Description
Introduction In this assignment, you will train a few attention-based neural machine translation models to translate words from English to Pig-Latin. Along the way, you’ll gain experience with several important concepts in NMT, including gated recurrent neural networks and attention. Pig Latin Pig Latin is a simple transformation of English based on the following rules (applied on a per-word basis): 1. If the first letter of a word is a consonant, then the letter is moved to the end of the word, and the letters “ay” are added to the end: team → eamtay. 2. If the first letter is a vowel, then the word is left unchanged and the letters “way” are added to the end: impress → impressway. 3. In addition, some consonant pairs, such as “sh”, are treated as a block and are moved to the end of the string together: shopping → oppingshay. To translate a whole sentence from English to Pig-Latin, we simply apply these rules to each word independently: i went shopping → iway entway oppingshay Goal: We would like a neural machine translation model to learn the rules of Pig-Latin implicitly, from (English, Pig-Latin) word pairs. Since the translation to Pig Latin involves moving characters around in a string, we will use character-level recurrent neural networks for our model. Because English and Pig-Latin are so similar in structure, the translation task is almost a copy task; the model must remember each character in the input, and recall the characters in a specific order to produce the output. This makes it an ideal task for understanding the capacity of NMT models. Setting Up We recommend that you use Colab(https://colab.research.google.com/) for the assignment, as all the assignment notebooks have been tested on Colab. From the assignment zip file, you will find one python notebook file: nmt.ipynb. To setup the Colab environment, just upload this notebook file using the upload tab at https://colab.research.google.com/. 2 CSC413/2516 Winter 2020 with Professor Jimmy Ba Programming Assignment 3 Data The data for this task consists of pairs of words {(s (i) , t(i) )} N i=1 where the source s (i) is an English word, and the target t (i) is its translation in Pig-Latin. The dataset is composed of unique words from the book “Sense and Sensibility,” by Jane Austen. The vocabulary consists of 29 tokens: the 26 standard alphabet letters (all lowercase), the dash symbol -, and two special tokens and that denote the start and end of a sequence, respectively. 3 The dataset contains 6387 unique (English, Pig-Latin) pairs in total; the first few examples are: { (the, ethay), (family, amilyfay), (of, ofway), … } In order to simplify the processing of mini-batches of words, the word pairs are grouped based on the lengths of the source and target. Thus, in each mini-batch the source words are all the same length, and the target words are all the same length. This simplifies the code, as we don’t have to worry about batches of variable-length sequences. Outline of Assignment Throughout the rest of the assignment, you will implement some attention-based neural machine translation models, and finally train the models and examine the results. You will first implement three main building blocks: Gated Recurrent Unit (GRU), Additive attention and Scaled dot-product attention. Using these building blocks, you will implement two encoders (RNN and transformer encoders) and three decoders (RNN, RNN+additive attention and transformer decoders). Using these, you will train three final models: • Part 1: (RNN encoder) + (RNN decoder) • Part 2: (RNN encoder) + (RNN decoder with additive attention) • Part 3: (Transformer encoder) + (Transformer decoder) • Part 4: BERT fine-tuning Deliverables Each section is followed by a checklist of deliverables to add in the assignment writeup. To also give a better sense of our expectations for the answers to the conceptual questions, we’ve put maximum sentence limits. You will not be graded for any additional sentences. 3Note that for the English-to-Pig-Latin task, the input and output sequences share the same vocabulary; this is not always the case for other translation tasks (i.e., between languages that use different alphabets). 3 CSC413/2516 Winter 2020 with Professor Jimmy Ba Programming Assignment 3 Part 1: Gated Recurrent Unit (GRU) [2pt] Translation is a sequence-to-sequence problem: in our case, both the input and output are sequences of characters. A common architecture used for seq-to-seq problems is the encoder-decoder model [2], composed of two RNNs, as follows: c a t a t c a y a t c a y Encoder Decoder Training Figure 1: Training the NMT encoder-decoder architecture. c a t a t c a y Encoder Decoder Generation Figure 2: Generating text with the NMT encoder-decoder architecture. The encoder RNN compresses the input sequence into a fixed-length vector, represented by the final hidden state hT . The decoder RNN conditions on this vector to produce the translation, character by character. Input characters are passed through an embedding layer before they are fed into the encoder RNN; in our model, we learn a 29 × 10 embedding matrix, where each of the 29 characters in the vocabulary is assigned a 10-dimensional embedding. At each time step, the decoder RNN outputs a vector of unnormalized log probabilities given by a linear transformation of the decoder hidden state. When these probabilities are normalized, they define a distribution over the vocabulary, indicating the most probable characters for that time step. The model is trained via a cross-entropy loss between the decoder distribution and ground-truth at each time step. The decoder produces a distribution over the output vocabulary conditioned on the previous 4 CSC413/2516 Winter 2020 with Professor Jimmy Ba Programming Assignment 3 hidden state and the output token in the previous timestep. A common practice used to train NMT models is to feed in the ground-truth token from the previous time step to condition the decoder output in the current step. This training procedure is known as “teacher-forcing” shown in Figure 1. At test time, we don’t have access to the ground-truth output sequence, so the decoder must condition its output on the token it generated in the previous time step, as shown in Figure 2. Lets begin with implementing common encoder models: the Gated Recurrent Unit and the transformer encoder. Open https://colab.research.google.com/drive/1rHYoCXb96INsxCSc1G4OmZhisnuabsIH on Colab and answer the following questions. 1. [1pt] The forward pass of a Gated Recurrent Unit is defined by the following equations: rt = σ(Wirxt + Whrht−1 + br) (1) zt = σ(Wizxt + Whzht−1 + bz) (2) gt = tanh(Winxt + rt (Whnht−1 + bg)) (3) ht = (1 − z) gt + z ht−1, (4) where is the element-wise multiplication. Although PyTorch has a GRU built in (nn.GRUCell), we’ll implement our own GRU cell from scratch, to better understand how it works. Complete the __init__ and forward methods of the MyGRUCell class, to implement the above equations. A template has been provided for the forward method. 2. [0pt] Run the cells including GRU-based encoder/decoder models. 3. [1pt] Train the RNN encoder/decoder model. We’ve provided implementations for recurrent encoder/decoder models using the GRU cell. (Make sure you have run all the relevant previous cells to load the training and utility functions.) By default, the script runs for 100 epochs. At the end of each epoch, the script prints training and validation losses, and the Pig-Latin translation of a fixed sentence, “the air conditioning is working”, so that you can see how the model improves qualitatively over time. The script also saves several items to the directory h20-bs64-rnn: • The best encoder and decoder model paramters, based on the validation loss. • A plot of the training and validation losses. After the training is complete, we will now use this model to translate the words in the next notebook cell using translate_sentence function. Try a few of your own words by changing the variable TEST_SENTENCE. Identify two distinct failure modes and briefly describe them. 5 CSC413/2516 Winter 2020 with Professor Jimmy Ba Programming Assignment 3 Deliverables Create a section in your report called Gated Recurrent Units. Add the following in this section: • A screenshot of your full MyGRUCell implementation. [1pt] • The training/validation loss plots. [0pts] • Your answer for the question in step 3. Make sure to add at least one input-output pair for each failure case you identify. Your answer should not exceed three sentences in total (excluding the failure cases you’ve identified. ) [1pt] 6 CSC413/2516 Winter 2020 with Professor Jimmy Ba Programming Assignment 3 Part 2: Additive Attention [2pt] Attention allows a model to look back over the input sequence, and focus on relevant input tokens when producing the corresponding output tokens. For our simple task, attention can help the model remember tokens from the input, e.g., focusing on the input letter c to produce the output letter c. The hidden states produced by the encoder while reading the input sequence, h enc 1 , . . . , henc T can be viewed as annotations of the input; each encoder hidden state h enc i captures information about the i th input token, along with some contextual information. At each time step, an attention-based decoder computes a weighting over the annotations, where the weight given to each one indicates its relevance in determining the current output token. In particular, at time step t, the decoder computes an attention weight α (t) i for each of the encoder hidden states h enc i . The attention weights are defined such that 0 ≤ α (t) i ≤ 1 and P i α (t) i = 1. α (t) i is a function of an encoder hidden state and the previous decoder hidden state, f(h dec t−1 , henc i ), where i ranges over the length of the input sequence. There are a few engineering choices for the possible function f. In this assignment, we will investigate two different attention models: 1) the additive attention using a two-layer MLP and 2) the scaled dot product attention, which measures the similarity between the two hidden states. To unify the interface across different attention modules, we consider attention as a function whose inputs are triple (queries, keys, values), denoted as (Q, K, V ). In the additive attention, we will learn the function f, parameterized as a two-layer fullyconnected network with a ReLU activation. This network produces unnormalized weights ˜α (t) i that are used to compute the final context vector. Decoder Hidden States Encoder Hidden States batch_size batch_size seq_len hidden_size hidden_size batch_size seq_len 1 Attention Weights Figure 3: Dimensions of the inputs, Decoder Hidden States (query), Encoder Hidden States (keys/values) and the attention weights (α (t) ). 