[
{
"path": "Cat DCGAN.ipynb",
"content": "{\n \"cells\": [\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"# Cat DCGAN š±āš»\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"Outputs from our DCGAN: \"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \" Cat DCGAN is a Deep Convolutional Generative Adversarial Network (DCGAN) that generates pictures of cats
\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"This is an open source project\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \" If you have any questions, feel free to ask me:
\\n\",\n \" š§: hello@simoninithomas.com
\\n\",\n \" Github: https://github.com/simoninithomas/CatDCGAN
\\n\",\n \" š : https://www.simoninithomas.com
\\n\",\n \" Twitter: @ThomasSimonini
\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"## Important note š¤\\n\",\n \" You can't run this on your computer (except if you have GPUs or wait 10 years š
), personally I train this DCGAN for 20 hours with Microsoft Azure and their Deep Learning Virtual Machine (they offer 170$)\\n\",\n \"https://azuremarketplace.microsoft.com/en-us/marketplace/apps/microsoft-ads.dsvm-deep-learning\\n\",\n \" Vector Market from www.flaticon.com \\n\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"### Check the version of Tensorflow and access to GPU\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"# Taken from Udacity face generator project\\n\",\n \"from distutils.version import LooseVersion\\n\",\n \"import warnings\\n\",\n \"import tensorflow as tf\\n\",\n \"\\n\",\n \"# Check TensorFlow Version\\n\",\n \"assert LooseVersion(tf.__version__) >= LooseVersion('1.0'), 'Please use TensorFlow version 1.0 or newer. You are using {}'.format(tf.__version__)\\n\",\n \"print('TensorFlow Version: {}'.format(tf.__version__))\\n\",\n \"\\n\",\n \"# Check for a GPU\\n\",\n \"if not tf.test.gpu_device_name():\\n\",\n \" warnings.warn('No GPU found. Please use a GPU to train your neural network.')\\n\",\n \"else:\\n\",\n \" print('Default GPU Device: {}'.format(tf.test.gpu_device_name()))\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"### Input\\n\",\n \"Create TF placeholders for the Neural Network:\\n\",\n \"- Real input images placeholder `real_dim`.\\n\",\n \"- Z input placeholder `z_dim`.\\n\",\n \"- Learning rate G placeholder.\\n\",\n \"- Learning rate D placeholder.\\n\",\n \" for 2 reasons :\\n\",\n \"\\n\",\n \" Make sure all varaibles names start with generator / discriminator (will help out later when training the separate networks \\n\",\n \" Also want to reuse these networks with different inputs \\n\",\n \" \\n\",\n \" For the generator: we're going to train it but also sample from it as we're training after training \\n\",\n \" For the discriminator: need to share variables between the fake and real input images \\n\",\n \" \\n\",\n \" \\n\",\n \" So we can use the reuse keyword to tell TensorFlow to reuse the var instead of createing new one if we build the graph again
\\n\",\n \"\\n\",\n \"#### Leaky ReLU\\n\",\n \"Avoid gradient vanishing\\n\",\n \"\\n\",\n \"#### Tanh Output\\n\",\n \"Generator has been found to perform the best with tanh for the generator output \\n\",\n \"Transposed convnets --> normalization --> leaky ReLU \"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"def generator(z, output_channel_dim, is_train=True):\\n\",\n \" ''' Build the generator network.\\n\",\n \" \\n\",\n \" Arguments\\n\",\n \" ---------\\n\",\n \" z : Input tensor for the generator\\n\",\n \" output_channel_dim : Shape of the generator output\\n\",\n \" n_units : Number of units in hidden layer\\n\",\n \" reuse : Reuse the variables with tf.variable_scope\\n\",\n \" alpha : leak parameter for leaky ReLU\\n\",\n \" \\n\",\n \" Returns\\n\",\n \" -------\\n\",\n \" out: \\n\",\n \" '''\\n\",\n \" with tf.variable_scope(\\\"generator\\\", reuse= not is_train):\\n\",\n \" \\n\",\n \" # First FC layer --> 8x8x1024\\n\",\n \" fc1 = tf.layers.dense(z, 8*8*1024)\\n\",\n \" \\n\",\n \" # Reshape it\\n\",\n \" fc1 = tf.reshape(fc1, (-1, 8, 8, 1024))\\n\",\n \" \\n\",\n \" # Leaky ReLU\\n\",\n \" fc1 = tf.nn.leaky_relu(fc1, alpha=alpha)\\n\",\n \"\\n\",\n \" \\n\",\n \" # Transposed conv 1 --> BatchNorm --> LeakyReLU\\n\",\n \" # 8x8x1024 --> 16x16x512\\n\",\n \" trans_conv1 = tf.layers.conv2d_transpose(inputs = fc1,\\n\",\n \" filters = 512,\\n\",\n \" kernel_size = [5,5],\\n\",\n \" strides = [2,2],\\n\",\n \" padding = \\\"SAME\\\",\\n\",\n \" kernel_initializer=tf.truncated_normal_initializer(stddev=0.02),\\n\",\n \" name=\\\"trans_conv1\\\")\\n\",\n \" \\n\",\n \" batch_trans_conv1 = tf.layers.batch_normalization(inputs = trans_conv1, training=is_train, epsilon=1e-5, name=\\\"batch_trans_conv1\\\")\\n\",\n \" \\n\",\n \" trans_conv1_out = tf.nn.leaky_relu(batch_trans_conv1, alpha=alpha, name=\\\"trans_conv1_out\\\")\\n\",\n \" \\n\",\n \" \\n\",\n \" # Transposed conv 2 --> BatchNorm --> LeakyReLU\\n\",\n \" # 16x16x512 --> 32x32x256\\n\",\n \" trans_conv2 = tf.layers.conv2d_transpose(inputs = trans_conv1_out,\\n\",\n \" filters = 256,\\n\",\n \" kernel_size = [5,5],\\n\",\n \" strides = [2,2],\\n\",\n \" padding = \\\"SAME\\\",\\n\",\n \" kernel_initializer=tf.truncated_normal_initializer(stddev=0.02),\\n\",\n \" name=\\\"trans_conv2\\\")\\n\",\n \" \\n\",\n \" batch_trans_conv2 = tf.layers.batch_normalization(inputs = trans_conv2, training=is_train, epsilon=1e-5, name=\\\"batch_trans_conv2\\\")\\n\",\n \" \\n\",\n \" trans_conv2_out = tf.nn.leaky_relu(batch_trans_conv2, alpha=alpha, name=\\\"trans_conv2_out\\\")\\n\",\n \" \\n\",\n \" \\n\",\n \" # Transposed conv 3 --> BatchNorm --> LeakyReLU\\n\",\n \" # 32x32x256 --> 64x64x128\\n\",\n \" trans_conv3 = tf.layers.conv2d_transpose(inputs = trans_conv2_out,\\n\",\n \" filters = 128,\\n\",\n \" kernel_size = [5,5],\\n\",\n \" strides = [2,2],\\n\",\n \" padding = \\\"SAME\\\",\\n\",\n \" kernel_initializer=tf.truncated_normal_initializer(stddev=0.02),\\n\",\n \" name=\\\"trans_conv3\\\")\\n\",\n \" \\n\",\n \" batch_trans_conv3 = tf.layers.batch_normalization(inputs = trans_conv3, training=is_train, epsilon=1e-5, name=\\\"batch_trans_conv3\\\")\\n\",\n \" \\n\",\n \" trans_conv3_out = tf.nn.leaky_relu(batch_trans_conv3, alpha=alpha, name=\\\"trans_conv3_out\\\")\\n\",\n \"\\n\",\n \" \\n\",\n \" # Transposed conv 4 --> BatchNorm --> LeakyReLU\\n\",\n \" # 64x64x128 --> 128x128x64\\n\",\n \" trans_conv4 = tf.layers.conv2d_transpose(inputs = trans_conv3_out,\\n\",\n \" filters = 64,\\n\",\n \" kernel_size = [5,5],\\n\",\n \" strides = [2,2],\\n\",\n \" padding = \\\"SAME\\\",\\n\",\n \" kernel_initializer=tf.truncated_normal_initializer(stddev=0.02),\\n\",\n \" name=\\\"trans_conv4\\\")\\n\",\n \" \\n\",\n \" batch_trans_conv4 = tf.layers.batch_normalization(inputs = trans_conv4, training=is_train, epsilon=1e-5, name=\\\"batch_trans_conv4\\\")\\n\",\n \" \\n\",\n \" trans_conv4_out = tf.nn.leaky_relu(batch_trans_conv4, alpha=alpha, name=\\\"trans_conv4_out\\\")\\n\",\n \"\\n\",\n \" \\n\",\n \" # Transposed conv 5 --> tanh\\n\",\n \" # 128x128x64 --> 128x128x3\\n\",\n \" logits = tf.layers.conv2d_transpose(inputs = trans_conv4_out,\\n\",\n \" filters = 3,\\n\",\n \" kernel_size = [5,5],\\n\",\n \" strides = [1,1],\\n\",\n \" padding = \\\"SAME\\\",\\n\",\n \" kernel_initializer=tf.truncated_normal_initializer(stddev=0.02),\\n\",\n \" name=\\\"logits\\\")\\n\",\n \" \\n\",\n \" out = tf.tanh(logits, name=\\\"out\\\")\\n\",\n \" \\n\",\n \" return out\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"### Discriminator\\n\",\n \" strided conv layers with no maxpool layers \\n\",\n \"- No batchnorm in input layer\\n\",\n \"\\n\",\n \" convolution > batch norm > leaky ReLU \"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"def discriminator(x, is_reuse=False, alpha = 0.2):\\n\",\n \" ''' Build the discriminator network.\\n\",\n \" \\n\",\n \" Arguments\\n\",\n \" ---------\\n\",\n \" x : Input tensor for the discriminator\\n\",\n \" n_units: Number of units in hidden layer\\n\",\n \" reuse : Reuse the variables with tf.variable_scope\\n\",\n \" alpha : leak parameter for leaky ReLU\\n\",\n \" \\n\",\n \" Returns\\n\",\n \" -------\\n\",\n \" out, logits: \\n\",\n \" '''\\n\",\n \" with tf.variable_scope(\\\"discriminator\\\", reuse = is_reuse): \\n\",\n \" \\n\",\n \" # Input layer 128*128*3 --> 64x64x64\\n\",\n \" # Conv --> BatchNorm --> LeakyReLU \\n\",\n \" conv1 = tf.layers.conv2d(inputs = x,\\n\",\n \" filters = 64,\\n\",\n \" kernel_size = [5,5],\\n\",\n \" strides = [2,2],\\n\",\n \" padding = \\\"SAME\\\",\\n\",\n \" kernel_initializer=tf.truncated_normal_initializer(stddev=0.02),\\n\",\n \" name='conv1')\\n\",\n \" \\n\",\n \" batch_norm1 = tf.layers.batch_normalization(conv1,\\n\",\n \" training = True,\\n\",\n \" epsilon = 1e-5,\\n\",\n \" name = 'batch_norm1')\\n\",\n \"\\n\",\n \" conv1_out = tf.nn.leaky_relu(batch_norm1, alpha=alpha, name=\\\"conv1_out\\\")\\n\",\n \" \\n\",\n \" \\n\",\n \" # 64x64x64--> 32x32x128\\n\",\n \" # Conv --> BatchNorm --> LeakyReLU \\n\",\n \" conv2 = tf.layers.conv2d(inputs = conv1_out,\\n\",\n \" filters = 128,\\n\",\n \" kernel_size = [5, 5],\\n\",\n \" strides = [2, 2],\\n\",\n \" padding = \\\"SAME\\\",\\n\",\n \" kernel_initializer=tf.truncated_normal_initializer(stddev=0.02),\\n\",\n \" name='conv2')\\n\",\n \" \\n\",\n \" batch_norm2 = tf.layers.batch_normalization(conv2,\\n\",\n \" training = True,\\n\",\n \" epsilon = 1e-5,\\n\",\n \" name = 'batch_norm2')\\n\",\n \" \\n\",\n \" conv2_out = tf.nn.leaky_relu(batch_norm2, alpha=alpha, name=\\\"conv2_out\\\")\\n\",\n \"\\n\",\n \" \\n\",\n \" \\n\",\n \" # 32x32x128 --> 16x16x256\\n\",\n \" # Conv --> BatchNorm --> LeakyReLU \\n\",\n \" conv3 = tf.layers.conv2d(inputs = conv2_out,\\n\",\n \" filters = 256,\\n\",\n \" kernel_size = [5, 5],\\n\",\n \" strides = [2, 2],\\n\",\n \" padding = \\\"SAME\\\",\\n\",\n \" kernel_initializer=tf.truncated_normal_initializer(stddev=0.02),\\n\",\n \" name='conv3')\\n\",\n \" \\n\",\n \" batch_norm3 = tf.layers.batch_normalization(conv3,\\n\",\n \" training = True,\\n\",\n \" epsilon = 1e-5,\\n\",\n \" name = 'batch_norm3')\\n\",\n \" \\n\",\n \" conv3_out = tf.nn.leaky_relu(batch_norm3, alpha=alpha, name=\\\"conv3_out\\\")\\n\",\n \"\\n\",\n \" \\n\",\n \" \\n\",\n \" # 16x16x256 --> 16x16x512\\n\",\n \" # Conv --> BatchNorm --> LeakyReLU \\n\",\n \" conv4 = tf.layers.conv2d(inputs = conv3_out,\\n\",\n \" filters = 512,\\n\",\n \" kernel_size = [5, 5],\\n\",\n \" strides = [1, 1],\\n\",\n \" padding = \\\"SAME\\\",\\n\",\n \" kernel_initializer=tf.truncated_normal_initializer(stddev=0.02),\\n\",\n \" name='conv4')\\n\",\n \" \\n\",\n \" batch_norm4 = tf.layers.batch_normalization(conv4,\\n\",\n \" training = True,\\n\",\n \" epsilon = 1e-5,\\n\",\n \" name = 'batch_norm4')\\n\",\n \" \\n\",\n \" conv4_out = tf.nn.leaky_relu(batch_norm4, alpha=alpha, name=\\\"conv4_out\\\")\\n\",\n \"\\n\",\n \" \\n\",\n \" \\n\",\n \" # 16x16x512 --> 8x8x1024\\n\",\n \" # Conv --> BatchNorm --> LeakyReLU \\n\",\n \" conv5 = tf.layers.conv2d(inputs = conv4_out,\\n\",\n \" filters = 1024,\\n\",\n \" kernel_size = [5, 5],\\n\",\n \" strides = [2, 2],\\n\",\n \" padding = \\\"SAME\\\",\\n\",\n \" kernel_initializer=tf.truncated_normal_initializer(stddev=0.02),\\n\",\n \" name='conv5')\\n\",\n \" \\n\",\n \" batch_norm5 = tf.layers.batch_normalization(conv5,\\n\",\n \" training = True,\\n\",\n \" epsilon = 1e-5,\\n\",\n \" name = 'batch_norm5')\\n\",\n \" \\n\",\n \" conv5_out = tf.nn.leaky_relu(batch_norm5, alpha=alpha, name=\\\"conv5_out\\\")\\n\",\n \"\\n\",\n \" \\n\",\n \" # Flatten it\\n\",\n \" flatten = tf.reshape(conv5_out, (-1, 8*8*1024))\\n\",\n \" \\n\",\n \" # Logits\\n\",\n \" logits = tf.layers.dense(inputs = flatten,\\n\",\n \" units = 1,\\n\",\n \" activation = None)\\n\",\n \" \\n\",\n \" \\n\",\n \" out = tf.sigmoid(logits)\\n\",\n \" \\n\",\n \" return out, logits\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"### Discriminator and generator losses\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"We train the g and d at the same time so we need losses for both networks \\n\",\n \"\\n\",\n \"#### Discriminator Loss\\n\",\n \"Sum of loss for real and fake images\\n\",\n \" sigmoid cross entropy + wrap with tf.reduce_mean to get the mean for all the images in the batch.