import tensorflow as tf import matplotlib.pyplot as plt import numpy as np import tensorflow as tf from pathlib import Path import os import PIL from tqdm.auto import tqdm import argparse from tensorflow.keras import layers from datasets import load_dataset from transformers import DefaultDataCollator from huggingface_hub import push_to_hub_keras def parse_args(args=None): parser = argparse.ArgumentParser() parser.add_argument("--dataset", type=str, default="mnist", help="Dataset to load from the HuggingFace hub.") parser.add_argument("--batch_size", type=int, default=128, help="Batch size to use during training") parser.add_argument("--number_of_examples_to_generate", type=int, default=4, help="Number of examples to be generated in inference mode") parser.add_argument( "--generator_hidden_size", type=int, default=28, help="Hidden size of the generator's feature maps.", ) parser.add_argument("--latent_dim", type=int, default=100, help="Dimensionality of the latent space.") parser.add_argument( "--discriminator_hidden_size", type=int, default=28, help="Hidden size of the discriminator's feature maps.", ) parser.add_argument( "--image_size", type=int, default=28, help="Spatial size to use when resizing images for training.", ) parser.add_argument( "--num_channels", type=int, default=3, help="Number of channels in the training images. For color images this is 3.", ) parser.add_argument("--num_epochs", type=int, default=5, help="number of epochs of training") parser.add_argument("--output_dir", type=Path, default=Path("./output"), help="Name of the directory to dump generated images during training.") parser.add_argument( "--push_to_hub", action="store_true", help="Whether to push the model to the HuggingFace hub after training.", ) parser.add_argument( "--model_name", default=None, type=str, help="Name of the model on the hub.", ) parser.add_argument( "--organization_name", default="huggan", type=str, help="Organization name to push to, in case args.push_to_hub is specified.", ) args = parser.parse_args() if args.push_to_hub: assert args.output_dir is not None, "Need an `output_dir` to create a repo when `--push_to_hub` is passed." assert args.model_name is not None, "Need a `model_name` to create a repo when `--push_to_hub` is passed." if args.output_dir is not None: os.makedirs(args.output_dir, exist_ok=True) return args def stack_generator_layers(model, units): model.add(layers.Conv2DTranspose(units, (4, 4), strides=2, padding='same', use_bias=False)) model.add(layers.BatchNormalization()) model.add(layers.LeakyReLU()) return model def create_generator(channel, hidden_size, latent_dim): generator = tf.keras.Sequential() generator.add(layers.Input((latent_dim,))) # generator.add(layers.Dense(hidden_size*4*7*7, use_bias=False, input_shape=(100,))) generator.add(layers.LeakyReLU()) generator.add(layers.Reshape((7, 7, hidden_size*4))) units = [hidden_size*2, hidden_size*1] for unit in units: generator = stack_generator_layers(generator, unit) generator.add(layers.Conv2DTranspose(args.num_channels, (4, 4), strides=1, padding='same', use_bias=False, activation='tanh')) return generator def stack_discriminator_layers(model, units, use_batch_norm=False, use_dropout=False): model.add(layers.Conv2D(units, (4, 4), strides=(2, 2), padding='same')) if use_batch_norm: model.add(layers.BatchNormalization()) if use_dropout: model.add(layers.Dropout(0.1)) model.add(layers.LeakyReLU()) return model def create_discriminator(channel, hidden_size, args): discriminator = tf.keras.Sequential() discriminator.add(layers.Input((args.image_size, args.image_size, args.num_channels))) discriminator = stack_discriminator_layers(discriminator, hidden_size, use_batch_norm = True, use_dropout = True) discriminator = stack_discriminator_layers(discriminator, hidden_size * 2) discriminator = stack_discriminator_layers(discriminator,True, hidden_size*4) discriminator = stack_discriminator_layers(discriminator,True, hidden_size*16) discriminator.add(layers.Flatten()) discriminator.add(layers.Dense(1)) return discriminator def discriminator_loss(real_image, generated_image): real_loss = cross_entropy(tf.ones_like(real_image), real_image) fake_loss = cross_entropy(tf.zeros_like(generated_image), generated_image) total_loss = real_loss + fake_loss return total_loss @tf.function def train_step(images): noise = tf.random.normal([128, 100]) with tf.GradientTape() as gen_tape, tf.GradientTape() as disc_tape: generated_images = generator(noise, training=True) real_image = discriminator(images, training=True) generated_image = discriminator(generated_images, training=True) # calculate loss inside train step gen_loss = cross_entropy(tf.ones_like(generated_image), generated_image) disc_loss = discriminator_loss(real_image, generated_image) gradients_of_generator = gen_tape.gradient(gen_loss, generator.trainable_variables) gradients_of_discriminator = disc_tape.gradient(disc_loss, discriminator.trainable_variables) generator_optimizer.apply_gradients(zip(gradients_of_generator, generator.trainable_variables)) discriminator_optimizer.apply_gradients(zip(gradients_of_discriminator, discriminator.trainable_variables)) def generate_and_save_images(model, epoch, test_input, output_dir, number_of_examples_to_generate): predictions = model(test_input, training=False) fig = plt.figure(figsize=(number_of_examples_to_generate*4, number_of_examples_to_generate*16)) for i in range(predictions.shape[0]): plt.subplot(1, number_of_examples_to_generate, i+1) if args.num_channels == 1: plt.imshow(predictions[i, :, :, :], cmap='gray') else: plt.imshow(predictions[i, :, :, :]) plt.axis('off') plt.savefig(f'{output_dir}/image_at_epoch_{epoch}.png') def train(dataset, epochs, output_dir, args): for epoch in range(epochs): print("Epoch:", epoch) for image_batch in tqdm(dataset): train_step(image_batch) generate_and_save_images(generator, epoch + 1, seed, output_dir, args.number_of_examples_to_generate) def preprocess(examples): images = (np.asarray(examples["image"]).astype('float32')- 127.5) / 127.5 images = np.expand_dims(images, -1) examples["pixel_values"] = images return examples def preprocess_images(dataset, args): data_collator = DefaultDataCollator(return_tensors="tf") processed_dataset = dataset.map(preprocess) tf_train_dataset = processed_dataset["train"].to_tf_dataset( columns=['pixel_values'], shuffle=True, batch_size=args.batch_size, collate_fn=data_collator) return tf_train_dataset if __name__ == "__main__": args = parse_args() print("Downloading dataset..") dataset = load_dataset(args.dataset) dataset= preprocess_images(dataset, args) print("Training model..") generator = create_generator(args.num_channels, args.generator_hidden_size, args.latent_dim) discriminator = create_discriminator(args.num_channels, args.discriminator_hidden_size, args) generator_optimizer = tf.keras.optimizers.Adam(1e-4) discriminator_optimizer = tf.keras.optimizers.Adam(1e-4) # create seed with dimensions of number of examples to generate and noise seed = tf.random.normal([args.number_of_examples_to_generate, args.latent_dim]) cross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True) train(dataset, args.num_epochs, args.output_dir, args) if args.push_to_hub is not None: push_to_hub_keras(generator, repo_path_or_name=f"{args.output_dir}/{args.model_name}",organization=args.organization_name)