Coursera 吴恩达 DeepLearning Sequence model 第二周作业 Emojify - v2（多分类问题，有彩蛋，性能超过预期）

自动表情符号

在这个练习中，首先完成基准模型Emojifier-V1，使用word embeddings求平均值。可以通过输入的句子输出一个表情符号。这个模型无法识别单词的顺序和复杂的句子结构。然后将会结合LSTM来建立一个更加复杂的模型Emojifier-V2。

我的总结

1. 实验的结果发现复杂的LSTM模型并没什么卵用，V1的综合性能比V2要好。除了作业的训练集和测试集，我另外用了两个极端测试例。这两个例子无论V1，V2都无法得到正确的分类。
   1) I could not agree anymore. label：happy
   2) this movie is not good and not enjoyable label：unhappy
2. 在训练集中添加上面两个例子，会导致训练结果变差，cost很大。因为对于word embedding求平均的方法，上面两个例子本来就无法得到合适的编码。如果硬添加到训练集中，属于outlier。
3. 添加learning rate decay功能，并增加epoch 数量，V1的training accuracy 和test accuracy都可以超过作业中的预期。分别达到98.49%和91.07%。‘not feeling happy’也可以猜对。
4. 尝试过从句子中删除stopword，并且修改过对应的predict函数，但结果不如预期。我的解释是首先，有一些stopword 是有用的，比如How dare you ask that，如果把how 和 that去掉，这个句子的意思就很难理解了。其次，stopword是在word embeddings之前出现的概念，针对的是stopword大量出现的场景且使用的是onehot编码。而现在stopword在通过大量真实语料训练得到了自己的word embeddings，并且在这种短句子中，stopword出现次数又小于等于1。那么它们的embeddings加上求平均，应该是有益无害，无需去除。
5. 从confusion matrix中看出，label 1和4，吃饭和运动，这两类比较不容易出错。label0, 2, 3，爱情，开心，不开心，这三类之间容易混淆。我觉得也是符合预期的。开心和不开心因为单词顺序的原因，单纯的embeddings求平均容易出现误判。
6. 下面训练集和测试集中出错的句子。可以看出，除了第一个长度为10的句子外，其他出错的情况，正确标签的概率还是比较高的，而且这些句子都不存在单词顺序问题。可以猜测如果用100d的word embeddings，这里的准确率应该还可以提升。在尝试过100d和200d的Glove之后，发现维数越高，训练集收敛速度越快，且准确率越高。100d和200d分别在400个和1000个epoch后达到收敛，准确率100%。但是测试集性能变得更差，100d和200d测试集准确率分别为82%和80%。测试集中错误大多是0和2之间，以及0和3之间，其他错误没有。这个准确率跟数据也有关系。

输入句子	标签	预测结果	softmax输出
I am so excited to see you after so long	2(happy)	3(unhappy)	[0.052, 0.0035, 0.093, 0.85, 0.0014]
I am looking for a date	0 (love)	3(unhappy)	[0.22 0.007 0.33 0.40 0.05

输入句子	标签	预测结果	softmax输出概率
work is hard	3(unhappy)	2(happy)	[0.0042 0.005 0.6025 0.386 0.0018]
I love taking breaks	0(love)	3(unhappy)	[0.25 0.00323232 0.045 0.7 0.0014 ]
My life is so boring	3(unhappy)	0(love)	[0.4 1.6e-05 0.21 0.39 8.6e-05]
will you be my valentine	2(happy)	0(love)	[0.66 0.0084 0.22 0.1 0.01]
is go away	3(unhappy)	1(sport)	[1.88-03 0.67 1.6e-06 0.32 5.5e-03]

第一部分的总结

1. 即使只有127个training examples，我们仍然能够得到一个相当好的模型。这得益于word vectors的泛化能力。
2. Emojify-V1会在以下句子上表现很差，比如“this movie is not good and not enjoyable”，因为它仅仅是将word-embeddings 向量做了一个平均，没有注意到词语的顺序，自然无法理解这些词语组合的意义。

第二部分总结

1. 如果NLP任务的训练集很小，那么使用word embeddings 对算法将有极大的帮助。Word embeddings能够让你的模型当测试集中出现了训练集中没有的单词时仍然正常工作。
2. 在Keras中训练时序模型需要以下几个重要细节
   1) 为了使用mini batch，序列需要补齐，这样所有mini batch 中的例子拥有同样的长度
   2) 通过一个预先训练好的值来初始化一个embedding 层。
   3) LSTM有一个return_sequence的flag，用来决定是否返回所有隐层或者最后一个隐层。
   4) LSTM()后使用dropout()可以帮助正则化。

注：本文涉及的图片集资料均整理翻译自Andrew Ng的Deep Learning 系列课程，版权归其所有。翻译整理水平有限，如有不妥的地方欢迎指出。如有侵权请联系删除。谢谢。

Emojify!

