I'm trying to predict age from a given picture. I built the model below but the problem is that I'm getting very large loss value with low accuracy while fitting the model.

I think the problem is choosing the wrong loss function (here mean_squared_error). What can be the problem here?

import tensorflow as tf
from tensorflow import keras

X = X.reshape(-1, image_size[0], image_size[1], 1)
model = keras.models.Sequential()

model.add(keras.layers.Conv2D(32, (5, 5), activation='relu', input_shape=X.shape[1:]))
model.add(keras.layers.MaxPooling2D((2, 2)))

model.add(keras.layers.Conv2D(32, (3, 3), activation='relu'))
model.add(keras.layers.MaxPooling2D(2, 2))

model.add(keras.layers.Conv2D(64, (3, 3), activation='relu'))

model.add(keras.layers.Flatten())
model.add(keras.layers.Dense(60, activation='relu'))
model.add(keras.layers.Dropout(0.4))
model.add(keras.layers.Dense(1, activation='softmax'))
model.compile(optimizer='adam', loss=keras.losses.mean_squared_error, metrics=['accuracy'])


model.fit(X, Y, epochs=170, shuffle=True, validation_split=0.1)
As another question, are my layers correct to predict a number for a given picture? This post was edited by Pranav B at June 11, 2019 5:42 PM IST

On the output layer, softmax and cross-entropy are used for classification. I am not too familiar with cross-entropy, but softmax essentially tries to hammer a bunch of numbers into something that looks like probabilities (although there's no good theoretical basis I know of that this actually makes more sense than simple normalization).

These 2 are not usually used in hidden layers.

Linear is also used in the output layer, when you want a real-numbered output (regression). It doesn't make sense to use linear activation in hidden layers, because then the whole system is linear, and you can collapse the whole thing into a 0-hidden-layer network.

@pranav accuracy is not an accurate metric for a regression problem, it's well defined for classification problems but not for this type of model. You should change that for something that's more apt for the problem, e.g., mean_absolute_error. Please check these references en.wikipedia.org/wiki/Mean_absolute_error and tensorflow.org/api_docs/python/tf/keras/metrics/…

# Import the libraries required in this example:
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers

inputs = keras.Input(shape=(784,), name="digits")
x = layers.Dense(64, activation="relu", name="dense_1")(inputs)
x = layers.Dense(64, activation="relu", name="dense_2")(x)
outputs = layers.Dense(10, activation="softmax", name="predictions")(x)

model = keras.Model(inputs=inputs, outputs=outputs)

(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()

# Preprocess the data (NumPy arrays):
x_train = x_train.reshape(60000, 784).astype("float32") / 255
x_test = x_test.reshape(10000, 784).astype("float32") / 255

y_train = y_train.astype("float32")
y_test = y_test.astype("float32")

# Allocate 10,000 samples for validation:
x_val = x_train[-10000:]
y_val = y_train[-10000:]
x_train = x_train[:-10000]
y_train = y_train[:-10000]

model.compile(
    optimizer=keras.optimizers.RMSprop(),  # Optimizer
    # Minimize loss:
    loss=keras.losses.SparseCategoricalCrossentropy(),
    # Monitor metrics:
    metrics=[keras.metrics.SparseCategoricalAccuracy()],
)

print("Fit model on training data")
history = model.fit(
    x_train,
    y_train,
    batch_size=64,
    epochs=2,
    # Validation of loss and metrics
    # at the end of each epoch:
    validation_data=(x_val, y_val),
)

history.history

print("Evaluate model on test data")
results = model.evaluate(x_test, y_test, batch_size=128)
print("test loss, test acc:", results)

# Generate a prediction using model.predict() 
# and calculate it's shape:
print("Generate a prediction")
prediction = model.predict(x_test[:1])
print("prediction shape:", prediction.shape)

import tensorflow as tf
import keras

this is a text, not a code

This post was edited by Raji Reddy A at November 13, 2019 5:47 PM IST

This is the problem: model.add(keras.layers.Dense(1, activation='softmax'))

For predicting real-valued data such as age, it is customary to set the activation as linear or in this case, you can probably use relu.

To illustrate, softmax will create a distribution over the output, in this case the output of the model will be always 1. Since the age is positive definite continuous number, you need an activation that has a range of real numbers >= 0 which the relu activation satisfies.

As another question , are my layers good to predict a number for given picture ?
The architecture seems reasonable so try changing the model's final activation first.