まず、ゼロから学習しその精度を調べます
モデル作成はKerasを用います
次はword2vecで事前に学習済みのものをファインチューニングしたもの
にする予定でしたが、ダウンロードに時間がかかっています
それで、先にGloVeで事前に学習済みのファインチューニングの結果を
先に、紹介します
umich si650 ゼロからの学習
Epoch 14/20
4960/4960 [==============================] - 9s 2ms/step - loss: 7.4957e-04 - acc: 0.9996 - val_loss: 0.0160 - val_acc: 0.9944
Epoch 15/20
4960/4960 [==============================] - 8s 2ms/step - loss: 9.4082e-04 - acc: 0.9996 - val_loss: 0.0156 - val_acc: 0.9944
Epoch 16/20
4960/4960 [==============================] - 8s 2ms/step - loss: 5.9232e-04 - acc: 0.9998 - val_loss: 0.0168 - val_acc: 0.9958
Epoch 17/20
4960/4960 [==============================] - 8s 2ms/step - loss: 8.8396e-04 - acc: 0.9998 - val_loss: 0.0167 - val_acc: 0.9939
Epoch 18/20
4960/4960 [==============================] - 8s 2ms/step - loss: 0.0010 - acc: 0.9998 - val_loss: 0.0162 - val_acc: 0.9953
Epoch 19/20
4960/4960 [==============================] - 8s 2ms/step - loss: 0.0010 - acc: 0.9998 - val_loss: 0.0168 - val_acc: 0.9939
Epoch 20/20
4960/4960 [==============================] - 8s 2ms/step - loss: 7.6745e-04 - acc: 0.9998 - val_loss: 0.0167 - val_acc: 0.9958
2126/2126 [==============================] - 1s 382us/step
Test score: 0.017, accuracy: 0.996
4960/4960 [==============================] - 9s 2ms/step - loss: 7.4957e-04 - acc: 0.9996 - val_loss: 0.0160 - val_acc: 0.9944
Epoch 15/20
4960/4960 [==============================] - 8s 2ms/step - loss: 9.4082e-04 - acc: 0.9996 - val_loss: 0.0156 - val_acc: 0.9944
Epoch 16/20
4960/4960 [==============================] - 8s 2ms/step - loss: 5.9232e-04 - acc: 0.9998 - val_loss: 0.0168 - val_acc: 0.9958
Epoch 17/20
4960/4960 [==============================] - 8s 2ms/step - loss: 8.8396e-04 - acc: 0.9998 - val_loss: 0.0167 - val_acc: 0.9939
Epoch 18/20
4960/4960 [==============================] - 8s 2ms/step - loss: 0.0010 - acc: 0.9998 - val_loss: 0.0162 - val_acc: 0.9953
Epoch 19/20
4960/4960 [==============================] - 8s 2ms/step - loss: 0.0010 - acc: 0.9998 - val_loss: 0.0168 - val_acc: 0.9939
Epoch 20/20
4960/4960 [==============================] - 8s 2ms/step - loss: 7.6745e-04 - acc: 0.9998 - val_loss: 0.0167 - val_acc: 0.9958
2126/2126 [==============================] - 1s 382us/step
Test score: 0.017, accuracy: 0.996
GloVe 事前学習済みのファインチューニング
4960/4960 [==============================] - 18s 4ms/step - loss: 0.0076 - acc: 0.9986 - val_loss: 0.0215 - val_acc: 0.9920
Epoch 4/10
4960/4960 [==============================] - 18s 4ms/step - loss: 0.0047 - acc: 0.9992 - val_loss: 0.0191 - val_acc: 0.9934
Epoch 5/10
4960/4960 [==============================] - 18s 4ms/step - loss: 0.0031 - acc: 0.9992 - val_loss: 0.0190 - val_acc: 0.9929
Epoch 6/10
4960/4960 [==============================] - 18s 4ms/step - loss: 0.0020 - acc: 0.9998 - val_loss: 0.0206 - val_acc: 0.9925
Epoch 7/10
4960/4960 [==============================] - 18s 4ms/step - loss: 0.0020 - acc: 0.9998 - val_loss: 0.0188 - val_acc: 0.9934
Epoch 8/10
4960/4960 [==============================] - 19s 4ms/step - loss: 0.0026 - acc: 0.9996 - val_loss: 0.0203 - val_acc: 0.9944
Epoch 9/10
4960/4960 [==============================] - 18s 4ms/step - loss: 0.0031 - acc: 0.9996 - val_loss: 0.0190 - val_acc: 0.9934
Epoch 10/10
4960/4960 [==============================] - 18s 4ms/step - loss: 0.0016 - acc: 0.9998 - val_loss: 0.0202 - val_acc: 0.9934
2126/2126 [==============================] - 2s 846us/step
Test score: 0.020, accuracy: 0.993
Epoch 4/10
4960/4960 [==============================] - 18s 4ms/step - loss: 0.0047 - acc: 0.9992 - val_loss: 0.0191 - val_acc: 0.9934
Epoch 5/10
4960/4960 [==============================] - 18s 4ms/step - loss: 0.0031 - acc: 0.9992 - val_loss: 0.0190 - val_acc: 0.9929
Epoch 6/10
4960/4960 [==============================] - 18s 4ms/step - loss: 0.0020 - acc: 0.9998 - val_loss: 0.0206 - val_acc: 0.9925
Epoch 7/10
4960/4960 [==============================] - 18s 4ms/step - loss: 0.0020 - acc: 0.9998 - val_loss: 0.0188 - val_acc: 0.9934
Epoch 8/10
4960/4960 [==============================] - 19s 4ms/step - loss: 0.0026 - acc: 0.9996 - val_loss: 0.0203 - val_acc: 0.9944
Epoch 9/10
4960/4960 [==============================] - 18s 4ms/step - loss: 0.0031 - acc: 0.9996 - val_loss: 0.0190 - val_acc: 0.9934
Epoch 10/10
4960/4960 [==============================] - 18s 4ms/step - loss: 0.0016 - acc: 0.9998 - val_loss: 0.0202 - val_acc: 0.9934
2126/2126 [==============================] - 2s 846us/step
Test score: 0.020, accuracy: 0.993