vector-entailment/ComputeSentenceClassificationExampleCostAndGrad.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
% Want to distribute this code? Have other questions? -> sbowman@stanford.edu
function [ cost, grad, embGrad, pred ] = ComputeSentenceClassificationExampleCostAndGrad(theta, thetaDecoder, dataPoint, separateWordFeatures, hyperParams, computeGradient)
% Compute cost, gradient, and predicted label for one example.

grad = [];
embGrad = [];

assert(~hyperParams.useLattices, 'Unbatched computation is not available for the LatticeNN. Use batching.');
assert(~hyperParams.gpu, 'Use batching for GPU computation.');

% Unpack theta
[ ~, ~, ...
    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;

if isfield(hyperParams, 'smallVecs') && hyperParams.smallVecs
    sigma = 0.25;
else
    sigma = 1;
end

if length(hyperParams.randomEmbeddingIndices) > 0
    wordFeatures(:, hyperParams.randomEmbeddingIndices) = ...
        fNormrnd(0, sigma, [EMBDIM, length(hyperParams.randomEmbeddingIndices)], ...
                 hyperParams.gpu, hyperParams.gpu && hyperParams.largeVocabMode);
    wordFeatures(1, hyperParams.randomEmbeddingIndices) = 1;
end

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

NUMTRANS = size(embeddingTransformMatrix, 3);

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

[ features, mask ] = Dropout(dataPoint.sentence.getFeatures(), hyperParams.topDropout, computeGradient, hyperParams.gpu);

% Run layers forward
if hyperParams.penultDim == 2 * hyperParams.dim
    % Hack for transfer learning: pad with zeros.
    features = [zeros(size(features), 'like', features); features];
end

extraInputs = zeros(hyperParams.penultDim, 1, hyperParams.topDepth);
extraInnerOutputs = zeros(hyperParams.penultDim, 1, hyperParams.topDepth - 1);
extraInputs(:, 1, 1) = features;
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(dataPoint.label(1));
    [ labelProbs, cost ] = ComputeSoftmaxLayer( ...
                        extraInputs(:,hyperParams.topDepth), softmaxMatrix, hyperParams, dataPoint.label', multiplier);
else
    [ labelProbs, cost ] = ComputeSoftmaxLayer( ...
                      extraInputs(:,hyperParams.topDepth), softmaxMatrix, hyperParams, dataPoint.label');
end


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

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

    if hyperParams.penultDim == 2 * hyperParams.dim
        % Hack for transfer learning: pad with zeros.
        extraDelta = extraDelta(hyperParams.dim + 1:end, :);
    end

    extraDelta = extraDelta .* mask;

    [ localWordFeatureGradients, ...
      localCompositionMatricesGradients, ...
      localCompositionMatrixGradients, ...
      localEmbeddingTransformMatrixGradients ] = ...
       dataPoint.sentence.getGradient(extraDelta, [], wordFeatures, ...
                            compositionMatrices, compositionMatrix, ...
                            embeddingTransformMatrix, ...
                            hyperParams.compNLDeriv, hyperParams);

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