static MTensor predictOnMTML()

in FBSDKCoreKit/FBSDKCoreKit/AppEvents/Internal/ML/FBSDKModelRuntime.hpp [266:332]

• 51 lines of code
• 1 McCabe index (conditional complexity)

  static MTensor predictOnMTML(const std::string task, const char *texts, const std::unordered_map<std::string, MTensor> &weights, const float *df)
  {
    MTensor dense_tensor = getDenseTensor(df);
    std::string final_layer_weight_key = task + ".weight";
    std::string final_layer_bias_key = task + ".bias";

    const MTensor &embed_t = weights.at("embed.weight");
    const MTensor &conv0w_t = weights.at("convs.0.weight");
    const MTensor &conv1w_t = weights.at("convs.1.weight");
    const MTensor &conv2w_t = weights.at("convs.2.weight");
    const MTensor &conv0b_t = weights.at("convs.0.bias");
    const MTensor &conv1b_t = weights.at("convs.1.bias");
    const MTensor &conv2b_t = weights.at("convs.2.bias");
    const MTensor &fc1w_t = weights.at("fc1.weight"); // (128, 190)
    const MTensor &fc1b_t = weights.at("fc1.bias"); // 128
    const MTensor &fc2w_t = weights.at("fc2.weight"); // (64, 128)
    const MTensor &fc2b_t = weights.at("fc2.bias"); // 64
    const MTensor &final_layer_weight_t = weights.at(final_layer_weight_key); // (2, 64) or (5, 64)
    const MTensor &final_layer_bias_t = weights.at(final_layer_bias_key); // 2 or 5

    const MTensor &convs_0_weight = transpose3D(conv0w_t);
    const MTensor &convs_1_weight = transpose3D(conv1w_t);
    const MTensor &convs_2_weight = transpose3D(conv2w_t);
    const MTensor &fc1_weight = transpose2D(fc1w_t);
    const MTensor &fc2_weight = transpose2D(fc2w_t);
    const MTensor &final_layer_weight = transpose2D(final_layer_weight_t);

    // embedding
    const MTensor &embed_x = embedding(texts, SEQ_LEN, embed_t);

    // conv0
    MTensor c0 = conv1D(embed_x, convs_0_weight); // (1, 126, 32)
    addmv(c0, conv0b_t);
    relu(c0);

    // conv1
    MTensor c1 = conv1D(c0, convs_1_weight); // (1, 124, 64)
    addmv(c1, conv1b_t);
    relu(c1);
    c1 = maxPool1D(c1, 2); // (1, 123, 64)

    // conv2
    MTensor c2 = conv1D(c1, convs_2_weight); // (1, 121, 64)
    addmv(c2, conv2b_t);
    relu(c2);

    // max pooling
    MTensor ca = maxPool1D(c0, c0.size(1));
    MTensor cb = maxPool1D(c1, c1.size(1));
    MTensor cc = maxPool1D(c2, c2.size(1));

    // concatenate
    flatten(ca, 1);
    flatten(cb, 1);
    flatten(cc, 1);
    std::vector<MTensor *> concat_tensors { &ca, &cb, &cc, &dense_tensor };
    const MTensor &concat = concatenate(concat_tensors);

    // dense + relu
    MTensor dense1_x = dense(concat, fc1_weight, fc1b_t);
    relu(dense1_x);
    MTensor dense2_x = dense(dense1_x, fc2_weight, fc2b_t);
    relu(dense2_x);
    MTensor final_layer_dense_x = dense(dense2_x, final_layer_weight, final_layer_bias_t);
    softmax(final_layer_dense_x);
    return final_layer_dense_x;
  }