Welcome to the second assignment of Week 2. You are going to use word vector representations to build an Emojifier.

Have you ever wanted to make your text messages more expressive? Your emojifier app will help you do that. So rather than writing “Congratulations on the promotion! Lets get coffee and talk. Love you!” the emojifier can automatically turn this into “Congratulations on the promotion! ? Lets get coffee and talk. ☕️ Love you! ❤️”

You will implement a model which inputs a sentence (such as “Let’s go see the baseball game tonight!”) and finds the most appropriate emoji to be used with this sentence (⚾️). In many emoji interfaces, you need to remember that ❤️ is the “heart” symbol rather than the “love” symbol. But using word vectors, you’ll see that even if your training set explicitly relates only a few words to a particular emoji, your algorithm will be able to generalize and associate words in the test set to the same emoji even if those words don’t even appear in the training set. This allows you to build an accurate classifier mapping from sentences to emojis, even using a small training set.

In this exercise, you’ll start with a baseline model (Emojifier-V1) using word embeddings, then build a more sophisticated model (Emojifier-V2) that further incorporates an LSTM.

Lets get started! Run the following cell to load the package you are going to use.

import numpy as np
from emo_utils import *
import emoji
import matplotlib.pyplot as plt

%matplotlib inline

1 - Baseline model: Emojifier-V1

1.1 - Dataset EMOJISET

Let’s start by building a simple baseline classifier.

You have a tiny dataset (X, Y) where:
- X contains 127 sentences (strings)
- Y contains a integer label between 0 and 4 corresponding to an emoji for each sentence

Figure 1: EMOJISET - a classification problem with 5 classes. A few examples of sentences are given here.

Let’s load the dataset using the code below. We split the dataset between training (127 examples) and testing (56 examples).

X_train, Y_train = read_csv('data/train_emoji.csv')
X_test, Y_test = read_csv('data/tesss.csv')

maxLen = len(max(X_train, key=len).split())
print(maxLen)

10

Run the following cell to print sentences from X_train and corresponding labels from Y_train. Change index to see different examples. Because of the font the iPython notebook uses, the heart emoji may be colored black rather than red.

index = 131
print(X_train[index], label_to_emoji(Y_train[index]))

great job ?

1.2 - Overview of the Emojifier-V1

In this part, you are going to implement a baseline model called “Emojifier-v1”.

Figure 2: Baseline model (Emojifier-V1).

The input of the model is a string corresponding to a sentence (e.g. “I love you). In the code, the output will be a probability vector of shape (1,5), that you then pass in an argmax layer to extract the index of the most likely emoji output.

To get our labels into a format suitable for training a softmax classifier, lets convert $Y$ from its current shape current shape $(m, 1)$ into a “one-hot representation” $(m, 5)$, where each row is a one-hot vector giving the label of one example, You can do so using this next code snipper. Here, Y_oh stands for “Y-one-hot” in the variable names Y_oh_train and Y_oh_test:

Y_oh_train = convert_to_one_hot(Y_train, C = 5)
Y_oh_test = convert_to_one_hot(Y_test, C = 5)

Let’s see what convert_to_one_hot() did. Feel free to change index to print out different values.

index = 50
print(Y_train[index], "is converted into one hot", Y_oh_train[index])

0 is converted into one hot [ 1.  0.  0.  0.  0.]

All the data is now ready to be fed into the Emojify-V1 model. Let’s implement the model!

1.3 - Implementing Emojifier-V1

As shown in Figure (2), the first step is to convert an input sentence into the word vector representation, which then get averaged together. Similar to the previous exercise, we will use pretrained 50-dimensional GloVe embeddings. Run the following cell to load the word_to_vec_map, which contains all the vector representations.

word_to_index, index_to_word, word_to_vec_map = read_glove_vecs('data/glove.6B.50d.txt')

You’ve loaded:
- word_to_index: dictionary mapping from words to their indices in the vocabulary (400,001 words, with the valid indices ranging from 0 to 400,000)
- index_to_word: dictionary mapping from indices to their corresponding words in the vocabulary
- word_to_vec_map: dictionary mapping words to their GloVe vector representation.

