2
$\begingroup$

During model training, I noticed various behaviour in between training and validation accuracy. I understand that 'The training set is used to train the model, while the validation set is only used to evaluate the model's performance...', but I'd like to know if there is any relationship between training and validation accuracy and if yes,

1) what exactly is happening when training and validation accuracy change during training and;

2) what do different behaviours imply

For instance, some believe there is overfitting problem if training > validation accuracy. What happens if one is greater than the other alternately, which is the case below?

Here is the code

inputs_1 = keras.Input(shape=(10081,1))

layer1 = Conv1D(64,14)(inputs_1)
layer2 = layers.MaxPool1D(5)(layer1)
layer3 = Conv1D(64, 14)(layer2)
layer4 = layers.GlobalMaxPooling1D()(layer3)

inputs_2 = keras.Input(shape=(104,))             
layer5 = layers.concatenate([layer4, inputs_2])
layer6 = Dense(128, activation='relu')(layer5)
layer7 = Dense(2, activation='softmax')(layer6)


model_2 = keras.models.Model(inputs = [inputs_1, inputs_2], output = [layer7])
model_2.summary()


X_train, X_test, y_train, y_test = train_test_split(df.iloc[:,0:10185], df[['Result_cat','Result_cat1']].values, test_size=0.2) 
X_train = X_train.to_numpy()
X_train = X_train.reshape([X_train.shape[0], X_train.shape[1], 1]) 
X_train_1 = X_train[:,0:10081,:]
X_train_2 = X_train[:,10081:10185,:].reshape(736,104)   


X_test = X_test.to_numpy()
X_test = X_test.reshape([X_test.shape[0], X_test.shape[1], 1]) 
X_test_1 = X_test[:,0:10081,:]
X_test_2 = X_test[:,10081:10185,:].reshape(185,104)    

adam = keras.optimizers.Adam(lr = 0.0005)
model_2.compile(loss = 'categorical_crossentropy', optimizer = adam, metrics = ['acc'])

history = model_2.fit([X_train_1,X_train_2], y_train, epochs = 120, batch_size = 256, validation_split = 0.2, callbacks = [keras.callbacks.EarlyStopping(monitor='val_loss', patience=20)])

model summary

/usr/local/lib/python3.6/dist-packages/ipykernel_launcher.py:15: UserWarning: Update your `Model` call to the Keras 2 API: `Model(inputs=[<tf.Tenso..., outputs=[<tf.Tenso...)`
  from ipykernel import kernelapp as app
Model: "model_3"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_5 (InputLayer)            (None, 10081, 1)     0                                            
__________________________________________________________________________________________________
conv1d_5 (Conv1D)               (None, 10068, 64)    960         input_5[0][0]                    
__________________________________________________________________________________________________
max_pooling1d_3 (MaxPooling1D)  (None, 2013, 64)     0           conv1d_5[0][0]                   
__________________________________________________________________________________________________
conv1d_6 (Conv1D)               (None, 2000, 64)     57408       max_pooling1d_3[0][0]            
__________________________________________________________________________________________________
global_max_pooling1d_3 (GlobalM (None, 64)           0           conv1d_6[0][0]                   
__________________________________________________________________________________________________
input_6 (InputLayer)            (None, 104)          0                                            
__________________________________________________________________________________________________
concatenate_3 (Concatenate)     (None, 168)          0           global_max_pooling1d_3[0][0]     
                                                                 input_6[0][0]                    
__________________________________________________________________________________________________
dense_5 (Dense)                 (None, 128)          21632       concatenate_3[0][0]              
__________________________________________________________________________________________________
dense_6 (Dense)                 (None, 2)            258         dense_5[0][0]                    
==================================================================================================
Total params: 80,258
Trainable params: 80,258
Non-trainable params: 0

