netGradientCheck

NeuralNetwork.netGradientCheck(batch_Y, exp_work, costs)

勾配の計算のチェック

引数:

• batch_Y (float[]) – 正解の出力
• exp_work (float[]) – 作業用データ
• costs (float[]) – ミニバッチ内のコストの配列

ソース

netGradientCheck(batch_Y, exp_work, costs){
    inGradientCheck = true;

    var last_layer = this.layers[this.layers.length - 1];
    var last_delta_y = new Float32Array(last_layer.deltaY.dt);

    for (var batch_idx = 0; batch_idx < miniBatchSize; batch_idx++){
        last_layer.deltaY.dt = new Float32Array(last_delta_y.length);

        for(var i = 0; i < last_layer.unitSize; i++){
            var k = batch_idx * last_layer.unitSize + i;
            last_layer.deltaY.dt[k] = last_delta_y[k];
        }

        for (var i = this.layers.length - 1; 1 <= i; i--) {
            this.layers[i].backpropagation();
        }

        for(var layer_idx = 0; layer_idx < this.layers.length; layer_idx++){

            console.log("勾配確認 %s", this.layers[layer_idx].constructor.name);
            this.layers[layer_idx].gradientCheck(batch_Y, exp_work, costs[batch_idx], batch_idx, layer_idx);
        }
    }

    inGradientCheck = false;
}