train_mixgan.py 文件源码

def __init__(self, args):

        self.args = args

        rng = np.random.RandomState(self.args.seed) # fixed random seeds
        theano_rng = MRG_RandomStreams(rng.randint(2 ** 15))
        lasagne.random.set_rng(np.random.RandomState(rng.randint(2 ** 15)))
        data_rng = np.random.RandomState(self.args.seed_data)

        ''' specify pre-trained generator E '''
        self.enc_layers = [LL.InputLayer(shape=(None, 3, 32, 32), input_var=None)]
        enc_layer_conv1 = dnn.Conv2DDNNLayer(self.enc_layers[-1], 64, (5,5), pad=0, stride=1, W=Normal(0.01), nonlinearity=nn.relu)
        self.enc_layers.append(enc_layer_conv1)
        enc_layer_pool1 = LL.MaxPool2DLayer(self.enc_layers[-1], pool_size=(2, 2))
        self.enc_layers.append(enc_layer_pool1)
        enc_layer_conv2 = dnn.Conv2DDNNLayer(self.enc_layers[-1], 128, (5,5), pad=0, stride=1, W=Normal(0.01), nonlinearity=nn.relu)
        self.enc_layers.append(enc_layer_conv2)
        enc_layer_pool2 = LL.MaxPool2DLayer(self.enc_layers[-1], pool_size=(2, 2))
        self.enc_layers.append(enc_layer_pool2)
        self.enc_layer_fc3 = LL.DenseLayer(self.enc_layers[-1], num_units=256, nonlinearity=T.nnet.relu)
        self.enc_layers.append(self.enc_layer_fc3)
        self.enc_layer_fc4 = LL.DenseLayer(self.enc_layers[-1], num_units=10, nonlinearity=T.nnet.softmax)
        self.enc_layers.append(self.enc_layer_fc4)


        ''' load pretrained weights for encoder '''
        weights_toload = np.load('pretrained/encoder.npz')
        weights_list_toload = [weights_toload['arr_{}'.format(k)] for k in range(len(weights_toload.files))]
        LL.set_all_param_values(self.enc_layers[-1], weights_list_toload)


        ''' input tensor variables '''
        #self.G_weights
        #self.D_weights
        self.dummy_input = T.scalar()
        self.G_layers = []
        self.z = theano_rng.uniform(size=(self.args.batch_size, self.args.z0dim))
        self.x = T.tensor4()
        self.meanx = T.tensor3()
        self.Gen_x = T.tensor4() 
        self.D_layers = []
        self.D_layer_adv = [] 
        self.D_layer_z_recon = []
        self.gen_lr = T.scalar() # learning rate
        self.disc_lr = T.scalar() # learning rate
        self.y = T.ivector()
        self.y_1hot = T.matrix()
        self.Gen_x_list = []
        self.y_recon_list = []
        self.mincost = T.scalar()
        #self.enc_layer_fc3 = self.get_enc_layer_fc3()

        self.real_fc3 = LL.get_output(self.enc_layer_fc3, self.x, deterministic=True)