and the training process

__________________________________________________________________________________________________
Train on 588 samples, validate on 148 samples
Epoch 1/120
588/588 [==============================] - 16s 26ms/step - loss: 5.6355 - acc: 0.4932 - val_loss: 4.1086 - val_acc: 0.6216
Epoch 2/120
588/588 [==============================] - 15s 25ms/step - loss: 4.5977 - acc: 0.5748 - val_loss: 3.8252 - val_acc: 0.4459
Epoch 3/120
588/588 [==============================] - 15s 25ms/step - loss: 4.3815 - acc: 0.4575 - val_loss: 2.4087 - val_acc: 0.6622
Epoch 4/120
588/588 [==============================] - 15s 25ms/step - loss: 3.7480 - acc: 0.6003 - val_loss: 2.0060 - val_acc: 0.6892
Epoch 5/120
588/588 [==============================] - 15s 25ms/step - loss: 3.3019 - acc: 0.5408 - val_loss: 2.3176 - val_acc: 0.5676
Epoch 6/120
588/588 [==============================] - 15s 25ms/step - loss: 3.1739 - acc: 0.5663 - val_loss: 2.2607 - val_acc: 0.6892
Epoch 7/120
588/588 [==============================] - 15s 25ms/step - loss: 3.2322 - acc: 0.6207 - val_loss: 1.8898 - val_acc: 0.7230
Epoch 8/120
588/588 [==============================] - 15s 25ms/step - loss: 2.9777 - acc: 0.6020 - val_loss: 1.8401 - val_acc: 0.7500
Epoch 9/120
588/588 [==============================] - 15s 25ms/step - loss: 2.8982 - acc: 0.6429 - val_loss: 1.8517 - val_acc: 0.7365
Epoch 10/120
588/588 [==============================] - 15s 25ms/step - loss: 2.8342 - acc: 0.6344 - val_loss: 1.7941 - val_acc: 0.7095
Epoch 11/120
588/588 [==============================] - 15s 25ms/step - loss: 2.7426 - acc: 0.6327 - val_loss: 1.8495 - val_acc: 0.7162
Epoch 12/120
588/588 [==============================] - 15s 25ms/step - loss: 2.7340 - acc: 0.6531 - val_loss: 1.7652 - val_acc: 0.7162
Epoch 13/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6680 - acc: 0.6616 - val_loss: 1.8097 - val_acc: 0.7365
Epoch 14/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6922 - acc: 0.6786 - val_loss: 1.7143 - val_acc: 0.7500
Epoch 15/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6161 - acc: 0.6786 - val_loss: 1.6960 - val_acc: 0.7568
Epoch 16/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6054 - acc: 0.6905 - val_loss: 1.6779 - val_acc: 0.7297
Epoch 17/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6072 - acc: 0.6684 - val_loss: 1.6750 - val_acc: 0.7703
Epoch 18/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5907 - acc: 0.6871 - val_loss: 1.6774 - val_acc: 0.7432
Epoch 19/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5854 - acc: 0.6718 - val_loss: 1.6609 - val_acc: 0.7770
Epoch 20/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5621 - acc: 0.6905 - val_loss: 1.6709 - val_acc: 0.7365
Epoch 21/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5515 - acc: 0.6854 - val_loss: 1.6904 - val_acc: 0.7703
Epoch 22/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5749 - acc: 0.6837 - val_loss: 1.6862 - val_acc: 0.7297
Epoch 23/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6212 - acc: 0.6514 - val_loss: 1.7215 - val_acc: 0.7568
Epoch 24/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6532 - acc: 0.6633 - val_loss: 1.7105 - val_acc: 0.7230
Epoch 25/120
588/588 [==============================] - 15s 25ms/step - loss: 2.7300 - acc: 0.6344 - val_loss: 1.6870 - val_acc: 0.7432
Epoch 26/120
588/588 [==============================] - 15s 25ms/step - loss: 2.7355 - acc: 0.6650 - val_loss: 1.6733 - val_acc: 0.7703
Epoch 27/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6336 - acc: 0.6650 - val_loss: 1.6572 - val_acc: 0.7297
Epoch 28/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6018 - acc: 0.6803 - val_loss: 1.7292 - val_acc: 0.7635
Epoch 29/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5448 - acc: 0.7143 - val_loss: 1.8065 - val_acc: 0.7095
Epoch 30/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5724 - acc: 0.6820 - val_loss: 1.8029 - val_acc: 0.7297
Epoch 31/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6622 - acc: 0.6650 - val_loss: 1.6594 - val_acc: 0.7568
Epoch 32/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6211 - acc: 0.6582 - val_loss: 1.6375 - val_acc: 0.7770
Epoch 33/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5911 - acc: 0.6854 - val_loss: 1.6964 - val_acc: 0.7500
Epoch 34/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5050 - acc: 0.7262 - val_loss: 1.8496 - val_acc: 0.6892
Epoch 35/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6012 - acc: 0.6752 - val_loss: 1.7443 - val_acc: 0.7432
