Refactor NN structure (NN_CONFIG_V2).
Refactor NN structure (NN_CONFIG_V2).
- Add struct FC_LAYER to make the NN configuration more clear.
- Add ReLU and sigmoid as optional activation functions.
- Add softmax cross entropy as optional loss function.
Add the pointer to layer outputs and gradients.
- We plan to add back propagation for the NN model to make it online
learning (adjust the model weight during the encoding process).
Implement forward prediction for NN_CONFIG_V2.
Add an experimental flag CONFIG_NN_V2 for the change.
Change-Id: Id5c038dbc4a12a248c43d841b80f99946ce7ae6e
diff --git a/av1/encoder/ml.c b/av1/encoder/ml.c
index 579900a..b5d8a16 100644
--- a/av1/encoder/ml.c
+++ b/av1/encoder/ml.c
@@ -57,6 +57,76 @@
}
}
+#if CONFIG_NN_V2
+// Applies the ReLu activation to one fc layer
+// output[i] = Max(input[i],0.0f)
+static float *nn_relu(const float *input, FC_LAYER *layer) {
+ for (int i = 0; i < layer->num_outputs; ++i) {
+ layer->output[i] = AOMMAX(input[i], 0.0f);
+ }
+
+ return layer->output;
+}
+
+// Applies the Sigmoid activation to one fc layer
+// output[i] = 1/(1+exp(input[i]))
+static float *nn_sigmoid(const float *input, FC_LAYER *layer) {
+ for (int i = 0; i < layer->num_outputs; ++i) {
+ const float tmp = AOMMIN(AOMMAX(input[i], -10.0f), 10.0f);
+ layer->output[i] = 1.0f / (1.0f + expf(-tmp));
+ }
+
+ return layer->output;
+}
+
+// Forward prediction in one fc layer, used in function av1_nn_predict_V2
+static float *nn_fc_forward(const float *input, FC_LAYER *layer) {
+ const float *weights = layer->weights;
+ const float *bias = layer->bias;
+ assert(layer->num_outputs < NN_MAX_NODES_PER_LAYER);
+ // fc
+ for (int node = 0; node < layer->num_outputs; ++node) {
+ float val = bias[node];
+ for (int i = 0; i < layer->num_inputs; ++i) val += weights[i] * input[i];
+ layer->output[node] = val;
+ weights += layer->num_inputs;
+ }
+
+ // activation
+ switch (layer->activation) {
+ case NONE: return layer->output;
+ case RELU: return nn_relu(layer->output, layer);
+ case SIGMOID: return nn_sigmoid(layer->output, layer);
+ case SOFTSIGN:
+ assert(0 && "Softsign has not been supported in NN."); // TO DO
+ return NULL;
+ default:
+ assert(0 && "Unknown activation"); // Unknown activation
+ return NULL;
+ }
+}
+
+void av1_nn_predict_v2(const float *feature, NN_CONFIG_V2 *nn_config,
+ float *output) {
+ const float *input_nodes = feature;
+
+ // Propagate the layers.
+ const int num_layers = nn_config->num_hidden_layers;
+ assert(num_layers <= NN_MAX_HIDDEN_LAYERS);
+ for (int i = 0; i < num_layers; ++i) {
+ input_nodes = nn_fc_forward(input_nodes, nn_config->layer + i);
+ assert(nn_config->layer[i + 1].num_inputs ==
+ nn_config->layer[i].num_outputs);
+ }
+
+ // Final layer
+ input_nodes = nn_fc_forward(input_nodes, nn_config->layer + num_layers);
+ assert(nn_config->layer[num_layers].num_outputs == nn_config->num_logits);
+ // Copy the final layer output
+ memcpy(output, input_nodes, sizeof(*input_nodes) * nn_config->num_logits);
+}
+#endif // CONFIG_NN_V2
+
void av1_nn_softmax(const float *input, float *output, int n) {
// Softmax function is invariant to adding the same constant
// to all input values, so we subtract the maximum input to avoid