7 CSC413/2516 Winter 2020 with Professor Jimmy Ba Programming Assignment 3 For the forward pass, you are given a batch of query of the current time step, which has dimension batch_size x hidden_size, and a batch of keys and values for each time step of the input sequence, both have dimension batch_size x seq_len x hidden_size. The goal is to obtain the context vector. We first compute the function f(Qt , K) for each query in the batch and all corresponding keys Ki , where i ranges over seq_len different values. You must do this in a vectorized fashion. Since f(Qt , Ki) is a scalar, the resulting tensor of attention weights should have dimension batch_size x seq_len x 1. Some of the important tensor dimensions in the AdditiveAttention module are visualized in Figure 3. The AdditiveAttention module should return both the context vector batch_size x 1 x hidden_size and the attention weights batch_size x seq_len x 1. 1. [1pt] Read how the provided forward methods of the AdditiveAttention class computes α˜ (t) i , α (t) i , ct . Write down the mathematical expression for these quantity as a function of W1, W2, b1, b2, Qt , Ki . (Hint: Take a look at the equations in Part 4.1 for the scaled dot product attention model.) α˜ (t) i = f(Qt , Ki) = α (t) i = ct = Here, ˜α (t) i is the unnormalized attention weights; α (t) i is the attention weights in between 0 and 1; ct is the final context vector. 2. [1pt] We will now apply the AdditiveAttention module to the RNN decoder. You are given a batch of decoder hidden states as the query, h dec t−1 , for time t − 1, which has dimension batch_size x hidden_size, and a batch of encoder hidden states as the keys and values, h enc = [h enc 1 , . . . , henc i , . . . ] (annotations), for each timestep in the input sequence, which has dimension batch_size x seq_len x hidden_size. Qt ← h dec t−1 , K ← h enc, V ← h enc We will use these as the inputs to the self.attention to obtain the context. The output context vector is concatenated with the input vector and passed into the decoder GRU cell at each time step, as shown in Figure 4. Fill in the forward method of the RNNAttentionDecoder class to implement the interface shown in Figure 4. There are three steps we will need to implement: (a) Get the embedding corresponding to the time step. (given) (b) Compute the context vector and the attention weights using self.attentionl. (implement) 8 CSC413/2516 Winter 2020 with Professor Jimmy Ba Programming Assignment 3 … + h1 enc α1 αT Figure 4: Computing a context vector with attention. (c) Concatenate the context vector with the current decoder input. (implement) (d) Feed the concatenation to the decoder GRU cell to obtain the new hidden state. (given) 3. [0pt] Now run the following cell to train a language model that has additive attention in its decoder. Find one training example where the decoder with attention performs better than the decoder without attention. Show the input/outputs of the model with attention, and the model without attention that you’ve trained in the previous section. 4. [0pt] How does the training speed compare? Why? Deliverables Create a section called Additive Attention. Add the following in this section: • Three equations for question 1. [1pt] • A screenshot of your RNNAttentionDecoder class implementation. [1pt] • Training/validation plots you’ve obtained in this section. [0 pts] • Answers to question 3. [0 pts] • Answer to question 4.[0 pts] 9 CSC413/2516 Winter 2020 with Professor Jimmy Ba Programming Assignment 3 Part 3: Scaled Dot Product Attention [4pt] 1. [0.5pt] In lecture, we learnt about Scaled Dot-product Attention used in the transformer models. The function f is a dot product between the linearly transformed query and keys using weight matrices Wq and Wk: α˜ (t) i = f(Qt , Ki) = (WqQt) T (WkKi) √ d , α (t) i = softmax(˜α (t) )i , ct = X T i=1 α (t) i WvVi , where, d is the dimension of the query and the Wv denotes weight matrix project the value to produce the final context vectors. Implement the scaled dot-product attention mechanism. Fill in the forward methods of the ScaledDotAttention class. Use the PyTorch torch.bmm (or @) to compute the dot product between the batched queries and the batched keys in the forward pass of the ScaledDotAttention class for the unnormalized attention weights. The following functions are useful in implementing models like this. You might find it useful to get familiar with how they work. (click to jump to the PyTorch documentation): • squeeze • unsqueeze • expand as • cat • view • bmm (or @) Your forward pass needs to work with both 2D query tensor (batch_size x (1) x hidden_size) and 3D query tensor (batch_size x k x hidden_size). 