Epoch 36/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5688 - acc: 0.6871 - val_loss: 1.6220 - val_acc: 0.7568
Epoch 37/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4843 - acc: 0.7279 - val_loss: 1.6166 - val_acc: 0.7905
Epoch 38/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4707 - acc: 0.7449 - val_loss: 1.6496 - val_acc: 0.7905
Epoch 39/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4683 - acc: 0.7109 - val_loss: 1.6641 - val_acc: 0.7432
Epoch 40/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4671 - acc: 0.7279 - val_loss: 1.6553 - val_acc: 0.7703
Epoch 41/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4479 - acc: 0.7347 - val_loss: 1.6302 - val_acc: 0.7973
Epoch 42/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4355 - acc: 0.7551 - val_loss: 1.6241 - val_acc: 0.7973
Epoch 43/120
588/588 [==============================] - 14s 25ms/step - loss: 2.4286 - acc: 0.7568 - val_loss: 1.6249 - val_acc: 0.7973
Epoch 44/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4250 - acc: 0.7585 - val_loss: 1.6248 - val_acc: 0.7770
Epoch 45/120
588/588 [==============================] - 14s 25ms/step - loss: 2.4198 - acc: 0.7517 - val_loss: 1.6212 - val_acc: 0.7703
Epoch 46/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4246 - acc: 0.7568 - val_loss: 1.6129 - val_acc: 0.7838
Epoch 47/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4237 - acc: 0.7517 - val_loss: 1.6166 - val_acc: 0.7973
Epoch 48/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4287 - acc: 0.7432 - val_loss: 1.6309 - val_acc: 0.8041
Epoch 49/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4179 - acc: 0.7381 - val_loss: 1.6271 - val_acc: 0.7838
Epoch 50/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4164 - acc: 0.7381 - val_loss: 1.6258 - val_acc: 0.7973
Epoch 51/120
588/588 [==============================] - 14s 24ms/step - loss: 2.1996 - acc: 0.7398 - val_loss: 1.3612 - val_acc: 0.7973
Epoch 52/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1387 - acc: 0.8265 - val_loss: 1.4811 - val_acc: 0.7973
Epoch 53/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1607 - acc: 0.8078 - val_loss: 1.5060 - val_acc: 0.7838
Epoch 54/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1783 - acc: 0.8129 - val_loss: 1.4878 - val_acc: 0.8176
Epoch 55/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1745 - acc: 0.8197 - val_loss: 1.4762 - val_acc: 0.8108
Epoch 56/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1764 - acc: 0.8129 - val_loss: 1.4631 - val_acc: 0.7905
Epoch 57/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1637 - acc: 0.8078 - val_loss: 1.4615 - val_acc: 0.7770
Epoch 58/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1563 - acc: 0.8112 - val_loss: 1.4487 - val_acc: 0.7703
Epoch 59/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1396 - acc: 0.8146 - val_loss: 1.4362 - val_acc: 0.7905
Epoch 60/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1240 - acc: 0.8316 - val_loss: 1.4333 - val_acc: 0.8041
Epoch 61/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1173 - acc: 0.8333 - val_loss: 1.4369 - val_acc: 0.8041
Epoch 62/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1228 - acc: 0.8384 - val_loss: 1.4393 - val_acc: 0.8041
Epoch 63/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1113 - acc: 0.8316 - val_loss: 1.4380 - val_acc: 0.8041
Epoch 64/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1102 - acc: 0.8452 - val_loss: 1.4217 - val_acc: 0.8041
Epoch 65/120
588/588 [==============================] - 15s 25ms/step - loss: 1.0961 - acc: 0.8469 - val_loss: 1.4129 - val_acc: 0.7973
Epoch 66/120
588/588 [==============================] - 15s 25ms/step - loss: 1.0903 - acc: 0.8537 - val_loss: 1.4019 - val_acc: 0.8041
Epoch 67/120
588/588 [==============================] - 15s 25ms/step - loss: 1.0890 - acc: 0.8503 - val_loss: 1.3850 - val_acc: 0.8176
Epoch 68/120
588/588 [==============================] - 15s 25ms/step - loss: 1.0878 - acc: 0.8520 - val_loss: 1.4035 - val_acc: 0.7635
Epoch 69/120
588/588 [==============================] - 15s 25ms/step - loss: 1.0984 - acc: 0.8469 - val_loss: 1.4060 - val_acc: 0.8041
Epoch 70/120
588/588 [==============================] - 15s 25ms/step - loss: 1.0893 - acc: 0.8418 - val_loss: 1.3981 - val_acc: 0.7973
Epoch 71/120
588/588 [==============================] - 15s 25ms/step - loss: 1.0876 - acc: 0.8605 - val_loss: 1.3951 - val_acc: 0.8041__________________________________________________________________________________________________
