def _reset_noise(self):
# noise for mixing
noise = self.gen_noise().reshape(shape= (self.noiseAll_size,) + (self.train.shape[1:]))
try: sigma = self.train.max(axis = 2) / float(self.SNR) / nd.array(noise[:self.train_size].asnumpy().std(axis = 2,dtype='float64'),ctx=ctx)
except: sigma = self.train.max(axis = -1) / float(self.SNR) / nd.array(noise[:self.train_size].asnumpy().std(axis = -1,dtype='float64'),ctx=ctx)
self.sigma = sigma
signal_train = nd.divide(self.train, sigma.reshape((self.train_size,self.num_channel,-1)))
data_train = signal_train + noise[:self.train_size]
try: sigma = self.test.max(axis = 2) / float(self.SNR) / nd.array(noise[-self.test_size:].asnumpy().std(axis = 2,dtype='float64'),ctx=ctx)
except: sigma = self.test.max(axis = -1) / float(self.SNR) / nd.array(noise[-self.test_size:].asnumpy().std(axis = -1,dtype='float64'),ctx=ctx)
signal_test = nd.divide(self.test, sigma.reshape((self.test_size,self.num_channel,-1)))
data_test = signal_test + noise[-self.test_size:]
# noise for pure conterpart
noise = self.gen_noise().reshape(shape= (self.noiseAll_size,) + (self.train.shape[1:]))
X_train = Normolise_nd(nd.concat(data_train, noise[:self.train_size], dim=0), self.num_channel)
try: dataset_train = gluon.data.ArrayDataset(X_train, self.y_train)
except: dataset_train = gluon.data.ArrayDataset(X_train, self.y_train)
self.train_data = gluon.data.DataLoader(dataset_train, self.batch_size, shuffle=True, last_batch='keep')
X_test = Normolise_nd(nd.concat(data_test, noise[-self.test_size:], dim=0), self.num_channel)
try: dataset_test = gluon.data.ArrayDataset(X_test, self.y_test)
except: dataset_test = gluon.data.ArrayDataset(X_test, self.y_test)
self.test_data = gluon.data.DataLoader(dataset_test, self.batch_size, shuffle=True, last_batch='keep')
def backward(self, req, out_grad, in_data, out_data, in_grad, aux):
pred = in_data[0]
ll = in_data[1]
out = nd.add(pred, ll)
out = nd.divide(ll, out)
# out = (ll/(pred+ll))**2
out = - nd.multiply(out, out)
self.assign(in_grad[0], req[0], out)
def preprocess_lab(lab):
l_chan, a_chan, b_chan = nd.split(lab, axis=2, num_outputs=3)
l_chan = nd.divide(l_chan, 50) - 1
a_chan = nd.divide(a_chan, 110)
b_chan = nd.divide(b_chan, 110)
return [l_chan, a_chan, b_chan]
import cv2
def forward(self, x):
norm = (x**2).sum(axis=1, keepdims=True).sqrt() + self.eps
x = F.divide(x, norm)
out = self.weight.data().expand_dims(axis=0).expand_dims(
axis=2).expand_dims(axis=3) * x
return out
import mxnet as mx
import cv2