useless comments in example & GAN

README.md

@@ -7,10 +7,11 @@ See some [examples](examples) to learn about the framework.
 You can actually train them and reproduce the performance... not just to see how to write code.
 
 + [DoReFa-Net: training binary / low bitwidth CNN](examples/DoReFa-Net)
-+ [InceptionV3 on ImageNet](examples/Inception/inceptionv3.py)
 + [ResNet for ImageNet/Cifar10/SVHN classification](examples/ResNet)
++ [InceptionV3 on ImageNet](examples/Inception/inceptionv3.py)
 + [Fully-convolutional Network for Holistically-Nested Edge Detection](examples/HED)
 + [Spatial Transformer Networks on MNIST addition](examples/SpatialTransformer)
++ [Generative Adversarial Networks](examples/GAN)
 + [Double DQN plays Atari games](examples/Atari2600)
 + [Asynchronous Advantage Actor-Critic(A3C) with demos on OpenAI Gym Atari games](examples/OpenAIGym)
 + [char-rnn language model](examples/char-rnn)

examples/Atari2600/DQN.py

@@ -181,7 +181,7 @@ def get_config():
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--load', help='load model')
     parser.add_argument('--task', help='task to perform',
             choices=['play', 'eval', 'train'], default='train')

examples/DisturbLabel/mnist-disturb.py

@@ -52,7 +52,7 @@ class Model(mnist_example.Model):
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--load', help='load model')
     parser.add_argument('--prob', help='disturb prob', type=float, required=True)
     args = parser.parse_args()

examples/DoReFa-Net/svhn-digit-dorefa.py

@@ -170,7 +170,7 @@ def get_config():
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='the GPU to use') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='the GPU to use')
     parser.add_argument('--load', help='load a checkpoint')
     parser.add_argument('--dorefa',
             help='number of bits for W,A,G, separated by comma. Defaults to \'1,2,4\'',

examples/GAN/GAN.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+# File: GAN.py
+# Author: Yuxin Wu <ppwwyyxxc@gmail.com>
+
+import tensorflow as tf
+import numpy as np
+from tensorpack import (QueueInputTrainerBase, TowerContext,
+        get_global_step_var)
+from tensorpack.tfutils.summary import summary_moving_average, add_moving_summary
+from tensorpack.dataflow import DataFlow
+
+class GANTrainer(QueueInputTrainerBase):
+    def __init__(self, config, g_vs_d=3):
+        super(GANTrainer, self).__init__(config)
+        self._build_enque_thread()
+        if g_vs_d > 1:
+            self._opt_g = g_vs_d
+            self._opt_d = 1
+        else:
+            self._opt_g = 1
+            self._opt_d = int(1.0 / g_vs_d)
+
+    def _setup(self):
+        with TowerContext(''):
+            actual_inputs = self._get_input_tensors_noreuse()
+            self.model.build_graph(actual_inputs)
+        self.gs_incr = tf.assign_add(get_global_step_var(), 1, name='global_step_incr')
+        self.g_min = self.config.optimizer.minimize(self.model.g_loss,
+                var_list=self.model.g_vars, name='g_op')
+        self.d_min = self.config.optimizer.minimize(self.model.d_loss,
+                var_list=self.model.d_vars)
+        self.d_min = tf.group(self.d_min, summary_moving_average(), name='d_op')
+
+    def run_step(self):
+        for _ in range(self._opt_g):
+            self.sess.run(self.g_min)
+        for _ in range(self._opt_d):
+            self.sess.run(self.d_min)
+        self.sess.run(self.gs_incr)
+
+class RandomZData(DataFlow):
+    def __init__(self, shape):
+        super(RandomZData, self).__init__()
+        self.shape = shape
+    def get_data(self):
+        while True:
+            yield [np.random.uniform(-1, 1, size=self.shape)]
+
+def build_GAN_losses(vecpos, vecneg):
+    sigmpos = tf.sigmoid(vecpos)
+    sigmneg = tf.sigmoid(vecneg)
+    tf.histogram_summary('sigmoid-pos', sigmpos)
+    tf.histogram_summary('sigmoid-neg', sigmneg)
+
+    d_loss_pos = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
+        vecpos, tf.ones_like(vecpos)), name='d_loss_pos')
+    d_loss_neg = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
+        vecneg, tf.zeros_like(vecneg)), name='d_loss_neg')
+
+    d_pos_acc = tf.reduce_mean(tf.cast(sigmpos > 0.5, tf.float32), name='pos_acc')
+    d_neg_acc = tf.reduce_mean(tf.cast(sigmneg < 0.5, tf.float32), name='neg_acc')
+
+    g_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
+        vecneg, tf.ones_like(vecneg)), name='g_loss')
+    d_loss = tf.add(d_loss_pos, d_loss_neg, name='d_loss')
+    add_moving_summary(d_loss_pos, d_loss_neg,
+            g_loss, d_loss,
+            d_pos_acc, d_neg_acc)
+    return g_loss, d_loss

