High val_loss and low val_accuracy while train is okay

TensorFlow General Discussion
1

Hello. Maybe abyone could give me some fast answer with this:

Epoch 2/50
20/20 [==============================] - 14s 646ms/step - loss: 1.5944 - accuracy: 0.3752 - val_loss: 252.2562 - val_accuracy: 0.2038 - lr: 0.0010
Epoch 3/50
20/20 [==============================] - 14s 651ms/step - loss: 1.3564 - accuracy: 0.4499 - val_loss: 87.6514 - val_accuracy: 0.2038 - lr: 0.0010
Epoch 4/50
20/20 [==============================] - 14s 650ms/step - loss: 1.2102 - accuracy: 0.4984 - val_loss: 89.7791 - val_accuracy: 0.2038 - lr: 0.0010
Epoch 5/50
20/20 [==============================] - 14s 651ms/step - loss: 1.0986 - accuracy: 0.5556 - val_loss: 33.2936 - val_accuracy: 0.2006 - lr: 0.0010
Epoch 6/50
20/20 [==============================] - 15s 679ms/step - loss: 1.0566 - accuracy: 0.5835 - val_loss: 117.8140 - val_accuracy: 0.2038 - lr: 0.0010
Epoch 7/50
20/20 [==============================] - 14s 664ms/step - loss: 1.0371 - accuracy: 0.5707 - val_loss: 115.3857 - val_accuracy: 0.2038 - lr: 0.0010
Epoch 8/50
20/20 [==============================] - 14s 654ms/step - loss: 0.9841 - accuracy: 0.6041 - val_loss: 36.1700 - val_accuracy: 0.2038 - lr: 0.0010
Epoch 9/50
20/20 [==============================] - 14s 661ms/step - loss: 0.9125 - accuracy: 0.6351 - val_loss: 42.5603 - val_accuracy: 0.2038 - lr: 0.0010
Epoch 10/50
20/20 [==============================] - 14s 655ms/step - loss: 0.8932 - accuracy: 0.6510 - val_loss: 128.5940 - val_accuracy: 0.2038 - lr: 0.0010
Epoch 11/50
20/20 [==============================] - 14s 659ms/step - loss: 0.8138 - accuracy: 0.6685 - val_loss: 51.7578 - val_accuracy: 0.2102 - lr: 5.0000e-04
Epoch 12/50
20/20 [==============================] - 14s 654ms/step - loss: 0.7900 - accuracy: 0.7003 - val_loss: 20.2659 - val_accuracy: 0.2962 - lr: 5.0000e-04
Epoch 13/50
20/20 [==============================] - 14s 660ms/step - loss: 0.7158 - accuracy: 0.7194 - val_loss: 48.9284 - val_accuracy: 0.1911 - lr: 5.0000e-04

Does it mean my dataset is too weak?

2

Your model is showing clear signs of overfitting - the training accuracy is improving (from 37% to 71%) while validation accuracy remains low (~20-29%). This indicates a gap between training and validation performance.

Here are quick fixes to try:

# Add regularization
model.add(tf.keras.layers.Dropout(0.3))

# Or reduce model complexity
model.add(tf.keras.layers.Dense(units=64, kernel_regularizer=tf.keras.regularizers.l2(0.01)))

# Increase data augmentation
data_augmentation = tf.keras.Sequential([
    tf.keras.layers.RandomFlip("horizontal"),
    tf.keras.layers.RandomRotation(0.2)
])

Focus on regularization techniques first.