shithub: opus

Info  •  Files  •  Log  •  Branches

Dump Conv1D (didn't check weight ordering at all)

--- a/dnn/dump_lpcnet.py

+++ b/dnn/dump_lpcnet.py

@@ -64,10 +64,7 @@

     weights = self.get_weights()

     printVector(f, weights[0], name + '_weights')

     printVector(f, weights[-1], name + '_bias')

-    if hasattr(self, 'activation'):

-        activation = self.activation.__name__.upper()

-    else:

-        activation = 'TANH'

+    activation = self.activation.__name__.upper()

     f.write('const DenseLayer {} = {{\n   {}_bias,\n   {}_weights,\n   {}, {}, ACTIVATION_{}\n}};\n\n'

             .format(name, name, name, weights[0].shape[0], weights[0].shape[1], activation))

     hf.write('#define {}_SIZE {}\n'.format(name.upper(), weights[0].shape[1]))

@@ -82,10 +79,7 @@

     printVector(f, weights[0], name + '_weights')

     printVector(f, weights[1], name + '_bias')

     printVector(f, weights[1], name + '_factor')

-    if hasattr(self, 'activation'):

-        activation = self.activation.__name__.upper()

-    else:

-        activation = 'TANH'

+    activation = self.activation.__name__.upper()

     f.write('const MDenseLayer {} = {{\n   {}_bias,\n   {}_weights,\n   {}_factor,\n   {}, {}, ACTIVATION_{}\n}};\n\n'

             .format(name, name, name, name, weights[0].shape[0], weights[0].shape[1], activation))

     hf.write('#define {}_SIZE {}\n'.format(name.upper(), weights[0].shape[0]))

@@ -92,6 +86,21 @@

     hf.write('extern const MDenseLayer {};\n\n'.format(name));

     return False

 MDense.dump_layer = dump_mdense_layer

+

+def dump_conv1d_layer(self, f, hf):

+    name = self.name

+    print("printing layer " + name + " of type " + self.__class__.__name__)

+    weights = self.get_weights()

+    printVector(f, weights[0], name + '_weights')

+    printVector(f, weights[-1], name + '_bias')

+    activation = self.activation.__name__.upper()

+    f.write('const Conv1DLayer {} = {{\n   {}_bias,\n   {}_weights,\n   {}, {}, {}, ACTIVATION_{}\n}};\n\n'

+            .format(name, name, name, weights[0].shape[1], weights[0].shape[0], weights[0].shape[2], activation))

+    hf.write('#define {}_SIZE {}\n'.format(name.upper(), weights[0].shape[1]))

+    hf.write('extern const Conv1DLayer {};\n\n'.format(name));

+    return False

+Conv1D.dump_layer = dump_conv1d_layer

+

 def dump_embedding_layer(self, f, hf):

     name = self.name

--- a/dnn/nnet.c

+++ b/dnn/nnet.c

@@ -128,6 +128,10 @@

    compute_activation(output, output, N, layer->activation);

 }

+void compute_mdense(const MDenseLayer *layer, float *output, const float *input)

+{

+}

+

 void compute_gru(const GRULayer *gru, float *state, const float *input)

 {

    int i;

@@ -146,8 +150,7 @@

       z[i] = gru->bias[i];

    gemm_accum(z, gru->input_weights, N, M, stride, input);

    gemm_accum(z, gru->recurrent_weights, N, N, stride, state);

-   for (i=0;i<N;i++)

-      z[i] = sigmoid_approx(z[i]);

+   compute_activation(z, z, N, ACTIVATION_SIGMOID);

    /* Compute reset gate. */

    for (i=0;i<N;i++)

@@ -154,8 +157,7 @@

       r[i] = gru->bias[N + i];

    gemm_accum(r, &gru->input_weights[N], N, M, stride, input);

    gemm_accum(r, &gru->recurrent_weights[N], N, N, stride, state);

-   for (i=0;i<N;i++)

-      r[i] = sigmoid_approx(r[i]);

+   compute_activation(r, r, N, ACTIVATION_SIGMOID);

    /* Compute output. */

    for (i=0;i<N;i++)

@@ -174,8 +176,9 @@

       gemm_accum(h, &gru->input_weights[2*N], N, M, stride, input);

       gemm_accum(h, &gru->recurrent_weights[2*N], N, N, stride, tmp);

    }

+   compute_activation(h, h, N, gru->activation);

    for (i=0;i<N;i++)

-      h[i] = z[i]*state[i] + (1-z[i])*tansig_approx(h[i]);

+      h[i] = z[i]*state[i] + (1-z[i])*h[i];

    for (i=0;i<N;i++)

       state[i] = h[i];

 }

--- a/dnn/nnet.h

+++ b/dnn/nnet.h

@@ -84,6 +84,8 @@

 void compute_dense(const DenseLayer *layer, float *output, const float *input);

+void compute_mdense(const MDenseLayer *layer, float *output, const float *input);

+

 void compute_gru(const GRULayer *gru, float *state, const float *input);

 void compute_conv1d(const Conv1DLayer *layer, float *output, float *mem, const float *input);

-- 

⑨