Train on 588 samples, validate on 148 samples
Epoch 1/120
588/588 [==============================] - 16s 26ms/step - loss: 5.6355 - acc: 0.4932 - val_loss: 4.1086 - val_acc: 0.6216
Epoch 2/120
588/588 [==============================] - 15s 25ms/step - loss: 4.5977 - acc: 0.5748 - val_loss: 3.8252 - val_acc: 0.4459
Epoch 3/120
588/588 [==============================] - 15s 25ms/step - loss: 4.3815 - acc: 0.4575 - val_loss: 2.4087 - val_acc: 0.6622
Epoch 4/120
588/588 [==============================] - 15s 25ms/step - loss: 3.7480 - acc: 0.6003 - val_loss: 2.0060 - val_acc: 0.6892
Epoch 5/120
588/588 [==============================] - 15s 25ms/step - loss: 3.3019 - acc: 0.5408 - val_loss: 2.3176 - val_acc: 0.5676
Epoch 6/120
588/588 [==============================] - 15s 25ms/step - loss: 3.1739 - acc: 0.5663 - val_loss: 2.2607 - val_acc: 0.6892
Epoch 7/120
588/588 [==============================] - 15s 25ms/step - loss: 3.2322 - acc: 0.6207 - val_loss: 1.8898 - val_acc: 0.7230
Epoch 8/120
588/588 [==============================] - 15s 25ms/step - loss: 2.9777 - acc: 0.6020 - val_loss: 1.8401 - val_acc: 0.7500
Epoch 9/120
588/588 [==============================] - 15s 25ms/step - loss: 2.8982 - acc: 0.6429 - val_loss: 1.8517 - val_acc: 0.7365
Epoch 10/120
588/588 [==============================] - 15s 25ms/step - loss: 2.8342 - acc: 0.6344 - val_loss: 1.7941 - val_acc: 0.7095
Epoch 11/120
588/588 [==============================] - 15s 25ms/step - loss: 2.7426 - acc: 0.6327 - val_loss: 1.8495 - val_acc: 0.7162
Epoch 12/120
588/588 [==============================] - 15s 25ms/step - loss: 2.7340 - acc: 0.6531 - val_loss: 1.7652 - val_acc: 0.7162
Epoch 13/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6680 - acc: 0.6616 - val_loss: 1.8097 - val_acc: 0.7365
Epoch 14/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6922 - acc: 0.6786 - val_loss: 1.7143 - val_acc: 0.7500
Epoch 15/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6161 - acc: 0.6786 - val_loss: 1.6960 - val_acc: 0.7568
Epoch 16/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6054 - acc: 0.6905 - val_loss: 1.6779 - val_acc: 0.7297
Epoch 17/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6072 - acc: 0.6684 - val_loss: 1.6750 - val_acc: 0.7703
Epoch 18/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5907 - acc: 0.6871 - val_loss: 1.6774 - val_acc: 0.7432
Epoch 19/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5854 - acc: 0.6718 - val_loss: 1.6609 - val_acc: 0.7770
Epoch 20/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5621 - acc: 0.6905 - val_loss: 1.6709 - val_acc: 0.7365
Epoch 21/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5515 - acc: 0.6854 - val_loss: 1.6904 - val_acc: 0.7703
Epoch 22/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5749 - acc: 0.6837 - val_loss: 1.6862 - val_acc: 0.7297
Epoch 23/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6212 - acc: 0.6514 - val_loss: 1.7215 - val_acc: 0.7568
Epoch 24/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6532 - acc: 0.6633 - val_loss: 1.7105 - val_acc: 0.7230
Epoch 25/120
588/588 [==============================] - 15s 25ms/step - loss: 2.7300 - acc: 0.6344 - val_loss: 1.6870 - val_acc: 0.7432
Epoch 26/120
588/588 [==============================] - 15s 25ms/step - loss: 2.7355 - acc: 0.6650 - val_loss: 1.6733 - val_acc: 0.7703
Epoch 27/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6336 - acc: 0.6650 - val_loss: 1.6572 - val_acc: 0.7297
Epoch 28/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6018 - acc: 0.6803 - val_loss: 1.7292 - val_acc: 0.7635
Epoch 29/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5448 - acc: 0.7143 - val_loss: 1.8065 - val_acc: 0.7095
Epoch 30/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5724 - acc: 0.6820 - val_loss: 1.8029 - val_acc: 0.7297
Epoch 31/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6622 - acc: 0.6650 - val_loss: 1.6594 - val_acc: 0.7568
Epoch 32/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6211 - acc: 0.6582 - val_loss: 1.6375 - val_acc: 0.7770
Epoch 33/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5911 - acc: 0.6854 - val_loss: 1.6964 - val_acc: 0.7500
Epoch 34/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5050 - acc: 0.7262 - val_loss: 1.8496 - val_acc: 0.6892
Epoch 35/120
588/588 [==============================] - 15s 25ms/step - loss: 2.6012 - acc: 0.6752 - val_loss: 1.7443 - val_acc: 0.7432