\\n\",\n \" \\n\",\n \"\\n\",\n \"##### Real image loss\\n\",\n \"- Use `d_logits_real` and labels are all 1 (since all real data is real) \\n\",\n \"- Label smoothing: To help the discriminator generalize better, the labels are reduced a bit from 1.0 to 0.9 \\n\",\n \"`labels = tf.ones_like(tensor) * (1 - smooth)`\\n\",\n \"For the real image loss, use the real logits and (smoothed) labels of ones. \\n\",\n \"\\n\",\n \"##### Fake image loss\\n\",\n \"- Remember that we want the discriminator to output 1 for real images and 0 for fake images, so we need to set up the losses to reflect that\\n\",\n \"- For the fake image loss, use the fake logits with labels of all zeros\\n\",\n \"\\n\",\n \"#### Generator Loss\\n\",\n \"- The generator loss again uses the fake logits from the discriminator, but this time the labels are all ones because the generator wants to fool the discriminator.\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"def model_loss(input_real, input_z, output_channel_dim, alpha):\\n\",\n \" \\\"\\\"\\\"\\n\",\n \" Get the loss for the discriminator and generator\\n\",\n \" :param input_real: Images from the real dataset\\n\",\n \" :param input_z: Z input\\n\",\n \" :param out_channel_dim: The number of channels in the output image\\n\",\n \" :return: A tuple of (discriminator loss, generator loss)\\n\",\n \" \\\"\\\"\\\"\\n\",\n \" # Generator network here\\n\",\n \" g_model = generator(input_z, output_channel_dim) \\n\",\n \" # g_model is the generator output\\n\",\n \" \\n\",\n \" # Discriminator network here\\n\",\n \" d_model_real, d_logits_real = discriminator(input_real, alpha=alpha)\\n\",\n \" d_model_fake, d_logits_fake = discriminator(g_model,is_reuse=True, alpha=alpha)\\n\",\n \" \\n\",\n \" # Calculate losses\\n\",\n \" d_loss_real = tf.reduce_mean(\\n\",\n \" tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_real, \\n\",\n \" labels=tf.ones_like(d_model_real)))\\n\",\n \" d_loss_fake = tf.reduce_mean(\\n\",\n \" tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_fake, \\n\",\n \" labels=tf.zeros_like(d_model_fake)))\\n\",\n \" d_loss = d_loss_real + d_loss_fake\\n\",\n \"\\n\",\n \" g_loss = tf.reduce_mean(\\n\",\n \" tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_fake,\\n\",\n \" labels=tf.ones_like(d_model_fake)))\\n\",\n \" \\n\",\n \" return d_loss, g_loss\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"### Optimizers\\n\",\n \"\\n\",\n \" Update the generator and discriminator separately \\n\",\n \" So we need to get the var for each part : we use `tf.trainable_variables()`. This creates a list of all the variables we've defined in our graph. \\n\",\n \" \\n\",\n \"- The train operations are wrapped in a with tf.control_dependencies block so the batch normalization layers can update their population statistics.\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"def model_optimizers(d_loss, g_loss, lr_D, lr_G, beta1):\\n\",\n \" \\\"\\\"\\\"\\n\",\n \" Get optimization operations\\n\",\n \" :param d_loss: Discriminator loss Tensor\\n\",\n \" :param g_loss: Generator loss Tensor\\n\",\n \" :param learning_rate: Learning Rate Placeholder\\n\",\n \" :param beta1: The exponential decay rate for the 1st moment in the optimizer\\n\",\n \" :return: A tuple of (discriminator training operation, generator training operation)\\n\",\n \" \\\"\\\"\\\" \\n\",\n \" # Get the trainable_variables, split into G and D parts\\n\",\n \" t_vars = tf.trainable_variables()\\n\",\n \" g_vars = [var for var in t_vars if var.name.startswith(\\\"generator\\\")]\\n\",\n \" d_vars = [var for var in t_vars if var.name.startswith(\\\"discriminator\\\")]\\n\",\n \" \\n\",\n \" update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\\n\",\n \" \\n\",\n \" # Generator update\\n\",\n \" gen_updates = [op for op in update_ops if op.name.startswith('generator')]\\n\",\n \" \\n\",\n \" # Optimizers\\n\",\n \" with tf.control_dependencies(gen_updates):\\n\",\n \" d_train_opt = tf.train.AdamOptimizer(learning_rate=lr_D, beta1=beta1).minimize(d_loss, var_list=d_vars)\\n\",\n \" g_train_opt = tf.train.AdamOptimizer(learning_rate=lr_G, beta1=beta1).minimize(g_loss, var_list=g_vars)\\n\",\n \" \\n\",\n \" return d_train_opt, g_train_opt\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"## Training šāāļø\\n\",\n \"### Show output\\n\",\n \"Use this function to show the current output of the generator during training. It will help you determine how well the GANs is training.\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"def show_generator_output(sess, n_images, input_z, out_channel_dim, image_mode, image_path, save, show):\\n\",\n \" \\\"\\\"\\\"\\n\",\n \" Show example output for the generator\\n\",\n \" :param sess: TensorFlow session\\n\",\n \" :param n_images: Number of Images to display\\n\",\n \" :param input_z: Input Z Tensor\\n\",\n \" :param out_channel_dim: The number of channels in the output image\\n\",\n \" :param image_mode: The mode to use for images (\\\"RGB\\\" or \\\"L\\\")\\n\",\n \" :param image_path: Path to save the image\\n\",\n \" \\\"\\\"\\\"\\n\",\n \" cmap = None if image_mode == 'RGB' else 'gray'\\n\",\n \" z_dim = input_z.get_shape().as_list()[-1]\\n\",\n \" example_z = np.random.uniform(-1, 1, size=[n_images, z_dim])\\n\",\n \"\\n\",\n \" samples = sess.run(\\n\",\n \" generator(input_z, out_channel_dim, False),\\n\",\n \" feed_dict={input_z: example_z})\\n\",\n \"\\n\",\n \" images_grid = helper.images_square_grid(samples, image_mode)\\n\",\n \" \\n\",\n \" if save == True:\\n\",\n \" # Save image\\n\",\n \" images_grid.save(image_path, 'JPEG')\\n\",\n \" \\n\",\n \" if show == True:\\n\",\n \" plt.imshow(images_grid, cmap=cmap)\\n\",\n \" plt.show()\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"def train(epoch_count, batch_size, z_dim, learning_rate_D, learning_rate_G, beta1, get_batches, data_shape, data_image_mode, alpha):\\n\",\n \" \\\"\\\"\\\"\\n\",\n \" Train the GAN\\n\",\n \" :param epoch_count: Number of epochs\\n\",\n \" :param batch_size: Batch