Vanishing gradient problem

The other possible gradient problem is when the gradients vanish, or go to zero. This is a much harder problem to solve because it is not as easy to detect. If the loss function does not improve on every step, is it because the gradients went to zero and thus didn't update the weights? Or is it because the model is not able to learn?

This problem occurs more often in RNN models when long memory is required (having long sentences).

In this exercise you will observe the problem on the IMDB data, with longer sentences selected. The data is loaded in X and y variables, as well as classes Sequential, SimpleRNN, Dense and matplotlib.pyplot as plt. The model was pre-trained with 100 epochs, its weights and history are stored on the file model_weights.h5 and variable history.

This exercise is part of the course

Recurrent Neural Networks (RNNs) for Language Modeling with Keras

Exercise instructions

Add a SimpleRNN layer to the model.
Load the pre-trained weights on the model using the method .load_weights().
Add the accuracy of the training data available on the attribute 'acc' to the plot.
Display the plot using the method .show().

Hands-on interactive exercise

Have a go at this exercise by completing this sample code.

# Create the model
model = Sequential()
model.add(____(units=600, input_shape=(None, 1)))
model.add(Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer='sgd', metrics=['accuracy'])

# Load pre-trained weights
model.____('model_weights.h5')

# Plot the accuracy x epoch graph
plt.plot(history.history[____])
plt.plot(history.history['val_acc'])
plt.legend(['train', 'val'], loc='upper left')
plt.____()

Edit and Run Code

This exercise is part of the course

Recurrent Neural Networks (RNNs) for Language Modeling with Keras

AdvancedSkill Level

4.8+

Start Course for Free

In this chapter, you will learn the foundations of Recurrent Neural Networks (RNN). Starting with some prerequisites, continuing to understanding how information flows through the network and finally seeing how to implement such models with Keras in the sentiment classification task.

Exercise 1: Introduction to the course Exercise 2: Comparing the number of parameter of RNN and ANN Exercise 3: Sentiment analysis Exercise 4: Sequence to sequence models Exercise 5: Introduction to language models Exercise 6: Getting used to text data Exercise 7: Preparing text data for model input Exercise 8: Transforming new text Exercise 9: Introduction to RNN inside Keras Exercise 10: Keras models Exercise 11: Keras preprocessing Exercise 12: Your first RNN model

You will learn about the vanishing and exploding gradient problems, often occurring in RNNs, and how to deal with them with the GRU and LSTM cells. Furthermore, you'll create embedding layers for language models and revisit the sentiment classification task.

Exercise 1: Vanishing and exploding gradients Exercise 2: Exploding gradient problem Exercise 3: Vanishing gradient problem

Current Exercise

Exercise 4: GRU and LSTM cells Exercise 5: GRU cells are better than simpleRNN Exercise 6: Stacking RNN layers Exercise 7: The Embedding layer Exercise 8: Number of parameters comparison Exercise 9: Transfer learning Exercise 10: Embeddings improves performance Exercise 11: Sentiment classification revisited Exercise 12: Better sentiment classification Exercise 13: Using the CNN layer

Next, in this chapter you will learn how to prepare data for the multi-class classification task, as well as the differences between multi-class classification and binary classification (sentiment analysis). Finally, you will learn how to create models and measure their performance with Keras.

Exercise 1: Data pre-processing Exercise 2: Prepare label vectors Exercise 3: Pre-process data Exercise 4: Transfer learning for language models Exercise 5: Transfer learning starting point Exercise 6: Word2Vec Exercise 7: Multi-class classification models Exercise 8: Exploring 20 News Groups dataset Exercise 9: Classifying news articles Exercise 10: Assessing the model's performance Exercise 11: Precision-Recall trade-off Exercise 12: Precision or Recall, that is the question Exercise 13: Performance on multi-class classification

This chapter introduces you to two applications of RNN models: Text Generation and Neural Machine Translation. You will learn how to prepare the text data to the format needed by the models. The Text Generation model is used for replicating a character's way of speech and will have some fun mimicking Sheldon from The Big Bang Theory. Neural Machine Translation is used for example by Google Translate in a much more complex model. In this chapter, you will create a model that translates Portuguese small phrases into English.

Exercise 1: Sequence to Sequence Models Exercise 2: Text generation examples Exercise 3: NMT example Exercise 4: The Text Generating Function Exercise 5: Predict next character Exercise 6: Generate sentence with context Exercise 7: Change the probability scale Exercise 8: Text Generation Models Exercise 9: Create vectors of sentences and next characters Exercise 10: Preparing the data for training Exercise 11: Creating the text generation model Exercise 12: Neural Machine Translation Exercise 13: Preparing the input text Exercise 14: Preparing the output text Exercise 15: Translate Portuguese to English Exercise 16: Congratulations!