Epoch 36/120
588/588 [==============================] - 15s 25ms/step - loss: 2.5688 - acc: 0.6871 - val_loss: 1.6220 - val_acc: 0.7568
Epoch 37/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4843 - acc: 0.7279 - val_loss: 1.6166 - val_acc: 0.7905
Epoch 38/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4707 - acc: 0.7449 - val_loss: 1.6496 - val_acc: 0.7905
Epoch 39/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4683 - acc: 0.7109 - val_loss: 1.6641 - val_acc: 0.7432
Epoch 40/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4671 - acc: 0.7279 - val_loss: 1.6553 - val_acc: 0.7703
Epoch 41/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4479 - acc: 0.7347 - val_loss: 1.6302 - val_acc: 0.7973
Epoch 42/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4355 - acc: 0.7551 - val_loss: 1.6241 - val_acc: 0.7973
Epoch 43/120
588/588 [==============================] - 14s 25ms/step - loss: 2.4286 - acc: 0.7568 - val_loss: 1.6249 - val_acc: 0.7973
Epoch 44/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4250 - acc: 0.7585 - val_loss: 1.6248 - val_acc: 0.7770
Epoch 45/120
588/588 [==============================] - 14s 25ms/step - loss: 2.4198 - acc: 0.7517 - val_loss: 1.6212 - val_acc: 0.7703
Epoch 46/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4246 - acc: 0.7568 - val_loss: 1.6129 - val_acc: 0.7838
Epoch 47/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4237 - acc: 0.7517 - val_loss: 1.6166 - val_acc: 0.7973
Epoch 48/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4287 - acc: 0.7432 - val_loss: 1.6309 - val_acc: 0.8041
Epoch 49/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4179 - acc: 0.7381 - val_loss: 1.6271 - val_acc: 0.7838
Epoch 50/120
588/588 [==============================] - 15s 25ms/step - loss: 2.4164 - acc: 0.7381 - val_loss: 1.6258 - val_acc: 0.7973
Epoch 51/120
588/588 [==============================] - 14s 24ms/step - loss: 2.1996 - acc: 0.7398 - val_loss: 1.3612 - val_acc: 0.7973
Epoch 52/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1387 - acc: 0.8265 - val_loss: 1.4811 - val_acc: 0.7973
Epoch 53/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1607 - acc: 0.8078 - val_loss: 1.5060 - val_acc: 0.7838
Epoch 54/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1783 - acc: 0.8129 - val_loss: 1.4878 - val_acc: 0.8176
Epoch 55/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1745 - acc: 0.8197 - val_loss: 1.4762 - val_acc: 0.8108
Epoch 56/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1764 - acc: 0.8129 - val_loss: 1.4631 - val_acc: 0.7905
Epoch 57/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1637 - acc: 0.8078 - val_loss: 1.4615 - val_acc: 0.7770
Epoch 58/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1563 - acc: 0.8112 - val_loss: 1.4487 - val_acc: 0.7703
Epoch 59/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1396 - acc: 0.8146 - val_loss: 1.4362 - val_acc: 0.7905
Epoch 60/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1240 - acc: 0.8316 - val_loss: 1.4333 - val_acc: 0.8041
Epoch 61/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1173 - acc: 0.8333 - val_loss: 1.4369 - val_acc: 0.8041
Epoch 62/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1228 - acc: 0.8384 - val_loss: 1.4393 - val_acc: 0.8041
Epoch 63/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1113 - acc: 0.8316 - val_loss: 1.4380 - val_acc: 0.8041
Epoch 64/120
588/588 [==============================] - 15s 25ms/step - loss: 1.1102 - acc: 0.8452 - val_loss: 1.4217 - val_acc: 0.8041
Epoch 65/120
588/588 [==============================] - 15s 25ms/step - loss: 1.0961 - acc: 0.8469 - val_loss: 1.4129 - val_acc: 0.7973
Epoch 66/120
588/588 [==============================] - 15s 25ms/step - loss: 1.0903 - acc: 0.8537 - val_loss: 1.4019 - val_acc: 0.8041
Epoch 67/120
588/588 [==============================] - 15s 25ms/step - loss: 1.0890 - acc: 0.8503 - val_loss: 1.3850 - val_acc: 0.8176
Epoch 68/120
588/588 [==============================] - 15s 25ms/step - loss: 1.0878 - acc: 0.8520 - val_loss: 1.4035 - val_acc: 0.7635
Epoch 69/120
588/588 [==============================] - 15s 25ms/step - loss: 1.0984 - acc: 0.8469 - val_loss: 1.4060 - val_acc: 0.8041
Epoch 70/120
588/588 [==============================] - 15s 25ms/step - loss: 1.0893 - acc: 0.8418 - val_loss: 1.3981 - val_acc: 0.7973
Epoch 71/120
588/588 [==============================] - 15s 25ms/step - loss: 1.0876 - acc: 0.8605 - val_loss: 1.3951 - val_acc: 0.8041

