
+
for fc in self.polarity_fcs
|
139 |
+
])
|
140 |
+
|
141 |
+
if input.ndim == 2:
|
142 |
+
polarities = polarities.transpose(0, 1)
|
143 |
+
return polarities
|
144 |
+
|
145 |
+
|
146 |
+
class SmartphoneBERT(nn.Module):
|
147 |
+
def __init__(
|
148 |
+
self,
|
149 |
+
vocab_size: int,
|
150 |
+
embed_dim: int = 768,
|
151 |
+
num_heads: int = 8,
|
152 |
+
num_encoders: int = 4,
|
153 |
+
encoder_dropout: float = 0.1,
|
154 |
+
fc_dropout: float =0.4,
|
155 |
+
fc_hidden_size: int = 128,
|
156 |
+
*args, **kwargs
|
157 |
+
):
|
158 |
+
super().__init__(*args, **kwargs)
|
159 |
+
self.embed = nn.Embedding(
|
160 |
+
num_embeddings=vocab_size,
|
161 |
+
embedding_dim=embed_dim,
|
162 |
+
padding_idx=0
|
163 |
+
)
|
164 |
+
self.encoder = nn.TransformerEncoder(
|
165 |
+
nn.TransformerEncoderLayer(
|
166 |
+
d_model=embed_dim,
|
167 |
+
nhead=num_heads,
|
168 |
+
dim_feedforward=embed_dim,
|
169 |
+
dropout=encoder_dropout,
|
170 |
+
batch_first=True
|
171 |
+
),
|
172 |
+
num_layers=num_encoders,
|
173 |
+
norm=nn.LayerNorm(embed_dim),
|
174 |
+
enable_nested_tensor=False
|
175 |
+
)
|
176 |
+
self.a_fc = AspectClassifier(
|
177 |
+
input_size=2*embed_dim,
|
178 |
+
dropout=fc_dropout,
|
179 |
+
hidden_size=fc_hidden_size
|
180 |
+
)
|
181 |
+
self.p_fc = PolarityClassifier(
|
182 |
+
input_size=2*embed_dim,
|
183 |
+
dropout=fc_dropout,
|
184 |
+
hidden_size=fc_hidden_size
|
185 |
+
)
|
186 |
+
|
187 |
+
def forward(
|
188 |
+
self,
|
189 |
+
input_ids: torch.Tensor,
|
190 |
+
attention_mask: torch.Tensor
|
191 |
+
):
|
192 |
+
padding_mask = ~attention_mask.bool()
|
193 |
+
x = self.embed(input_ids)
|
194 |
+
x = self.encoder(x, src_key_padding_mask=padding_mask)
|
195 |
+
x[padding_mask] = 0
|
196 |
+
x = torch.cat([
|
197 |
+
x[..., 0, :],
|
198 |
+
torch.mean(x, dim=-2)
|
199 |
+
], dim=-1)
|
200 |
+
|
201 |
+
a_logits = self.a_fc(x)
|
202 |
+
p_logits = self.p_fc(x)
|
203 |
+
return a_logits, p_logits
|