Neural Style Transfer.rar_styletransfer. dll_tensorflow 风格迁移_tra

import tensorflow as tf from tensorflow.contrib import eager as tfe tf.enable_eager_execution() import matplotlib.pyplot as plt import matplotlib as mpl mpl.rcParams['figure.figsize'] = (10, 10) mpl.rcParams['axes.grid'] = False import os import numpy as np import time from PIL import Image def get_images(): img_dir = './nst' if not os.path.exists(img_dir): os.makedirs(img_dir) file_urls = ['https://upload.wikimedia.org/wikipedia/commons/d/d7/Green_Sea_Turtle_grazing_seagrass.jpg', 'https://upload.wikimedia.org/wikipedia/commons/0/0a/The_Great_Wave_off_Kanagawa.jpg', 'https://upload.wikimedia.org/wikipedia/commons/b/b4/Vassily_Kandinsky%2C_1913_-_Composition_7.jpg', 'https://upload.wikimedia.org/wikipedia/commons/0/00/Tuebingen_Neckarfront.jpg', 'https://upload.wikimedia.org/wikipedia/commons/6/68/Pillars_of_creation_2014_HST_WFC3-UVIS_full-res_denoised.jpg', 'https://upload.wikimedia.org/wikipedia/commons/thumb/e/ea/Van_Gogh_-_Starry_Night_-_Google_Art_Project.jpg/1024px-Van_Gogh_-_Starry_Night_-_Google_Art_Project.jpg'] filenames = ['Green_Sea_Turtle_grazing_seagrass.jpg', 'The_Great_Wave_off_Kanagawa.jpg', 'Vassily_Kandinsky%2C_1913_-_Composition_7.jpg', 'Tuebingen_Neckarfront.jpg', 'Pillars_of_creation_2014_HST_WFC3-UVIS_full-res_denoised.jpg', '1024px-Van_Gogh_-_Starry_Night_-_Google_Art_Project.jpg'] file_paths = [] for file, url in zip(filenames, file_urls): file_paths.append(tf.keras.utils.get_file(fname=file, cache_subdir=os.path.abspath(img_dir), origin=url)) return file_paths def load_img(file_path): image = Image.open(file_path) scale = 512 / max(image.size) image = image.resize((round(scale * image.size[0]), round(scale * image.size[1])), Image.ANTIALIAS) image = tf.keras.preprocessing.image.img_to_array(image) image = np.expand_dims(image, axis=0) return image def load_and_process_img(file_path): return tf.keras.applications.vgg19.preprocess_input(load_img(file_path)) def deprocess_img(image): if len(image.shape) == 4: image = np.squeeze(image, axis=0) if len(image.shape) == 3: image[:, :, 0] += 103.939 image[:, :, 1] += 116.779 image[:, :, 2] += 123.68 else: raise ValueError("Invalid input to deprocessing image") image=image[:,:,::-1] image = np.clip(image, 0, 255).astype('uint8') return image def get_model(): content_layers = ['block5_conv2'] style_layers = ['block1_conv1', 'block2_conv1', 'block3_conv1', 'block4_conv1', 'block5_conv1' ] num_content_layers = len(content_layers) num_style_layers = len(style_layers) # print(num_content_layers,num_style_layers) vgg = tf.keras.applications.vgg19.VGG19(include_top=False, weights='imagenet') vgg.trainable = False style_outputs = [vgg.get_layer(name).output for name in style_layers] content_outputs = [vgg.get_layer(name).output for name in content_layers] model_outputs = style_outputs + content_outputs return tf.keras.models.Model(vgg.input, model_outputs) def get_content_loss(base_content, target): return tf.reduce_mean(tf.square(base_content - target)) def gram_matrix(input_tensor): channels = int(input_tensor.shape[-1]) a = tf.reshape(input_tensor, shape=[-1, channels]) n = tf.shape(a)[0] gram = tf.matmul(a, a, transpose_a=True) return gram / tf.cast(n, tf.float32) def get_style_loss(base_style, gram_target): height, width, channel = base_style.get_shape().as_list() gram_style = gram_matrix(base_style) return tf.reduce_mean(tf.square(gram_style - gram_target)) def get_feature_representations(model, content_path, style_path): num_style_layers = 5 content_image = load_and_process_img(content_path) style_image = load_and_process_img(style_path) style_outputs = model(style_image) content_outputs = model(content_image) style_features = [style_layer[0] for style_layer in style_outputs[:num_style_layers]] content_features = [content_layer[0] for content_layer in content_outputs[num_style_layers:]] return style_features, content_features def compute_loss(model, loss_weights, init_image, gram_style_features, content_features): num_content_layers = 1 num_style_layers = 5 style_weights, content_weights = loss_weights model_outputs = model(init_image) style_output_features = model_outputs[:num_style_layers] content_output_features = model_outputs[num_style_layers:] style_score = 0 content_score = 0 for base_style, gram_target in zip(style_output_features, gram_style_features): style_score += get_style_loss(base_style[0], gram_target) for base_content, target in zip(content_output_features, content_features): content_score += get_content_loss(base_content[0], target) style_score /= num_style_layers content_score /= num_content_layers content_score *= content_weights style_score *= style_weights loss = content_score + style_score return loss, style_score, content_score def compute_gradient(cfg): with tf.GradientTape() as t: loss = compute_loss(**cfg) gradients = t.gradient(loss[0], cfg['init_image']) return gradients, loss def run_style_transfer(content_path, style_path, max_iters=1000, content_weights=1e3, style_weights=1e-2): model = get_model() model.trainable = False style_features, content_features = get_feature_representations(model, content_path, style_path) gram_style_features = [gram_matrix(style_feature) for style_feature in style_features] init_image = load_and_process_img(content_path) init_image = tfe.Variable(init_image, dtype=tf.float32) optimizer = tf.train.AdamOptimizer(learning_rate=5, beta1=0.99, epsilon=1e-1) best_loss, best_img = float('inf'), None loss_weights = [style_weights, content_weights] cfg = { 'model': model, 'loss_weights': loss_weights, 'init_image': init_image, 'gram_style_features': gram_style_features, 'content_features': content_features } rows = 2 cols = 5 display_interval = max_iters / rows / cols imgs = [] norm_mean = tf.constant([103.939, 116.779, 123.68], dtype=tf.float32) min_value = -norm_mean max_value = 255 - norm_mean global_start = time.time() start = time.time() for i in range(max_iters): gradients, loss = compute_gradient(cfg) optimizer.apply_gradients([(gradients, init_image)]) init_image.assign(tf.clip_by_value(init_image, min_value, max_value)) if loss[0] < best_loss: best_loss = loss[0] best_img = deprocess_img(init_image.numpy()) if (i + 1) % display_interval == 0: plot_img = deprocess_img(init_image.numpy()) imgs.append(plot_img) print('Iteration: {}'.format(i + 1)) print('Total loss: {:.4e}, ' 'style loss: {:.4e}, ' 'content loss: {:.4e}, ' 'time: {:.4f}s'.format(loss[0], loss[1], loss[2], time.time() - start)) start = time.time() print('Total time: {:.4f}s'.format(time.time() - global_start)) plt.figure(figsize=(14, 4)) for i, img in enumerate(imgs): plt.subplot(rows, cols, i + 1) plt.imshow(img) plt.axis('off') plt.show() return best_img, best_loss if __name__ == "__main__": image_paths = get_images() for i in range(len(image_paths)): for j in