Run the following cell to check if it works.

word = "cucumber"
index = 289846
print("the index of", word, "in the vocabulary is", word_to_index[word])
print("the", str(index) + "th word in the vocabulary is", index_to_word[index])
print(word_to_vec_map[word])

the index of cucumber in the vocabulary is 113317
the 289846th word in the vocabulary is potatos
[ 0.68224  -0.31608  -0.95201   0.47108   0.56571   0.13151   0.22457
  0.094995 -1.3237   -0.51545  -0.39337   0.88488   0.93826   0.22931
  0.088624 -0.53908   0.23396   0.73245  -0.019123 -0.26552  -0.40433
 -1.5832    1.1316    0.4419   -0.48218   0.4828    0.14938   1.1245
  1.0159   -0.50213   0.83831  -0.31303   0.083242  1.7161    0.15024
  1.0324   -1.5005    0.62348   0.54508  -0.88484   0.53279  -0.085119
  0.02141  -0.56629   1.1463    0.6464    0.78318  -0.067662  0.22884
 -0.042453]

Exercise: Implement sentence_to_avg(). You will need to carry out two steps:
1. Convert every sentence to lower-case, then split the sentence into a list of words. X.lower() and X.split() might be useful.
2. For each word in the sentence, access its GloVe representation. Then, average all these values.

# GRADED FUNCTION: sentence_to_avg

def sentence_to_avg(sentence, word_to_vec_map):
    """
    Converts a sentence (string) into a list of words (strings). Extracts the GloVe representation of each word
    and averages its value into a single vector encoding the meaning of the sentence.

    Arguments:
    sentence -- string, one training example from X
    word_to_vec_map -- dictionary mapping every word in a vocabulary into its 50-dimensional vector representation

    Returns:
    avg -- average vector encoding information about the sentence, numpy-array of shape (50,)
    """

    ### START CODE HERE ###
    # Step 1: Split sentence into list of lower case words (≈ 1 line)
    words = sentence.split()

    # Initialize the average word vector, should have the same shape as your word vectors.
    avg = np.zeros([50,])

    # Step 2: average the word vectors. You can loop over the words in the list "words".
    for w in words:
        avg += word_to_vec_map[w.lower()]
    avg = avg/len(words)

    ### END CODE HERE ###

    return avg

avg = sentence_to_avg("Morrocan couscous is my favorite dish", word_to_vec_map)
print("avg = ", avg)

avg =  [-0.008005    0.56370833 -0.50427333  0.258865    0.55131103  0.03104983
 -0.21013718  0.16893933 -0.09590267  0.141784   -0.15708967  0.18525867
  0.6495785   0.38371117  0.21102167  0.11301667  0.02613967  0.26037767
  0.05820667 -0.01578167 -0.12078833 -0.02471267  0.4128455   0.5152061
  0.38756167 -0.898661   -0.535145    0.33501167  0.68806933 -0.2156265
  1.797155    0.10476933 -0.36775333  0.750785    0.10282583  0.348925
 -0.27262833  0.66768    -0.10706167 -0.283635    0.59580117  0.28747333
 -0.3366635   0.23393817  0.34349183  0.178405    0.1166155  -0.076433
  0.1445417   0.09808667]

Expected Output:

**avg= **

[-0.008005 0.56370833 -0.50427333 0.258865 0.55131103 0.03104983 -0.21013718 0.16893933 -0.09590267 0.141784 -0.15708967 0.18525867 0.6495785 0.38371117 0.21102167 0.11301667 0.02613967 0.26037767 0.05820667 -0.01578167 -0.12078833 -0.02471267 0.4128455 0.5152061 0.38756167 -0.898661 -0.535145 0.33501167 0.68806933 -0.2156265 1.797155 0.10476933 -0.36775333 0.750785 0.10282583 0.348925 -0.27262833 0.66768 -0.10706167 -0.283635 0.59580117 0.28747333 -0.3366635 0.23393817 0.34349183 0.178405 0.1166155 -0.076433 0.1445417 0.09808667]

Model

You now have all the pieces to finish implementing the model() function. After using sentence_to_avg() you need to pass the average through forward propagation, compute the cost, and then backpropagate to update the softmax’s parameters.

Exercise: Implement the model() function described in Figure (2). Assuming here that $Yoh$ (“Y one hot”) is the one-hot encoding of the output labels, the equations you need to implement in the forward pass and to compute the cross-entropy cost are:

$$z^{(i)} = W . avg^{(i)} + b$$
$$a^{(i)} = softmax(z^{(i)})$$
$$\mathcal{L}^{(i)} = - \sum_{k = 0}^{n_y - 1} Yoh^{(i)}_k * log(a^{(i)}_k)$$

It is possible to come up with a more efficient vectorized implementation. But since we are using a for-loop to convert the sentences one at a time into the avg^{(i)} representation anyway, let’s not bother this time.

We provided you a function softmax().

# GRADED FUNCTION: model

def model(X, Y, word_to_vec_map, learning_rate = 0.005, num_iterations = 4000):
    """
    Model to train word vector representations in numpy.