examples/GAN/README.md

+# Deep Convolutional Generative Adversarial Networks
+
+Reproduce DCGAN following the setup in [dcgan.torch](https://github.com/soumith/dcgan.torch).
+
+More results to come.
+
+See the docstring in each executable script for usage.

examples/GAN/celebA.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+# File: celebA.py
+# Author: Yuxin Wu <ppwwyyxxc@gmail.com>
+
+import numpy as np
+import tensorflow as tf
+import glob
+import os, sys
+import argparse
+import cv2
+
+from tensorpack import *
+from tensorpack.utils.viz import build_patch_list
+from tensorpack.utils.viz import dump_dataflow_images
+from tensorpack.tfutils.summary import add_moving_summary, summary_moving_average
+import tensorpack.tfutils.symbolic_functions as symbf
+from GAN import GANTrainer, RandomZData, build_GAN_losses
+
+"""
+DCGAN on CelebA dataset.
+1. Download the 'aligned&cropped' version of CelebA dataset.
+2. Start training:
+    ./celebA.py --data /path/to/image_align_celeba/
+3. Visualize samples of a trained model:
+    ./celebA.py --load model.tfmodel --sample
+"""
+
+SHAPE = 64
+BATCH = 128
+
+class Model(ModelDesc):
+    def _get_input_vars(self):
+        return [InputVar(tf.float32, (None, SHAPE, SHAPE, 3), 'input') ]
+
+    def generator(self, z):
+        """ return a image generated from z"""
+        l = FullyConnected('fc0', z, 64 * 8 * 4 * 4, nl=tf.identity)
+        l = tf.reshape(l, [-1, 4, 4, 64*8])
+        l = BNReLU(l)
+        with argscope(Deconv2D, nl=BNReLU, kernel_shape=5, stride=2):
+            l = Deconv2D('deconv1', l, [8, 8, 64 * 4])
+            l = Deconv2D('deconv2', l, [16, 16, 64 * 2])
+            l = Deconv2D('deconv3', l, [32, 32, 64])
+            l = Deconv2D('deconv4', l, [64, 64, 3], nl=tf.identity)
+            l = tf.tanh(l, name='gen')
+        return l
+
+    def discriminator(self, imgs):
+        """ return a (b, 1) logits"""
+        with argscope(Conv2D, nl=tf.identity, kernel_shape=5, stride=2), \
+                argscope(LeakyReLU, alpha=0.2):
+            l = (LinearWrap(imgs)
+                .Conv2D('conv0', 64)
+                .LeakyReLU('lr0')
+                .Conv2D('conv1', 64*2)
+                .BatchNorm('bn1')
+                .LeakyReLU('lr1')
+                .Conv2D('conv2', 64*4)
+                .BatchNorm('bn2')
+                .LeakyReLU('lr2')
+                .Conv2D('conv3', 64*8)
+                .BatchNorm('bn3')
+                .LeakyReLU('lr3')
+                .FullyConnected('fct', 1, nl=tf.identity)())
+        return l
+
+    def _build_graph(self, input_vars):
+        image_pos = input_vars[0]
+        image_pos = image_pos / 128.0 - 1
+        z = tf.random_uniform(tf.pack([tf.shape(image_pos)[0], 100]), -1, 1, name='z')
+        z.set_shape([None, 100])    # issue#5680
+
+        with argscope([Conv2D, Deconv2D, FullyConnected],
+                W_init=tf.truncated_normal_initializer(stddev=0.02)):
+            with tf.variable_scope('gen'):
+                image_gen = self.generator(z)
+                tf.image_summary('gen', image_gen, max_images=30)
+            with tf.variable_scope('discrim'):
+                vecpos = self.discriminator(image_pos)
