[python] Deep-Learning Nan loss reasons

Perhaps too general a question, but can anyone explain what would cause a Convolutional Neural Network to diverge?

Specifics:

I am using Tensorflow's iris_training model with some of my own data and keep getting

ERROR:tensorflow:Model diverged with loss = NaN.

Traceback...

tensorflow.contrib.learn.python.learn.monitors.NanLossDuringTrainingError: NaN loss during training.

Traceback originated with line:

 tf.contrib.learn.DNNClassifier(feature_columns=feature_columns,
                                        hidden_units=[300, 300, 300],
                                        #optimizer=tf.train.ProximalAdagradOptimizer(learning_rate=0.001, l1_regularization_strength=0.00001),                                                          
                                        n_classes=11,
                                        model_dir="/tmp/iris_model")

I've tried adjusting the optimizer, using a zero for learning rate, and using no optimizer. Any insights into network layers, data size, etc is appreciated.

I'd like to plug in some (shallow) reasons I have experienced as follows:

1. we may have updated our dictionary(for NLP tasks) but the model and the prepared data used a different one.
2. we may have reprocessed our data(binary tf_record) but we loaded the old model. The reprocessed data may conflict with the previous one.
3. we may should train the model from scratch but we forgot to delete the checkpoints and the model loaded the latest parameters automatically.

Hope that helps.

If you're training for cross entropy, you want to add a small number like 1e-8 to your output probability.

Because log(0) is negative infinity, when your model trained enough the output distribution will be very skewed, for instance say I'm doing a 4 class output, in the beginning my probability looks like

0.25 0.25 0.25 0.25

but toward the end the probability will probably look like

1.0 0 0 0

And you take a cross entropy of this distribution everything will explode. The fix is to artifitially add a small number to all the terms to prevent this.

If using integers as targets, makes sure they aren't symmetrical at 0.

I.e., don't use classes -1, 0, 1. Use instead 0, 1, 2.

In my case I got NAN when setting distant integer LABELs. ie:

• Labels [0..100] the training was ok,
• Labels [0..100] plus one additional label 8000, then I got NANs.

So, not use a very distant Label.

EDIT You can see the effect in the following simple code:

from keras.models import Sequential
from keras.layers import Dense, Activation
import numpy as np

X=np.random.random(size=(20,5))
y=np.random.randint(0,high=5, size=(20,1))

model = Sequential([
            Dense(10, input_dim=X.shape[1]),
            Activation('relu'),
            Dense(5),
            Activation('softmax')
            ])
model.compile(optimizer = "Adam", loss = "sparse_categorical_crossentropy", metrics = ["accuracy"] )

print('fit model with labels in range 0..5')
history = model.fit(X, y, epochs= 5 )

X = np.vstack( (X, np.random.random(size=(1,5))))
y = np.vstack( ( y, [[8000]]))
print('fit model with labels in range 0..5 plus 8000')
history = model.fit(X, y, epochs= 5 )

The result shows the NANs after adding the label 8000:

fit model with labels in range 0..5
Epoch 1/5
20/20 [==============================] - 0s 25ms/step - loss: 1.8345 - acc: 0.1500
Epoch 2/5
20/20 [==============================] - 0s 150us/step - loss: 1.8312 - acc: 0.1500
Epoch 3/5
20/20 [==============================] - 0s 151us/step - loss: 1.8273 - acc: 0.1500
Epoch 4/5
20/20 [==============================] - 0s 198us/step - loss: 1.8233 - acc: 0.1500
Epoch 5/5
20/20 [==============================] - 0s 151us/step - loss: 1.8192 - acc: 0.1500
fit model with labels in range 0..5 plus 8000
Epoch 1/5
21/21 [==============================] - 0s 142us/step - loss: nan - acc: 0.1429
Epoch 2/5
21/21 [==============================] - 0s 238us/step - loss: nan - acc: 0.2381
Epoch 3/5
21/21 [==============================] - 0s 191us/step - loss: nan - acc: 0.2381
Epoch 4/5
21/21 [==============================] - 0s 191us/step - loss: nan - acc: 0.2381
Epoch 5/5
21/21 [==============================] - 0s 188us/step - loss: nan - acc: 0.2381

If you'd like to gather more information on the error and if the error occurs in the first few iterations, I suggest you run the experiment in CPU-only mode (no GPUs). The error message will be much more specific.

Source: https://github.com/tensorflow/tensor2tensor/issues/574

Although most of the points are already discussed. But I would like to highlight again one more reason for NaN which is missing.

tf.estimator.DNNClassifier(
    hidden_units, feature_columns, model_dir=None, n_classes=2, weight_column=None,
    label_vocabulary=None, optimizer='Adagrad', activation_fn=tf.nn.relu,
    dropout=None, config=None, warm_start_from=None,
    loss_reduction=losses_utils.ReductionV2.SUM_OVER_BATCH_SIZE, batch_norm=False
)

By default activation function is "Relu". It could be possible that intermediate layer's generating a negative value and "Relu" convert it into the 0. Which gradually stops training.

I observed the "LeakyRelu" able to solve such problems.

Regularization can help. For a classifier, there is a good case for activity regularization, whether it is binary or a multi-class classifier. For a regressor, kernel regularization might be more appropriate.

The reason for nan, inf or -inf often comes from the fact that division by 0.0 in TensorFlow doesn't result in a division by zero exception. It could result in a nan, inf or -inf "value". In your training data you might have 0.0 and thus in your loss function it could happen that you perform a division by 0.0.

a = tf.constant([2., 0., -2.])
b = tf.constant([0., 0., 0.])
c = tf.constant([1., 1., 1.])
print((a / b) + c)

Output is the following tensor:

tf.Tensor([ inf  nan -inf], shape=(3,), dtype=float32)

Adding a small eplison (e.g., 1e-5) often does the trick. Additionally, since TensorFlow 2 the opteration tf.math.division_no_nan is defined.