    Arguments:
    X -- input data, numpy array of sentences as strings, of shape (m, 1)
    Y -- labels, numpy array of integers between 0 and 7, numpy-array of shape (m, 1)
    word_to_vec_map -- dictionary mapping every word in a vocabulary into its 50-dimensional vector representation
    learning_rate -- learning_rate for the stochastic gradient descent algorithm
    num_iterations -- number of iterations

    Returns:
    pred -- vector of predictions, numpy-array of shape (m, 1)
    W -- weight matrix of the softmax layer, of shape (n_y, n_h)
    b -- bias of the softmax layer, of shape (n_y,)
    """

    np.random.seed(1)

    # Define number of training examples
    m = Y.shape[0]                          # number of training examples
    n_y = 5                                 # number of classes  
    n_h = 50                                # dimensions of the GloVe vectors 

    # Initialize parameters using Xavier initialization
    W = np.random.randn(n_y, n_h) / np.sqrt(n_h)
    b = np.zeros((n_y,))

    # Convert Y to Y_onehot with n_y classes
    Y_oh = convert_to_one_hot(Y, C = n_y) 
    avg = 0;

    # Optimization loop
    for t in range(num_iterations):                       # Loop over the number of iterations
        for i in range(m):                                # Loop over the training examples

            ### START CODE HERE ### (≈ 4 lines of code)
            # Average the word vectors of the words from the i'th training example
            avg = sentence_to_avg(X[i], word_to_vec_map)

            # Forward propagate the avg through the softmax layer
            z = np.dot(W,avg) + b
            a = softmax(z)

            # Compute cost using the i'th training label's one hot representation and "A" (the output of the softmax)
            cost = -np.dot(Y_oh[i],np.log(a)) 
            ### END CODE HERE ###

            # Compute gradients 
            dz = a - Y_oh[i]
            dW = np.dot(dz.reshape(n_y,1), avg.reshape(1, n_h))
            db = dz

            # Update parameters with Stochastic Gradient Descent
            if(t > 500):
                decay = 1/(t/500)
            else:
                decay = 1
            W = W - learning_rate * decay * dW
            b = b - learning_rate * decay * db

        if t % 100 == 0:
            print("Epoch: " + str(t) + " --- cost = " + str(cost))
            pred = predict(X, Y, W, b, word_to_vec_map)

    return pred, W, b

print(X_train.shape)
print(Y_train.shape)
print(np.eye(5)[Y_train.reshape(-1)].shape)

(133,)
(133,)
(133, 5)

Run the next cell to train your model and learn the softmax parameters (W,b).

pred, W, b = model(X_train, Y_train, word_to_vec_map)
#print(pred)

Epoch: 0 --- cost = 1.007899235090512
Accuracy: 0.3157894736842105
Epoch: 100 --- cost = 0.09642903966197412
Accuracy: 0.8947368421052632
Epoch: 200 --- cost = 0.053997178964793696
Accuracy: 0.9473684210526315
Epoch: 300 --- cost = 0.036345944698977915
Accuracy: 0.9473684210526315
Epoch: 400 --- cost = 0.02697821509931659
Accuracy: 0.9548872180451128
......
Epoch: 3900 --- cost = 0.006559565738278793
Accuracy: 0.9849624060150376

Expected Output (on a subset of iterations):

**Epoch: 0**	cost = 1.95204988128	Accuracy: 0.348484848485
**Epoch: 100**	cost = 0.0797181872601	Accuracy: 0.931818181818
**Epoch: 200**	cost = 0.0445636924368	Accuracy: 0.954545454545
**Epoch: 300**	cost = 0.0343226737879	Accuracy: 0.969696969697

Great! Your model has pretty high accuracy on the training set. Lets now see how it does on the test set.

1.4 - Examining test set performance

print("Training set:")
pred_train = predict(X_train, Y_train, W, b, word_to_vec_map)
print('Test set:')
pred_test = predict(X_test, Y_test, W, b, word_to_vec_map)

Training set:
Accuracy: 0.9849624060150376
Test set:
Accuracy: 0.9107142857142857

Expected Output:

**Train set accuracy**	97.7
**Test set accuracy**	85.7

Random guessing would have had 20% accuracy given that there are 5 classes. This is pretty good performance after training on only 127 examples.

In the training set, the algorithm saw the sentence “I love you” with the label ❤️. You can check however that the word “adore” does not appear in the training set. Nonetheless, lets see what happens if you write “I adore you.”

X_my_sentences = np.array(["i adore you", "i love you", "funny lol", "lets play with a ball", "food is ready", "not feeling happy"])
Y_my_labels = np.array([[0], [0], [2], [1], [4],[3]])

pred = predict(X_my_sentences, Y_my_labels , W, b, word_to_vec_map)
print_predictions(X_my_sentences, pred)

Accuracy: 1.0

i adore you ❤️
i love you ❤️
funny lol ?
lets play with a ball ⚾
food is ready ?
not feeling happy ?