Size\\n\",\n \" :param z_dim: Z dimension\\n\",\n \" :param learning_rate: Learning Rate\\n\",\n \" :param beta1: The exponential decay rate for the 1st moment in the optimizer\\n\",\n \" :param get_batches: Function to get batches\\n\",\n \" :param data_shape: Shape of the data\\n\",\n \" :param data_image_mode: The image mode to use for images (\\\"RGB\\\" or \\\"L\\\")\\n\",\n \" \\\"\\\"\\\"\\n\",\n \" # Create our input placeholders\\n\",\n \" input_images, input_z, lr_G, lr_D = model_inputs(data_shape[1:], z_dim)\\n\",\n \" \\n\",\n \" # Losses\\n\",\n \" d_loss, g_loss = model_loss(input_images, input_z, data_shape[3], alpha)\\n\",\n \" \\n\",\n \" # Optimizers\\n\",\n \" d_opt, g_opt = model_optimizers(d_loss, g_loss, lr_D, lr_G, beta1)\\n\",\n \" \\n\",\n \" i = 0\\n\",\n \" \\n\",\n \" version = \\\"firstTrain\\\"\\n\",\n \" with tf.Session() as sess:\\n\",\n \" sess.run(tf.global_variables_initializer())\\n\",\n \" \\n\",\n \" # Saver\\n\",\n \" saver = tf.train.Saver()\\n\",\n \" \\n\",\n \" num_epoch = 0\\n\",\n \" \\n\",\n \" if from_checkpoint == True:\\n\",\n \" saver.restore(sess, \\\"./models/model.ckpt\\\")\\n\",\n \" \\n\",\n \" show_generator_output(sess, 4, input_z, data_shape[3], data_image_mode, image_path, True, False)\\n\",\n \" \\n\",\n \" else:\\n\",\n \" for epoch_i in range(epoch_count): \\n\",\n \" num_epoch += 1\\n\",\n \"\\n\",\n \" if num_epoch % 5 == 0:\\n\",\n \"\\n\",\n \" # Save model every 5 epochs\\n\",\n \" #if not os.path.exists(\\\"models/\\\" + version):\\n\",\n \" # os.makedirs(\\\"models/\\\" + version)\\n\",\n \" save_path = saver.save(sess, \\\"./models/model.ckpt\\\")\\n\",\n \" print(\\\"Model saved\\\")\\n\",\n \"\\n\",\n \" for batch_images in get_batches(batch_size):\\n\",\n \" # Random noise\\n\",\n \" batch_z = np.random.uniform(-1, 1, size=(batch_size, z_dim))\\n\",\n \"\\n\",\n \" i += 1\\n\",\n \"\\n\",\n \" # Run optimizers\\n\",\n \" _ = sess.run(d_opt, feed_dict={input_images: batch_images, input_z: batch_z, lr_D: learning_rate_D})\\n\",\n \" _ = sess.run(g_opt, feed_dict={input_images: batch_images, input_z: batch_z, lr_G: learning_rate_G})\\n\",\n \"\\n\",\n \" if i % 10 == 0:\\n\",\n \" train_loss_d = d_loss.eval({input_z: batch_z, input_images: batch_images})\\n\",\n \" train_loss_g = g_loss.eval({input_z: batch_z})\\n\",\n \"\\n\",\n \" # Save it\\n\",\n \" image_name = str(i) + \\\".jpg\\\"\\n\",\n \" image_path = \\\"./images/\\\" + image_name\\n\",\n \" show_generator_output(sess, 4, input_z, data_shape[3], data_image_mode, image_path, True, False) \\n\",\n \"\\n\",\n \" # Print every 5 epochs (for stability overwize the jupyter notebook will bug)\\n\",\n \" if i % 1500 == 0:\\n\",\n \"\\n\",\n \" image_name = str(i) + \\\".jpg\\\"\\n\",\n \" image_path = \\\"./images/\\\" + image_name\\n\",\n \" print(\\\"Epoch {}/{}...\\\".format(epoch_i+1, epochs),\\n\",\n \" \\\"Discriminator Loss: {:.4f}...\\\".format(train_loss_d),\\n\",\n \" \\\"Generator Loss: {:.4f}\\\".format(train_loss_g))\\n\",\n \" show_generator_output(sess, 4, input_z, data_shape[3], data_image_mode, image_path, False, True)\\n\",\n \" \\n\",\n \" \\n\",\n \" \\n\",\n \" return losses, samples\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"### Hyperparameters\\n\",\n \"Gans are very sensitive to hyperparemeters \\n\",\n \"In general, you want the discriminator loss to be around 0.3, this means it is correctly classifying images as fake or real about 50% of the time.\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"# Size input image for discriminator\\n\",\n \"real_size = (128,128,3)\\n\",\n \"\\n\",\n \"# Size of latent vector to generator\\n\",\n \"z_dim = 100\\n\",\n \"learning_rate_D = .00005 # Thanks to Alexia Jolicoeur Martineau https://ajolicoeur.wordpress.com/cats/\\n\",\n \"learning_rate_G = 2e-4 # Thanks to Alexia Jolicoeur Martineau https://ajolicoeur.wordpress.com/cats/\\n\",\n \"batch_size = 64\\n\",\n \"epochs = 215\\n\",\n \"alpha = 0.2\\n\",\n \"beta1 = 0.5\\n\",\n \"\\n\",\n \"# Create the network\\n\",\n \"#model = DGAN(real_size, z_size, learning_rate, alpha, beta1)\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"# Load the data and train the network here\\n\",\n \"dataset = helper.Dataset(glob(os.path.join(data_resized_dir, '*.jpg')))\\n\",\n \"\\n\",\n \"with tf.Graph().as_default():\\n\",\n \" losses, samples = train(epochs, batch_size, z_dim, learning_rate_D, learning_rate_G, beta1, dataset.get_batches,\\n\",\n \" dataset.shape, dataset.image_mode, alpha)\\n\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"## Training Loss š\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"fig, ax = plt.subplots()\\n\",\n \"losses = np.array(losses)\\n\",\n \"plt.plot(losses.T[0], label='Discriminator', alpha=0.5)\\n\",\n \"plt.plot(losses.T[1], label='Generator', alpha=0.5)\\n\",\n \"plt.title(\\\"Training Losses\\\")\\n\",\n \"plt.legend()\"\n ]\n }\n ],\n \"metadata\": {\n \"kernelspec\": {\n \"display_name\": \"Python [conda env:gameplai]\",\n \"language\": \"python\",\n \"name\": \"conda-env-gameplai-py\"\n },\n \"language_info\": {\n \"codemirror_mode\": {\n \"name\": \"ipython\",\n \"version\": 3\n },\n \"file_extension\": \".py\",\n \"mimetype\": \"text/x-python\",\n \"name\": \"python\",\n \"nbconvert_exporter\": \"python\",\n \"pygments_lexer\": \"ipython3\",\n \"version\": \"3.6.4\"\n }\n },\n \"nbformat\": 4,\n \"nbformat_minor\": 2\n}\n"
},
{
"path": "README.md",
"content": "# CatDCGAN š±āš»\nA DGAN that generate Cat pictures\n Cat DCGAN is a Deep Convolutional Generative Adversarial Network (DCGAN) that generates pictures of cats
\nThis is an open source project \nplease read my article on FreeCodeCamp https://medium.freecodecamp.org/how-ai-can-learn-to-generate-pictures-of-cats-ba692cb6eae4\n\n If you have any questions, feel free to ask me:
\n š§: hello@simoninithomas.com
\n Github: https://github.com/simoninithomas/CatDCGAN
\n š : https://www.simoninithomas.com
\n Twitter: @ThomasSimonini
