CODEWITHSUNDEEP
python-3.xnumpytensorflowkerasneural-network
Shun7_natural
Shun7_natural

Reputation: 133

Why is my BP neural network's accuracy too low?

The assignment is recognizing handwritten digits by BP neural network. I have tried add hidden layer, increase units and change activation, momentum, but the result shows:

Test Acc: 0.11142061281337047

the splitting dataset X:(1797,64) Y(1797,) 

np.random.shuffle(X)
np.random.shuffle(y)
offset1=int(len(X)*0.6)
offset2=int(len(X)*0.8)
offset3=len(X)-1
X_train, y_train = X[0:offset1,:], y[0:offset1]
X_valid, y_valid = X[offset1:offset2,:], y[offset1:offset2]
X_test, y_test = X[offset2:offset3,:], y[offset2:offset3]

And my neural network:

def create_network():
  # TODO
  net = tf.keras.Sequential()
  net.add(tf.keras.layers.Flatten())
  net.add(tf.keras.layers.Dense(64, activation='relu'))
  net.add(tf.keras.layers.Dense(64, activation='relu'))
  net.add(tf.keras.layers.Dense(10, activation='softmax'))
  net.compile(
      optimizer='adam',
      #tf.optimizers.SGD(learning_rate=0.01, momentum=0.0, nesterov=False), 
      loss="sparse_categorical_crossentropy", 
      metrics=["accuracy"])
  return net


def train_network(network,
                  X_train,
                  y_train,
                  X_valid,
                  y_valid,
                  n_epoch=32,
                  batch_size=64):
  n_iter = 0
  network.fit(x=X_train, y=y_train, epochs=n_epoch, validation_data=(X_valid,y_valid))
  #opt = tf.optimizers.SGD(learning_rate=0.01, momentum=0.0, nesterov=False) #SGD=Stochastic Gradient Descent



net = create_network()
train_network(net, X_train, y_train, X_valid, y_valid)

# evaluate on test set:
y_test_pred = net(X_test)
y_test_pred = np.argmax(y_test_pred.numpy(), axis=1)
print("Test Acc:", accuracy_score(y_true=y_test, y_pred=y_test_pred))

Upvotes: 0

Views: 92

Answers (2)

Nebiyou Yismaw
Nebiyou Yismaw

Reputation: 800

Assuming your model and training are correct, one issue I see is, You are shuffling x & y separately. np.random.shuffle will modify both arrays in place. After you do that the training points will no longer be correctly matched with the correct labels. Hence, your training and prediction won't work as expected. You can do something like,

# shuffle the ids
shuffled_ids = np.arange(len(x)))
np.random.shuffle(shuffled_ids)
# index x & y with the same shuffled numbers, 
# so the correct relationship is maintained 
shuf_x = x[shuffled_ids]
shuf_y = y[shuffled_ids]

Upvotes: 0

Mark
Mark

Reputation: 36

@NebiyouTen, I agree with what you said. But in fact np.random.shuffle doesn't return anything. Maybe you need to do some modifications:

shuffled_ids = np.arange(len(x))
np.random.shuffle(shuffled_ids)
shuf_x = x[shuffled_ids]
shuf_y = y[shuffled_ids]

And here is an alternative way to shuffle the data:

random_state = np.random.get_state()
np.random.shuffle(x)
np.random.set_state(random_state)
np.random.shuffle(y)

Upvotes: 1

Related Questions