Amazing! Because adore has a similar embedding as love, the algorithm has generalized correctly even to a word it has never seen before. Words such as heart, dear, beloved or adore have embedding vectors similar to love, and so might work too—feel free to modify the inputs above and try out a variety of input sentences. How well does it work?

Note though that it doesn’t get “not feeling happy” correct. This algorithm ignores word ordering, so is not good at understanding phrases like “not happy.”

Printing the confusion matrix can also help understand which classes are more difficult for your model. A confusion matrix shows how often an example whose label is one class (“actual” class) is mislabeled by the algorithm with a different class (“predicted” class).

print(Y_test.shape)
print('           '+ label_to_emoji(0)+ '    ' + label_to_emoji(1) + '    ' +  label_to_emoji(2)+ '    ' + label_to_emoji(3)+'   ' + label_to_emoji(4))
print(pd.crosstab(Y_test, pred_test.reshape(56,), rownames=['Actual'], colnames=['Predicted'], margins=True))
plot_confusion_matrix(Y_test, pred_test)

(56,)
           ❤️    ⚾    ?    ?   ?
Predicted  0.0  1.0  2.0  3.0  4.0  All
Actual                                 
0            6    0    0    1    0    7
1            0    8    0    0    0    8
2            2    0   16    0    0   18
3            1    1    2   12    0   16
4            0    0    1    0    6    7
All          9    9   19   13    6   56

What you should remember from this part:
- Even with a 127 training examples, you can get a reasonably good model for Emojifying. This is due to the generalization power word vectors gives you.
- Emojify-V1 will perform poorly on sentences such as “This movie is not good and not enjoyable” because it doesn’t understand combinations of words–it just averages all the words’ embedding vectors together, without paying attention to the ordering of words. You will build a better algorithm in the next part.

第一部分的总结
1. 即使只有127个training examples，我们仍然能够得到一个相当好的模型。这得益于word vectors的泛化能力。
2. Emojify-V1会在以下句子上表现很差，比如“this movie is not good and not enjoyable”，因为它仅仅是将word embeddings 向量做了一个平均，没有注意到词语的顺序，自然无法理解这些词语组合的意义。

2 - Emojifier-V2: Using LSTMs in Keras:

Let’s build an LSTM model that takes as input word sequences. This model will be able to take word ordering into account. Emojifier-V2 will continue to use pre-trained word embeddings to represent words, but will feed them into an LSTM, whose job it is to predict the most appropriate emoji.

Run the following cell to load the Keras packages.

import numpy as np
np.random.seed(0)
from keras.models import Model
from keras.layers import Dense, Input, Dropout, LSTM, Activation
from keras.layers.embeddings import Embedding
from keras.preprocessing import sequence
from keras.initializers import glorot_uniform
np.random.seed(1)

Using TensorFlow backend.

2.1 - Overview of the model

Here is the Emojifier-v2 you will implement:

Figure 3: Emojifier-V2. A 2-layer LSTM sequence classifier.

2.2 Keras and mini-batching

In this exercise, we want to train Keras using mini-batches. However, most deep learning frameworks require that all sequences in the same mini-batch have the same length. This is what allows vectorization to work: If you had a 3-word sentence and a 4-word sentence, then the computations needed for them are different (one takes 3 steps of an LSTM, one takes 4 steps) so it’s just not possible to do them both at the same time.

The common solution to this is to use padding. Specifically, set a maximum sequence length, and pad all sequences to the same length. For example, of the maximum sequence length is 20, we could pad every sentence with “0”s so that each input sentence is of length 20. Thus, a sentence “i love you” would be represented as $(e_{i}, e_{love}, e_{you}, \vec{0}, \vec{0}, \ldots, \vec{0})$. In this example, any sentences longer than 20 words would have to be truncated. One simple way to choose the maximum sequence length is to just pick the length of the longest sentence in the training set.

2.3 - The Embedding layer

In Keras, the embedding matrix is represented as a “layer”, and maps positive integers (indices corresponding to words) into dense vectors of fixed size (the embedding vectors). It can be trained or initialized with a pretrained embedding. In this part, you will learn how to create an Embedding() layer in Keras, initialize it with the GloVe 50-dimensional vectors loaded earlier in the notebook. Because our training set is quite small, we will not update the word embeddings but will instead leave their values fixed. But in the code below, we’ll show you how Keras allows you to either train or leave fixed this layer.

The Embedding() layer takes an integer matrix of size (batch size, max input length) as input. This corresponds to sentences converted into lists of indices (integers), as shown in the figure below.

Figure 4: Embedding layer. This example shows the propagation of two examples through the embedding layer. Both have been zero-padded to a length of max_len=5. The final dimension of the representation is (2,max_len,50) because the word embeddings we are using are 50 dimensional.

