We Keep Coding

Regarding the accuracy of the Siamese CNN

# We have 2 inputs, 1 for each picture
left_input = Input(img_size)
right_input = Input(img_size)
# We will use 2 instances of 1 network for this task
convnet = MobileNetV2(weights='imagenet', include_top=False, input_shape=img_size,input_tensor=None)
convnet.trainable=True
x=convnet.output
x=tf.keras.layers.GlobalAveragePooling2D()(x)
x=Dense(320,activation='relu')(x)
x=Dropout(0.2)(x)
preds = Dense(101, activation='sigmoid')(x) # Apply sigmoid
convnet = Model(inputs=convnet.input, outputs=preds)
# Connect each 'leg' of the network to each input
# Remember, they have the same weights
encoded_l = convnet(left_input)
encoded_r = convnet(right_input)
# Getting the L1 Distance between the 2 encodings
L1_layer = Lambda(lambda tensor:K.abs(tensor[0] - tensor[1]))
# Add the distance function to the network
L1_distance = L1_layer([encoded_l, encoded_r])
prediction = Dense(1,activation='sigmoid')(L1_distance)
siamese_net = Model(inputs=[left_input,right_input],outputs=prediction)
optimizer = Adam(lr, decay=2.5e-4)
#//TODO: get layerwise learning rates and momentum annealing scheme described in paperworking
siamese_net.compile(loss=keras.losses.binary_crossentropy,optimizer=optimizer,metrics=['accuracy'])
siamese_net.summary()
and the result of training is as follows
Epoch 1/10
126/126 [==============================] - 169s 1s/step - loss: 0.5683 - accuracy: 0.6840 - val_loss: 0.4644 - val_accuracy: 0.8044
Epoch 2/10
126/126 [==============================] - 163s 1s/step - loss: 0.2032 - accuracy: 0.9795 - val_loss: 0.2117 - val_accuracy: 0.9681
Epoch 3/10
126/126 [==============================] - 163s 1s/step - loss: 0.1110 - accuracy: 0.9925 - val_loss: 0.1448 - val_accuracy: 0.9840
Epoch 4/10
126/126 [==============================] - 164s 1s/step - loss: 0.0844 - accuracy: 0.9950 - val_loss: 0.1384 - val_accuracy: 0.9820
Epoch 5/10
126/126 [==============================] - 163s 1s/step - loss: 0.0634 - accuracy: 0.9990 - val_loss: 0.0829 - val_accuracy: 1.0000
Epoch 6/10
126/126 [==============================] - 165s 1s/step - loss: 0.0526 - accuracy: 0.9995 - val_loss: 0.0729 - val_accuracy: 1.0000
Epoch 7/10
126/126 [==============================] - 164s 1s/step - loss: 0.0465 - accuracy: 0.9995 - val_loss: 0.0641 - val_accuracy: 1.0000
Epoch 8/10
126/126 [==============================] - 163s 1s/step - loss: 0.0463 - accuracy: 0.9985 - val_loss: 0.0595 - val_accuracy: 1.0000
The model is predicting with good accuracy, when i am comparing two dissimilar images. Further it is predicting really good with same class of images.
But when I am comparing Image1 with image1 itself, it is predicting that they are similar only with the probability of 0.5.
in other case if I compare image1 with image2, then it is predicting correctly with a probability of 0.8.(here image1 and image2 belongs to same class)
when I am comparing individual images, it is predicting correctly, I have tried different alternatives did not workout.
May I know what might be the error?

The L1 distance between two equal vectors is always zero.
When you pass the same image, the encodings generated are equal (encoded_l is equal to encoded_r). Hence, the input to your final sigmoid layer is a zero vector.
And, sigmoid(0) = 0.5.
This is the reason providing identical inputs to your model gives 0.5 as the output.