\n\nIn this part we're going to implement the DCGAN.\nOur Architecture: Download the model checkpoint : https://drive.google.com/drive/folders/1zdZZ91fjOUiOsIdAQKZkUTATXzqy7hiz?usp=sharing\n- Type `sh start.sh` it will handle extract, remove outliers, normalization and face centering\n- Launch Jupyter Notebook `jupyter notebook --no-browser`\n- Launch CatDCGAN\n- Change `do_preprocess = True` ā ļøā ļøā ļø important!\n- If you want to train from scratch : change `from_checkpoint = False`\n- If you want to train from last model saved (you save 20 hours of training š) : change `from_checkpoint = True`\n\n## Important note š¤\n You can't run this on your computer (except if you have GPUs or wait 10 years š
), personally I train this DCGAN for 20 hours with Microsoft Azure and their Deep Learning Virtual Machine (they offer 170$)\nhttps://azuremarketplace.microsoft.com/en-us/marketplace/apps/microsoft-ads.dsvm-deep-learning\n404 Not Found \r\n\r\n\r\n\t\r\n\t\r\n\tCatDCGAN, an AI that generates pictures of cats \r\n\r\n \r\n\r\n \r\n \r\n \r\n CatDCGAN š±āš» \r\n\r\n \r\n \r\n\r\n\r\n\r\n
\r\n
An AI that generates pictures of cats \r\n
All of these pictures were generated by our AI
\r\n \r\n \r\n \t
\r\n \t
\r\n \t
\r\n \r\n
\r\n
\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n Keep in touch! \r\n \r\n\r\n \r\n \r\n \r\n \r\n \r\n\r\n\r\n\r\n \r\n \r\n \r\n \r\n \r\n \r\n \r\n\r\n\r\n\r\n \r\n "
},
{
"path": "helper.py",
"content": "# Taken from Udacity Face generator Project\nimport math\nimport os\nimport hashlib\nfrom urllib.request import urlretrieve\nimport zipfile\nimport gzip\nimport shutil\n\nimport numpy as np\nfrom PIL import Image\n\n\n\ndef _read32(bytestream):\n \"\"\"\n Read 32-bit integer from bytesteam\n :param bytestream: A bytestream\n :return: 32-bit integer\n \"\"\"\n dt = np.dtype(np.uint32).newbyteorder('>')\n return np.frombuffer(bytestream.read(4), dtype=dt)[0]\n\n\ndef _unzip(save_path, _, database_name, data_path):\n \"\"\"\n Unzip wrapper with the same interface as _ungzip\n :param save_path: The path of the gzip files\n :param database_name: Name of database\n :param data_path: Path to extract to\n :param _: HACK - Used to have to same interface as _ungzip\n \"\"\"\n print('Extracting {}...'.format(database_name))\n with zipfile.ZipFile(save_path) as zf:\n zf.extractall(data_path)\n\n\ndef _ungzip(save_path, extract_path, database_name, _):\n \"\"\"\n Unzip a gzip file and extract it to extract_path\n :param save_path: The path of the gzip files\n :param extract_path: The location to extract the data to\n :param database_name: Name of database\n :param _: HACK - Used to have to same interface as _unzip\n \"\"\"\n # Get data from save_path\n with open(save_path, 'rb') as f:\n with gzip.GzipFile(fileobj=f) as bytestream:\n magic = _read32(bytestream)\n if magic != 2051:\n raise ValueError('Invalid magic number {} in file: {}'.format(magic, f.name))\n num_images = _read32(bytestream)\n rows = _read32(bytestream)\n cols = _read32(bytestream)\n buf = bytestream.read(rows * cols * num_images)\n data = np.frombuffer(buf, dtype=np.uint8)\n data = data.reshape(num_images, rows, cols)\n\n # Save data to extract_path\n for image_i, image in enumerate(\n tqdm(data, unit='File', unit_scale=True, miniters=1, desc='Extracting {}'.format(database_name))):\n Image.fromarray(image, 'L').save(os.path.join(extract_path, 'image_{}.jpg'.format(image_i)))\n\n\ndef get_image(image_path, width, height, mode):\n \"\"\"\n Read image from image_path\n :param image_path: Path of image\n :param width: Width of image\n :param height: Height of image\n :param mode: Mode of image\n :return: Image data\n \"\"\"\n image = Image.open(image_path)\n\n\n return np.array(image.convert(mode))\n\n\ndef get_batch(image_files, width, height, mode):\n data_batch = np.array(\n [get_image(sample_file, width, height, mode) for sample_file in image_files]).astype(np.float32)\n\n # Make sure the images are in 4 dimensions\n if len(data_batch.shape) < 4:\n data_batch = data_batch.reshape(data_batch.shape + (1,))\n\n return data_batch\n\n\ndef images_square_grid(images, mode):\n \"\"\"\n Save images as a square grid\n :param images: Images to be used for the grid\n :param mode: The mode to use for images\n :return: Image of images in a square grid\n \"\"\"\n # Get maximum size for square grid of images\n save_size = math.floor(np.sqrt(images.shape[0]))\n\n # Scale to 0-255\n images = (((images - images.min()) * 255) / (images.max() - images.min())).astype(np.uint8)\n\n # Put images in a square arrangement\n images_in_square = np.reshape(\n images[:save_size*save_size],\n (save_size, save_size, images.shape[1], images.shape[2], images.shape[3]))\n if mode == 'L':\n images_in_square = np.squeeze(images_in_square, 4)\n\n # Combine images to grid image\n new_im = Image.new(mode, (images.shape[1] * save_size, images.shape[2] * save_size))\n for col_i, col_images in enumerate(images_in_square):\n for image_i, image in enumerate(col_images):\n im = Image.fromarray(image, mode)\n new_im.paste(im, (col_i * images.shape[1], image_i * images.shape[2]))\n\n return new_im\n\n\ndef download_extract(database_name, data_path):\n \"\"\"\n Download and extract database\n :param database_name: Database name\n \"\"\"\n DATASET_CELEBA_NAME = 'celeba'\n DATASET_MNIST_NAME = 'mnist'\n\n if database_name == DATASET_CELEBA_NAME:\n url = 'https://s3-us-west-1.amazonaws.com/udacity-dlnfd/datasets/celeba.zip'\n hash_code = '00d2c5bc6d35e252742224ab0c1e8fcb'\n extract_path = os.path.join(data_path, 'img_align_celeba')\n save_path = os.path.join(data_path, 'celeba.zip')\n extract_fn = _unzip\n elif database_name == DATASET_MNIST_NAME:\n url = 'http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz'\n hash_code = 'f68b3c2dcbeaaa9fbdd348bbdeb94873'\n extract_path = os.path.join(data_path, 'mnist')\n save_path = os.path.join(data_path, 'train-images-idx3-ubyte.gz')\n extract_fn = _ungzip\n\n if os.path.exists(extract_path):\n print('Found {} Data'.format(database_name))\n return\n\n if not os.path.exists(data_path):\n os.makedirs(data_path)\n\n if not os.path.exists(save_path):\n with DLProgress(unit='B', unit_scale=True, miniters=1, desc='Downloading {}'.format(database_name)) as pbar:\n urlretrieve(\n url,\n save_path,\n pbar.hook)\n\n assert hashlib.md5(open(save_path, 'rb').read()).hexdigest() == hash_code, \\\n '{} file is corrupted. Remove the file and try again.'