public static Tensor Apply(Tensor input, Dictionary <string, Tensor> weights)
{
var layers = new Stack <Tensor>();
layers.Push(Tensor.Conv2D(
input,
weights["generator/encoder_1/conv2d/kernel"],
weights["generator/encoder_1/conv2d/bias"]
));
for (var i = 2; i <= 8; i++)
{
var scope = "generator/encoder_" + i;
layers.Push(
Tensor.BatchNorm(
Tensor.Conv2D(
Tensor.LeakyRelu(layers.Peek(), 0.2f),
weights[scope + "/conv2d/kernel"],
weights[scope + "/conv2d/bias"]
),
weights[scope + "/batch_normalization/gamma"],
weights[scope + "/batch_normalization/beta"]
)
);
}
var decoding = layers.Pop();
for (var i = 8; i >= 2; i--)
{
var scope = "generator/decoder_" + i;
if (i < 8)
{
decoding = Tensor.Concat(decoding, layers.Pop());
}
decoding =
Tensor.BatchNorm(
Tensor.Deconv2D(
Tensor.Relu(decoding),
weights[scope + "/conv2d_transpose/kernel"],
weights[scope + "/conv2d_transpose/bias"]
),
weights[scope + "/batch_normalization/gamma"],
weights[scope + "/batch_normalization/beta"]
);
}
return(Tensor.Tanh(
Tensor.Deconv2D(
Tensor.Relu(Tensor.Concat(decoding, layers.Pop())),
weights["generator/decoder_1/conv2d_transpose/kernel"],
weights["generator/decoder_1/conv2d_transpose/bias"]
)
));
}
Apply(Tensor input, Dictionary <string, Tensor> weights, Tensor output)
{
var layers = new Stack <Tensor>();
layers.Push(Tensor.Conv2D(
input,
weights["generator/encoder_1/conv2d/kernel"],
weights["generator/encoder_1/conv2d/bias"]
));
yield return(_conv2DCosts[0]);
for (var i = 2; i <= 8; i++)
{
var scope = "generator/encoder_" + i;
var kernel = weights[scope + "/conv2d/kernel"];
var bias = weights[scope + "/conv2d/bias"];
var beta = weights[scope + "/batch_normalization/beta"];
var gamma = weights[scope + "/batch_normalization/gamma"];
var rect = Tensor.LeakyRelu(layers.Peek(), 0.2f);
var conv = Tensor.Conv2D(rect, kernel, bias);
var norm = Tensor.BatchNorm(conv, gamma, beta);
layers.Push(norm);
rect.Dispose();
conv.Dispose();
yield return(_conv2DCosts[i - 1]);
}
var decoding = layers.Pop();
for (var i = 8; i >= 2; i--)
{
var scope = "generator/decoder_" + i;
var kernel = weights[scope + "/conv2d_transpose/kernel"];
var bias = weights[scope + "/conv2d_transpose/bias"];
var beta = weights[scope + "/batch_normalization/beta"];
var gamma = weights[scope + "/batch_normalization/gamma"];
if (i < 8)
{
var prev = decoding;
var skip = layers.Pop();
decoding = Tensor.Concat(prev, skip);
prev.Dispose();
skip.Dispose();
}
var rect = Tensor.Relu(decoding);
var conv = Tensor.Deconv2D(rect, kernel, bias);
var norm = Tensor.BatchNorm(conv, gamma, beta);
rect.Dispose();
conv.Dispose();
decoding.Dispose();
decoding = norm;
yield return(_deconv2DCosts[i - 1]);
}
{
var kernel = weights["generator/decoder_1/conv2d_transpose/kernel"];
var bias = weights["generator/decoder_1/conv2d_transpose/bias"];
var skip = layers.Pop();
var join = Tensor.Concat(decoding, skip);
var rect = Tensor.Relu(join);
var conv = Tensor.Deconv2D(rect, kernel, bias);
Tensor.Tanh(conv, output);
decoding.Dispose();
skip.Dispose();
join.Dispose();
rect.Dispose();
conv.Dispose();
yield return(_deconv2DCosts[0]);
}
}