Notice how at first acc is lower than val_acc and later is greater than val_acc. Can someone please shed some light what could be happening here? Thank you

$\endgroup$
0
$\begingroup$

very interesting questions:

1) what exactly is happening when training and validation accuracy change during training

  • The accuracy change after every batch computation. You have 588 batches, so loss will be computed after each one of these batches (let's say each batch have 8 images). However, the accuracy you see in the progress bar it is the accuracy of the current batch averaged with the accuracy of all the previous batches so far. See keras.utils.generic_utils.Progbar.
  • The val_acc is computed only at the end of one epoch and it is computed with all your validation dataset at once (considering it as a single batch, so if you have 100 images for validation it will compute accuracy as a single batch of 100 images)

2) what do different behaviours imply

  • The acc and val_acc normally differ from each other due to different split sizes.

    • Try same experiment with validation_split=0.01 and validation_split=0.4 and you will see how both accuracy and val_acc will change.
    • Normally the greater the validation split, the more similar both metrics will be since the validation split will be big enough to be representative (let's say it has cats and dogs, not only cats), taking into account that you need enough data to train correctly. This explains why in some cases the val_acc is higher than accuracy and vice versa.
  • Overfitting only occurs when the graph fashion or tendency changes and val_acc starts to drop and accuracy keeping increasing. This means that your model can not do any better with the validation dataset (non previously seen images).

I work with loss and val_loss which are highly correlated with accuracy. Normally the loss is the inverse, so interpret the comments above in the inverse sense (sorry about the confusion but I'm taking this example from my current experiments) I hope it helps:

enter image description here

There are 2 experiments, orange and grey.

  • In both experiments, val_loss is always slightly higher than loss (because of my current validation split which it happens to be also 0.2, but normally is 0.01 and val_loss is even higher).

  • On both experiments the loss trend is linearly decreasing, this is because gradient descent works and the loss functions is well defined and it converges.

  • Orange experiment is overfitting from epoch 20 onwards because, the val_loss won't drop any more and, on the contrary, it start increasing.

  • Grey experiment is just right, both loss and val_loss are still decreasing, and although the val_loss might be greater than loss it is not overfitting because it is still decreasing. So that is why it is still training :)

Complex concepts here, I hope I was able to explain myself clearly! Cheers

| improve this answer | |
$\endgroup$
  • $\begingroup$ Thank you for the detailed explanation! Just a few questions: 1) If loss is computed after each batch and val_acc is computed only at the end of one epoch, does val_acc/accuracy reflect the accuracy of the entire model or only the batch? 2) Acc vs loss: when one is increasing, does the other necessarily have to decrease proportionally? Thank you $\endgroup$ – nilsinelabore Dec 4 '19 at 12:21
  • $\begingroup$ 1) In Keras after computing the accuracy for the batch, it is averaged with the accuracy from all the previous batches, so yes, it takes them all into account and hence it is representative for the accuracy of your model. Same with val_acc (but with only one batch) 2) In 99% of cases yes, but it depends on how you define loss and accuracy (in standard keras metrics they are inversely proportional) $\endgroup$ – JVGD Dec 4 '19 at 13:33
  • $\begingroup$ I edited the response to add this $\endgroup$ – JVGD Dec 4 '19 at 13:41
  • $\begingroup$ Thank you for the amazing answer! $\endgroup$ – nilsinelabore Dec 4 '19 at 21:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.