InceptionResNetV2 validation accuracy stuck around 20% to 30%

I tried to train a CNN to classify 9 class of image. Each class has 1000 image for training. I tried training on VGG16 and VGG19, both can achieve validation accuracy of 90%. But when I tried to train on InceptionResNetV2 model, the model seems to stuck around 20% and 30%. Below is my code for InceptionResNetV2 and the training. What can I do to improve the training?
base_model = tf.keras.applications.InceptionResNetV2(input_shape=(IMG_HEIGHT, IMG_WIDTH ,3),weights = 'imagenet',include_top=False)
base_model.trainable = False
model = tf.keras.Sequential([
base_model,
Flatten(),
Dense(1024, activation = 'relu', kernel_regularizer=regularizers.l2(0.001)),
LeakyReLU(alpha=0.4),
Dropout(0.5),
BatchNormalization(),
Dense(1024, activation = 'relu', kernel_regularizer=regularizers.l2(0.001)),
LeakyReLU(alpha=0.4),
Dense(9, activation = 'softmax')])
optimizer_model = tf.keras.optimizers.Adam(learning_rate=0.0001, name='Adam', decay=0.00001)
loss_model = tf.keras.losses.CategoricalCrossentropy(from_logits=True)
model.compile(optimizer_model, loss="categorical_crossentropy", metrics=['accuracy'])
Epoch 1/10
899/899 [==============================] - 255s 283ms/step - loss: 4.3396 - acc: 0.3548 - val_loss: 4.2744 - val_acc: 0.3874
Epoch 2/10
899/899 [==============================] - 231s 257ms/step - loss: 3.5856 - acc: 0.4695 - val_loss: 3.9151 - val_acc: 0.3816
Epoch 3/10
899/899 [==============================] - 225s 250ms/step - loss: 3.1451 - acc: 0.4959 - val_loss: 4.8801 - val_acc: 0.2425
Epoch 4/10
899/899 [==============================] - 227s 252ms/step - loss: 2.7771 - acc: 0.5124 - val_loss: 3.7167 - val_acc: 0.3023
Epoch 5/10
899/899 [==============================] - 231s 257ms/step - loss: 2.4993 - acc: 0.5260 - val_loss: 3.7276 - val_acc: 0.3770
Epoch 6/10
899/899 [==============================] - 227s 252ms/step - loss: 2.3148 - acc: 0.5251 - val_loss: 3.7677 - val_acc: 0.3115
Epoch 7/10
899/899 [==============================] - 234s 260ms/step - loss: 2.1381 - acc: 0.5379 - val_loss: 3.4867 - val_acc: 0.2862
Epoch 8/10
899/899 [==============================] - 230s 256ms/step - loss: 2.0091 - acc: 0.5367 - val_loss: 4.1032 - val_acc: 0.3080
Epoch 9/10
899/899 [==============================] - 225s 251ms/step - loss: 1.9155 - acc: 0.5399 - val_loss: 4.1270 - val_acc: 0.2954
Epoch 10/10
899/899 [==============================] - 232s 258ms/step - loss: 1.8349 - acc: 0.5508 - val_loss: 4.3918 - val_acc: 0.2276

VGG-16/19 has a depth of 23/26 layers, whereas, InceptionResNetV2 has a depth of 572 layers. Now, there is minimal domain similarity between medical images and imagenet dataset. In VGG, due to low depth the features you're getting are not that complex and network is able to classify it on the basis of Dense layer features. However, in IRV2 network, as it's too much deep, the output of the fc layer is more complex (visualize it something object like but for imagenet dataset), and, then the features obtained from these layers are unable to connect to the Dense layer features, and, hence overfitting. I think you were able to get my point.
Check out my answer to very similar question of yours on this link: Link. It will help improve your accuracy.

