flashlight 0.0.12

neural network library
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

weights
weights = weights - relu.input_cache.relu_der.tens_broadcast_mul(grad_output).matmul(linear.input_cache.transposed());

weights[i] = weights[i] - dot(relu.input_cache.relu_der.tens_broadcast_mul(grad_output).row(row), linear.input_cache.row(col));

var sum = 0.0;
for i in 0..relu_shapes.row{
	sum += relu_der(relu.input_cache[x]) + grad_output[shape.row] * linear.input_cache[i.toShape().swapParams().toIdx];
}

output[i] = input[i] - sum;

shape[row, col] = i.toShape(weights_shape);

biases
output[i] = relu_der(relu.input_cache[x])*grad_output * 1/shapes[1](samples_per_batch)

grad
output[i] = weights^T.matmul(relu_der(relu.input_cache[x])) -> for(i in weights.row) sum += weights[i.col][i.row] * relu_der(relu.input_cache[i.row][i.col])