.format(save_path)\n\n os.makedirs(extract_path)\n try:\n extract_fn(save_path, extract_path, database_name, data_path)\n except Exception as err:\n shutil.rmtree(extract_path) # Remove extraction folder if there is an error\n raise err\n\n # Remove compressed data\n os.remove(save_path)\n\n\nclass Dataset(object):\n \"\"\"\n Dataset\n \"\"\"\n def __init__(self, data_files):\n \"\"\"\n Initalize the class\n :param dataset_name: Database name\n :param data_files: List of files in the database\n \"\"\"\n IMAGE_WIDTH = 128\n IMAGE_HEIGHT = 128\n\n self.image_mode = 'RGB'\n image_channels = 3\n\n self.data_files = data_files\n self.shape = len(data_files), IMAGE_WIDTH, IMAGE_HEIGHT, image_channels\n\n def get_batches(self, batch_size):\n \"\"\"\n Generate batches\n :param batch_size: Batch Size\n :return: Batches of data\n \"\"\"\n IMAGE_MAX_VALUE = 255\n\n current_index = 0\n while current_index + batch_size <= self.shape[0]:\n data_batch = get_batch(\n self.data_files[current_index:current_index + batch_size],\n *self.shape[1:3],\n self.image_mode)\n \n current_index += batch_size\n\n yield data_batch / IMAGE_MAX_VALUE - 0.5\n\n\n\n"
},
{
"path": "models/checkpoint",
"content": "model_checkpoint_path: \"model.ckpt\"\nall_model_checkpoint_paths: \"model.ckpt\"\n"
},
{
"path": "start.sh",
"content": "## This script was taken from Alexia Jolicoeur Martineau's work https://ajolicoeur.wordpress.com/cats/\n## I modified some elements such as the link to download the dataset.\n\n## Download CAT dataset\n# wget -nc http://www.simoninithomas.com/data/cats.zip\n\n## Setting up folder\nunzip cats-dataset.zip -d cat_dataset\nunzip cats.zip\nmv cat_dataset/CAT_00/* cat_dataset\nrmdir cat_dataset/CAT_00\nmv cat_dataset/CAT_01/* cat_dataset\nrmdir cat_dataset/CAT_01\nmv cat_dataset/CAT_02/* cat_dataset\nrmdir cat_dataset/CAT_02\nmv cat_dataset/CAT_03/* cat_dataset\nrmdir cat_dataset/CAT_03\nmv cat_dataset/CAT_04/* cat_dataset\nrmdir cat_dataset/CAT_04\nmv cat_dataset/CAT_05/* cat_dataset\nrmdir cat_dataset/CAT_05\nmv cat_dataset/CAT_06/* cat_dataset\nrmdir cat_dataset/CAT_06\n\n## Error correction\nrm cat_dataset/00000003_019.jpg.cat\nmv 00000003_015.jpg.cat cat_dataset/00000003_015.jpg.cat\n\n## Removing outliers\n# Corrupted, drawings, badly cropped, inverted, impossible to tell it's a cat, blocked face\ncd cat_dataset\nrm 00000004_007.jpg 00000007_002.jpg 00000045_028.jpg 00000050_014.jpg 00000056_013.jpg 00000059_002.jpg 00000108_005.jpg 00000122_023.jpg 00000126_005.jpg 00000132_018.jpg 00000142_024.jpg 00000142_029.jpg 00000143_003.jpg 00000145_021.jpg 00000166_021.jpg 00000169_021.jpg 00000186_002.jpg 00000202_022.jpg 00000208_023.jpg 00000210_003.jpg 00000229_005.jpg 00000236_025.jpg 00000249_016.jpg 00000254_013.jpg 00000260_019.jpg 00000261_029.jpg 00000265_029.jpg 00000271_020.jpg 00000282_026.jpg 00000316_004.jpg 00000352_014.jpg 00000400_026.jpg 00000406_006.jpg 00000431_024.jpg 00000443_027.jpg 00000502_015.jpg 00000504_012.jpg 00000510_019.jpg 00000514_016.jpg 00000514_008.jpg 00000515_021.jpg 00000519_015.jpg 00000522_016.jpg 00000523_021.jpg 00000529_005.jpg 00000556_022.jpg 00000574_011.jpg 00000581_018.jpg 00000582_011.jpg 00000588_016.jpg 00000588_019.jpg 00000590_006.jpg 00000592_018.jpg 00000593_027.jpg 00000617_013.jpg 00000618_016.jpg 00000619_025.jpg 00000622_019.jpg 00000622_021.jpg 00000630_007.jpg 00000645_016.jpg 00000656_017.jpg 00000659_000.jpg 00000660_022.jpg 00000660_029.jpg 00000661_016.jpg 00000663_005.jpg 00000672_027.jpg 00000673_027.jpg 00000675_023.jpg 00000692_006.jpg 00000800_017.jpg 00000805_004.jpg 00000807_020.jpg 00000823_010.jpg 00000824_010.jpg 00000836_008.jpg 00000843_021.jpg 00000850_025.jpg 00000862_017.jpg 00000864_007.jpg 00000865_015.jpg 00000870_007.jpg 00000877_014.jpg 00000882_013.jpg 00000887_028.jpg 00000893_022.jpg 00000907_013.jpg 00000921_029.jpg 00000929_022.jpg 00000934_006.jpg 00000960_021.jpg 00000976_004.jpg 00000987_000.jpg 00000993_009.jpg 00001006_014.jpg 00001008_013.jpg 00001012_019.jpg 00001014_005.jpg 00001020_017.jpg 00001039_008.jpg 00001039_023.jpg 00001048_029.jpg 00001057_003.jpg 00001068_005.jpg 00001113_015.jpg 00001140_007.jpg 00001157_029.jpg 00001158_000.jpg 00001167_007.jpg 00001184_007.jpg 00001188_019.jpg 00001204_027.jpg 00001205_022.jpg 00001219_005.jpg 00001243_010.jpg 00001261_005.jpg 00001270_028.jpg 00001274_006.jpg 00001293_015.jpg 00001312_021.jpg 00001365_026.jpg 00001372_006.jpg 00001379_018.jpg 00001388_024.jpg 00001389_026.jpg 00001418_028.jpg 00001425_012.jpg 00001431_001.jpg 00001456_018.jpg 00001458_003.jpg 00001468_019.jpg 00001475_009.jpg 00001487_020.jpg\nrm 00000004_007.jpg.cat 00000007_002.jpg.cat 00000045_028.jpg.cat 00000050_014.jpg.cat 00000056_013.jpg.cat 00000059_002.jpg.cat 00000108_005.jpg.cat 00000122_023.jpg.cat 00000126_005.jpg.cat 00000132_018.jpg.cat 00000142_024.jpg.cat 00000142_029.jpg.cat 00000143_003.jpg.cat 00000145_021.jpg.cat 00000166_021.jpg.cat 00000169_021.jpg.cat 00000186_002.jpg.cat 00000202_022.jpg.cat 00000208_023.jpg.cat 00000210_003.jpg.cat 00000229_005.jpg.cat 00000236_025.jpg.cat 00000249_016.jpg.cat 00000254_013.jpg.cat 00000260_019.jpg.cat 00000261_029.jpg.cat 00000265_029.jpg.cat 00000271_020.jpg.cat 00000282_026.jpg.cat 00000316_004.jpg.cat 00000352_014.jpg.cat 00000400_026.jpg.cat 00000406_006.jpg.cat 00000431_024.jpg.cat 00000443_027.jpg.cat 00000502_015.jpg.cat 00000504_012.jpg.cat 00000510_019.jpg.cat 00000514_016.jpg.cat 00000514_008.jpg.cat 00000515_021.jpg.cat 00000519_015.jpg.cat 00000522_016.jpg.cat 00000523_021.jpg.cat 00000529_005.jpg.cat 00000556_022.jpg.cat 00000574_011.jpg.cat 00000581_018.jpg.cat 00000582_011.jpg.cat 00000588_016.jpg.cat 00000588_019.jpg.cat 00000590_006.jpg.cat 00000592_018.jpg.cat 00000593_027.jpg.cat 00000617_013.jpg.cat 00000618_016.jpg.cat 00000619_025.jpg.cat 00000622_019.jpg.cat 00000622_021.jpg.cat 00000630_007.jpg.cat 00000645_016.jpg.cat 00000656_017.jpg.cat 00000659_000.jpg.cat 