2. [0.5pt] Implement the causal scaled dot-product attention mechanism. Fill in the forward method in the CausalScaledDotAttention class. It will be mostly the same as the ScaledDotAttention class. The additional computation is to mask out the attention to the future time steps. You will need to add self.neg_inf to some of the entries in the unnormalized attention weights. You may find torch.tril handy for this part. 10 CSC413/2516 Winter 2020 with Professor Jimmy Ba Programming Assignment 3 Figure 5: The transformer architecture. [3] 11 CSC413/2516 Winter 2020 with Professor Jimmy Ba Programming Assignment 3 3. [0.5pt] We will now use ScaledDotAttention as the building blocks for a simplified transformer [3] encoder. The encoder looks like the left half of Figure 5. The encoder consists of three components (already provided): • Positional encoding: Without any additional modifications, self attention is permutationequivariant. To encode the position of each word, we add to its embedding a constant vector that depends on its position: pth word embedding = input embedding + positional encoding(p) We follow the same positional encoding methodology described in [3]. That is we use sine and cosine functions: PE(pos, 2i) = sin pos 100002i/dmodel (5) PE(pos, 2i + 1) = cos pos 100002i/dmodel (6) Since we always use the same positional encodings throughout the training, we pregenerate all those we’ll need while constructing this class (before training) and keep reusing them throughout the training. • A ScaledDotAttention operation. • A following MLP. Now, complete the forward method of TransformerEncoder. Most of the code is given, except for two lines with . . . in them. Complete these lines. 4. [0.5pt] The decoder, in addition to all the components the encoder has, also requires a CausalScaledDotAttention component. Take a look at Figure 5. The transformer solves the translation problem using layers of attention modules. In each layer, we first apply the CausalScaledDotAttention self-attention to the decoder inputs followed by ScaledDotAttention attention module to the encoder annotations, similar to the attention decoder from the previous question. The output of the attention layers are fed into an hidden layer using ReLU activation. The final output of the last transformer layer are passed to the self.out to compute the word prediction. To improve the optimization, we add residual connections between the attention layers and ReLU layers. Now, complete the forward method of TransformerDecoder. Again, most of the code is given to you – fill in the two lines that have . . . . 5. [1pt] Now, train the language model with transformer based encoder/decoder. How do the translation results compare to the previous decoders? Write a short, qualitative analysis. 12 CSC413/2516 Winter 2020 with Professor Jimmy Ba Programming Assignment 3 6. [1pt] Modify the transformer decoder __init__ to use non-causal attention for both self attention and encoder attention. What do you observe when training this modified transformer? How do the results compare with the causal model? Why? 7. [0pt] What are the advantages and disadvantages of using additive attention vs scaled dotproduct attention? List one advantage and one disadvantage for each method. Deliverables Create a section in your report called Scaled Dot Product Attention. Add the following: • Screenshots of your ScaledDotProduct, CausalScaledDotProduct, TransformerEncoder and TransformerDecoder implementations. Highlight the lines you’ve added. [2pt] • Training/validation plots you’ve generated. Your response to question 5. Your analysis should not exceed three sentences (excluding the failure cases you’ve identified). [1pt] • Your response to question 6. Your response should not exceed three sentences. [1pt] • Your response to question 7. [0pt] 13 CSC413/2516 Winter 2020 with Professor Jimmy Ba Programming Assignment 3 Part 4: BERT for arithmetic sentiment analysis [2pt] In this section, we will learn how to use a pre-trained BERT model to determine whether an verbal numerical expression is negative (label 0), zero (label 1), or positive (label 2). For example, “eight minus ten” is negative so the output of our sentence classifier should output label index 0. We start by explaining what BERT is, and how we can add a classifier on top of the pre-trained BERT to perform sentiment analysis for verbal numerical expressions. Most code is given to you in the notebook https://colab.research.google.com/drive/1QMGZsQ5u7JWuXiwvOhaH_OUd8Cn8E3aw Your task is to slightly modify the sentence classifier layer, make plots, report performances, and think about inference examples to test the model. Please carefully review the background for BERT before starting to answer the questions. The Hugging Face transformers library, used in this tutorial, has more than 20k stars on github due to its ease of use, and will be very