The largest integer (i.e. word index) in the input should be no larger than the vocabulary size. The layer outputs an array of shape (batch size, max input length, dimension of word vectors).

The first step is to convert all your training sentences into lists of indices, and then zero-pad all these lists so that their length is the length of the longest sentence.

Exercise: Implement the function below to convert X (array of sentences as strings) into an array of indices corresponding to words in the sentences. The output shape should be such that it can be given to Embedding() (described in Figure 4).

# GRADED FUNCTION: sentences_to_indices

def sentences_to_indices(X, word_to_index, max_len):
    """
    Converts an array of sentences (strings) into an array of indices corresponding to words in the sentences.
    The output shape should be such that it can be given to `Embedding()` (described in Figure 4). 

    Arguments:
    X -- array of sentences (strings), of shape (m, 1)
    word_to_index -- a dictionary containing the each word mapped to its index
    max_len -- maximum number of words in a sentence. You can assume every sentence in X is no longer than this. 

    Returns:
    X_indices -- array of indices corresponding to words in the sentences from X, of shape (m, max_len)
    """

    m = X.shape[0]                                   # number of training examples

    ### START CODE HERE ###
    # Initialize X_indices as a numpy matrix of zeros and the correct shape (≈ 1 line)
    X_indices = np.zeros([m,max_len])

    for i in range(m):                               # loop over training examples

        # Convert the ith training sentence in lower case and split is into words. You should get a list of words.
        sentence_words = X[i].split()

        # Initialize j to 0
        j = 0

        # Loop over the words of sentence_words
        for w in sentence_words:
            # Set the (i,j)th entry of X_indices to the index of the correct word.
            X_indices[i, j] = word_to_index[w.lower()]
            # Increment j to j + 1
            j = j+1

    ### END CODE HERE ###

    return X_indices

Run the following cell to check what sentences_to_indices() does, and check your results.

X1 = np.array(["funny lol", "lets play baseball", "food is ready for you"])
X1_indices = sentences_to_indices(X1,word_to_index, max_len = 5)
print("X1 =", X1)
print("X1_indices =", X1_indices)

X1 = ['funny lol' 'lets play baseball' 'food is ready for you']
X1_indices = [[ 155345.  225122.       0.       0.       0.]
 [ 220930.  286375.   69714.       0.       0.]
 [ 151204.  192973.  302254.  151349.  394475.]]

Expected Output:

**X1 =**	[‘funny lol’ ‘lets play football’ ‘food is ready for you’]
**X1_indices =**	[[ 155345. 225122. 0. 0. 0.] [ 220930. 286375. 151266. 0. 0.] [ 151204. 192973. 302254. 151349. 394475.]]

Let’s build the Embedding() layer in Keras, using pre-trained word vectors. After this layer is built, you will pass the output of sentences_to_indices() to it as an input, and the Embedding() layer will return the word embeddings for a sentence.

Exercise: Implement pretrained_embedding_layer(). You will need to carry out the following steps:
1. Initialize the embedding matrix as a numpy array of zeroes with the correct shape.
2. Fill in the embedding matrix with all the word embeddings extracted from word_to_vec_map.
3. Define Keras embedding layer. Use Embedding(). Be sure to make this layer non-trainable, by setting trainable = False when calling Embedding(). If you were to set trainable = True, then it will allow the optimization algorithm to modify the values of the word embeddings.
4. Set the embedding weights to be equal to the embedding matrix

# GRADED FUNCTION: pretrained_embedding_layer

def pretrained_embedding_layer(word_to_vec_map, word_to_index):
    """
    Creates a Keras Embedding() layer and loads in pre-trained GloVe 50-dimensional vectors.

    Arguments:
    word_to_vec_map -- dictionary mapping words to their GloVe vector representation.
    word_to_index -- dictionary mapping from words to their indices in the vocabulary (400,001 words)

    Returns:
    embedding_layer -- pretrained layer Keras instance
    """

    vocab_len = len(word_to_index) + 1                  # adding 1 to fit Keras embedding (requirement)
    emb_dim = word_to_vec_map["cucumber"].shape[0]      # define dimensionality of your GloVe word vectors (= 50)

    ### START CODE HERE ###
    # Initialize the embedding matrix as a numpy array of zeros of shape (vocab_len, dimensions of word vectors = emb_dim)
    emb_matrix = np.zeros([vocab_len,emb_dim])

    # Set each row "index" of the embedding matrix to be the word vector representation of the "index"th word of the vocabulary
    for word, index in word_to_index.items():
        emb_matrix[index, :] = word_to_vec_map[word]

    # Define Keras embedding layer with the correct output/input sizes, make it trainable. Use Embedding(...). Make sure to set trainable=False. 
    embedding_layer = Embedding(vocab_len,emb_dim,trainable = False)
    ### END CODE HERE ###