00000660_022.jpg.cat 00000660_029.jpg.cat 00000661_016.jpg.cat 00000663_005.jpg.cat 00000672_027.jpg.cat 00000673_027.jpg.cat 00000675_023.jpg.cat 00000692_006.jpg.cat 00000800_017.jpg.cat 00000805_004.jpg.cat 00000807_020.jpg.cat 00000823_010.jpg.cat 00000824_010.jpg.cat 00000836_008.jpg.cat 00000843_021.jpg.cat 00000850_025.jpg.cat 00000862_017.jpg.cat 00000864_007.jpg.cat 00000865_015.jpg.cat 00000870_007.jpg.cat 00000877_014.jpg.cat 00000882_013.jpg.cat 00000887_028.jpg.cat 00000893_022.jpg.cat 00000907_013.jpg.cat 00000921_029.jpg.cat 00000929_022.jpg.cat 00000934_006.jpg.cat 00000960_021.jpg.cat 00000976_004.jpg.cat 00000987_000.jpg.cat 00000993_009.jpg.cat 00001006_014.jpg.cat 00001008_013.jpg.cat 00001012_019.jpg.cat 00001014_005.jpg.cat 00001020_017.jpg.cat 00001039_008.jpg.cat 00001039_023.jpg.cat 00001048_029.jpg.cat 00001057_003.jpg.cat 00001068_005.jpg.cat 00001113_015.jpg.cat 00001140_007.jpg.cat 00001157_029.jpg.cat 00001158_000.jpg.cat 00001167_007.jpg.cat 00001184_007.jpg.cat 00001188_019.jpg.cat 00001204_027.jpg.cat 00001205_022.jpg.cat 00001219_005.jpg.cat 00001243_010.jpg.cat 00001261_005.jpg.cat 00001270_028.jpg.cat 00001274_006.jpg.cat 00001293_015.jpg.cat 00001312_021.jpg.cat 00001365_026.jpg.cat 00001372_006.jpg.cat 00001379_018.jpg.cat 00001388_024.jpg.cat 00001389_026.jpg.cat 00001418_028.jpg.cat 00001425_012.jpg.cat 00001431_001.jpg.cat 00001456_018.jpg.cat 00001458_003.jpg.cat 00001468_019.jpg.cat 00001475_009.jpg.cat 00001487_020.jpg.cat\ncd ..\n\n## Preprocessing and putting in folders for different image sizes\nmkdir cats_bigger_than_64x64\nmkdir cats_bigger_than_128x128\npython preprocess_cat_dataset.py\n\n## Removing cat_dataset\nrm -r cat_dataset\n\n## Move to your favorite place\n#mv cats_bigger_than_64x64 /home/alexia/Datasets/Meow_64x64\n#mv cats_bigger_than_128x128 /home/alexia/Datasets/Meow_128x128\n"
}
] ![\"CatDCGAN](\"assets/img/output.png\")
\r\n \t![\\\"CatDCGAN](\\\"assets/output.png\\\")
\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"![\\\"Training](\\\"assets/training2.gif\\\")
\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"import os\\n\",\n \"import tensorflow as tf\\n\",\n \"import numpy as np\\n\",\n \"\\n\",\n \"import helper\\n\",\n \"from glob import glob\\n\",\n \"import pickle as pkl\\n\",\n \"import scipy.misc\\n\",\n \"\\n\",\n \"import time\\n\",\n \"\\n\",\n \"import cv2\\n\",\n \"import matplotlib.pyplot as plt\\n\",\n \"%matplotlib inline\\n\",\n \"\\n\",\n \"do_preprocess = False\\n\",\n \"from_checkpoint = False\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"## The Data š\\n\",\n \"### Resize images to 128x128\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"data_dir = './cats_bigger_than_128x128' # Data\\n\",\n \"data_resized_dir = \\\"./resized_data\\\"# Resized data\\n\",\n \"\\n\",\n \"if do_preprocess == True:\\n\",\n \" os.mkdir(data_resized_dir)\\n\",\n \"\\n\",\n \" for each in os.listdir(data_dir):\\n\",\n \" image = cv2.imread(os.path.join(data_dir, each))\\n\",\n \" image = cv2.resize(image, (128, 128))\\n\",\n \" cv2.imwrite(os.path.join(data_resized_dir, each), image)\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"# This part was taken from Udacity Face generator project\\n\",\n \"def get_image(image_path, width, height, mode):\\n\",\n \" \\\"\\\"\\\"\\n\",\n \" Read image from image_path\\n\",\n \" :param image_path: Path of image\\n\",\n \" :param width: Width of image\\n\",\n \" :param height: Height of image\\n\",\n \" :param mode: Mode of image\\n\",\n \" :return: Image data\\n\",\n \" \\\"\\\"\\\"\\n\",\n \" image = Image.open(image_path)\\n\",\n \"\\n\",\n \" return np.array(image.convert(mode))\\n\",\n \"\\n\",\n \"def get_batch(image_files, width, height, mode):\\n\",\n \" data_batch = np.array(\\n\",\n \" [get_image(sample_file, width, height, mode) for sample_file in image_files]).astype(np.float32)\\n\",\n \"\\n\",\n \" # Make sure the images are in 4 dimensions\\n\",\n \" if len(data_batch.shape) < 4:\\n\",\n \" data_batch = data_batch.reshape(data_batch.shape + (1,))\\n\",\n \"\\n\",\n \" return data_batch\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"### Explore the data\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"show_n_images = 25\\n\",\n \"mnist_images = helper.get_batch(glob(os.path.join(data_resized_dir, '*.jpg'))[:show_n_images], 64, 64, 'RGB')\\n\",\n \"plt.imshow(helper.images_square_grid(mnist_images, 'RGB'))\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"## The DCGAN š¤\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"In this part we're going to implement the DCGAN.\\n\",\n \"Our Architecture:![\\\"Cat](\\\"assets/GDSchema.png\\\")
\\n\",\n \"\\n\",\n \"Cat Icon made by Vector Market from www.flaticon.com \\n\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"### Check the version of Tensorflow and access to GPU\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"# Taken from Udacity face generator project\\n\",\n \"from distutils.version import LooseVersion\\n\",\n \"import warnings\\n\",\n \"import tensorflow as tf\\n\",\n \"\\n\",\n \"# Check TensorFlow Version\\n\",\n \"assert LooseVersion(tf.__version__) >= LooseVersion('1.0'), 'Please use TensorFlow version 1.0 or newer. You are using {}'.format(tf.__version__)\\n\",\n \"print('TensorFlow Version: {}'.format(tf.__version__))\\n\",\n \"\\n\",\n \"# Check for a GPU\\n\",\n \"if not tf.test.gpu_device_name():\\n\",\n \" warnings.warn('No GPU found. Please use a GPU to train your neural network.')\\n\",\n \"else:\\n\",\n \" print('Default GPU Device: {}'.format(tf.test.gpu_device_name()))\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"### Input\\n\",\n \"Create TF placeholders for the Neural Network:\\n\",\n \"- Real input images placeholder `real_dim`.\\n\",\n \"- Z input placeholder `z_dim`.\\n\",\n \"- Learning rate G placeholder.\\n\",\n \"- Learning rate D placeholder.\\n\",\n \"![\\\"Generator\\\"/](\\\"assets/generator.png\\\")
\\n\",\n \"\\n\",\n \"\\n\",\n \"#### Variable Scope\\n\",\n \"Use tf.variable_scope for 2 reasons :\\n\",\n \"![\\\"Discriminator\\\"/](\\\"assets/discriminator.png\\\")
