search

amansrivastava17 / lstm-siamese-text-similarity

⚛️ It is keras based implementation of siamese architecture using lstm encoders to compute text similarity
MIT License
282 stars 88 forks source link

Why is the accuracy of the training set lower than the validation set? #10

Closed Kiteflyingee closed 5 years ago

Kiteflyingee commented 5 years ago

Hi, I use sample_data.csv to train the model,but i get lower behavior on training data than validation set. I am confused.

Epoch 1/200 450/450 [==============================] - 5s 11ms/step - loss: 0.8987 - acc: 0.4933 - val_loss: 0.7807 - val_acc: 0.4286 Epoch 2/200 450/450 [==============================] - 1s 2ms/step - loss: 0.7921 - acc: 0.5356 - val_loss: 0.6995 - val_acc: 0.5306 Epoch 3/200 450/450 [==============================] - 1s 2ms/step - loss: 0.7451 - acc: 0.5644 - val_loss: 0.6261 - val_acc: 0.5918 Epoch 4/200 450/450 [==============================] - 1s 2ms/step - loss: 0.6697 - acc: 0.6178 - val_loss: 0.5605 - val_acc: 0.7143 Epoch 5/200

amansrivastava17 commented 5 years ago

@Kiteflyingee Since this dataset is so small, this problem might occur so often. I suggest you use quora full dataset on Kaggle to get higher accuracy.

https://www.kaggle.com/quora/question-pairs-dataset

Kiteflyingee commented 5 years ago

@amansrivastava17 Thanks, I'm trying to extend my dataset!