useful for your research or projects in the future. Background for BERT: Bidirectional Encoder Representations from Transformers (BERT) [1], as the name suggests, is a language model based on the Transformer [3] encoder architecture that has been pre-trained on a large dataset of unlabeled sentences from Wikipedia and BookCorpus [4]. Given a sequence of tokens representing sentence(s), BERT outputs a “contextualized representation” vector for each of the token. Now, suppose we are given some down-stream tasks, such as sentence classification or question-answering. We can take the BERT model, add a small layer on top of the BERT representation(s), and then fine-tune the added parameters and BERT parameters on the downstream dataset, which is typically much smaller than the data used to pre-train BERT. In traditional language modeling task, the objective is to maximize the log likelihood of predicting the current word (or token) in the sentence, given the previous words (to the left of current work) as context. This is called the “autoregressive model”. In BERT, however, we wish to predict the current word given both the words before and after (i.e. to the left and to the right) of the sentence–hence “bidirectional”. To be able to attend from both directions, BERT uses the encoder Transformer, which does not apply any attention masking unlike the decoder. We briefly describe how BERT is pre-trained. BERT has 2 task objectives for pre-training: (1) Masked Language Modeling (Masked LM), and (2) Next Sentence Prediction (NSP). The input to the model is a sequence of tokens of the form: [CLS] Sentence A [SEP] Sentence B, where [CLS] (“class”) and [SEP] (“separator”) are special tokens. In Masked LM, some percentage of the input tokens are converted into [MASK] tokens, and the objective is to use the final layer representation for that masked token to predict the correct word that was masked out4 . For 4The full training detail is slightly more complicated, but conceptually similar. 14 CSC413/2516 Winter 2020 with Professor Jimmy Ba Programming Assignment 3 BERT BERT E[CLS] E1 E[SEP] … EN E1 ’ … EM ’ C T1 T[SEP] … TN T1 ’ … TM ’ [CLS] Tok 1 [SEP] … Tok N Tok 1 … TokM Question Paragraph Start/End Span BERT E[CLS] E1 E[SEP] … EN E1 ’ … EM ’ C T1 T[SEP] … TN T1 ’ … TM ’ [CLS] Tok 1 [SEP] … Tok N Tok 1 … TokM Masked Sentence A Masked Sentence B Pre-training Fine-Tuning NSP Mask LM Mask LM Unlabeled Sentence A and B Pair SQuAD Question Answer Pair MNLI NER Figure 6: Overall pre-training and fine-tuning for BERT. Reproduced from BERT paper [1] NSP, the task is to use the contextualized representation for the [CLS] token to perform binary classification for whether sentence A and sentence B are consecutive sentences in the unlabeled dataset. See Figure 6 for the conceptual picture of BERT pre-training and fine-tuning. In this assignment, you will be fine-tuning BERT on a single sentence classification task (see below about the dataset). Figure 7 illustrates the architecture for fine-tuning on this task. We prepend the tokenized sentence with the [CLS] token, then feed the sequence into BERT. We then take the contextualized [CLS] token representation at the last layer of BERT and add either a softmax layer on top corresponding to the number of output classes in the task. Alternatively, we can have fully connected hidden layers before the softmax layer for more expressivity for harder tasks. Then, both the new layers and the entire BERT parameters are trained end to end on the task for a few epochs. 15 CSC413/2516 Winter 2020 with Professor Jimmy Ba Programming Assignment 3 %(57 (>6(3@ 1 ( ¶ (0¶ &>6(3@ 1 7 ¶ 70¶ >(3@ 7RN 1 7RN 7RN 0 4XHVWDJUDSK %(57 (>&/6@ ( ( (1 & 7 7 71 6LQJOH6HQWHQFH %(57 >&/6@ 7RN 7RN 7RN1 (>&/6@ ( ( (1 & 7 7 71 6LQJOH6HQWHQFH 2 %3(5 2 ( >&/6@ ( (>6(3 &OD (1 7>6(3@ 71 7 70¶ W(QG6SDQ &ODVV /DEHO %(57 >&/6@ 7RN >&/6@ 7RN 7RN >6(3@ 7RN 1 7RN 0 6HQWHQFH 6HQWHQFH Figue-tuning BERT for single sentence classification by adding a layer on top of the contextualized [CLS] token representation. Reproduced from BERT paper [1] Dataset Description The verbal arithmetic dataset contains pairs of input sentence and label. The label is tertiary. Label 0, 1, 2 mean the input expressions are evaluated as “negative” , “zero”, and “positive” respectively. Note that the size of training dataset is 640 and the size of test dataset is 160. In our dataset, we only have sentences with three word tokens as the input, similar to the examples shown below: Input expression Label Label meaning eighteen minus eighteen 1 “zero” four plus seven 2 “positive” four minus ten 0 “negative” Questions: 1. [0pt] Classifier layer. Open the notebook https://colab.research.google.com/drive/ 1QMGZsQ5u7JWuXiwvOhaH_OUd8Cn8E3aw, we have provided two example BERT classes: BertCSC413 Linear and BertCSC413 MLP Example that both add a classifier for classification. In this part, you need to make your own BertCSC413 MLP class by, for example, modifying the provided examples: change the number of layers; change the number of hidden neurons; or try a different activation. 16