{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.10.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"none","dataSources":[{"sourceId":21755,"databundleVersionId":1475600,"sourceType":"competition"},{"sourceId":15523526,"sourceType":"kernelVersion"},{"sourceId":57786,"sourceType":"modelInstanceVersion","modelInstanceId":48458},{"sourceId":57798,"sourceType":"modelInstanceVersion","modelInstanceId":48464},{"sourceId":58577,"sourceType":"modelInstanceVersion","modelInstanceId":49072},{"sourceId":58748,"sourceType":"modelInstanceVersion","modelInstanceId":49189}],"dockerImageVersionId":30699,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"# This Python 3 environment comes with many helpful analytics libraries installed\n# It is defined by the kaggle/python Docker image: https://github.com/kaggle/docker-python\n# For example, here's several helpful packages to load\n\nimport numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\n\n# Input data files are available in the read-only \"../input/\" directory\n# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory\n\nimport os\nfor dirname, _, filenames in os.walk('/kaggle/input'):\n    for filename in filenames:\n        print(os.path.join(dirname, filename))\n\n# You can write up to 20GB to the current directory (/kaggle/working/) that gets preserved as output when you create a version using \"Save & Run All\" \n# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of the current session","metadata":{"_uuid":"64d6e400-4ce8-4269-ab17-338d03aebd2d","_cell_guid":"2cba9bb8-17ab-4632-923b-60f8743cbbef","collapsed":false,"jupyter":{"outputs_hidden":false},"execution":{"iopub.status.busy":"2024-05-31T00:28:23.809625Z","iopub.execute_input":"2024-05-31T00:28:23.809972Z","iopub.status.idle":"2024-05-31T00:28:29.595022Z","shell.execute_reply.started":"2024-05-31T00:28:23.809944Z","shell.execute_reply":"2024-05-31T00:28:29.59391Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"---","metadata":{}},{"cell_type":"markdown","source":"# Loading the datasets\n","metadata":{}},{"cell_type":"code","source":"import tensorflow as tf\nimport os\nimport cv2\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n# Define the image size\nIMG_SIZE = 256\nBATCH_SIZE = 1  # CycleGAN typically uses batch size of 1\n\n# Function to preprocess input image\ndef preprocess_image(image):\n    image = tf.image.resize(image, [IMG_SIZE, IMG_SIZE])\n    image = (image - 127.5) / 127.5  # Normalize to [-1, 1]\n    return image\n\n# Function to parse images from directory\ndef load_images(image_dir):\n    image_files = [os.path.join(image_dir, filename) for filename in os.listdir(image_dir)]\n    dataset = tf.data.Dataset.from_tensor_slices(image_files)\n    dataset = dataset.map(lambda x: preprocess_image(tf.image.decode_jpeg(tf.io.read_file(x), channels=3)))\n    dataset = dataset.shuffle(buffer_size=1024).batch(BATCH_SIZE)\n    return dataset\n\n# Function to parse TFRecord files\ndef parse_tfrecord_fn(example):\n    feature_description = {\n        'image': tf.io.FixedLenFeature([], tf.string),\n    }\n    example = tf.io.parse_single_example(example, feature_description)\n    image = tf.io.decode_jpeg(example['image'], channels=3)\n    image = preprocess_image(image)\n    return image\n\n# Load dataset from TFRecord files\ndef load_tfrecord_dataset(tfrecord_files):\n    raw_dataset = tf.data.TFRecordDataset(tfrecord_files)\n    parsed_dataset = raw_dataset.map(parse_tfrecord_fn)\n    dataset = parsed_dataset.shuffle(buffer_size=1024).batch(BATCH_SIZE)\n    return dataset\n\n# Paths to the directories\nreal_photos_dir = '/kaggle/input/gan-getting-started/photo_jpg/'\nmonet_tfrecord_dir = '/kaggle/input/gan-getting-started/monet_tfrec/'\n\n# List of TFRecord files with full paths\ntfrecord_files = [os.path.join(monet_tfrecord_dir, filename) for filename in os.listdir(monet_tfrecord_dir)]\n\n# Load the datasets\nreal_photos_dataset = load_images(real_photos_dir)\nmonet_paintings_dataset = load_tfrecord_dataset(tfrecord_files)\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:29.597111Z","iopub.execute_input":"2024-05-31T00:28:29.597911Z","iopub.status.idle":"2024-05-31T00:28:41.723885Z","shell.execute_reply.started":"2024-05-31T00:28:29.597876Z","shell.execute_reply":"2024-05-31T00:28:41.72291Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Building the first CycleGAN model\n","metadata":{}},{"cell_type":"code","source":"from tensorflow.keras import layers, Model\n\n# Define the generator\ndef build_generator():\n    inputs = layers.Input(shape=(IMG_SIZE, IMG_SIZE, 3))\n\n    down_stack = [\n        layers.Conv2D(64, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', use_bias=False),\n        layers.LeakyReLU(),\n        layers.Conv2D(128, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', use_bias=False),\n        layers.BatchNormalization(),\n        layers.LeakyReLU(),\n        layers.Conv2D(256, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', use_bias=False),\n        layers.BatchNormalization(),\n        layers.LeakyReLU()\n    ]\n\n    up_stack = [\n        layers.Conv2DTranspose(128, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', use_bias=False),\n        layers.BatchNormalization(),\n        layers.ReLU(),\n        layers.Conv2DTranspose(64, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', use_bias=False),\n        layers.BatchNormalization(),\n        layers.ReLU()\n    ]\n\n    last = layers.Conv2DTranspose(3, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', activation='tanh')\n\n    x = inputs\n    for down in down_stack:\n        x = down(x)\n    for up in up_stack:\n        x = up(x)\n    outputs = last(x)\n\n    return Model(inputs, outputs)\n\n# Define the discriminator\ndef build_discriminator():\n    inputs = layers.Input(shape=(IMG_SIZE, IMG_SIZE, 3))\n\n    x = layers.Conv2D(64, (4, 4), strides=2, padding='same', kernel_initializer='he_normal')(inputs)\n    x = layers.LeakyReLU()(x)\n    x = layers.Conv2D(128, (4, 4), strides=2, padding='same', kernel_initializer='he_normal')(x)\n    x = layers.BatchNormalization()(x)\n    x = layers.LeakyReLU()(x)\n    x = layers.Conv2D(256, (4, 4), strides=2, padding='same', kernel_initializer='he_normal')(x)\n    x = layers.BatchNormalization()(x)\n    x = layers.LeakyReLU()(x)\n    x = layers.Conv2D(512, (4, 4), strides=2, padding='same', kernel_initializer='he_normal')(x)\n    x = layers.BatchNormalization()(x)\n    x = layers.LeakyReLU()(x)\n    x = layers.Flatten()(x)\n    x = layers.Dense(1)(x)\n\n    return Model(inputs, x)\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:41.725023Z","iopub.execute_input":"2024-05-31T00:28:41.725285Z","iopub.status.idle":"2024-05-31T00:28:41.740084Z","shell.execute_reply.started":"2024-05-31T00:28:41.725262Z","shell.execute_reply":"2024-05-31T00:28:41.739206Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# CycleGAN Training\n","metadata":{}},{"cell_type":"code","source":"from IPython.display import clear_output\n\n# Loss function for both generator and discriminator\ncross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)\n\n# Discriminator loss\ndef discriminator_loss(real_output, fake_output):\n    real_loss = cross_entropy(tf.ones_like(real_output), real_output)\n    fake_loss = cross_entropy(tf.zeros_like(fake_output), fake_output)\n    total_loss = real_loss + fake_loss\n    return total_loss\n\n# Generator loss\ndef generator_loss(fake_output):\n    return cross_entropy(tf.ones_like(fake_output), fake_output)\n\n# Cycle consistency loss\ndef calc_cycle_loss(real_image, cycled_image):\n    loss = tf.reduce_mean(tf.abs(real_image - cycled_image))\n    return loss\n\n# Identity