vector-entailment/ComputeEntailmentExampleCostAndGrad.m at master · sleepinyourhat/vector-entailment · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
% Want to distribute this code? Have other questions? -> sbowman@stanford.edu
function [ cost, grad, embGrad, pred ] = ComputeEntailmentExampleCostAndGrad(theta, thetaDecoder, dataPoint, separateWordFeatures, hyperParams, computeGradient)
% Compute cost, gradient, and predicted label for one example.

grad = [];
embGrad = [];

assert(~hyperParams.useLattices, 'Use batched computation for the LatticeNN.');
assert(~hyperParams.gpu, 'Use batching for GPU computation.');

% Unpack theta
[ mergeMatrices, mergeMatrix ...
    softmaxMatrix, trainedWordFeatures, compositionMatrices,...
    compositionMatrix, ~, classifierExtraMatrix, ...
    embeddingTransformMatrix ] ...
    = stack2param(theta, thetaDecoder);

if hyperParams.trainWords && ~hyperParams.largeVocabMode
    wordFeatures = trainedWordFeatures;
else
    wordFeatures = separateWordFeatures;
end

DIM = hyperParams.dim;
EMBDIM = hyperParams.embeddingDim;

% Set the number of mposition functions
if hyperParams.useSumming
    NUMCOMP = 0;
elseif ~hyperParams.untied
    NUMCOMP = 1;
else
    NUMCOMP = 3;
end

NUMTRANS = size(embeddingTransformMatrix, 3);

left = dataPoint.left;
right = dataPoint.right;
trueLabel = dataPoint.label;

% Run the trees/sequences forward
left.updateFeatures(wordFeatures, compositionMatrices, ...
        compositionMatrix, embeddingTransformMatrix, hyperParams.compNL, computeGradient, hyperParams);
right.updateFeatures(wordFeatures, compositionMatrices, ...
        compositionMatrix, embeddingTransformMatrix, hyperParams.compNL, computeGradient, hyperParams);

leftFeatures = left.getFeatures();
rightFeatures = right.getFeatures();

[ leftFeatures, leftMask ] = Dropout(leftFeatures, hyperParams.topDropout, computeGradient, hyperParams.gpu);
[ rightFeatures, rightMask ] = Dropout(rightFeatures, hyperParams.topDropout, computeGradient, hyperParams.gpu);

% Compute classification tensor layer (or plain RNN layer)
if hyperParams.useThirdOrderMerge
    mergeOutput = ComputeTensorLayer(leftFeatures, ...
        rightFeatures, mergeMatrices, mergeMatrix, hyperParams.classNL);
else
    mergeOutput = ComputeRNNLayer(leftFeatures, rightFeatures, ...
        mergeMatrix, hyperParams.classNL);
end

% Run layers forward
extraInputs = zeros(hyperParams.penultDim, 1, hyperParams.topDepth);
extraInnerOutputs = zeros(hyperParams.penultDim, 1, hyperParams.topDepth - 1);
extraInputs(:, 1, 1) = mergeOutput;
for layer = 1:(hyperParams.topDepth - 1)
    extraInnerOutputs(:, 1, layer) = classifierExtraMatrix(:, :, layer) * [1; extraInputs(:, 1, layer)];
    extraInputs(:, 1, layer + 1) = hyperParams.classNL(extraInnerOutputs(:, 1, layer));
end

if ~isempty(hyperParams.labelCostMultipliers)
    multiplier = hyperParams.labelCostMultipliers(trueLabel(1));
    [ labelProbs, cost ] = ComputeSoftmaxLayer( ...
                        extraInputs(:,hyperParams.topDepth), softmaxMatrix, hyperParams, trueLabel', multiplier);
else
    [ labelProbs, cost ] = ComputeSoftmaxLayer( ...
                      extraInputs(:,hyperParams.topDepth), softmaxMatrix, hyperParams, trueLabel');
end


% Produce gradient
if nargout > 1 && (nargin < 6 || computeGradient)
    if ~isempty(hyperParams.labelCostMultipliers)
        [ localSoftmaxGradient, softmaxDelta ] = ...
            ComputeSoftmaxClassificationGradients( ...
              softmaxMatrix, labelProbs, trueLabel', ...
              extraInputs(:,hyperParams.topDepth), hyperParams, multiplier);
    else
        [ localSoftmaxGradient, softmaxDelta ] = ...
            ComputeSoftmaxClassificationGradients( ...
              softmaxMatrix, labelProbs, trueLabel', ...
              extraInputs(:,hyperParams.topDepth), hyperParams);
    end

