| | |
| | | return l; |
| | | } |
| | | |
| | | void softmax_array(float *input, int n, float *output) |
| | | void softmax_array(float *input, int n, float temp, float *output) |
| | | { |
| | | int i; |
| | | float sum = 0; |
| | |
| | | if(input[i] > largest) largest = input[i]; |
| | | } |
| | | for(i = 0; i < n; ++i){ |
| | | sum += exp(input[i]-largest); |
| | | sum += exp(input[i]/temp-largest/temp); |
| | | } |
| | | if(sum) sum = largest+log(sum); |
| | | if(sum) sum = largest/temp+log(sum); |
| | | else sum = largest-100; |
| | | for(i = 0; i < n; ++i){ |
| | | output[i] = exp(input[i]-sum); |
| | | output[i] = exp(input[i]/temp-sum); |
| | | } |
| | | } |
| | | |
| | |
| | | int inputs = l.inputs / l.groups; |
| | | int batch = l.batch * l.groups; |
| | | for(b = 0; b < batch; ++b){ |
| | | softmax_array(state.input+b*inputs, inputs, l.output+b*inputs); |
| | | softmax_array(state.input+b*inputs, inputs, l.temperature, l.output+b*inputs); |
| | | } |
| | | } |
| | | |
| | |
| | | { |
| | | int i; |
| | | for(i = 0; i < l.inputs*l.batch; ++i){ |
| | | state.delta[i] = l.delta[i]; |
| | | state.delta[i] += l.delta[i]; |
| | | } |
| | | } |
| | | |