Run this notebook online:\ |Binder| or Colab: |Colab| .. |Binder| image:: https://mybinder.org/badge_logo.svg :target: https://mybinder.org/v2/gh/deepjavalibrary/d2l-java/master?filepath=chapter_attention-mechanisms/multihead-attention.ipynb .. |Colab| image:: https://colab.research.google.com/assets/colab-badge.svg :target: https://colab.research.google.com/github/deepjavalibrary/d2l-java/blob/colab/chapter_attention-mechanisms/multihead-attention.ipynb .. _sec_multihead-attention: Multi-Head Attention ==================== In practice, given the same set of queries, keys, and values we may want our model to combine knowledge from different behaviors of the same attention mechanism, such as capturing dependencies of various ranges (e.g., shorter-range vs. longer-range) within a sequence. Thus, it may be beneficial to allow our attention mechanism to jointly use different representation subspaces of queries, keys, and values. To this end, instead of performing a single attention pooling, queries, keys, and values can be transformed with :math:`h` independently learned linear projections. Then these :math:`h` projected queries, keys, and values are fed into attention pooling in parallel. In the end, :math:`h` attention pooling outputs are concatenated and transformed with another learned linear projection to produce the final output. This design is called *multi-head attention*, where each of the :math:`h` attention pooling outputs is a *head* :cite:`Vaswani.Shazeer.Parmar.ea.2017`. Using fully-connected layers to perform learnable linear transformations, :numref:`fig_multi-head-attention` describes multi-head attention. |Multi-head attention, where multiple heads are concatenated then linearly transformed.| .. _fig_multi-head-attention: Model ----- Before providing the implementation of multi-head attention, let us formalize this model mathematically. Given a query :math:`\mathbf{q} \in \mathbb{R}^{d_q}`, a key :math:`\mathbf{k} \in \mathbb{R}^{d_k}`, and a value :math:`\mathbf{v} \in \mathbb{R}^{d_v}`, each attention head :math:`\mathbf{h}_i` (:math:`i = 1, \ldots, h`) is computed as .. math:: \mathbf{h}_i = f(\mathbf W_i^{(q)}\mathbf q, \mathbf W_i^{(k)}\mathbf k,\mathbf W_i^{(v)}\mathbf v) \in \mathbb R^{p_v}, where learnable parameters :math:`\mathbf W_i^{(q)}\in\mathbb R^{p_q\times d_q}`, :math:`\mathbf W_i^{(k)}\in\mathbb R^{p_k\times d_k}` and :math:`\mathbf W_i^{(v)}\in\mathbb R^{p_v\times d_v}`, and :math:`f` is attention pooling, such as additive attention and scaled dot-product attention in :numref:`sec_attention-scoring-functions`. The multi-head attention output is another linear transformation via learnable parameters :math:`\mathbf W_o\in\mathbb R^{p_o\times h p_v}` of the concatenation of :math:`h` heads: .. math:: \mathbf W_o \begin{bmatrix}\mathbf h_1\\\vdots\\\mathbf h_h\end{bmatrix} \in \mathbb{R}^{p_o}. Based on this design, each head may attend to different parts of the input. More sophisticated functions than the simple weighted average can be expressed. .. |Multi-head attention, where multiple heads are concatenated then linearly transformed.| image:: https://raw.githubusercontent.com/d2l-ai/d2l-en/master/img/multi-head-attention.svg .. code:: java %load ../utils/djl-imports %load ../utils/plot-utils %load ../utils/Functions.java %load ../utils/PlotUtils.java %load ../utils/attention/Chap10Utils.java %load ../utils/attention/DotProductAttention.java %load ../utils/attention/MultiHeadAttention.java %load ../utils/attention/PositionalEncoding.java .. code:: java NDManager manager = NDManager.newBaseManager(); To allow for parallel computation of multiple heads, the below ``MultiHeadAttention`` class uses two transposition functions as defined below. Specifically, the ``transposeOutput`` function reverses the operation of the ``transposeQkv`` function. .. code:: java public static NDArray transposeQkv(NDArray X, int numHeads) { // Shape of input `X`: // (`batchSize`, no. of queries or key-value pairs, `numHiddens`). // Shape of output `X`: // (`batchSize`, no. of queries or key-value pairs, `numHeads`, // `numHiddens` / `numHeads`) X = X.reshape(X.getShape().get(0), X.getShape().get(1), numHeads, -1); // Shape of output `X`: // (`batchSize`, `numHeads`, no. of queries or key-value pairs, // `numHiddens` / `numHeads`) X = X.transpose(0, 2, 1, 3); // Shape of `output`: // (`batchSize` * `numHeads`, no. of queries or key-value pairs, // `numHiddens` / `numHeads`) return X.reshape(-1, X.getShape().get(2), X.getShape().get(3)); } public static NDArray transposeOutput(NDArray X, int numHeads) { X = X.reshape(-1, numHeads, X.getShape().get(1), X.getShape().get(2)); X = X.transpose(0, 2, 1, 3); return X.reshape(X.getShape().get(0), X.getShape().get(1), -1); } Implementation -------------- In our implementation, we choose the scaled dot-product attention for each head of the multi-head attention. To avoid significant growth of computational cost and parameterization cost, we set :math:`p_q = p_k = p_v = p_o / h`. Note that :math:`h` heads can be computed in parallel if we set the number of outputs of linear transformations for the query, key, and value to :math:`p_q h = p_k h = p_v h = p_o`. In the following implementation, :math:`p_o` is