I have an autoencoder to regenerate input images. The image shape is (10, 1308608). 1308608 is 4*644*508.
class AutoEncoder(chainer.Chain):
def __init__(self, input_size, n_filters, n_units, filter_size,
activation):
self.activation = activation#{'relu': F.relu, 'sigmoid':
F.sigmoid}[activation]
self.n_filters = n_filters
self.n_units = n_units
self.filter_size = filter_size
self.dim1 = input_size - filter_size + 1
super(AutoEncoder, self).__init__(
conv1 = L.Convolution2D(1, n_filters, filter_size),
lenc1 = L.Linear(n_filters*self.dim1*self.dim1, n_units),
ldec1 = L.Linear(n_units, n_filters*self.dim1*self.dim1),
deconv1 = L.Deconvolution2D(n_filters, 1, filter_size)
)
def __call__(self, x):
h1 = self.activation(self.conv1(x))
h2 = F.dropout(self.activation(self.lenc1(h1)))
h3 = F.reshape(self.activation(self.ldec1(h2)), (x.data.shape[0],
self.n_filters, self.dim1, self.dim1))
h4 = self.activation(self.deconv1(h3))
return h4
class Regression(chainer.Chain):
def __init__(self, predictor):
super(Regression, self).__init__(predictor=predictor)
def __call__(self, x, t):
y = self.predictor(x)
self.loss = F.mean_squared_error(y, t)
report({'loss': self.loss}, self)
return self.loss
def dump(self, x):
return self.predictor(x, False)
For mnist of size (1,28,28) they have used input_size=28, n_filters = 10, n_units = 20, filter_size = 9. I want to understand how to calculate n_filters, n_units, filter_size according to the input_size.
You can refer Chainer official document for convolution_2d.
Output height can be calculated as
hO = (hI+2 * hP − hK) / sY+1
where
same applies for width.