support upsampling is possible?

[shawnlee103]

support upsampling is possible?
or
how to gen AE model with Convolution ?

[attractivechaos]

KANN doesn't support unpooling and deconvolution. Implementing unpooling may be easy, but implementing deconvolution efficiently is not trivial. I am occupied recently. I can fix confirmed bugs, but won't have time to implement deconvolution unfortunately. Sorry.

[shawnlee103]

could support unpooling only?

[shawnlee103]

I try to implement upSampling2d, but seems not work.
Any suggestion?
int kad_op_upSampling2d(kad_node_t *p, int action)
{
conv_conf_t aux = (conv_conf_t)p->ptr;
kad_node_t *q = p->child[0];
if (action == KAD_SYNC_DIM) {
if (q->n_d != 4) return -1;
p->n_d = 4;
p->d[0] = q->d[0], p->d[1] = q->d[1], p->d[2] = up_pooling_out_size(q->d[2], &aux[0]), p->d[3] = up_pooling_out_size(q->d[3], &aux[1]);
}
else if (action == KAD_ALLOC) {
p->gtmp = realloc(p->gtmp, kad_len(p) * sizeof(int));
}
else if (action == KAD_FORWARD) {
int rest = 1, len, t, i;
int f = (int)p->gtmp;
len = kad_len(p);
for (i = 0; i < len; ++i) p->x[i] = -FLT_MAX;
for (i = 0; i < p->n_d - 2; ++i) rest = p->d[i];
for (t = 0; t < rest; ++t) {
int q_i, j, k, l, q_row = q->d[p->n_d - 2], q_col = q->d[p->n_d - 1];
for (q_i = 0; q_i < q_row; ++q_i) {
int q_u = (t * q_row + q_i) * q_col;
for (k = 0; k < aux[0].kernel_size; ++k) {
int p_v0, p_v_end, p_i = q_i * aux[0].kernel_size + k - aux[0].pad[0];
if (p_i < 0 || p_i >= p->d[p->n_d - 2]) continue;
p_v0 = (t * p->d[p->n_d - 2] + p_i) * p->d[p->n_d - 1];
p_v_end = p_v0 + p->d[p->n_d - 1];
for (j = 0; j < q_col ; j++)
for (l = 0; l < aux[1].kernel_size; ++l) {
p->x[p_v0 + j * aux[1].kernel_size + l] = q->x[q_u + j];
f[p_v0 + j * aux[1].kernel_size + l] = q_u + j;
}
} / ~k /
} / ~i */
}
}
else if (action == KAD_BACKWARD) {
int i, len, f = (int)p->gtmp;
len = kad_len(p);
for (i = 0; i < len; ++i) {
q->g[f[i]] += p->g[i];
}
}
return 0;
}

and i want to try the AE model with mnist
static kann_t *model_gen(int n_in, int n_hidden, float i_dropout)
{
kad_node_t *t, *x;
x = kad_feed(4, 1, 1, 28, 28), x->ext_flag |= KANN_F_IN | KANN_F_TRUTH;
t = kad_relu(kann_layer_conv2d(x, 4, 3, 3, 1, 1, 2, 2)); // 3x3 kernel; 1x1 stride; 1x1 padding
t = kad_max2d(t, 2, 2, 2, 2, 0, 0); // 2x2 kernel; 2x2 stride; 0x0 padding
t = kad_relu(kann_layer_conv2d(t, 4, 3, 3, 1, 1, 2, 2)); // 3x3 kernel; 1x1 stride; 1x1 padding
t = kad_max2d(t, 2, 2, 2, 2, 0, 0); // 2x2 kernel; 2x2 stride; 0x0 padding

t = kad_relu(kann_layer_conv2d(t, 4, 3, 3, 1, 1, 2, 2)); // 3x3 kernel; 1x1 stride; 1x1 padding
t = kad_upSampling2d(t, 2, 2); // 2x2 kernel
t = kad_relu(kann_layer_conv2d(t, 4, 3, 3, 1, 1, 2, 2)); // 3x3 kernel; 1x1 stride; 1x1 padding
t = kad_upSampling2d(t, 2, 2); // 2x2 kernel
t = kad_sigm(kann_layer_conv2d(t, 1, 3, 3, 1, 1, 2, 2)), t->ext_flag = KANN_F_OUT; // 3x3 kernel; 1x1 stride; 1x1 padding
t = kad_mse(t, x), t->ext_flag = KANN_F_COST;
return kann_new(t, 0);

}