prefetch keep dataflow size. conv2d use_bisa

example_cifar10.py

@@ -18,7 +18,8 @@ from tensorpack.dataflow import *
 from tensorpack.dataflow import imgaug
 
 """
-CIFAR10 90% validation accuracy after 100k step
+CIFAR10 90% validation accuracy after 100k step.
+91% after 160k step
 """
 
 BATCH_SIZE = 128
@@ -128,7 +129,7 @@ def get_config():
         learning_rate=1e-2,
         global_step=get_global_step_var(),
         decay_steps=dataset_train.size() * 30,
-        decay_rate=0.7, staircase=True, name='learning_rate')
+        decay_rate=0.5, staircase=True, name='learning_rate')
     tf.scalar_summary('learning_rate', lr)
 
     return TrainConfig(

example_svhn_digit.py

@@ -17,6 +17,10 @@ from tensorpack.utils.summary import *
 from tensorpack.dataflow import *
 from tensorpack.dataflow import imgaug
 
+"""
+SVHN convnet.
+About 2.9% validation error after 70 epoch.
+"""
 
 class Model(ModelDesc):
     def _get_input_vars(self):
@@ -27,7 +31,7 @@ class Model(ModelDesc):
         image, label = input_vars
         keep_prob = tf.constant(0.5 if is_training else 1.0)
 
-        image = image / 255.0
+        image = image / 128.0 - 1
 
         nl = lambda x, name: tf.abs(tf.tanh(x), name=name)
         l = Conv2D('conv1', image, 24, 5, padding='VALID', nl=nl)
@@ -76,14 +80,17 @@ def get_config():
 
     augmentors = [
         imgaug.Resize((40, 40)),
-        imgaug.BrightnessAdd(63),
-        imgaug.Contrast((0.2,1.8)),
+        imgaug.BrightnessAdd(30),
+        imgaug.Contrast((0.5,1.5)),
+        imgaug.GaussianDeform(
+            [(0.2, 0.2), (0.2, 0.8), (0.8,0.8), (0.8,0.2)],
+            (40,40), 0.2, 3),
     ]
     train = AugmentImageComponent(train, augmentors)
     train = BatchData(train, 128)
-    nr_proc = 2
-    train = PrefetchData(train, 3, nr_proc)
-    step_per_epoch = train.size() / nr_proc
+    nr_proc = 5
+    train = PrefetchData(train, 5, nr_proc)
+    step_per_epoch = train.size()
 
     augmentors = [
         imgaug.Resize((40, 40)),
@@ -91,14 +98,13 @@ def get_config():
     test = AugmentImageComponent(test, augmentors)
     test = BatchData(test, 128, remainder=True)
 
-    sess_config = get_default_sess_config()
-    sess_config.gpu_options.per_process_gpu_memory_fraction = 0.5
+    sess_config = get_default_sess_config(0.8)
 
     lr = tf.train.exponential_decay(
-        learning_rate=1e-4,
+        learning_rate=1e-3,
         global_step=get_global_step_var(),
-        decay_steps=train.size() * 50,
-        decay_rate=0.7, staircase=True, name='learning_rate')
+        decay_steps=train.size() * 30,
+        decay_rate=0.5, staircase=True, name='learning_rate')
     tf.scalar_summary('learning_rate', lr)
 
     return TrainConfig(
@@ -112,7 +118,7 @@ def get_config():
         session_config=sess_config,
         model=Model(),
         step_per_epoch=step_per_epoch,
-        max_epoch=100,
+        max_epoch=350,
     )
 
 if __name__ == '__main__':

tensorpack/dataflow/imgaug/crop.py

@@ -7,7 +7,8 @@ from .base import ImageAugmentor
 import numpy as np
 from abc import abstractmethod
 