loss\ndef identity_loss(real_image, same_image):\n    loss = tf.reduce_mean(tf.abs(real_image - same_image))\n    return loss\n\n# Optimizers\ngenerator_g_optimizer = tf.keras.optimizers.Adam(2e-10, beta_1=0.5)\ngenerator_f_optimizer = tf.keras.optimizers.Adam(2e-10, beta_1=0.5)\ndiscriminator_x_optimizer = tf.keras.optimizers.Adam(2e-4, beta_1=0.5)\ndiscriminator_y_optimizer = tf.keras.optimizers.Adam(2e-4, beta_1=0.5)\n\n# Instantiate the models\ngenerator_g = build_generator()  # Real photo to Monet\ngenerator_f = build_generator()  # Monet to real photo\ndiscriminator_x = build_discriminator()  # Discriminator for Monet\ndiscriminator_y = build_discriminator()  # Discriminator for real photo\n\n@tf.function\ndef train_step(real_x, real_y):\n    with tf.GradientTape(persistent=True) as tape:\n        # Generator G translates X -> Y\n        fake_y = generator_g(real_x, training=True)\n        cycled_x = generator_f(fake_y, training=True)\n\n        # Generator F translates Y -> X\n        fake_x = generator_f(real_y, training=True)\n        cycled_y = generator_g(fake_x, training=True)\n\n        # Same images to identity loss\n        same_x = generator_f(real_x, training=True)\n        same_y = generator_g(real_y, training=True)\n\n        # Discriminator outputs\n        disc_real_x = discriminator_x(real_x, training=True)\n        disc_real_y = discriminator_y(real_y, training=True)\n        disc_fake_x = discriminator_x(fake_x, training=True)\n        disc_fake_y = discriminator_y(fake_y, training=True)\n\n        # Generator adversarial loss\n        gen_g_loss = generator_loss(disc_fake_y)\n        gen_f_loss = generator_loss(disc_fake_x)\n\n        # Cycle consistency loss\n        total_cycle_loss = calc_cycle_loss(real_x, cycled_x) + calc_cycle_loss(real_y, cycled_y)\n\n        # Total generator loss\n        total_gen_g_loss = gen_g_loss + total_cycle_loss + identity_loss(real_y, same_y)\n        total_gen_f_loss = gen_f_loss + total_cycle_loss + identity_loss(real_x, same_x)\n\n        # Discriminator loss\n        disc_x_loss = discriminator_loss(disc_real_x, disc_fake_x)\n        disc_y_loss = discriminator_loss(disc_real_y, disc_fake_y)\n\n    # Calculate the gradients\n    generator_g_gradients = tape.gradient(total_gen_g_loss, generator_g.trainable_variables)\n    generator_f_gradients = tape.gradient(total_gen_f_loss, generator_f.trainable_variables)\n    discriminator_x_gradients = tape.gradient(disc_x_loss, discriminator_x.trainable_variables)\n    discriminator_y_gradients = tape.gradient(disc_y_loss, discriminator_y.trainable_variables)\n\n    # Apply the gradients\n    generator_g_optimizer.apply_gradients(zip(generator_g_gradients, generator_g.trainable_variables))\n    generator_f_optimizer.apply_gradients(zip(generator_f_gradients, generator_f.trainable_variables))\n    discriminator_x_optimizer.apply_gradients(zip(discriminator_x_gradients, discriminator_x.trainable_variables))\n    discriminator_y_optimizer.apply_gradients(zip(discriminator_y_gradients, discriminator_y.trainable_variables))\n\ncheckpoint_dir = './checkpoints'\ncheckpoint_prefix = os.path.join(checkpoint_dir, \"ckpt\")\ncheckpoint = tf.train.Checkpoint(generator_g=generator_g,\n                                 generator_f=generator_f,\n                                 discriminator_x=discriminator_x,\n                                 discriminator_y=discriminator_y,\n                                 generator_g_optimizer=generator_g_optimizer,\n                                 generator_f_optimizer=generator_f_optimizer,\n                                 discriminator_x_optimizer=discriminator_x_optimizer,\n                                 discriminator_y_optimizer=discriminator_y_optimizer)\n\n# Training function\nimport time\ndef generate_images(model, test_input):\n    prediction = model(test_input, training=True)\n    plt.figure(figsize=(12, 12))\n\n    display_list = [test_input[0], prediction[0]]\n    title = ['Input Image', 'Generated Image']\n\n    for i in range(2):\n        plt.subplot(1, 2, i+1)\n        plt.title(title[i])\n        plt.imshow((display_list[i] * 127.5 + 127.5).numpy().astype(np.uint8))\n        plt.axis('off')\n    plt.show()\n\ndef train(real_dataset, monet_dataset, epochs):\n    for epoch in range(epochs):\n        start = time.time()\n\n        n = 0\n        for real_x, real_y in tf.data.Dataset.zip((real_dataset, monet_dataset)):\n            train_step(real_x, real_y)\n            if n % 10 == 0:\n                print('.', end='', flush=True)\n            n += 1\n\n        print ('Time for epoch {} is {} sec'.format(epoch + 1, time.time()-start))\n\n        # Produce images for the GIF as we go\n#         clear_output(wait=True)\n        generate_images(generator_g, sample_real_image)\n        # Save the model every 5 epochs\n        if (epoch + 1) % 20 == 0:\n            print('checkpoint_prefix')\n            checkpoint.save(file_prefix=checkpoint_prefix)\n\n# Sample images\nsample_real_image = next(iter(real_photos_dataset))\n\n\n# # Restore the latest checkpoint\n# latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)\n# if latest_checkpoint:\n#     checkpoint.restore('./checkpoints/ckpt-16')\n#     print(f'Restored from {latest_checkpoint}')\n# else:\n#     print('Starting training from scratch.')\n\n# Train the CycleGAN\nEPOCHS = 1\ntrain(real_photos_dataset, monet_paintings_dataset, EPOCHS)\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:41.741278Z","iopub.execute_input":"2024-05-31T00:28:41.741534Z","iopub.status.idle":"2024-05-31T00:28:41.773297Z","shell.execute_reply.started":"2024-05-31T00:28:41.741512Z","shell.execute_reply":"2024-05-31T00:28:41.772408Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Prepare last checkpoint for download","metadata":{}},{"cell_type":"markdown","source":"* After the 150 epochs the model seems to start overfitting:\n    - we can see that the generator is able to recreate the correct shape fo the image\n    - my approach would be to consider the generator at that state and try rfine the training from that point","metadata":{}},{"cell_type":"code","source":"# Restore the latest checkpoint\nlatest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)\nif latest_checkpoint:\n    checkpoint.restore(latest_checkpoint)\n    print(f'Restored from {latest_checkpoint}')\nelse:\n    print('Starting training from scratch.')","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:41.776183Z","iopub.execute_input":"2024-05-31T00:28:41.776756Z","iopub.status.idle":"2024-05-31T00:28:41.786486Z","shell.execute_reply.started":"2024-05-31T00:28:41.77673Z","shell.execute_reply":"2024-05-31T00:28:41.78568Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import shutil\n\n\nshutil.make_archive(\"/kaggle/working/new\", 'zip', latest_checkpoint)","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:41.787327Z","iopub.execute_input":"2024-05-31T00:28:41.78763Z","iopub.status.idle":"2024-05-31T00:28:41.796306Z","shell.execute_reply.started":"2024-05-31T00:28:41.787601Z","shell.execute_reply":"2024-05-31T00:28:41.795533Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Saving the models","metadata":{}},{"cell_type":"code","source":"Save the generator model\ngenerator_g.save('/kaggle/working/model_/generator_model_latest.h5')\ngenerator_f.save('/kaggle/working/model_/generator_model_inverse_latest.h5')\n\n\n# Save the discriminator model\ndiscriminator_x.save('/kaggle/working/model_/discriminator_model_latest.h5')\ndiscriminator_y.save('/kaggle/working/model_/discriminator_model_inverse_latest.h5')\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:41.797336Z","iopub.execute_input":"2024-05-31T00:28:41.797589Z","iopub.status.idle":"2024-05-31T00:28:41.80682Z","shell.execute_reply.started":"2024-05-31T00:28:41.797568Z","shell.execute_reply":"2024-05-31T00:28:41.806061Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import