モデルマージ
Browse files- merge-pytorch-model.py +50 -0
merge-pytorch-model.py
ADDED
|
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import torch.nn as nn
|
| 3 |
+
from transformers import AutoModel, AutoTokenizer, XLMRobertaModel
|
| 4 |
+
|
| 5 |
+
# カスタムモデルの定義
|
| 6 |
+
class CustomXLMRobertaModel(XLMRobertaModel):
|
| 7 |
+
def __init__(self, config):
|
| 8 |
+
super(CustomXLMRobertaModel, self).__init__(config)
|
| 9 |
+
self.sparse_linear = SparseLinear(config.hidden_size, 1) # 適切な出力次元を設定
|
| 10 |
+
|
| 11 |
+
def forward(self, *args, **kwargs):
|
| 12 |
+
outputs = super(CustomXLMRobertaModel, self).forward(*args, **kwargs)
|
| 13 |
+
dense_embeddings = outputs.last_hidden_state
|
| 14 |
+
sparse_embeddings = self.sparse_linear(dense_embeddings)
|
| 15 |
+
return sparse_embeddings
|
| 16 |
+
|
| 17 |
+
# カスタムレイヤーの定義
|
| 18 |
+
class SparseLinear(nn.Module):
|
| 19 |
+
def __init__(self, input_dim, output_dim):
|
| 20 |
+
super(SparseLinear, self).__init__()
|
| 21 |
+
self.linear = nn.Linear(input_dim, output_dim)
|
| 22 |
+
|
| 23 |
+
def forward(self, x):
|
| 24 |
+
return self.linear(x)
|
| 25 |
+
|
| 26 |
+
# モデルとトークナイザーのロード
|
| 27 |
+
model_name = "." # ローカルディレクトリを指定
|
| 28 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 29 |
+
config = AutoModel.from_pretrained(model_name).config
|
| 30 |
+
model = CustomXLMRobertaModel.from_pretrained(model_name, config=config)
|
| 31 |
+
|
| 32 |
+
# カスタムレイヤーのインスタンスを作成
|
| 33 |
+
input_dim = 1024 # Denseベクトルの次元(モデルのhidden_sizeに合わせる)
|
| 34 |
+
output_dim = 1 # Sparseベクトルの次元(保存された重みに合わせる)
|
| 35 |
+
sparse_linear = SparseLinear(input_dim, output_dim)
|
| 36 |
+
|
| 37 |
+
# Sparse線形変換のロード
|
| 38 |
+
sparse_linear_path = "sparse_linear.pt"
|
| 39 |
+
sparse_linear_state_dict = torch.load(sparse_linear_path, weights_only=True)
|
| 40 |
+
|
| 41 |
+
# state_dictのキーを変換
|
| 42 |
+
sparse_linear_state_dict = {
|
| 43 |
+
f"linear.{key}": value for key, value in sparse_linear_state_dict.items()
|
| 44 |
+
}
|
| 45 |
+
|
| 46 |
+
# カスタムレイヤーにstate_dictをロード
|
| 47 |
+
sparse_linear.load_state_dict(sparse_linear_state_dict)
|
| 48 |
+
|
| 49 |
+
# カスタムレイヤーの重みを元のモデルの重みに追加
|
| 50 |
+
model.sparse_linear.load_state_dict(sparse_linear.state_dict())
|