+            with tf.variable_scope('discrim', reuse=True):
+                vecneg = self.discriminator(image_gen)
+
+        self.g_loss, self.d_loss = build_GAN_losses(vecpos, vecneg)
+        all_vars = tf.trainable_variables()
+        self.g_vars = [v for v in all_vars if v.name.startswith('gen/')]
+        self.d_vars = [v for v in all_vars if v.name.startswith('discrim/')]
+
+def get_data():
+    datadir = args.data
+    imgs = glob.glob(datadir + '/*.jpg')
+    ds = ImageFromFile(imgs, channel=3, shuffle=True)
+    augs = [ imgaug.CenterCrop(110), imgaug.Resize(64) ]
+    ds = AugmentImageComponent(ds, augs)
+    ds = BatchData(ds, BATCH)
+    ds = PrefetchDataZMQ(ds, 1)
+    return ds
+
+def get_config():
+    logger.auto_set_dir()
+    dataset = get_data()
+    lr = symbolic_functions.get_scalar_var('learning_rate', 2e-4, summary=True)
+    return TrainConfig(
+        dataset=dataset,
+        optimizer=tf.train.AdamOptimizer(lr, beta1=0.5, epsilon=1e-3),
+        callbacks=Callbacks([
+            StatPrinter(), ModelSaver(),
+            ScheduledHyperParamSetter('learning_rate', [(200, 1e-4)])
+        ]),
+        session_config=get_default_sess_config(0.5),
+        model=Model(),
+        step_per_epoch=300,
+        max_epoch=500,
+    )
+
+def sample(model_path):
+    pred = PredictConfig(
+       session_init=get_model_loader(model_path),
+       model=Model(),
+       input_names=['z'],
+       output_names=['gen/gen'])
+    pred = SimpleDatasetPredictor(pred, RandomZData((128, 100)))
+    for o in pred.get_result():
+        o = o[0] + 1
+        o = o * 128.0
+        o = o[:,:,:,::-1]
+        viz = next(build_patch_list(o, nr_row=10, nr_col=10))
+        cv2.imshow("", viz)
+        cv2.waitKey()
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
+    parser.add_argument('--load', help='load model')
+    parser.add_argument('--sample', action='store_true', help='run sampling')
+    parser.add_argument('--data', help='`image_align_celeba` directory of the celebA dataset')
+    global args
+    args = parser.parse_args()
+    if args.gpu:
+        os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
+    if args.sample:
+        sample(args.load)
+    else:
+        config = get_config()
+        if args.load:
+            config.session_init = SaverRestore(args.load)
+        GANTrainer(config).train()
+

examples/HED/hed.py

@@ -200,7 +200,7 @@ def run(model_path, image_path, output):
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--load', help='load model')
     parser.add_argument('--view', help='view dataset', action='store_true')
     parser.add_argument('--run', help='run model on images')

examples/Inception/inception-bn.py

@@ -175,7 +175,7 @@ def get_config():
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--load', help='load model')
     parser.add_argument('--data', help='ImageNet data root directory', required=True)
     args = parser.parse_args()

examples/Inception/inceptionv3.py

@@ -281,7 +281,7 @@ def get_config():
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--data', help='ILSVRC dataset dir')
     parser.add_argument('--load', help='load model')
     args = parser.parse_args()