shutil\n\n\nshutil.make_archive(\"/kaggle/working/final\", 'zip', '/kaggle/working/model')","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:41.807886Z","iopub.execute_input":"2024-05-31T00:28:41.808246Z","iopub.status.idle":"2024-05-31T00:28:41.817086Z","shell.execute_reply.started":"2024-05-31T00:28:41.808217Z","shell.execute_reply":"2024-05-31T00:28:41.816351Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Visualization","metadata":{}},{"cell_type":"code","source":"def generate_images(model, test_input):\n    prediction = model(test_input, training=True)\n    plt.figure(figsize=(12, 12))\n\n    display_list = [test_input[0], prediction[0]]\n    title = ['Input Image', 'Generated Image']\n\n    for i in range(2):\n        plt.subplot(1, 2, i+1)\n        plt.title(title[i])\n        plt.imshow((display_list[i] * 127.5 + 127.5).numpy().astype(np.uint8))\n        plt.axis('off')\n    plt.show()\n    return  prediction ","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:41.818108Z","iopub.execute_input":"2024-05-31T00:28:41.81837Z","iopub.status.idle":"2024-05-31T00:28:41.825991Z","shell.execute_reply.started":"2024-05-31T00:28:41.818349Z","shell.execute_reply":"2024-05-31T00:28:41.825165Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Generate Monet-style paintings for the sample photos\nfor _ in range(5):\n    sample_real_image = next(iter(real_photos_dataset))\n\n#     clear_output(wait=True)\n\n    p = generate_images(generator_g, sample_real_image)\n    generate_images(generator_f, p)","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:41.827125Z","iopub.execute_input":"2024-05-31T00:28:41.827382Z","iopub.status.idle":"2024-05-31T00:28:41.836768Z","shell.execute_reply.started":"2024-05-31T00:28:41.827361Z","shell.execute_reply":"2024-05-31T00:28:41.835764Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# (p * 127.5 + 127.5).numpy().astype(np.uint8)","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:41.837823Z","iopub.execute_input":"2024-05-31T00:28:41.8381Z","iopub.status.idle":"2024-05-31T00:28:41.8491Z","shell.execute_reply.started":"2024-05-31T00:28:41.838077Z","shell.execute_reply":"2024-05-31T00:28:41.848211Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Load model","metadata":{}},{"cell_type":"code","source":"# # Load the saved generator model\n# loaded_generator = tf.keras.models.load_model('/kaggle/input/cyclegan/keras/enhancedcolors/1/generator_model_enhanced.h5')\n# loaded_generator = tf.keras.models.load_model('/kaggle/input/cyclegan/keras/enhancedcolors/1/generator_model_enhanced.h5')\n\n# # Load the saved discriminator model\n# loaded_discriminator = tf.keras.models.load_model('/kaggle/input/cyclegan/keras/1_first_train/1/discriminator_model.h5')\n# loaded_discriminator = tf.keras.models.load_model('/kaggle/input/cyclegan/keras/1_first_train/1/discriminator_model.h5')\n\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:41.850256Z","iopub.execute_input":"2024-05-31T00:28:41.850607Z","iopub.status.idle":"2024-05-31T00:28:42.645503Z","shell.execute_reply.started":"2024-05-31T00:28:41.850577Z","shell.execute_reply":"2024-05-31T00:28:42.644269Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# import os \n# len(os.listdir('/kaggle/input/gan-getting-started/photo_jpg'))\n# #creating a tmp file\n# os.mkdir(\"../images\")","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:42.647211Z","iopub.execute_input":"2024-05-31T00:28:42.647589Z","iopub.status.idle":"2024-05-31T00:28:42.651777Z","shell.execute_reply.started":"2024-05-31T00:28:42.647549Z","shell.execute_reply":"2024-05-31T00:28:42.650614Z"},"_kg_hide-output":true,"_kg_hide-input":true,"jupyter":{"source_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# !ls ../images ","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:42.657978Z","iopub.execute_input":"2024-05-31T00:28:42.658658Z","iopub.status.idle":"2024-05-31T00:28:42.662353Z","shell.execute_reply.started":"2024-05-31T00:28:42.658609Z","shell.execute_reply":"2024-05-31T00:28:42.661445Z"},"jupyter":{"source_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# import os \n# len(os.listdir('/kaggle/input/gan-getting-started/photo_jpg'))\n# #creating a tmp file\n# os.mkdir(\"../images\")\n# import zipfile\n\n# num_images_ = len(os.listdir('/kaggle/input/gan-getting-started/photo_jpg'))\n\n# # Define a function to generate and save images\n# def generate_and_save_images(model, dataset, num_images):\n#     # Generate Monet-style paintings for the specified number of images\n#     for num in range(num_images):\n#         # Get a batch of real images from the dataset\n#         sample_real_images = next(iter(dataset))\n#         # Generate Monet-style paintings for the batch of real images\n#         generated_images = model(sample_real_images, training=True)\n#         # Convert the pixel values to the range [0, 255] and cast to uint8\n#         generated_images = (generated_images.numpy() * 127.5 + 127.5).astype(np.uint8)\n#         # Save each generated image to the zip file\n#         for i, image in enumerate(generated_images):\n#             image_filename = f'../images/image_{num * num_images + i}.jpg'\n#             # Write the image to the zip file\n#             cv2.imwrite(image_filename, cv2.cvtColor(image, cv2.COLOR_RGB2BGR))\n            \n#         if num%500==0:\n#             print(f'We have preprocessed {num} images')\n\n# # Generate and save images\n# generate_and_save_images(loaded_generator, real_photos_dataset,num_images_)\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:42.663468Z","iopub.execute_input":"2024-05-31T00:28:42.66378Z","iopub.status.idle":"2024-05-31T00:28:42.67308Z","shell.execute_reply.started":"2024-05-31T00:28:42.663757Z","shell.execute_reply":"2024-05-31T00:28:42.672293Z"},"jupyter":{"source_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# import shutil\n\n\n# shutil.make_archive(\"/kaggle/working/images\", 'zip', \"/kaggle/images\")","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:42.674137Z","iopub.execute_input":"2024-05-31T00:28:42.674845Z","iopub.status.idle":"2024-05-31T00:28:42.685957Z","shell.execute_reply.started":"2024-05-31T00:28:42.674814Z","shell.execute_reply":"2024-05-31T00:28:42.685113Z"},"jupyter":{"source_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"---","metadata":{}},{"cell_type":"markdown","source":"# Second CycleGAN\n- Using the old Model as a pretrained And Creating a new arch for the generator while keeping the same discriminator","metadata":{}},{"cell_type":"markdown","source":"## New Arch","metadata":{}},{"cell_type":"code","source":"from tensorflow.keras import layers, Model\n\ndef build_generator_new():\n    inputs = layers.Input(shape=(IMG_SIZE, IMG_SIZE, 3))\n\n    down_stack = [\n        layers.Conv2D(64, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', use_bias=False),\n        layers.LeakyReLU(),\n        layers.Conv2D(128, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', use_bias=False),\n        layers.BatchNormalization(),\n        layers.LeakyReLU(),\n        layers.Conv2D(256, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', use_bias=False),\n        layers.BatchNormalization(),\n        layers.LeakyReLU(),\n        layers.Conv2D(512, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', use_bias=False),\n        layers.BatchNormalization(),\n        layers.LeakyReLU(),\n        