    % Compute gradients for extra top layers
    [ localExtraMatrixGradients, extraDelta ] = ...
          ComputeExtraClassifierGradients(classifierExtraMatrix,...
            softmaxDelta, extraInputs, hyperParams.classNLDeriv);

    if hyperParams.useThirdOrderMerge
        % Compute gradients for classification tensor layer
        [ localMergeMatricesGradients, localMergeMatrixGradients, ...
            MergeDeltaLeft, MergeDeltaRight ] = ...
            ComputeTensorLayerGradients(leftFeatures, rightFeatures, ...
              mergeMatrices, mergeMatrix, ...
              extraDelta, hyperParams.classNLDeriv, mergeOutput);
    else
         % Compute gradients for classification first layer
         localMergeMatricesGradients = zeros(0, 0, 0);
         [ localMergeMatrixGradients, MergeDeltaLeft, ...
            MergeDeltaRight ] = ...
          ComputeRNNLayerGradients(leftFeatures, rightFeatures, ...
              mergeMatrix, ...
              extraDelta, hyperParams.classNLDeriv, mergeOutput);
    end

    MergeDeltaLeft = MergeDeltaLeft .* leftMask;
    MergeDeltaRight = MergeDeltaRight .* rightMask;

    [ localWordFeatureGradients, ...
      localCompositionMatricesGradients, ...
      localCompositionMatrixGradients, ...
      localEmbeddingTransformMatrixGradients ] = ...
       left.getGradient(MergeDeltaLeft, [], wordFeatures, ...
                            compositionMatrices, compositionMatrix, ...
                            embeddingTransformMatrix, ...
                            hyperParams.compNLDeriv, hyperParams);

    [ rightWordGradients, ...
      rightCompositionMatricesGradients, ...
      rightCompositionMatrixGradients, ...
      rightEmbeddingTransformMatrixGradients ] = ...
       right.getGradient(MergeDeltaRight, [], wordFeatures, ...
                            compositionMatrices, compositionMatrix, ...
                            embeddingTransformMatrix, ...
                            hyperParams.compNLDeriv, hyperParams);
    if hyperParams.trainWords
      localWordFeatureGradients = localWordFeatureGradients ...
          + rightWordGradients;
    end
    localCompositionMatricesGradients = localCompositionMatricesGradients...
        + rightCompositionMatricesGradients;
    localCompositionMatrixGradients = localCompositionMatrixGradients...
        + rightCompositionMatrixGradients;
    localEmbeddingTransformMatrixGradients = localEmbeddingTransformMatrixGradients...
        + rightEmbeddingTransformMatrixGradients;

    % Pack up gradients
    if hyperParams.largeVocabMode
      grad = param2stack(localMergeMatricesGradients, ...
          localMergeMatrixGradients, ...
          localSoftmaxGradient, ...
          [], localCompositionMatricesGradients, ...
          localCompositionMatrixGradients, ...
          [], ...
          localExtraMatrixGradients, ...
          localEmbeddingTransformMatrixGradients);
      embGrad = localWordFeatureGradients;
    else
      [ grad, dec ] = param2stack(localMergeMatricesGradients, ...
          localMergeMatrixGradients, ...
          localSoftmaxGradient, ...
          localWordFeatureGradients, localCompositionMatricesGradients, ...
          localCompositionMatrixGradients, ...
          [], ...
          localExtraMatrixGradients, ...
          localEmbeddingTransformMatrixGradients);
      embGrad = [];
    end

    % Clip the gradient.
    if hyperParams.clipGradients
        gradNorm = norm(grad);
        if gradNorm > hyperParams.maxGradNorm
            grad = grad .* (hyperParams.maxGradNorm ./ gradNorm);
        end
    end
end

% Compute prediction. Note: This will be in integer, indexing into whichever class set was used
% for this example.
if nargout > 3
    [ ~, pred ] = max(labelProbs);
end

end