specified via the argument ``numHiddens``. .. code:: java public static class MultiHeadAttention extends AbstractBlock { private int numHeads; public DotProductAttention attention; private Linear W_k; private Linear W_q; private Linear W_v; private Linear W_o; public MultiHeadAttention(int numHiddens, int numHeads, float dropout, boolean useBias) { this.numHeads = numHeads; attention = new DotProductAttention(dropout); W_q = Linear.builder().setUnits(numHiddens).optBias(useBias).build(); addChildBlock("W_q", W_q); W_k = Linear.builder().setUnits(numHiddens).optBias(useBias).build(); addChildBlock("W_k", W_k); W_v = Linear.builder().setUnits(numHiddens).optBias(useBias).build(); addChildBlock("W_v", W_v); W_o = Linear.builder().setUnits(numHiddens).optBias(useBias).build(); addChildBlock("W_o", W_o); Dropout dropout1 = Dropout.builder().optRate(dropout).build(); addChildBlock("dropout", dropout1); } @Override protected NDList forwardInternal( ParameterStore ps, NDList inputs, boolean training, PairList params) { // Shape of `queries`, `keys`, or `values`: // (`batchSize`, no. of queries or key-value pairs, `numHiddens`) // Shape of `validLens`: // (`batchSize`,) or (`batchSize`, no. of queries) // After transposing, shape of output `queries`, `keys`, or `values`: // (`batchSize` * `numHeads`, no. of queries or key-value pairs, // `numHiddens` / `numHeads`) NDArray queries = inputs.get(0); NDArray keys = inputs.get(1); NDArray values = inputs.get(2); NDArray validLens = inputs.get(3); // On axis 0, copy the first item (scalar or vector) for // `numHeads` times, then copy the next item, and so on validLens = validLens.repeat(0, numHeads); queries = transposeQkv( W_q.forward(ps, new NDList(queries), training, params).get(0), numHeads); keys = transposeQkv( W_k.forward(ps, new NDList(keys), training, params).get(0), numHeads); values = transposeQkv( W_v.forward(ps, new NDList(values), training, params).get(0), numHeads); // Shape of `output`: (`batchSize` * `numHeads`, no. of queries, // `numHiddens` / `numHeads`) NDArray output = attention .forward( ps, new NDList(queries, keys, values, validLens), training, params) .get(0); // Shape of `outputConcat`: // (`batchSize`, no. of queries, `numHiddens`) NDArray outputConcat = transposeOutput(output, numHeads); return new NDList(W_o.forward(ps, new NDList(outputConcat), training, params).get(0)); } @Override public Shape[] getOutputShapes(Shape[] inputShapes) { throw new UnsupportedOperationException("Not implemented"); } @Override public void initializeChildBlocks( NDManager manager, DataType dataType, Shape... inputShapes) { try (NDManager sub = manager.newSubManager()) { NDArray queries = sub.zeros(inputShapes[0], dataType); NDArray keys = sub.zeros(inputShapes[1], dataType); NDArray values = sub.zeros(inputShapes[2], dataType); NDArray validLens = sub.zeros(inputShapes[3], dataType); validLens = validLens.repeat(0, numHeads); ParameterStore ps = new ParameterStore(sub, false); W_q.initialize(manager, dataType, queries.getShape()); W_k.initialize(manager, dataType, keys.getShape()); W_v.initialize(manager, dataType, values.getShape()); queries = transposeQkv(W_q.forward(ps, new NDList(queries), false).get(0), numHeads); keys = transposeQkv(W_k.forward(ps, new NDList(keys), false).get(0), numHeads); values = transposeQkv(W_v.forward(ps, new NDList(values), false).get(0), numHeads); NDList list = new NDList(queries, keys, values, validLens); attention.initialize(sub, dataType, list.getShapes()); NDArray output = attention.forward(ps, list, false).head(); NDArray outputConcat = Chap10Utils.transposeOutput(output, numHeads); W_o.initialize(manager, dataType, outputConcat.getShape()); } } } Let us test our implemented ``MultiHeadAttention`` class using a toy example where keys and values are the same. As a result, the shape of the multi-head attention output is (``batchSize``, ``numQueries``, ``numHiddens``). .. code:: java int numHiddens = 100; int numHeads = 5; MultiHeadAttention attention = new MultiHeadAttention(numHiddens, numHeads, 0.5f, false); .. code:: java int batchSize = 2; int numQueries = 4; int numKvpairs = 6; NDArray validLens = manager.create(new float[] {3, 2}); NDArray X = manager.ones(new Shape(batchSize, numQueries, numHiddens)); NDArray Y = manager.ones(new Shape(batchSize, numKvpairs, numHiddens)); ParameterStore ps = new ParameterStore(manager, false); NDList input = new NDList(X, Y, Y, validLens); attention.initialize(manager, DataType.FLOAT32, input.getShapes()); NDList result = attention.forward(ps, input, false); result.get(0).getShape(); .. parsed-literal:: :class: output (2, 4, 100) Summary ------- - Multi-head attention combines knowledge of the same attention pooling via different representation subspaces of queries, keys, and values. - To compute multiple heads of multi-head attention in parallel, proper tensor manipulation is needed. Exercises --------- 1. Visualize attention weights of multiple heads in this experiment. 2. Suppose that we have a trained model based on multi-head attention and we want to prune least important attention heads to increase the prediction speed. How can we design experiments to measure the importance of an attention head?