layers.Conv2D(512, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', use_bias=False),\n        layers.BatchNormalization(),\n        layers.LeakyReLU()\n    ]\n\n    up_stack = [\n        layers.Conv2DTranspose(256, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', use_bias=False),\n        layers.BatchNormalization(),\n        layers.ReLU(),\n        layers.Conv2DTranspose(128, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', use_bias=False),\n        layers.BatchNormalization(),\n        layers.ReLU(),\n        layers.Conv2DTranspose(64, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', use_bias=False),\n        layers.BatchNormalization(),\n        layers.ReLU(),\n        layers.Conv2DTranspose(32, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', use_bias=False),\n        layers.BatchNormalization(),\n        layers.ReLU()\n    ]\n\n    last = layers.Conv2DTranspose(3, (4, 4), strides=2, padding='same', kernel_initializer='he_normal', activation='tanh')\n\n    x = inputs\n    for down in down_stack:\n        x = down(x)\n    for up in up_stack:\n        x = up(x)\n    outputs = last(x)\n\n    return Model(inputs, outputs)\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:42.688718Z","iopub.execute_input":"2024-05-31T00:28:42.688957Z","iopub.status.idle":"2024-05-31T00:28:42.704108Z","shell.execute_reply.started":"2024-05-31T00:28:42.688937Z","shell.execute_reply":"2024-05-31T00:28:42.70336Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def build_discriminator_new():\n    inputs = layers.Input(shape=(IMG_SIZE, IMG_SIZE, 3))\n    x = layers.Conv2D(64, (4, 4), strides=2, padding='same', kernel_initializer='he_normal')(inputs)\n\n    x = layers.Conv2D(64, (4, 4), strides=2, padding='same', kernel_initializer='he_normal')(inputs)\n    x = layers.LeakyReLU()(x)\n    x = layers.Conv2D(128, (4, 4), strides=2, padding='same', kernel_initializer='he_normal')(x)\n    x = layers.BatchNormalization()(x)\n    x = layers.LeakyReLU()(x)\n    x = layers.Conv2D(256, (4, 4), strides=2, padding='same', kernel_initializer='he_normal')(x)\n    x = layers.BatchNormalization()(x)\n    x = layers.LeakyReLU()(x)\n    x = layers.Conv2D(512, (4, 4), strides=2, padding='same', kernel_initializer='he_normal')(x)\n    x = layers.BatchNormalization()(x)\n    x = layers.LeakyReLU()(x)\n    x = layers.Conv2D(512, (4, 4), strides=2, padding='same', kernel_initializer='he_normal')(x)\n    x = layers.BatchNormalization()(x)\n    x = layers.LeakyReLU()(x)\n    x = layers.Flatten()(x)\n    x = layers.Dense(1)(x)\n\n    return Model(inputs, x)\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:42.705204Z","iopub.execute_input":"2024-05-31T00:28:42.705474Z","iopub.status.idle":"2024-05-31T00:28:42.718405Z","shell.execute_reply.started":"2024-05-31T00:28:42.705452Z","shell.execute_reply":"2024-05-31T00:28:42.717687Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Loading the previously trained CycleGAN models","metadata":{}},{"cell_type":"code","source":"# Load the saved generator model\nexisting_generator_g = tf.keras.models.load_model('/kaggle/input/cyclegan/keras/latest/1/generator_model_latest.h5')\nexisting_generator_f = tf.keras.models.load_model('/kaggle/input/cyclegan/keras/latest/1/generator_model_inverse_latest.h5')\n\n# Load the saved discriminator model\nexisting_discriminator_x = tf.keras.models.load_model('/kaggle/input/cyclegan/keras/latest/1/discriminator_model_latest.h5')\nexisting_discriminator_y = tf.keras.models.load_model('/kaggle/input/cyclegan/keras/latest/1/discriminator_model_inverse_latest.h5')\n\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:42.719288Z","iopub.execute_input":"2024-05-31T00:28:42.719601Z","iopub.status.idle":"2024-05-31T00:28:43.439297Z","shell.execute_reply.started":"2024-05-31T00:28:42.719578Z","shell.execute_reply":"2024-05-31T00:28:43.438535Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# # Define the existing models\n# existing_generator_g = generator_g\n# existing_generator_f = generator_f\n# existing_discriminator_x = discriminator_x\n# existing_discriminator_y = discriminator_y\n\n# Freeze the weights of the existing models\nfor model in [existing_generator_g, existing_generator_f]:\n    for layer in model.layers:\n        layer.trainable = False","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:43.440488Z","iopub.execute_input":"2024-05-31T00:28:43.440921Z","iopub.status.idle":"2024-05-31T00:28:43.44698Z","shell.execute_reply.started":"2024-05-31T00:28:43.440886Z","shell.execute_reply":"2024-05-31T00:28:43.44599Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Freeze the weights of the existing models\nfor model in [existing_discriminator_x, existing_discriminator_y]:\n    for layer in model.layers:\n        layer.trainable = False","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:43.448068Z","iopub.execute_input":"2024-05-31T00:28:43.448888Z","iopub.status.idle":"2024-05-31T00:28:43.458693Z","shell.execute_reply.started":"2024-05-31T00:28:43.448862Z","shell.execute_reply":"2024-05-31T00:28:43.457842Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Combining the the pretained generator with the new trainable architecture","metadata":{}},{"cell_type":"code","source":"second_generator_g = build_generator_new()\nsecond_generator_f = build_generator_new()\n\ncombined_generator_g = Model(existing_generator_g.input, [second_generator_g(existing_generator_g.output)])\ncombined_generator_f = Model(existing_generator_f.input, [second_generator_g(existing_generator_f.output)])\n\ncombined_generator_g, combined_generator_f","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:43.459676Z","iopub.execute_input":"2024-05-31T00:28:43.45993Z","iopub.status.idle":"2024-05-31T00:28:43.916892Z","shell.execute_reply.started":"2024-05-31T00:28:43.459909Z","shell.execute_reply":"2024-05-31T00:28:43.916038Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Training the new model","metadata":{}},{"cell_type":"code","source":"from IPython.display import clear_output\n\n# Loss function for both generator and discriminator\ncross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)\n\n# Discriminator loss\ndef discriminator_loss(real_output, fake_output):\n    real_loss = cross_entropy(tf.ones_like(real_output), real_output)\n    fake_loss = cross_entropy(tf.zeros_like(fake_output), fake_output)\n    total_loss = real_loss + fake_loss\n    return total_loss\n\n# Generator loss\ndef generator_loss(fake_output):\n    return cross_entropy(tf.ones_like(fake_output), fake_output)\n\n# Cycle consistency loss\ndef calc_cycle_loss(real_image, cycled_image):\n    loss = tf.reduce_mean(tf.abs(real_image - cycled_image))\n    return loss\n\n# Identity loss\ndef identity_loss(real_image, same_image):\n    loss = tf.reduce_mean(tf.abs(real_image - same_image))\n    return loss\n\n# Optimizers\ngenerator_g_optimizer = tf.keras.optimizers.Adam(2e-4, beta_1=0.5)\ngenerator_f_optimizer = tf.keras.optimizers.Adam(2e-4, beta_1=0.5)\ndiscriminator_x_optimizer = tf.keras.optimizers.Adam(2e-2, beta_1=0.5)\ndiscriminator_y_optimizer = tf.keras.optimizers.Adam(2e-2, beta_1=0.5)\n\n\n@tf.function\ndef train_step(real_x, real_y):\n    with tf.GradientTape(persistent=True) as tape:\n        # Generator G translates X -> Y\n        fake_y = combined_generator_g(real_x, training=True)\n        cycled_x = combined_generator_f(fake_y, training=True)\n\n        # Generator F translates Y -> X\n        fake_x = combined_generator_f(real_y, training=True)\n        cycled_y = combined_generator_g(fake_x, training=True)\n\n        # Same images to identity loss\n        same_x = combined_generator_f(real_x, training=True)\n        same_y = combined_generator_g(real_y, training=True)\n\n        # Discriminator outputs\n        disc_real_x = existing_discriminator_x(real_x, training=True)\n        disc_real_y = existing_discriminator_y(real_y, training=True)\n        disc_fake_x = existing_discriminator_x(fake_x, training=True)\n        disc_fake_y = existing_discriminator_y(fake_y, training=True)\n\n        # Generator adversarial loss\n        gen_g_loss = generator_loss(disc_fake_y)\n        gen_f_loss = generator_loss(disc_fake_x)\n\n        # Cycle consistency loss\n        total_cycle_loss = calc_cycle_loss(real_x, cycled_x) + calc_cycle_loss(real_y, cycled_y)\n\n        # Total generator loss\n        total_gen_g_loss = gen_g_loss + total_cycle_loss + identity_loss(real_y, same_y)\n        total_gen_f_loss = gen_f_loss + total_cycle_loss + identity_loss(real_x, same_x)\n\n        # Discriminator loss\n        disc_x_loss = discriminator_loss(disc_real_x, disc_fake_x)\n        disc_y_loss = discriminator_loss(disc_real_y, disc_fake_y)\n\n    # Calculate the gradients\n    generator_g_gradients = tape.gradient(total_gen_g_loss, combined_generator_g.trainable_variables)\n    generator_f_gradients = tape.gradient(total_gen_f_loss, combined_generator_f.trainable_variables)\n#     discriminator_x_gradients = tape.gradient(disc_x_loss, existing_discriminator_x.trainable_variables)\n#     discriminator_y_gradients = tape.gradient(disc_y_loss, existing_discriminator_y.trainable_variables)\n\n    # Apply the gradients\n    generator_g_optimizer.apply_gradients(zip(generator_g_gradients, combined_generator_g.trainable_variables))\n    generator_f_optimizer.apply_gradients(zip(generator_f_gradients, combined_generator_f.trainable_variables))\n#     discriminator_x_optimizer.apply_gradients(zip(discriminator_x_gradients, existing_discriminator_x.trainable_variables))\n#     discriminator_y_optimizer.apply_gradients(zip(discriminator_y_gradients, existing_discriminator_y.trainable_variables))\n\ncheckpoint_dir = './checkpoints_combined'\ncheckpoint_prefix = os.path.join(checkpoint_dir, \"ckpt\")\ncheckpoint = tf.train.Checkpoint(generator_g=combined_generator_g,\n                                 generator_f=combined_generator_f,\n                                 discriminator_x=existing_discriminator_x,\n                                 discriminator_y=existing_discriminator_y,\n                                 generator_g_optimizer=generator_g_optimizer,\n                                 generator_f_optimizer=generator_f_optimizer,\n                                 discriminator_x_optimizer=discriminator_x_optimizer,\n                                 discriminator_y_optimizer=discriminator_y_optimizer)\n\n# Training function\nimport time\ndef generate_images(model, test_input):\n    prediction = model(test_input, training=True)\n    plt.figure(figsize=(12, 12))\n\n    display_list = [test_input[0], prediction[0]]\n    title = ['Input Image', 'Generated Image']\n\n    for i in range(2):\n        plt.subplot(1, 2, i+1)\n        plt.title(title[i])\n        plt.imshow((display_list[i] * 127.5 + 127.5).numpy().astype(np.uint8))\n        plt.axis('off')\n    plt.show()\n\ndef train(real_dataset, monet_dataset, epochs):\n    for epoch in range(epochs):\n        start = time.time()\n\n        n = 0\n        for real_x, real_y in tf.data.Dataset.zip((real_dataset, monet_dataset)):\n            train_step(real_x, real_y)\n            if n % 10 == 0:\n                print('.', end='', flush=True)\n            n += 1\n\n        print ('Time for epoch {} is {} sec'.format(epoch + 1, time.time()-start))\n\n        # Produce images for the GIF as we go\n#         clear_output(wait=True)\n        generate_images(combined_generator_g, sample_real_image)\n        # Save the model every 5 epochs\n        if (epoch + 1) % 20 == 0:\n            print('checkpoint_prefix')\n            checkpoint.save(file_prefix=checkpoint_prefix)\n\n# Sample images\nsample_real_image = next(iter(real_photos_dataset))\n\n\n# # Restore the latest checkpoint\n# latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)\n# if latest_checkpoint:\n#     checkpoint.restore(latest_checkpoint)\n#     print(f'Restored from {latest_checkpoint}')\n# else:\n#     print('Starting training from scratch.')\n\n# Train the CycleGAN\nEPOCHS = 200\ntrain(real_photos_dataset, monet_paintings_dataset, EPOCHS)\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:28:43.918456Z","iopub.execute_input":"2024-05-31T00:28:43.918752Z","iopub.status.idle":"2024-05-31T00:29:37.68046Z","shell.execute_reply.started":"2024-05-31T00:28:43.918727Z","shell.execute_reply":"2024-05-31T00:29:37.679554Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### Visualize","metadata":{}},{"cell_type":"code","source":"def generate_images(model, test_input):\n    prediction = model(test_input, training=True)\n    plt.figure(figsize=(12, 12))\n\n    display_list = [test_input[0], prediction[0]]\n    title = ['Input Image', 'Generated Image']\n\n    for i in range(2):\n        plt.subplot(1, 2, i+1)\n        plt.title(title[i])\n        plt.imshow((display_list[i] * 127.5 + 127.5).numpy().astype(np.uint8))\n        plt.axis('off')\n    plt.show()\n    return  prediction ","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:29:37.681664Z","iopub.execute_input":"2024-05-31T00:29:37.681981Z","iopub.status.idle":"2024-05-31T00:29:37.689958Z","shell.execute_reply.started":"2024-05-31T00:29:37.681952Z","shell.execute_reply":"2024-05-31T00:29:37.68905Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Generate Monet-style paintings for the sample photos\nfor _ in range(5):\n    sample_real_image = next(iter(real_photos_dataset))\n\n#     clear_output(wait=True)\n\n    p = generate_images(combined_generator_g, sample_real_image)\n    generate_images(combined_generator_f, p)","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:29:37.691141Z","iopub.execute_input":"2024-05-31T00:29:37.691423Z","iopub.status.idle":"2024-05-31T00:29:45.04288Z","shell.execute_reply.started":"2024-05-31T00:29:37.691399Z","shell.execute_reply":"2024-05-31T00:29:45.041862Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Saving the generations and preparing them for submission","metadata":{}},{"cell_type":"code","source":"import os \nlen(os.listdir('/kaggle/input/gan-getting-started/photo_jpg'))\n#creating a tmp file\nos.mkdir(\"../images\")\nimport zipfile\n\nnum_images_ = len(os.listdir('/kaggle/input/gan-getting-started/photo_jpg'))\n\n# Define a function to generate and save images\ndef generate_and_save_images(model, dataset, num_images):\n    # Generate Monet-style paintings for the specified number of images\n    for num in range(num_images):\n        # Get a batch of real images from the dataset\n        sample_real_images = next(iter(dataset))\n        # Generate Monet-style paintings for the batch of real images\n        generated_images = model(sample_real_images, training=True)\n        # Convert the pixel values to the range [0, 255] and cast to uint8\n        generated_images = (generated_images.numpy() * 127.5 + 127.5).astype(np.uint8)\n        # Save each generated image to the zip file\n        for i, image in enumerate(generated_images):\n            image_filename = f'../images/image_{num * num_images + i}.jpg'\n            # Write the image to the zip file\n            cv2.imwrite(image_filename, cv2.cvtColor(image, cv2.COLOR_RGB2BGR))\n            \n        if num%500==0:\n            print(f'We have preprocessed {num} images')\n\n# Generate and save