examples/OpenAIGym/run-atari.py

@@ -76,7 +76,7 @@ def do_submit(output):
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--load', help='load model', required=True)
     parser.add_argument('--env', help='environment name', required=True)
     parser.add_argument('--episode', help='number of episodes to run',

examples/OpenAIGym/train-atari.py

@@ -210,7 +210,7 @@ def get_config():
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--load', help='load model')
     parser.add_argument('--env', help='env', required=True)
     parser.add_argument('--task', help='task to perform',

examples/ResNet/cifar10-resnet.py

@@ -160,7 +160,7 @@ def get_config():
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--load', help='load model')
     args = parser.parse_args()
 

examples/ResNet/load-resnet.py

@@ -189,7 +189,7 @@ def name_conversion(caffe_layer_name):
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--load', required=True,
                         help='.npy model file generated by tensorpack.utils.loadcaffe')
     parser.add_argument('-d', '--depth', help='resnet depth', required=True, type=int, choices=[50, 101, 152])

examples/ResNet/svhn-resnet.py

@@ -83,7 +83,7 @@ def get_config():
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--load', help='load model')
     args = parser.parse_args()
 

examples/SpatialTransformer/mnist-addition.py

@@ -163,7 +163,7 @@ def get_config():
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--load', help='load model')
     parser.add_argument('--view', action='store_true')
     args = parser.parse_args()

examples/char-rnn/char-rnn.py

@@ -170,7 +170,7 @@ def sample(path, start, length):
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--load', help='load model')
     subparsers = parser.add_subparsers(title='command', dest='command')
     parser_sample = subparsers.add_parser('sample', help='sample a trained model')

examples/cifar-convnet.py

@@ -133,7 +133,7 @@ def get_config(cifar_classnum):
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--load', help='load model')
     parser.add_argument('--classnum', help='10 for cifar10 or 100 for cifar100',
                         type=int, default=10)

examples/load-alexnet.py

@@ -77,7 +77,7 @@ def run_test(path, input):
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--load',
                         help='.npy model file generated by tensorpack.utils.loadcaffe',
                         required=True)

examples/mnist-convnet.py

@@ -125,7 +125,7 @@ def get_config():
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--load', help='load model')
     args = parser.parse_args()
     if args.gpu:

examples/svhn-digit-convnet.py

@@ -109,7 +109,7 @@ def get_config():
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.') # nargs='*' in multi mode
+    parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
     parser.add_argument('--load', help='load model')
     args = parser.parse_args()
 

tensorpack/dataflow/image.py

@@ -14,7 +14,7 @@ __all__ = ['ImageFromFile', 'AugmentImageComponent', 'AugmentImageComponents']
 class ImageFromFile(RNGDataFlow):
     def __init__(self, files, channel=3, resize=None, shuffle=False):
         """
-        Generate rgb images from list of files
+        Generate RGB images from list of files
         :param files: list of file paths
         :param channel: 1 or 3 channel
         :param resize: a (h, w) tuple. If given, will force a resize

tensorpack/utils/viz.py

@@ -46,12 +46,13 @@ def minnone(x, y):
     return min(x, y)
 
 def build_patch_list(patch_list,
-        nr_row=None, nr_col=None, border=5,
+        nr_row=None, nr_col=None, border=None,
         max_width=1000, max_height=1000,
         shuffle=False, bgcolor=255):
     """
     This is a generator.
     patch_list: bhw or bhwc
+    :param border: defaults to 0.1 * max(image_width, image_height)
     """
     patch_list = np.asarray(patch_list)
     if patch_list.ndim == 3:
@@ -60,6 +61,8 @@ def build_patch_list(patch_list,
     if shuffle:
         np.random.shuffle(patch_list)
     ph, pw = patch_list.shape[1:3]
+    if border is None:
+        border = int(0.1 * max(ph, pw))
     mh, mw = max(max_height, ph + border), max(max_width, pw + border)
     if nr_row is None:
         nr_row = minnone(nr_row, max_height / (ph + border))