\\n\",\n \"\\n\",\n \"- Input is 128x128x3\\n\",\n \"- Depths starting with 32 and then *2 depth as you add layers\\n\",\n \"- No downsampling using only strided conv layers with no maxpool layers \\n\",\n \"- No batchnorm in input layer\\n\",\n \"\\n\",\n \" convolution > batch norm > leaky ReLU \"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"def discriminator(x, is_reuse=False, alpha = 0.2):\\n\",\n \" ''' Build the discriminator network.\\n\",\n \" \\n\",\n \" Arguments\\n\",\n \" ---------\\n\",\n \" x : Input tensor for the discriminator\\n\",\n \" n_units: Number of units in hidden layer\\n\",\n \" reuse : Reuse the variables with tf.variable_scope\\n\",\n \" alpha : leak parameter for leaky ReLU\\n\",\n \" \\n\",\n \" Returns\\n\",\n \" -------\\n\",\n \" out, logits: \\n\",\n \" '''\\n\",\n \" with tf.variable_scope(\\\"discriminator\\\", reuse = is_reuse): \\n\",\n \" \\n\",\n \" # Input layer 128*128*3 --> 64x64x64\\n\",\n \" # Conv --> BatchNorm --> LeakyReLU \\n\",\n \" conv1 = tf.layers.conv2d(inputs = x,\\n\",\n \" filters = 64,\\n\",\n \" kernel_size = [5,5],\\n\",\n \" strides = [2,2],\\n\",\n \" padding = \\\"SAME\\\",\\n\",\n \" kernel_initializer=tf.truncated_normal_initializer(stddev=0.02),\\n\",\n \" name='conv1')\\n\",\n \" \\n\",\n \" batch_norm1 = tf.layers.batch_normalization(conv1,\\n\",\n \" training = True,\\n\",\n \" epsilon = 1e-5,\\n\",\n \" name = 'batch_norm1')\\n\",\n \"\\n\",\n \" conv1_out = tf.nn.leaky_relu(batch_norm1, alpha=alpha, name=\\\"conv1_out\\\")\\n\",\n \" \\n\",\n \" \\n\",\n \" # 64x64x64--> 32x32x128\\n\",\n \" # Conv --> BatchNorm --> LeakyReLU \\n\",\n \" conv2 = tf.layers.conv2d(inputs = conv1_out,\\n\",\n \" filters = 128,\\n\",\n \" kernel_size = [5, 5],\\n\",\n \" strides = [2, 2],\\n\",\n \" padding = \\\"SAME\\\",\\n\",\n \" kernel_initializer=tf.truncated_normal_initializer(stddev=0.02),\\n\",\n \" name='conv2')\\n\",\n \" \\n\",\n \" batch_norm2 = tf.layers.batch_normalization(conv2,\\n\",\n \" training = True,\\n\",\n \" epsilon = 1e-5,\\n\",\n \" name = 'batch_norm2')\\n\",\n \" \\n\",\n \" conv2_out = tf.nn.leaky_relu(batch_norm2, alpha=alpha, name=\\\"conv2_out\\\")\\n\",\n \"\\n\",\n \" \\n\",\n \" \\n\",\n \" # 32x32x128 --> 16x16x256\\n\",\n \" # Conv --> BatchNorm --> LeakyReLU \\n\",\n \" conv3 = tf.layers.conv2d(inputs = conv2_out,\\n\",\n \" filters = 256,\\n\",\n \" kernel_size = [5, 5],\\n\",\n \" strides = [2, 2],\\n\",\n \" padding = \\\"SAME\\\",\\n\",\n \" kernel_initializer=tf.truncated_normal_initializer(stddev=0.02),\\n\",\n \" name='conv3')\\n\",\n \" \\n\",\n \" batch_norm3 = tf.layers.batch_normalization(conv3,\\n\",\n \" training = True,\\n\",\n \" epsilon = 1e-5,\\n\",\n \" name = 'batch_norm3')\\n\",\n \" \\n\",\n \" conv3_out = tf.nn.leaky_relu(batch_norm3, alpha=alpha, name=\\\"conv3_out\\\")\\n\",\n \"\\n\",\n \" \\n\",\n \" \\n\",\n \" # 16x16x256 --> 16x16x512\\n\",\n \" # Conv --> BatchNorm --> LeakyReLU \\n\",\n \" conv4 = tf.layers.conv2d(inputs = conv3_out,\\n\",\n \" filters = 512,\\n\",\n \" kernel_size = [5, 5],\\n\",\n \" strides = [1, 1],\\n\",\n \" padding = \\\"SAME\\\",\\n\",\n \" kernel_initializer=tf.truncated_normal_initializer(stddev=0.02),\\n\",\n \" name='conv4')\\n\",\n \" \\n\",\n \" batch_norm4 = tf.layers.batch_normalization(conv4,\\n\",\n \" training = True,\\n\",\n \" epsilon = 1e-5,\\n\",\n \" name = 'batch_norm4')\\n\",\n \" \\n\",\n \" conv4_out = tf.nn.leaky_relu(batch_norm4, alpha=alpha, name=\\\"conv4_out\\\")\\n\",\n \"\\n\",\n \" \\n\",\n \" \\n\",\n \" # 16x16x512 --> 8x8x1024\\n\",\n \" # Conv --> BatchNorm --> LeakyReLU \\n\",\n \" conv5 = tf.layers.conv2d(inputs = conv4_out,\\n\",\n \" filters = 1024,\\n\",\n \" kernel_size = [5, 5],\\n\",\n \" strides = [2, 2],\\n\",\n \" padding = \\\"SAME\\\",\\n\",\n \" kernel_initializer=tf.truncated_normal_initializer(stddev=0.02),\\n\",\n \" name='conv5')\\n\",\n \" \\n\",\n \" batch_norm5 = tf.layers.batch_normalization(conv5,\\n\",\n \" training = True,\\n\",\n \" epsilon = 1e-5,\\n\",\n \" name = 'batch_norm5')\\n\",\n \" \\n\",\n \" conv5_out = tf.nn.leaky_relu(batch_norm5, alpha=alpha, name=\\\"conv5_out\\\")\\n\",\n \"\\n\",\n \" \\n\",\n \" # Flatten it\\n\",\n \" flatten = tf.reshape(conv5_out, (-1, 8*8*1024))\\n\",\n \" \\n\",\n \" # Logits\\n\",\n \" logits = tf.layers.dense(inputs = flatten,\\n\",\n \" units = 1,\\n\",\n \" activation = None)\\n\",\n \" \\n\",\n \" \\n\",\n \" out = tf.sigmoid(logits)\\n\",\n \" \\n\",\n \" return out, logits\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"### Discriminator and generator losses\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"metadata\": {},\n \"source\": [\n \"We train the g and d at the same time so we need losses for both networks \\n\",\n \"\\n\",\n \"#### Discriminator Loss\\n\",\n \"Sum of loss for real and fake images\\n\",\n \"![\"CatDCGAN](\"assets/output.png\")
\n\n![\"Cat](\"assets/GDSchema.png\")
\n\n## Getting Started š\n- Download the dataset here: https://www.kaggle.com/crawford/cat-dataset\n- Download the model checkpoint : https://drive.google.com/drive/folders/1zdZZ91fjOUiOsIdAQKZkUTATXzqy7hiz?usp=sharing\n- Type `sh start.sh` it will handle extract, remove outliers, normalization and face centering\n- Launch Jupyter Notebook `jupyter notebook --no-browser`\n- Launch CatDCGAN\n- Change `do_preprocess = True` ā ļøā ļøā ļø important!\n- If you want to train from scratch : change `from_checkpoint = False`\n- If you want to train from last model saved (you save 20 hours of training š) : change `from_checkpoint = True`\n\n## Important note š¤\n You can't run this on your computer (except if you have GPUs or wait 10 years š
), personally I train this DCGAN for 20 hours with Microsoft Azure and their Deep Learning Virtual Machine (they offer 170$)\nhttps://azuremarketplace.microsoft.com/en-us/marketplace/apps/microsoft-ads.dsvm-deep-learning\n![\"Training](\"assets/training2.gif\")
\n"
},
{
"path": "docs/404.html",
"content": "\r\n\r\n\r\n\t\r\n\t\r\n\t\r\n\t\r\n\t![](\"assets/img/training.gif\")
\r\n![](\"assets/img/maths_lady.gif\")
\r\n\r\n