images\ngenerate_and_save_images(combined_generator_g, real_photos_dataset,num_images_)","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:29:45.044229Z","iopub.execute_input":"2024-05-31T00:29:45.044945Z","iopub.status.idle":"2024-05-31T00:30:14.69421Z","shell.execute_reply.started":"2024-05-31T00:29:45.04491Z","shell.execute_reply":"2024-05-31T00:30:14.692929Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import shutil\n\n\nshutil.make_archive(\"/kaggle/working/images\", 'zip', \"/kaggle/images\")","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:30:14.695118Z","iopub.status.idle":"2024-05-31T00:30:14.695483Z","shell.execute_reply.started":"2024-05-31T00:30:14.695312Z","shell.execute_reply":"2024-05-31T00:30:14.695327Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Feel free to explore the other approachs below","metadata":{}},{"cell_type":"markdown","source":"## Including The MIFID loss/metric","metadata":{}},{"cell_type":"code","source":"# import os\n# import zipfile\n# import numpy as np\n# import cv2\n# import tensorflow as tf\n# import matplotlib.pyplot as plt\n# from tensorflow.keras.applications.inception_v3 import InceptionV3, preprocess_input\n# from scipy.linalg import sqrtm\n# from sklearn.metrics.pairwise import cosine_distances\n# from tqdm import tqdm\n# from PIL import Image\n# from IPython.display import clear_output\n\n# # Model Parameters\n# model_params = {\n#     'Inception': {\n#         'name': 'Inception',\n#         'imsize': 256,\n#         'output_layer': 'mixed10',\n#         'input_layer': 'input_1',\n#         'output_shape': 2048,\n#         'cosine_distance_eps': 0.1\n#     }\n# }\n\n# # Load the InceptionV3 model for feature extraction\n# inception_model = InceptionV3(include_top=False, pooling='avg', input_shape=(256, 256, 3))\n\n# # Define MiFID calculation functions\n# def calculate_fid(real_images, generated_images):\n#     real_images_resized = tf.image.resize(real_images, (256, 256))\n#     generated_images_resized = tf.image.resize(generated_images, (256, 256))\n\n#     real_images_preprocessed = preprocess_input(real_images_resized.numpy())\n#     generated_images_preprocessed = preprocess_input(generated_images_resized.numpy())\n\n#     real_features = inception_model.predict(real_images_preprocessed)\n#     generated_features = inception_model.predict(generated_images_preprocessed)\n\n#     mu_real, sigma_real = np.mean(real_features, axis=0), np.cov(real_features, rowvar=False)\n#     mu_generated, sigma_generated = np.mean(generated_features, axis=0), np.cov(generated_features, rowvar=False)\n\n#     diff = mu_real - mu_generated\n#     covmean, _ = sqrtm(sigma_real.dot(sigma_generated), disp=False)\n#     if np.iscomplexobj(covmean):\n#         covmean = covmean.real\n\n#     fid = np.sum(diff**2) + np.trace(sigma_real + sigma_generated - 2 * covmean)\n#     return fid\n\n# def calculate_mifid(real_images, generated_images):\n#     fid = calculate_fid(real_images, generated_images)\n\n#     real_features = inception_model.predict(preprocess_input(tf.image.resize(real_images, (256, 256)).numpy()))\n#     generated_features = inception_model.predict(preprocess_input(tf.image.resize(generated_images, (256, 256)).numpy()))\n\n#     min_cosine_distances = []\n#     for gen_feature in generated_features:\n#         distances = cosine_distances([gen_feature], real_features)\n#         min_distance = np.min(distances)\n#         min_cosine_distances.append(min_distance)\n\n#     memorization_distance = np.mean(min_cosine_distances)\n#     if memorization_distance < model_params['Inception']['cosine_distance_eps']:\n#         memorization_distance = model_params['Inception']['cosine_distance_eps']\n\n#     mifid = fid * (1 / memorization_distance)\n#     return mifid\n\n# # Loss function for both generator and discriminator\n# cross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)\n\n# # Discriminator loss\n# def discriminator_loss(real_output, fake_output):\n#     real_loss = cross_entropy(tf.ones_like(real_output), real_output)\n#     fake_loss = cross_entropy(tf.zeros_like(fake_output), fake_output)\n#     total_loss = real_loss + fake_loss\n#     return total_loss\n\n# # Generator loss\n# def generator_loss(fake_output):\n#     return cross_entropy(tf.ones_like(fake_output), fake_output)\n\n# # Cycle consistency loss\n# def calc_cycle_loss(real_image, cycled_image):\n#     loss = tf.reduce_mean(tf.abs(real_image - cycled_image))\n#     return loss\n\n# # Identity loss\n# def identity_loss(real_image, same_image):\n#     loss = tf.reduce_mean(tf.abs(real_image - same_image))\n#     return loss\n\n# # Optimizers\n# generator_g_optimizer = tf.keras.optimizers.Adam(2e-4, beta_1=0.5)\n# generator_f_optimizer = tf.keras.optimizers.Adam(2e-4, beta_1=0.5)\n# discriminator_x_optimizer = tf.keras.optimizers.Adam(2e-4, beta_1=0.5)\n# discriminator_y_optimizer = tf.keras.optimizers.Adam(2e-4, beta_1=0.5)\n\n# # Instantiate the models\n# generator_g = build_generator()  # Real photo to Monet\n# generator_f = build_generator()  # Monet to real photo\n# discriminator_x = build_discriminator()  # Discriminator for Monet\n# discriminator_y = build_discriminator()  # Discriminator for real photo\n\n# @tf.function\n# def train_step(real_x, real_y, epoch, total_epochs, initial_noise=0.1, final_noise=0.01):\n#     with tf.GradientTape(persistent=True) as tape:\n#         #adding noise progressivly\n#         def get_noise_schedule(epoch, total_epochs, initial_noise, final_noise):\n#             return initial_noise + (final_noise - initial_noise) * ((total_epochs - epoch) / total_epochs)\n        \n#         noise = get_noise_schedule(epoch, total_epochs, initial_noise, final_noise)\n\n#         # Add noise to real images\n#         real_x = real_x + noise * tf.random.normal(real_x.shape)\n#         fake_y = real_y + noise * tf.random.normal(real_y.shape)\n        \n#         # Generator G translates X -> Y\n#         fake_y = generator_g(real_x, training=True)\n#         cycled_x = generator_f(fake_y, training=True)\n\n#         # Generator F translates Y -> X\n#         fake_x = generator_f(real_y, training=True)\n#         cycled_y = generator_g(fake_x, training=True)\n\n#         # Same images to identity loss\n#         same_x = generator_f(real_x, training=True)\n#         same_y = generator_g(real_y, training=True)\n\n#         # Discriminator outputs\n#         disc_real_x = discriminator_x(real_x, training=True)\n#         disc_real_y = discriminator_y(real_y, training=True)\n#         disc_fake_x = discriminator_x(fake_x, training=True)\n#         disc_fake_y = discriminator_y(fake_y, training=True)\n\n#         # Generator adversarial loss\n#         gen_g_loss = generator_loss(disc_fake_y)\n#         gen_f_loss = generator_loss(disc_fake_x)\n\n#         # Cycle consistency loss\n#         total_cycle_loss = calc_cycle_loss(real_x, cycled_x) + calc_cycle_loss(real_y, cycled_y)\n\n#         # Total generator loss\n#         total_gen_g_loss = gen_g_loss + total_cycle_loss + identity_loss(real_y, same_y)\n#         total_gen_f_loss = gen_f_loss + total_cycle_loss + identity_loss(real_x, same_x)\n\n#         # Discriminator loss\n#         disc_x_loss = discriminator_loss(disc_real_x, disc_fake_x)\n#         disc_y_loss = discriminator_loss(disc_real_y, disc_fake_y)\n\n#     # Calculate the gradients\n#     generator_g_gradients = tape.gradient(total_gen_g_loss, generator_g.trainable_variables)\n#     generator_f_gradients = tape.gradient(total_gen_f_loss, generator_f.trainable_variables)\n#     discriminator_x_gradients = tape.gradient(disc_x_loss, discriminator_x.trainable_variables)\n#     discriminator_y_gradients = tape.gradient(disc_y_loss, discriminator_y.trainable_variables)\n\n#     # Apply the gradients\n#     generator_g_optimizer.apply_gradients(zip(generator_g_gradients, generator_g.trainable_variables))\n#     