    # Build the embedding layer, it is required before setting the weights of the embedding layer. Do not modify the "None".
    embedding_layer.build((None,))

    # Set the weights of the embedding layer to the embedding matrix. Your layer is now pretrained.
    embedding_layer.set_weights([emb_matrix])

    return embedding_layer

embedding_layer = pretrained_embedding_layer(word_to_vec_map, word_to_index)
print("weights[0][1][3] =", embedding_layer.get_weights()[0][1][3])

weights[0][1][3] = -0.3403

Expected Output:

**weights[0][1][3] =**

-0.3403

2.3 Building the Emojifier-V2

Lets now build the Emojifier-V2 model. You will do so using the embedding layer you have built, and feed its output to an LSTM network.

Figure 3: Emojifier-v2. A 2-layer LSTM sequence classifier.

Exercise: Implement Emojify_V2(), which builds a Keras graph of the architecture shown in Figure 3. The model takes as input an array of sentences of shape (m, max_len, ) defined by input_shape. It should output a softmax probability vector of shape (m, C = 5). You may need Input(shape = ..., dtype = '...'), LSTM(), Dropout(), Dense(), and Activation().

# GRADED FUNCTION: Emojify_V2

def Emojify_V2(input_shape, word_to_vec_map, word_to_index):
    """
    Function creating the Emojify-v2 model's graph.

    Arguments:
    input_shape -- shape of the input, usually (max_len,)
    word_to_vec_map -- dictionary mapping every word in a vocabulary into its 50-dimensional vector representation
    word_to_index -- dictionary mapping from words to their indices in the vocabulary (400,001 words)

    Returns:
    model -- a model instance in Keras
    """

    ### START CODE HERE ###
    # Define sentence_indices as the input of the graph, it should be of shape input_shape and dtype 'int32' (as it contains indices).
    sentence_indices = Input(shape=input_shape, dtype=np.int32)

    # Create the embedding layer pretrained with GloVe Vectors (≈1 line)
    embedding_layer = pretrained_embedding_layer(word_to_vec_map, word_to_index)
    # Propagate sentence_indices through your embedding layer, you get back the embeddings
    embeddings = embedding_layer(sentence_indices)   

    # Propagate the embeddings through an LSTM layer with 128-dimensional hidden state
    # Be careful, the returned output should be a batch of sequences.
    X = LSTM(128,return_sequences=True)(embeddings)
    # Add dropout with a probability of 0.5
    X = Dropout(0.5)(X)
    # Propagate X trough another LSTM layer with 128-dimensional hidden state
    # Be careful, the returned output should be a single hidden state, not a batch of sequences.
    X = LSTM(128)(X)
    # Add dropout with a probability of 0.5
    X = Dropout(0.5)(X)
    # Propagate X through a Dense layer with softmax activation to get back a batch of 5-dimensional vectors.
    X = Dense(5, activation='softmax')(X)
    # Add a softmax activation
    X = Activation('softmax')(X)

    # Create Model instance which converts sentence_indices into X.
    model = Model(sentence_indices, X)

    ### END CODE HERE ###

    return model

Run the following cell to create your model and check its summary. Because all sentences in the dataset are less than 10 words, we chose max_len = 10. You should see your architecture, it uses “20,223,927” parameters, of which 20,000,050 (the word embeddings) are non-trainable, and the remaining 223,877 are. Because our vocabulary size has 400,001 words (with valid indices from 0 to 400,000) there are 400,001*50 = 20,000,050 non-trainable parameters.

model = Emojify_V2((maxLen,), word_to_vec_map, word_to_index)
model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_2 (InputLayer)         (None, 10)                0         
_________________________________________________________________
embedding_3 (Embedding)      (None, 10, 50)            20000050  
_________________________________________________________________
lstm_3 (LSTM)                (None, 10, 128)           91648     
_________________________________________________________________
dropout_3 (Dropout)          (None, 10, 128)           0         
_________________________________________________________________
lstm_4 (LSTM)                (None, 128)               131584    
_________________________________________________________________
dropout_4 (Dropout)          (None, 128)               0         
_________________________________________________________________
dense_1 (Dense)              (None, 5)                 645       
_________________________________________________________________
activation_1 (Activation)    (None, 5)                 0         
=================================================================
Total params: 20,223,927
Trainable params: 223,877
Non-trainable params: 20,000,050
_________________________________________________________________

As usual, after creating your model in Keras, you need to compile it and define what loss, optimizer and metrics your are want to use. Compile your model using categorical_crossentropy loss, adam optimizer and ['accuracy'] metrics:

model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

It’s time to train your model. Your Emojifier-V2 model takes as input an array of shape (m, max_len) and outputs probability vectors of shape (m, number of classes). We thus have to convert X_train (array of sentences as strings) to X_train_indices (array of sentences as list of word indices), and Y_train (labels as indices) to Y_train_oh (labels as one-hot vectors).

