arch['flows'] = [LogisticFlowLayer]
arch['batch_norm'] = False
model_dict['z1y->z2'] = OrderedDict([('arch', arch)])
arch = OrderedDict()
arch['hidden'] = [20]
arch['gain'] = np.sqrt(2)
arch['nonlin'] = lasagne.nonlinearities.tanh
arch['num_output'] = num_features
arch['sample_layer'] = SampleLayer
arch['flows'] = []
arch['batch_norm'] = False
model_dict['yz2->_z1'] = OrderedDict([('arch', arch)])
# set result directory path
res_out = 'examples/results/crism/' + timeStamp().format("")
# construct the semi^2-supervised deep generative model
m = SSDGM(num_features,
num_output,
variational=True,
model_dict=model_dict,
prior_x=prior_x,
prior_y=prior_y,
prior_z2=prior_z2,
loss_x=L2,
loss_y=KL,
coeff_x=1,
coeff_y=1e-2,
coeff_x_dis=1,
coeff_y_dis=1,
arch['flows'] = [LogisticFlowLayer]
arch['batch_norm'] = False
model_dict['z1y->z2'] = OrderedDict([('arch', arch)])
arch = OrderedDict()
arch['hidden'] = [20]
arch['gain'] = np.sqrt(2)
arch['nonlin'] = lasagne.nonlinearities.tanh
arch['num_output'] = num_features
arch['sample_layer'] = SampleLayer
arch['flows'] = []
arch['batch_norm'] = False
model_dict['yz2->_z1'] = OrderedDict([('arch', arch)])
# set result directory path
res_out = 'examples/results/gridsearch/' + timeStamp().format("")
# construct the semi^2-supervised deep generative model
m = SSDGM(num_features,
num_output,
model_dict=model_dict,
variational=True,
prior_x=prior_x,
prior_y=prior_y,
prior_z2=prior_z2,
loss_x=L2,
loss_y=KL,
coeff_x=1,
coeff_y=1e-2,
coeff_x_dis=1,
coeff_y_dis=1e3,
arch['flows'] = [LogisticFlowLayer]
arch['batch_norm'] = False
model_dict['z1y->z2'] = OrderedDict([('arch',arch)])
arch = OrderedDict()
arch['hidden'] = [250,500]
arch['gain'] = np.sqrt(2)
arch['nonlin'] = lasagne.nonlinearities.tanh
arch['num_output'] = num_features
arch['sample_layer'] = BernoulliSampleLayer
arch['flows'] = []
arch['batch_norm'] = False
model_dict['yz2->_z1'] = OrderedDict([('arch',arch)])
# set result directory path
res_out='examples/results/mnist/'+timeStamp().format("")
# construct the semi^2-supervised deep generative model
m = SSDGM(num_features,num_output,variational=True,model_dict=model_dict,eq_samples=1,iw_samples=1,
prior_x=prior_x,prior_y=prior_y,prior_z2=prior_z2,loss_x=L2,loss_y=KL,
coeff_x=1e-1,coeff_y=1e-1,coeff_x_dis=1,coeff_y_dis=1e-2,coeff_x_prob=1e-1,coeff_y_prob=0,
num_epochs=1000,eval_freq=100,lr=1e-2,
batch_size_Xy_train=10000,batch_size_X__train=10000,batch_size__y_train=10000,
batch_size_Xy_eval=10000,batch_size_X__eval=10000,batch_size__y_eval=10000,
res_out=res_out)
# fit the model
m.fit(verbose=True,debug=True,**data)
# auto-set title and plot results (saved to res_out)
title = 'M2'
arch['flows'] = [LogisticFlowLayer]
arch['batch_norm'] = False
model_dict['z1y->z2'] = OrderedDict([('arch', arch)])
arch = OrderedDict()
arch['hidden'] = [50]
arch['gain'] = np.sqrt(2)
arch['nonlin'] = lasagne.nonlinearities.tanh
arch['num_output'] = num_features
arch['sample_layer'] = SampleLayer
arch['flows'] = [SoftplusFlowLayer]
arch['batch_norm'] = False
model_dict['yz2->_z1'] = OrderedDict([('arch', arch)])
# set result directory path
res_out = 'examples/results/raman/' + timeStamp().format("")
# construct the semi^2-supervised deep generative model
m = SSDGM(num_features,
num_output,
variational=True,
model_dict=model_dict,
prior_x=prior_x,
prior_y=prior_y,
prior_z2=prior_z2,
loss_x=L2,
loss_y=KL,
coeff_x=1e-2,
coeff_y=0,
coeff_x_dis=1,
coeff_y_dis=0,
arch['flows'] = [LogisticFlowLayer]
arch['batch_norm'] = False
model_dict['z1y->z2'] = OrderedDict([('arch', arch)])
arch = OrderedDict()
arch['hidden'] = [50]
arch['gain'] = np.sqrt(2)
arch['nonlin'] = lasagne.nonlinearities.tanh
arch['num_output'] = num_features
arch['sample_layer'] = SampleLayer
arch['flows'] = [SoftplusFlowLayer]
arch['batch_norm'] = False
model_dict['yz2->_z1'] = OrderedDict([('arch', arch)])
# set result directory path
res_out = 'examples/results/libs/' + timeStamp().format("")
# construct the semi^2-supervised deep generative model
m = SSDGM(num_features,
num_output,
model_dict=model_dict,
prior_x=prior_x,
prior_y=prior_y,
prior_z2=prior_z2,
loss_x=L2,
loss_y=KL,
coeff_x=1e-2,
coeff_y=1e-4,
coeff_x_dis=10,
coeff_y_dis=1e-4,
coeff_x_prob=0,