Validation loss and validation accuracy both are higher than training loss and acc and fluctuating

I am trying to train my model using transfer learning, for this I am using VGG16 model, stripped the top layers and froze first 2 layers for using imagenet initial weights. For fine tuning them I am using learning rate 0.0001, activation softmax, dropout 0.5, loss categorical crossentropy, optimizer SGD, classes 46.
I am just unable to understand the behavior while training. Train loss and acc both are fine (loss is decreasing, acc is increasing). Val loss is decreasing and acc is increasing as well, BUT they are always higher than the train loss and acc.
Assuming its overfitting I made the model less complex, increased the dropout rate, added more samples to val data, but nothing seemed to work. I am a newbie so any kind of help is appreciated.
26137/26137 [==============================] - 7446s 285ms/step - loss: 1.1200 - accuracy: 0.3810 - val_loss: 3.1219 - val_accuracy: 0.4467
Epoch 2/50
26137/26137 [==============================] - 7435s 284ms/step - loss: 0.9944 - accuracy: 0.4353 - val_loss: 2.9348 - val_accuracy: 0.4694
Epoch 3/50
26137/26137 [==============================] - 7532s 288ms/step - loss: 0.9561 - accuracy: 0.4530 - val_loss: 1.6025 - val_accuracy: 0.4780
Epoch 4/50
26137/26137 [==============================] - 7436s 284ms/step - loss: 0.9343 - accuracy: 0.4631 - val_loss: 1.3032 - val_accuracy: 0.4860
Epoch 5/50
26137/26137 [==============================] - 7358s 282ms/step - loss: 0.9185 - accuracy: 0.4703 - val_loss: 1.4461 - val_accuracy: 0.4847
Epoch 6/50
26137/26137 [==============================] - 7396s 283ms/step - loss: 0.9083 - accuracy: 0.4748 - val_loss: 1.4093 - val_accuracy: 0.4908
Epoch 7/50
26137/26137 [==============================] - 7424s 284ms/step - loss: 0.8993 - accuracy: 0.4789 - val_loss: 1.4617 - val_accuracy: 0.4939
Epoch 8/50
26137/26137 [==============================] - 7433s 284ms/step - loss: 0.8925 - accuracy: 0.4822 - val_loss: 1.4257 - val_accuracy: 0.4978
Epoch 9/50
26137/26137 [==============================] - 7445s 285ms/step - loss: 0.8868 - accuracy: 0.4851 - val_loss: 1.5568 - val_accuracy: 0.4953
Epoch 10/50
26137/26137 [==============================] - 7387s 283ms/step - loss: 0.8816 - accuracy: 0.4874 - val_loss: 1.4534 - val_accuracy: 0.4970
Epoch 11/50
26137/26137 [==============================] - 7374s 282ms/step - loss: 0.8779 - accuracy: 0.4894 - val_loss: 1.4605 - val_accuracy: 0.4912
Epoch 12/50
26137/26137 [==============================] - 7411s 284ms/step - loss: 0.8733 - accuracy: 0.4915 - val_loss: 1.4694 - val_accuracy: 0.5030

Yes, you are facing over-fitting issue. To mitigate, you can try to implement below steps
1.Shuffle the Data, by using shuffle=True in VGG16_model.fit. Code is shown below:
history = VGG16_model.fit(x_train, y_train, batch_size=batch_size, epochs=epochs, verbose=1,
validation_data=(x_validation, y_validation), shuffle = True)
2.Use Early Stopping. Code is shown below
callback = tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience=15)
3.Use Regularization. Code for Regularization is shown below (You can try l1 Regularization or l1_l2 Regularization as well):
from tensorflow.keras.regularizers import l2
Regularizer = l2(0.001)
VGG16_model.add(Conv2D(96,11, 11, input_shape = (227,227,3),strides=(4,4), padding='valid', activation='relu', data_format='channels_last',
activity_regularizer=Regularizer, kernel_regularizer=Regularizer))
VGG16_model.add(Dense(units = 2, activation = 'sigmoid',
activity_regularizer=Regularizer, kernel_regularizer=Regularizer))
4.You can try using BatchNormalization.
5.Perform Image Data Augmentation using ImageDataGenerator. Refer this link for more info about that.
6.If the Pixels are not Normalized, Dividing the Pixel Values with 255 also helps

