#!/usr/bin/env python # coding: utf-8 # Deep Learning Models -- A collection of various deep learning architectures, models, and tips for TensorFlow and PyTorch in Jupyter Notebooks. # - Author: Sebastian Raschka # - GitHub Repository: https://github.com/rasbt/deeplearning-models # In[1]: get_ipython().run_line_magic('load_ext', 'watermark') get_ipython().run_line_magic('watermark', "-a 'Sebastian Raschka' -v -p torch") # - Runs on CPU or GPU (if available) # # Model Zoo -- Generative Adversarial Networks (GAN) # Implementation of a standard GAN. # ## Imports # In[2]: import time import numpy as np import torch import torch.nn.functional as F from torchvision import datasets from torchvision import transforms import torch.nn as nn from torch.utils.data import DataLoader if torch.cuda.is_available(): torch.backends.cudnn.deterministic = True # ## Settings and Dataset # In[3]: ########################## ### SETTINGS ########################## # Device device = torch.device("cuda:2" if torch.cuda.is_available() else "cpu") # Hyperparameters random_seed = 123 generator_learning_rate = 0.001 discriminator_learning_rate = 0.001 NUM_EPOCHS = 100 BATCH_SIZE = 128 LATENT_DIM = 75 IMG_SHAPE = (1, 28, 28) IMG_SIZE = 1 for x in IMG_SHAPE: IMG_SIZE *= x ########################## ### MNIST DATASET ########################## # Note transforms.ToTensor() scales input images # to 0-1 range train_dataset = datasets.MNIST(root='data', train=True, transform=transforms.ToTensor(), download=True) test_dataset = datasets.MNIST(root='data', train=False, transform=transforms.ToTensor()) train_loader = DataLoader(dataset=train_dataset, batch_size=BATCH_SIZE, shuffle=True) test_loader = DataLoader(dataset=test_dataset, batch_size=BATCH_SIZE, shuffle=False) # Checking the dataset for images, labels in train_loader: print('Image batch dimensions:', images.shape) print('Image label dimensions:', labels.shape) break # ## Model # In[4]: ########################## ### MODEL ########################## class GAN(torch.nn.Module): def __init__(self): super(GAN, self).__init__() self.generator = nn.Sequential( nn.Linear(LATENT_DIM, 128), nn.LeakyReLU(inplace=True), nn.Dropout(p=0.5), nn.Linear(128, IMG_SIZE), nn.Tanh() ) self.discriminator = nn.Sequential( nn.Linear(IMG_SIZE, 128), nn.LeakyReLU(inplace=True), nn.Dropout(p=0.5), nn.Linear(128, 1), nn.Sigmoid() ) def generator_forward(self, z): img = self.generator(z) return img def discriminator_forward(self, img): pred = model.discriminator(img) return pred.view(-1) # In[5]: torch.manual_seed(random_seed) model = GAN() model = model.to(device) optim_gener = torch.optim.Adam(model.generator.parameters(), lr=generator_learning_rate) optim_discr = torch.optim.Adam(model.discriminator.parameters(), lr=discriminator_learning_rate) # ## Training # In[6]: start_time = time.time() discr_costs = [] gener_costs = [] for epoch in range(NUM_EPOCHS): model = model.train() for batch_idx, (features, targets) in enumerate(train_loader): features = (features - 0.5)*2. features = features.view(-1, IMG_SIZE).to(device) targets = targets.to(device) valid = torch.ones(targets.size(0)).float().to(device) fake = torch.zeros(targets.size(0)).float().to(device) ### FORWARD AND BACK PROP # -------------------------- # Train Generator # -------------------------- # Make new images z = torch.zeros((targets.size(0), LATENT_DIM)).uniform_(-1.0, 1.0).to(device) generated_features = model.generator_forward(z) # Loss for fooling the discriminator discr_pred = model.discriminator_forward(generated_features) gener_loss = F.binary_cross_entropy(discr_pred, valid) optim_gener.zero_grad() gener_loss.backward() optim_gener.step() # -------------------------- # Train Discriminator # -------------------------- discr_pred_real = model.discriminator_forward(features.view(-1, IMG_SIZE)) real_loss = F.binary_cross_entropy(discr_pred_real, valid) discr_pred_fake = model.discriminator_forward(generated_features.detach()) fake_loss = F.binary_cross_entropy(discr_pred_fake, fake) discr_loss = 0.5*(real_loss + fake_loss) optim_discr.zero_grad() discr_loss.backward() optim_discr.step() discr_costs.append(discr_loss) gener_costs.append(gener_loss) ### LOGGING if not batch_idx % 100: print ('Epoch: %03d/%03d | Batch %03d/%03d | Gen/Dis Loss: %.4f/%.4f' %(epoch+1, NUM_EPOCHS, batch_idx, len(train_loader), gener_loss, discr_loss)) print('Time elapsed: %.2f min' % ((time.time() - start_time)/60)) print('Total Training Time: %.2f min' % ((time.time() - start_time)/60)) # ## Evaluation # In[7]: get_ipython().run_line_magic('matplotlib', 'inline') import matplotlib.pyplot as plt # In[8]: ax1 = plt.subplot(1, 1, 1) ax1.plot(range(len(gener_costs)), gener_costs, label='Generator loss') ax1.plot(range(len(discr_costs)), discr_costs, label='Discriminator loss') ax1.set_xlabel('Iterations') ax1.set_ylabel('Loss') ax1.legend() ################### # Set scond x-axis ax2 = ax1.twiny() newlabel = list(range(NUM_EPOCHS+1)) iter_per_epoch = len(train_loader) newpos = [e*iter_per_epoch for e in newlabel] ax2.set_xticklabels(newlabel[::10]) ax2.set_xticks(newpos[::10]) ax2.xaxis.set_ticks_position('bottom') ax2.xaxis.set_label_position('bottom') ax2.spines['bottom'].set_position(('outward', 45)) ax2.set_xlabel('Epochs') ax2.set_xlim(ax1.get_xlim()) ################### plt.show() # In[9]: ########################## ### VISUALIZATION ########################## model.eval() # Make new images z = torch.zeros((5, LATENT_DIM)).uniform_(-1.0, 1.0).to(device) generated_features = model.generator_forward(z) imgs = generated_features.view(-1, 28, 28) fig, axes = plt.subplots(nrows=1, ncols=5, figsize=(20, 2.5)) for i, ax in enumerate(axes): axes[i].imshow(imgs[i].to(torch.device('cpu')).detach(), cmap='binary')