generator_f_optimizer.apply_gradients(zip(generator_f_gradients, generator_f.trainable_variables))\n#     discriminator_x_optimizer.apply_gradients(zip(discriminator_x_gradients, discriminator_x.trainable_variables))\n#     discriminator_y_optimizer.apply_gradients(zip(discriminator_y_gradients, discriminator_y.trainable_variables))\n\n# # Set up checkpointing\n# checkpoint_dir = './checkpoints_noisy'\n# checkpoint_prefix = os.path.join(checkpoint_dir, \"ckpt\")\n# checkpoint = tf.train.Checkpoint(generator_g=generator_g,\n#                                  generator_f=generator_f,\n#                                  discriminator_x=discriminator_x,\n#                                  discriminator_y=discriminator_y,\n#                                  generator_g_optimizer=generator_g_optimizer,\n#                                  generator_f_optimizer=generator_f_optimizer,\n#                                  discriminator_x_optimizer=discriminator_x_optimizer,\n#                                  discriminator_y_optimizer=discriminator_y_optimizer)\n# # # Restore the latest checkpoint\n# # latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)\n# # if latest_checkpoint:\n# #     checkpoint.restore(latest_checkpoint)\n# #     print(f'Restored from {latest_checkpoint}')\n# # else:\n# #     print('Starting training from scratch.')\n    \n# # Training function\n# import time\n\n# def generate_images(model, test_input):\n#     prediction = model(test_input, training=True)\n#     plt.figure(figsize=(12, 12))\n\n#     display_list = [test_input[0], prediction[0]]\n#     title = ['Input Image', 'Generated Image']\n\n#     for i in range(2):\n#         plt.subplot(1, 2, i + 1)\n#         plt.title(title[i])\n#         plt.imshow((display_list[i] * 127.5 + 127.5).numpy().astype(np.uint8))\n#         plt.axis('off')\n#     plt.show()\n\n# def train(real_dataset, monet_dataset, epochs):\n#     for epoch in range(epochs):\n#         start = time.time()\n\n#         n = 0\n#         for real_x, real_y in tf.data.Dataset.zip((real_dataset, monet_dataset)):\n#             train_step(real_x, real_y, epoch, total_epochs=epochs)\n#             if n % 10 == 0:\n#                 print('.', end='', flush=True)\n#             n += 1\n\n#         print('Time for epoch {} is {} sec'.format(epoch + 1, time.time() - start))\n\n#         # Produce images for the GIF as we go\n# #         clear_output(wait=True)\n#         generate_images(generator_g, sample_real_image)\n\n#         # Calculate MiFID\n#         real_images = []\n#         generated_images = []\n        \n#         for real_x, real_y in tf.data.Dataset.zip((real_dataset, monet_dataset)).take(100):\n#             real_images.append(real_x)\n#             generated_images.append(generator_g(real_x, training=False))\n        \n#         real_images = tf.concat(real_images, axis=0)\n#         generated_images = tf.concat(generated_images, axis=0)\n        \n#         mifid = calculate_mifid(real_images, generated_images)\n#         print('Epoch:', epoch + 1, 'MiFID:', mifid)\n\n#         # Save the model every 5 epochs\n#         if (epoch + 1) % 10 == 0:\n#             checkpoint.save(file_prefix=checkpoint_prefix)\n\n# # Load datasets and start training\n# BUFFER_SIZE = 1000\n# BATCH_SIZE = 1\n# EPOCHS = 100\n\n# sample_real_image = next(iter(real_photos_dataset.take(1)))\n\n# train(real_photos_dataset, monet_paintings_dataset, epochs=EPOCHS)\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:30:14.697606Z","iopub.status.idle":"2024-05-31T00:30:14.698008Z","shell.execute_reply.started":"2024-05-31T00:30:14.697825Z","shell.execute_reply":"2024-05-31T00:30:14.697841Z"},"jupyter":{"source_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# # Restore the latest checkpoint\n# latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)\n# if latest_checkpoint:\n#     checkpoint.restore(latest_checkpoint)\n#     print(f'Restored from {latest_checkpoint}')\n# else:\n#     print('Starting training from scratch.')\n# Restored from ./checkpoints/ckpt-35\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:30:14.699078Z","iopub.status.idle":"2024-05-31T00:30:14.699393Z","shell.execute_reply.started":"2024-05-31T00:30:14.699239Z","shell.execute_reply":"2024-05-31T00:30:14.699252Z"},"jupyter":{"source_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Possible noise tuning additions","metadata":{}},{"cell_type":"code","source":"def get_noise_schedule(epoch, total_epochs, initial_noise, final_noise):\n    return initial_noise + (final_noise - initial_noise) * (epoch / total_epochs)\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:30:14.700732Z","iopub.status.idle":"2024-05-31T00:30:14.701054Z","shell.execute_reply.started":"2024-05-31T00:30:14.700894Z","shell.execute_reply":"2024-05-31T00:30:14.700907Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def generate_multiscale_noise(image, scales, noise_level):\n    noises = []\n    for scale in scales:\n        scale = tf.cast(scale, tf.int32)  # Ensure scale is an integer\n        noise = tf.random.normal(shape=tf.shape(image) * scale, mean=0.0, stddev=noise_level)\n        noises.append(tf.image.resize(noise, size=tf.shape(image)[1:3]))\n    return tf.add_n(noises)\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:30:14.702419Z","iopub.status.idle":"2024-05-31T00:30:14.702755Z","shell.execute_reply.started":"2024-05-31T00:30:14.70257Z","shell.execute_reply":"2024-05-31T00:30:14.702582Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class NoisyGenerator(tf.keras.Model):\n    def __init__(self, generator, noise_level):\n        super(NoisyGenerator, self).__init__()\n        self.generator = generator\n        self.noise_level = noise_level\n\n    def call(self, inputs, training=True):\n        if training:\n            inputs = inputs + generate_multiscale_noise(inputs, noise_level=self.noise_level)\n            x = self.generator(inputs, training=training)\n            x = x + generate_multiscale_noise(x, noise_level=self.noise_level)\n        else:\n            x = self.generator(inputs, training=training)\n        return x\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:30:14.704247Z","iopub.status.idle":"2024-05-31T00:30:14.704554Z","shell.execute_reply.started":"2024-05-31T00:30:14.704401Z","shell.execute_reply":"2024-05-31T00:30:14.704414Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def adjust_noise_level(current_noise_level, diversity_measure, target_diversity, max_noise, min_noise, adjustment_rate=0.05):\n    if diversity_measure < target_diversity:\n        current_noise_level = min(max_noise, current_noise_level + adjustment_rate)\n    else:\n        current_noise_level = max(min_noise, current_noise_level - adjustment_rate)\n    return current_noise_level\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:30:14.706309Z","iopub.status.idle":"2024-05-31T00:30:14.706615Z","shell.execute_reply.started":"2024-05-31T00:30:14.706462Z","shell.execute_reply":"2024-05-31T00:30:14.706475Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Straight GAN","metadata":{}},{"cell_type":"markdown","source":"* Logic:\n    1. I want my generator to become good at converting Real images to Monet paintings\n    2. I want my discriminator to become good at detecting Monet paintings and distinguish them from fake ones.\n    3. We will end up with a competitive model that outputs genreations that discriminator isnt able to distinguish from.