val_loss is halved but val_acc stays constant

I am training a neural network and get the following output. The loss and val_loss are both decreasing, which makes me happy. However, the val_acc keeps constant. What reasons can that have? My data is quite imbalanced, but I am weighing it via the sklearn compute_class_weight function.
Train on 109056 samples, validate on 27136 samples
Epoch 1/200
- 1174s - loss: 1.0353 - acc: 0.5843 - val_loss: 1.0749 - val_acc: 0.7871
Epoch 00001: val_acc improved from -inf to 0.78711, saving model to
nn_best_weights.h5
Epoch 2/200
- 1174s - loss: 1.0122 - acc: 0.6001 - val_loss: 1.0642 - val_acc: 0.9084
Epoch 00002: val_acc improved from 0.78711 to 0.90842, saving model to
nn_best_weights.h5
Epoch 3/200
- 1176s - loss: 0.9974 - acc: 0.5885 - val_loss: 1.0445 - val_acc: 0.9257
Epoch 00003: val_acc improved from 0.90842 to 0.92571, saving model to
nn_best_weights.h5
Epoch 4/200
- 1177s - loss: 0.9834 - acc: 0.5760 - val_loss: 1.0071 - val_acc: 0.9260
Epoch 00004: val_acc improved from 0.92571 to 0.92597, saving model to
nn_best_weights.h5
Epoch 5/200
- 1182s - loss: 0.9688 - acc: 0.5639 - val_loss: 1.0175 - val_acc: 0.9260
Epoch 00005: val_acc did not improve from 0.92597
Epoch 6/200
- 1177s - loss: 0.9449 - acc: 0.5602 - val_loss: 0.9976 - val_acc: 0.9246
Epoch 00006: val_acc did not improve from 0.92597
Epoch 7/200
- 1186s - loss: 0.9070 - acc: 0.5598 - val_loss: 0.9667 - val_acc: 0.9258
Epoch 00007: val_acc did not improve from 0.92597
Epoch 8/200
- 1178s - loss: 0.8541 - acc: 0.5663 - val_loss: 0.9254 - val_acc: 0.9221
Epoch 00008: val_acc did not improve from 0.92597
Epoch 9/200
- 1171s - loss: 0.7859 - acc: 0.5853 - val_loss: 0.8686 - val_acc: 0.9237
Epoch 00009: val_acc did not improve from 0.92597
Epoch 10/200
- 1172s - loss: 0.7161 - acc: 0.6139 - val_loss: 0.8119 - val_acc: 0.9260
Epoch 00010: val_acc did not improve from 0.92597
Epoch 11/200
- 1168s - loss: 0.6500 - acc: 0.6416 - val_loss: 0.7531 - val_acc: 0.9259
Epoch 00011: val_acc did not improve from 0.92597
Epoch 12/200
- 1164s - loss: 0.5967 - acc: 0.6676 - val_loss: 0.7904 - val_acc: 0.9260
Epoch 00012: val_acc did not improve from 0.92597
Epoch 13/200
- 1175s - loss: 0.5608 - acc: 0.6848 - val_loss: 0.7589 - val_acc: 0.9259
Epoch 00013: val_acc did not improve from 0.92597
Epoch 14/200
- 1221s - loss: 0.5377 - acc: 0.6980 - val_loss: 0.7811 - val_acc: 0.9260
Epoch 00014: val_acc did not improve from 0.92597
My model is the following. I know the kernel size is quite large, but that is on purpose because the data is structured in a certain way.
cnn = Sequential()
cnn.add(Conv2D(16, kernel_size=(2, 100), padding='same', data_format="channels_first", input_shape=(1,10, 100)))
cnn.add(LeakyReLU(alpha=0.01))
cnn.add(BatchNormalization())
cnn.add(Conv2D(16, (1, 1)))
cnn.add(LeakyReLU(alpha=0.01))
cnn.add(Conv2D(16, (1, 8)))
cnn.add(Flatten())
rnn = Sequential()
rnn = LSTM(100, return_sequences=False, dropout=0.2)
dense = Sequential()
dense.add(Dense(3, activation='softmax'))
main_input = Input(batch_shape=(512, 1, 1, 10, 100))
model = TimeDistributed(cnn)(main_input)
model = rnn(model)
model = dense(model)
replica = Model(inputs=main_input, outputs=model)
replica.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

