def imageThree(): INPUT = 'noise' input_depth = 1 num_iter = 3001 show_every = 50 figsize = 8 reg_noise_std = 0.00 param_noise = True if 'skip' in NET_TYPE: depth = int(NET_TYPE[-1]) net = skip(input_depth, image_np.shape[0], num_channels_down = [16, 32, 64, 128, 128, 128][:depth], num_channels_up = [16, 32, 64, 128, 128, 128][:depth], num_channels_skip = [0, 0, 0, 0, 0, 0][:depth], filter_size_up = 3,filter_size_down = 5, filter_skip_size=1, upsample_mode='nearest', # downsample_mode='avg', need1x1_up=False, need_sigmoid=True, need_bias=True, pad=pad, act_fun='LeakyReLU').type(dtype) LR = 0.01 elif NET_TYPE == 'UNET': net = UNet(num_input_channels=input_depth, num_output_channels=3, feature_scale=8, more_layers=1, concat_x=False, upsample_mode='deconv', pad='zero', norm_layer=torch.nn.InstanceNorm2d, need_sigmoid=True, need_bias=True) LR = 0.001 param_noise = False elif NET_TYPE == 'ResNet': net = ResNet(input_depth, image_np.shape[0], 8, 32, need_sigmoid=True, act_fun='LeakyReLU') LR = 0.001 param_noise = False else: assert False net = net.type(dtype) net_input = get_noise(input_depth, INPUT, image_np.shape[1:]).type(dtype) return net_input, net, LR, num_iter, param_noise, reg_noise_std, show_every, figsize
def imageFive(): INPUT = 'meshgrid' input_depth = 2 LR = 0.01 num_iter = 5001 param_noise = False show_every = 50 figsize = 5 reg_noise_std = 0.03 net = skip(input_depth, image_np.shape[0], num_channels_down = [128] * 5, num_channels_up = [128] * 5, num_channels_skip = [0] * 5, upsample_mode='nearest', filter_skip_size=1, filter_size_up=3, filter_size_down=3, need_sigmoid=True, need_bias=True, pad=pad, act_fun='LeakyReLU').type(dtype) net = net.type(dtype) net_input = get_noise(input_depth, INPUT, image_np.shape[1:]).type(dtype) return net_input, net, LR, num_iter, param_noise, reg_noise_std, show_every, figsize
INPUT = 'noise' input_depth = 32 OPTIMIZER = 'adam' OPT_OVER = 'net' # imagen en color num_iter = 1000 LR = 0.01 reg_noise_std = 0.00 # devuelve un modelo net = skip(input_depth, image_np.shape[0], num_channels_down = [16, 32, 64, 128, 128], num_channels_up = [16, 32, 64, 128, 128], num_channels_skip = [0, 0, 0, 0, 0], filter_size_down = 3, filter_size_up = 3, filter_skip_size=1, upsample_mode='bilinear', downsample_mode='avg', need_sigmoid=True, need_bias=True, pad=pad).type(dtype) # Perdida mse = torch.nn.MSELoss().type(dtype) image_var = np_to_torch(image_np).type(dtype) # tensor cuda # tensor cuda net_input = get_noise(input_depth, INPUT, image_np.shape[1:]).type(dtype).detach() # Datos del modelo print(torch_summarize(net))