-__all__ = ['RandomCrop', 'CenterCrop', 'FixedCrop', 'CenterPaste']
+__all__ = ['RandomCrop', 'CenterCrop', 'FixedCrop', 'CenterPaste',
+           'ConstantBackgroundFiller']
 
 class RandomCrop(ImageAugmentor):
     """ Randomly crop the image into a smaller one """

tensorpack/dataflow/prefetch.py

@@ -34,14 +34,15 @@ class PrefetchProcess(multiprocessing.Process):
 class PrefetchData(DataFlow):
     def __init__(self, ds, nr_prefetch, nr_proc=1):
         """
-        use multiprocess, will duplicate ds by nr_proc times
+        use multiprocess
         """
         self.ds = ds
+        self._size = self.ds.size()
         self.nr_proc = nr_proc
         self.nr_prefetch = nr_prefetch
 
     def size(self):
-        return self.ds.size() * self.nr_proc
+        return self._size
 
     def get_data(self):
         queue = multiprocessing.Queue(self.nr_prefetch)
@@ -50,6 +51,7 @@ class PrefetchData(DataFlow):
         [x.start() for x in procs]
 
         end_cnt = 0
+        tot_cnt = 0
         try:
             while True:
                 dp = queue.get()
@@ -58,7 +60,10 @@ class PrefetchData(DataFlow):
                     if end_cnt == self.nr_proc:
                         break
                     continue
+                tot_cnt += 1
                 yield dp
+                if tot_cnt == self._size:
+                    break
         finally:
             queue.close()
             [x.terminate() for x in procs]

tensorpack/models/batch_norm.py

@@ -14,7 +14,7 @@ __all__ = ['BatchNorm']
 # http://stackoverflow.com/questions/33949786/how-could-i-use-batch-normalization-in-tensorflow
 # TF batch_norm only works for 4D tensor right now: #804
 @layer_register()
-def BatchNorm(x, is_training, gamma_init=1.0):
+def BatchNorm(x, is_training=True, gamma_init=1.0):
     """
     Batch normalization layer as described in:
     Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
@@ -54,12 +54,10 @@ def BatchNorm(x, is_training, gamma_init=1.0):
 
     if is_training:
         with tf.control_dependencies([ema_apply_op]):
-            mean, var = tf.identity(batch_mean), tf.identity(batch_var)
+            return tf.nn.batch_norm_with_global_normalization(
+                x, batch_mean, batch_var, beta, gamma, EPS, True)
     else:
         batch = tf.cast(tf.shape(x)[0], tf.float32)
         mean, var = ema_mean, ema_var * batch / (batch - 1) # unbiased variance estimator
-
-    normed = tf.nn.batch_norm_with_global_normalization(
+        return tf.nn.batch_norm_with_global_normalization(
             x, mean, var, beta, gamma, EPS, True)
-    return normed
-

tensorpack/models/conv2d.py

@@ -14,12 +14,13 @@ __all__ = ['Conv2D']
 def Conv2D(x, out_channel, kernel_shape,
            padding='SAME', stride=1,
            W_init=None, b_init=None,
-           nl=tf.nn.relu, split=1):
+           nl=tf.nn.relu, split=1, use_bias=True):
     """
     kernel_shape: (h, w) or a int
     stride: (h, w) or a int
     padding: 'valid' or 'same'
     split: split channels. used in Alexnet
+    use_bias: whether to use bias
     """
     in_shape = x.get_shape().as_list()
     num_in = np.prod(in_shape[1:])
@@ -39,6 +40,7 @@ def Conv2D(x, out_channel, kernel_shape,
         b_init = tf.constant_initializer()
 
     W = tf.get_variable('W', filter_shape, initializer=W_init)
+    if use_bias:
         b = tf.get_variable('b', [out_channel], initializer=b_init)
 
     if split == 1:
@@ -49,6 +51,6 @@ def Conv2D(x, out_channel, kernel_shape,
         outputs = [tf.nn.conv2d(i, k, stride, padding)
                    for i, k in zip(inputs, kernels)]
         conv = tf.concat(3, outputs)
-    return nl(tf.nn.bias_add(conv, b), name='output')
+    return nl(tf.nn.bias_add(conv, b) if use_bias else conv, name='output')