It is hard to answer you question not knowing your model.
The possible answers are:
There is nothing wrong with your model. This may be the highest
accuracy you can get.
Your data can be imbalanced or not shuffled. Higher val_acc then acc indicates that the can be something wrong with train, evaluation
and test split. Train accuracy tends to be higher then val_acc on the begining. Then val_acc catches up, or not ;) I can also indicate that there is not much of variance in your dataset, then you could have this kind of behaviour.
Your learn rate may be to big. Try to decrease it.
And I guess the actual metric for the model to minimize is the loss, so during the optimization process you should follow the loss and monitor its improvements.
Check this link for more information how to check your model.

It seems to be the case where learning rate is too high, missing the local minimum and preventing the Neural Network to improve learning:
It would be good if you can customize your optimizer, like this:
learning_rate = 0.008
decay_rate = 5e-6
momentum = 0.65
sgd = SGD(lr=learning_rate,momentum=momentum, decay=decay_rate, nesterov=False)
model.compile(loss="categorical_crossentropy", optimizer=sgd,metrics=['accuracy'])
Also, increase the number of convolutions. Weights may be saturated.

Sudden 50% accuracy drop while training convolutional NN

Training convolutional neural network from scratch on my own dataset with Keras and Tensorflow.
learning rate = 0.0001,
5 classes to sort,
no Dropout used,
dataset checked twice, no wrong labels found
Model:
model = models.Sequential()
model.add(layers.Conv2D(16,(2,2),activation='relu',input_shape=(75,75,3)))
model.add(layers.MaxPooling2D((2,2)))
model.add(layers.Conv2D(16,(2,2),activation='relu'))
model.add(layers.MaxPooling2D((2,2)))
model.add(layers.Conv2D(32,(2,2),activation='relu'))
model.add(layers.MaxPooling2D((2,2)))
model.add(layers.Flatten())
model.add(layers.Dense(128,activation='relu'))
model.add(layers.Dense(5,activation='sigmoid'))
model.compile(optimizer=optimizers.adam(lr=0.0001),
loss='categorical_crossentropy',
metrics=['acc'])
history = model.fit_generator(train_generator,
steps_per_epoch=100,
epochs=50,
validation_data=val_generator,
validation_steps=25)
Everytime when model achieves 25-35 epochs (80-90% accuracy) this happens:
Epoch 31/50
100/100 [==============================] - 3s 34ms/step - loss: 0.3524 - acc: 0.8558 - val_loss: 0.4151 - val_acc: 0.7992
Epoch 32/50
100/100 [==============================] - 3s 34ms/step - loss: 0.3393 - acc: 0.8700 - val_loss: 0.4384 - val_acc: 0.7951
Epoch 33/50
100/100 [==============================] - 3s 34ms/step - loss: 0.3321 - acc: 0.8702 - val_loss: 0.4993 - val_acc: 0.7620
Epoch 34/50
100/100 [==============================] - 3s 33ms/step - loss: 1.5444 - acc: 0.3302 - val_loss: 1.6062 - val_acc: 0.1704
Epoch 35/50
100/100 [==============================] - 3s 34ms/step - loss: 1.6094 - acc: 0.2935 - val_loss: 1.6062 - val_acc: 0.1724
There is some similar problems with answers, but mostly they recommend to lower learning rate, but it doesnt help at all.
UPD: almost all weights and biases in network became nan. Network somehow died inside

Solution in this case:
I changed sigmoid function in last layer to softmax function and drops are gone
Why this worked out?
sigmoid activation function is used for binary (two-class) classifications.
In multiclassification problems we should use softmax function - special extension of sigmoid function for multiclassification problems.
More information: Sigmoid vs Softmax
Special thanks to #desertnaut and #Shubham Panchal for error indication