\n","metadata":{}},{"cell_type":"code","source":"# import os\n# import zipfile\n# import numpy as np\n# import cv2\n# import tensorflow as tf\n# import matplotlib.pyplot as plt\n# from tensorflow.keras.applications.inception_v3 import InceptionV3, preprocess_input\n# from scipy.linalg import sqrtm\n# from sklearn.metrics.pairwise import cosine_distances\n# from tqdm import tqdm\n# from PIL import Image\n# from IPython.display import clear_output\n\n# # Model Parameters\n# model_params = {\n#     'Inception': {\n#         'name': 'Inception',\n#         'imsize': 256,\n#         'output_layer': 'mixed10',\n#         'input_layer': 'input_1',\n#         'output_shape': 2048,\n#         'cosine_distance_eps': 0.1\n#     }\n# }\n\n# # Load the InceptionV3 model for feature extraction\n# inception_model = InceptionV3(include_top=False, pooling='avg', input_shape=(256, 256, 3))\n\n# # Define MiFID calculation functions\n# def calculate_fid(real_images, generated_images):\n#     real_images_resized = tf.image.resize(real_images, (256, 256))\n#     generated_images_resized = tf.image.resize(generated_images, (256, 256))\n\n#     real_images_preprocessed = preprocess_input(real_images_resized.numpy())\n#     generated_images_preprocessed = preprocess_input(generated_images_resized.numpy())\n\n#     real_features = inception_model.predict(real_images_preprocessed)\n#     generated_features = inception_model.predict(generated_images_preprocessed)\n\n#     mu_real, sigma_real = np.mean(real_features, axis=0), np.cov(real_features, rowvar=False)\n#     mu_generated, sigma_generated = np.mean(generated_features, axis=0), np.cov(generated_features, rowvar=False)\n\n#     diff = mu_real - mu_generated\n#     covmean, _ = sqrtm(sigma_real.dot(sigma_generated), disp=False)\n#     if np.iscomplexobj(covmean):\n#         covmean = covmean.real\n\n#     fid = np.sum(diff**2) + np.trace(sigma_real + sigma_generated - 2 * covmean)\n#     return fid\n\n# def calculate_mifid(real_images, generated_images):\n#     fid = calculate_fid(real_images, generated_images)\n\n#     real_features = inception_model.predict(preprocess_input(tf.image.resize(real_images, (256, 256)).numpy()))\n#     generated_features = inception_model.predict(preprocess_input(tf.image.resize(generated_images, (256, 256)).numpy()))\n\n#     min_cosine_distances = []\n#     for gen_feature in generated_features:\n#         distances = cosine_distances([gen_feature], real_features)\n#         min_distance = np.min(distances)\n#         min_cosine_distances.append(min_distance)\n\n#     memorization_distance = np.mean(min_cosine_distances)\n#     if memorization_distance < model_params['Inception']['cosine_distance_eps']:\n#         memorization_distance = model_params['Inception']['cosine_distance_eps']\n\n#     mifid = fid * (1 / memorization_distance)\n#     return mifid\n\n# # Loss function for both generator and discriminator\n# cross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)\n\n# # Discriminator loss\n# def discriminator_loss(real_output, fake_output):\n#     real_loss = cross_entropy(tf.ones_like(real_output), real_output)\n#     fake_loss = cross_entropy(tf.zeros_like(fake_output), fake_output)\n#     total_loss = real_loss + fake_loss\n#     return total_loss\n\n# # Generator loss\n# def generator_loss(fake_output):\n#     return cross_entropy(tf.ones_like(fake_output), fake_output)\n\n# # Optimizers\n# generator_g_optimizer = tf.keras.optimizers.Adam(2e-4, beta_1=0.5)\n# discriminator_x_optimizer = tf.keras.optimizers.Adam(2e-4, beta_1=0.5)\n\n# # Instantiate the models\n# generator_g = build_generator()  # Real photo to Monet\n# discriminator_x = build_discriminator()  # Discriminator for Monet\n\n# @tf.function\n# def train_step(real_x, real_y, epoch, total_epochs, initial_noise=0.1, final_noise=0.01):\n#     with tf.GradientTape(persistent=True) as tape:\n#         # Adding noise progressively\n# #         def get_noise_schedule(epoch, total_epochs, initial_noise, final_noise):\n# #             return initial_noise + (final_noise - initial_noise) * ((total_epochs - epoch) / total_epochs)\n        \n# #         noise = get_noise_schedule(epoch, total_epochs, initial_noise, final_noise)\n\n# #         # Add noise to real images\n# #         real_x = real_x + noise * tf.random.normal(real_x.shape)\n\n#         # Generate fake images\n#         fake_y = generator_g(real_x, training=True)\n        \n#         # Discriminator outputs\n#         disc_real_x = discriminator_x(real_x, training=True)\n#         disc_fake_y = discriminator_x(fake_y, training=True)\n\n#         # Generator adversarial loss\n#         gen_g_loss = generator_loss(disc_fake_y)\n\n#         # Discriminator loss\n#         disc_x_loss = discriminator_loss(disc_real_x, disc_fake_y)\n\n#         # Total loss\n#         total_loss = gen_g_loss + disc_x_loss\n\n#     # Calculate the gradients\n#     generator_g_gradients = tape.gradient(total_loss, generator_g.trainable_variables)\n#     discriminator_x_gradients = tape.gradient(total_loss, discriminator_x.trainable_variables)\n\n#     # Apply the gradients\n#     generator_g_optimizer.apply_gradients(zip(generator_g_gradients, generator_g.trainable_variables))\n#     discriminator_x_optimizer.apply_gradients(zip(discriminator_x_gradients, discriminator_x.trainable_variables))\n\n#     return gen_g_loss, disc_x_loss\n\n\n# def train(real_dataset, monet_dataset, epochs):\n#     for epoch in range(epochs):\n#         start = time.time()\n\n#         gen_losses = []\n#         disc_losses = []\n\n#         n = 0\n#         for real_x, real_y in tf.data.Dataset.zip((real_dataset, monet_dataset)):\n#             gen_loss, disc_loss = train_step(real_x, real_y, epoch, total_epochs=epochs)\n#             gen_losses.append(gen_loss)\n#             disc_losses.append(disc_loss)\n#             if n % 10 == 0:\n#                 print('.', end='', flush=True)\n#             n += 1\n\n#         print('Time for epoch {} is {} sec'.format(epoch + 1, time.time() - start))\n\n#         # Generate images for visualization\n#         generate_images(generator_g, sample_real_image)\n\n#         # Calculate MiFID\n#         real_images = []\n#         generated_images = []\n        \n#         for real_x, real_y in tf.data.Dataset.zip((real_dataset, monet_dataset)).take(100):\n#             real_images.append(real_x)\n#             generated_images.append(generator_g(real_x, training=False))\n        \n#         real_images = tf.concat(real_images, axis=0)\n#         generated_images = tf.concat(generated_images, axis=0)\n        \n#         mifid = calculate_mifid(real_images, generated_images)\n#         print('Epoch:', epoch + 1, 'MiFID:', mifid)\n\n#         # Save the model every 5 epochs\n#         if (epoch + 1) % 5 == 0:\n#             checkpoint.save(file_prefix=checkpoint_prefix)\n\n#         # Print average losses for the epoch\n#         avg_gen_loss = np.mean(gen_losses)\n#         avg_disc_loss = np.mean(disc_losses)\n#         print('Epoch:', epoch + 1, 'Average Generator Loss:', avg_gen_loss, 'Average Discriminator Loss:', avg_disc_loss)\n\n\n# # Load datasets and start training\n# BUFFER_SIZE = 1000\n# BATCH_SIZE = 1\n# EPOCHS = 100\n\n# sample_real_image = next(iter(real_photos_dataset.take(1)))\n\n# train(real_photos_dataset, monet_paintings_dataset, epochs=EPOCHS)\n","metadata":{"execution":{"iopub.status.busy":"2024-05-31T00:30:14.71004Z","iopub.status.idle":"2024-05-31T00:30:14.710516Z","shell.execute_reply.started":"2024-05-31T00:30:14.710292Z","shell.execute_reply":"2024-05-31T00:30:14.710321Z"},"jupyter":{"source_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# I hope this notebook Was helpful for you to either start or explore new approachs. \n# Feel Free to Upvote !","metadata":{}},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}