X_train_indices = sentences_to_indices(X_train, word_to_index, maxLen)
Y_train_oh = convert_to_one_hot(Y_train, C = 5)

Fit the Keras model on X_train_indices and Y_train_oh. We will use epochs = 50 and batch_size = 32.

model.fit(X_train_indices, Y_train_oh, epochs = 50, batch_size = 32, shuffle=True)

Your model should perform close to 100% accuracy on the training set. The exact accuracy you get may be a little different. Run the following cell to evaluate your model on the test set.

X_test_indices = sentences_to_indices(X_test, word_to_index, max_len = maxLen)
Y_test_oh = convert_to_one_hot(Y_test, C = 5)
loss, acc = model.evaluate(X_test_indices, Y_test_oh)
print()
print("Test accuracy = ", acc)

32/56 [================>.............] - ETA: 0s
Test accuracy =  0.839285714286

You should get a test accuracy between 80% and 95%. Run the cell below to see the mislabelled examples.

# This code allows you to see the mislabelled examples
C = 5
y_test_oh = np.eye(C)[Y_test.reshape(-1)]
X_test_indices = sentences_to_indices(X_test, word_to_index, maxLen)
pred = model.predict(X_test_indices)
for i in range(len(X_test)):
    x = X_test_indices
    num = np.argmax(pred[i])
    if(num != Y_test[i]):
        print('Expected emoji:'+ label_to_emoji(Y_test[i]) + ' prediction: '+ X_test[i] + label_to_emoji(num).strip())

Expected emoji:? prediction: she got me a nice present  ❤️
Expected emoji:? prediction: This girl is messing with me   ❤️
Expected emoji:? prediction: work is horrible   ?
Expected emoji:? prediction: any suggestions for dinner ?
Expected emoji:? prediction: you brighten my day    ?
Expected emoji:? prediction: she is a bully ❤️
Expected emoji:? prediction: My life is so boring   ❤️
Expected emoji:? prediction: will you be my valentine   ?
Expected emoji:? prediction: go away    ⚾

Now you can try it on your own example. Write your own sentence below.

# Change the sentence below to see your prediction. Make sure all the words are in the Glove embeddings.  
x_test = np.array(['I can not agree more'])
X_test_indices = sentences_to_indices(x_test, word_to_index, maxLen)
print(x_test[0] +' '+  label_to_emoji(np.argmax(model.predict(X_test_indices))))

I can not agree more ?

Previously, Emojify-V1 model did not correctly label “not feeling happy,” but our implementation of Emojiy-V2 got it right. (Keras’ outputs are slightly random each time, so you may not have obtained the same result.) The current model still isn’t very robust at understanding negation (like “not happy”) because the training set is small and so doesn’t have a lot of examples of negation. But if the training set were larger, the LSTM model would be much better than the Emojify-V1 model at understanding such complex sentences.

Congratulations!

You have completed this notebook! ❤️❤️❤️

What you should remember:
- If you have an NLP task where the training set is small, using word embeddings can help your algorithm significantly. Word embeddings allow your model to work on words in the test set that may not even have appeared in your training set.
- Training sequence models in Keras (and in most other deep learning frameworks) requires a few important details:
- To use mini-batches, the sequences need to be padded so that all the examples in a mini-batch have the same length.
- An Embedding() layer can be initialized with pretrained values. These values can be either fixed or trained further on your dataset. If however your labeled dataset is small, it’s usually not worth trying to train a large pre-trained set of embeddings.
- LSTM() has a flag called return_sequences to decide if you would like to return every hidden states or only the last one.
- You can use Dropout() right after LSTM() to regularize your network.

如果NLP任务的训练集很小，那么使用word embeddings 对算法将有极大的帮助。Word embeddings能够让你的模型当测试集中出现了训练集中没有的单词时仍然正常工作。
在Keras中训练时序模型需要以下几个重要细节
1. 为了使用mini batch，序列需要补齐，这样所有mini batch 中的例子拥有同样的长度
2. 通过一个预先训练好的值来初始化一个embedding 层。
3. LSTM有一个return_sequence的flag，用来决定是否返回所有隐层或者最后一个隐层。
4. LSTM()后使用dropout()可以帮助正则化。

Congratulations on finishing this assignment and building an Emojifier. We hope you’re happy with what you’ve accomplished in this notebook!

注：本文涉及的图片集资料均整理翻译自Andrew Ng的Deep Learning 系列课程，版权归其所有。翻译整理水平有限，如有不妥的地方欢迎指出。如有侵权请联系删除。谢谢。