seq.add(link.Linear(500, 500)) seq.add(function.Activation("leaky_relu")) seq.add(link.Linear(500, 500, use_weightnorm=True)) seq.add(function.Activation("relu")) seq.add(link.BatchNormalization(500)) seq.add(link.Linear(500, 500)) seq.add(function.Activation("sigmoid")) seq.add(link.Linear(500, 500, use_weightnorm=True)) seq.add(function.Activation("softmax")) seq.add(link.BatchNormalization(500)) seq.add(link.Linear(500, 500)) seq.add(function.Activation("softplus")) seq.add(link.Linear(500, 500, use_weightnorm=True)) seq.add(function.Activation("tanh")) seq.add(link.Linear(500, 10)) seq.build() y = seq(x) print y.data.shape # Conv test x = np.random.normal(scale=1, size=(2, 3, 96, 96)).astype(np.float32) x = Variable(x) seq = Sequential(weight_initializer="GlorotNormal", weight_init_std=0.05) seq.add(link.Convolution2D(3, 64, ksize=4, stride=2, pad=0)) seq.add(link.BatchNormalization(64)) seq.add(function.Activation("relu")) seq.add(link.Convolution2D(64, 128, ksize=4, stride=2, pad=0)) seq.add(link.BatchNormalization(128)) seq.add(function.Activation("relu"))
# discriminator.add(BatchNormalization(1000)) discriminator.add(Linear(None, 500, use_weightnorm=config.use_weightnorm)) discriminator.add(gaussian_noise(std=0.5)) discriminator.add(Activation(config.nonlinearity)) # discriminator.add(BatchNormalization(500)) discriminator.add(Linear(None, 250, use_weightnorm=config.use_weightnorm)) discriminator.add(gaussian_noise(std=0.5)) discriminator.add(Activation(config.nonlinearity)) # discriminator.add(BatchNormalization(250)) if config.use_minibatch_discrimination: discriminator.add( MinibatchDiscrimination(None, num_kernels=50, ndim_kernel=5)) discriminator.add( Linear(None, config.ndim_output, use_weightnorm=config.use_weightnorm)) # no need to add softmax() here discriminator.build() params = { "config": config.to_dict(), "model": discriminator.to_dict(), } with open(discriminator_sequence_filename, "w") as f: json.dump(params, f, indent=4, sort_keys=True, separators=(',', ': ')) discriminator_params = params # specify generator generator_sequence_filename = args.model_dir + "/generator.json" if os.path.isfile(generator_sequence_filename):