so much need to commit

Joseph Redmon

2016-05-06 c7b10ceadb1a78e7480d281444a31ae2a7dc1b05

 src/cost_layer.c

@@ -10,7 +10,9 @@
COST_TYPE get_cost_type(char *s)
{
    if (strcmp(s, "sse")==0) return SSE;
    fprintf(stderr, "Couldn't find activation function %s, going with SSE\n", s);
    if (strcmp(s, "masked")==0) return MASKED;
    if (strcmp(s, "smooth")==0) return SMOOTH;
    fprintf(stderr, "Couldn't find cost type %s, going with SSE\n", s);
    return SSE;
}

@@ -19,65 +21,103 @@
    switch(a){
        case SSE:
            return "sse";
        case MASKED:
            return "masked";
        case SMOOTH:
            return "smooth";
    }
    return "sse";
}

cost_layer *make_cost_layer(int batch, int inputs, COST_TYPE type)
cost_layer make_cost_layer(int batch, int inputs, COST_TYPE cost_type, float scale)
{
    fprintf(stderr, "Cost Layer: %d inputs\n", inputs);
    cost_layer *layer = calloc(1, sizeof(cost_layer));
    layer->batch = batch;
    layer->inputs = inputs;
    layer->type = type;
    layer->delta = calloc(inputs*batch, sizeof(float));
    layer->output = calloc(1, sizeof(float));
    cost_layer l = {0};
    l.type = COST;

    l.scale = scale;
    l.batch = batch;
    l.inputs = inputs;
    l.outputs = inputs;
    l.cost_type = cost_type;
    l.delta = calloc(inputs*batch, sizeof(float));
    l.output = calloc(inputs*batch, sizeof(float));
    l.cost = calloc(1, sizeof(float));
    #ifdef GPU
    layer->delta_gpu = cuda_make_array(layer->delta, inputs*batch);
    l.delta_gpu = cuda_make_array(l.output, inputs*batch);
    l.output_gpu = cuda_make_array(l.delta, inputs*batch);
    #endif
    return layer;
    return l;
}

void forward_cost_layer(cost_layer layer, network_state state)
void resize_cost_layer(cost_layer *l, int inputs)
{
    l->inputs = inputs;
    l->outputs = inputs;
    l->delta = realloc(l->delta, inputs*l->batch*sizeof(float));
    l->output = realloc(l->output, inputs*l->batch*sizeof(float));
#ifdef GPU
    cuda_free(l->delta_gpu);
    cuda_free(l->output_gpu);
    l->delta_gpu = cuda_make_array(l->delta, inputs*l->batch);
    l->output_gpu = cuda_make_array(l->output, inputs*l->batch);
#endif
}

void forward_cost_layer(cost_layer l, network_state state)
{
    if (!state.truth) return;
    copy_cpu(layer.batch*layer.inputs, state.truth, 1, layer.delta, 1);
    axpy_cpu(layer.batch*layer.inputs, -1, state.input, 1, layer.delta, 1);
    *(layer.output) = dot_cpu(layer.batch*layer.inputs, layer.delta, 1, layer.delta, 1);
    //printf("cost: %f\n", *layer.output);
    if(l.cost_type == MASKED){
        int i;
        for(i = 0; i < l.batch*l.inputs; ++i){
            if(state.truth[i] == SECRET_NUM) state.input[i] = SECRET_NUM;
        }
    }
    if(l.cost_type == SMOOTH){
        smooth_l1_cpu(l.batch*l.inputs, state.input, state.truth, l.delta, l.output);
    } else {
        l2_cpu(l.batch*l.inputs, state.input, state.truth, l.delta, l.output);
    }
    l.cost[0] = sum_array(l.output, l.batch*l.inputs);
}

void backward_cost_layer(const cost_layer layer, network_state state)
void backward_cost_layer(const cost_layer l, network_state state)
{
    copy_cpu(layer.batch*layer.inputs, layer.delta, 1, state.delta, 1);
    axpy_cpu(l.batch*l.inputs, l.scale, l.delta, 1, state.delta, 1);
}

#ifdef GPU

void pull_cost_layer(cost_layer layer)
void pull_cost_layer(cost_layer l)
{
    cuda_pull_array(layer.delta_gpu, layer.delta, layer.batch*layer.inputs);
    cuda_pull_array(l.delta_gpu, l.delta, l.batch*l.inputs);
}

void push_cost_layer(cost_layer layer)
void push_cost_layer(cost_layer l)
{
    cuda_push_array(layer.delta_gpu, layer.delta, layer.batch*layer.inputs);
    cuda_push_array(l.delta_gpu, l.delta, l.batch*l.inputs);
}

void forward_cost_layer_gpu(cost_layer layer, network_state state)
void forward_cost_layer_gpu(cost_layer l, network_state state)
{
    if (!state.truth) return;
    
    copy_ongpu(layer.batch*layer.inputs, state.truth, 1, layer.delta_gpu, 1);
    axpy_ongpu(layer.batch*layer.inputs, -1, state.input, 1, layer.delta_gpu, 1);
    if (l.cost_type == MASKED) {
        mask_ongpu(l.batch*l.inputs, state.input, SECRET_NUM, state.truth);
    }

    cuda_pull_array(layer.delta_gpu, layer.delta, layer.batch*layer.inputs);
    *(layer.output) = dot_cpu(layer.batch*layer.inputs, layer.delta, 1, layer.delta, 1);
    if(l.cost_type == SMOOTH){
        smooth_l1_gpu(l.batch*l.inputs, state.input, state.truth, l.delta_gpu, l.output_gpu);
    } else {
        l2_gpu(l.batch*l.inputs, state.input, state.truth, l.delta_gpu, l.output_gpu);
    }

    cuda_pull_array(l.output_gpu, l.output, l.batch*l.inputs);
    l.cost[0] = sum_array(l.output, l.batch*l.inputs);
}

void backward_cost_layer_gpu(const cost_layer layer, network_state state)
void backward_cost_layer_gpu(const cost_layer l, network_state state)
{
    copy_ongpu(layer.batch*layer.inputs, layer.delta_gpu, 1, state.delta, 1);
    axpy_ongpu(l.batch*l.inputs, l.scale, l.delta_gpu, 1, state.delta, 1);
}
#endif