after_train

example_cifar10.py

 #!/usr/bin/env python2
 # -*- coding: UTF-8 -*-
-# File: loyaltry.py
+# File: example_cifar10.py
 # Author: Yuxin Wu <ppwwyyxx@gmail.com>
 
 import tensorflow as tf
@@ -16,10 +16,15 @@ from tensorpack.utils.symbolic_functions import *
 from tensorpack.utils.summary import *
 from tensorpack.dataflow import *
 from tensorpack.dataflow import imgaug
-from cifar10 import cifar10
+
+"""
+This config follows the same preprocessing/model/hyperparemeters as in
+tensorflow cifar10 examples. (https://tensorflow.googlesource.com/tensorflow/+/master/tensorflow/models/image/cifar10/)
+But it's faster.
+"""
 
 BATCH_SIZE = 128
-MIN_AFTER_DEQUEUE = 20000   # a large number, as in the official example
+MIN_AFTER_DEQUEUE = int(50000 * 0.4)
 CAPACITY = MIN_AFTER_DEQUEUE + 3 * BATCH_SIZE
 
 def get_model(inputs, is_training):
@@ -50,7 +55,7 @@ def get_model(inputs, is_training):
     l = FullyConnected('fc1', l, out_dim=192,
                        W_init=tf.truncated_normal_initializer(stddev=0.04),
                        b_init=tf.constant_initializer(0.1))
-    ## fc will have activation summary by default. disable this for the output layer
+    # fc will have activation summary by default. disable for the output layer
     logits = FullyConnected('linear', l, out_dim=10, summary_activation=False,
                             nl=tf.identity,
                             W_init=tf.truncated_normal_initializer(stddev=1.0/192))
@@ -91,14 +96,14 @@ def get_config():
         Flip(horiz=True),
         BrightnessAdd(63),
         Contrast((0.2,1.8)),
-        PerImageWhitening(all_channel=True)
+        MeanVarianceNormalize(all_channel=True)
     ]
     dataset_train = AugmentImageComponent(dataset_train, augmentors)
     dataset_train = BatchData(dataset_train, 128)
 
     augmentors = [
         CenterCrop((24, 24)),
-        PerImageWhitening(all_channel=True)
+        MeanVarianceNormalize(all_channel=True)
     ]
     dataset_test = dataset.Cifar10('test')
     dataset_test = AugmentImageComponent(dataset_test, augmentors)
@@ -107,7 +112,6 @@ def get_config():
 
     sess_config = get_default_sess_config()
     sess_config.gpu_options.per_process_gpu_memory_fraction = 0.5
-    sess_config.device_count['GPU'] = 2
 
     # prepare model
     input_vars = [
@@ -150,6 +154,8 @@ if __name__ == '__main__':
     args = parser.parse_args()
     if args.gpu:
         os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
+    else:
+        os.environ['CUDA_VISIBLE_DEVICES'] = '0'
 
     with tf.Graph().as_default():
         with tf.device('/cpu:0'):

tensorpack/callbacks/base.py

@@ -31,13 +31,15 @@ class Callback(object):
         Called before starting iterative training
         """
 
-    def trigger_step(self, inputs, outputs, cost):
+    def after_train(self):
+        """
+        Called after training
+        """
+
+    def trigger_step(self):
         """
         Callback to be triggered after every step (every backpropagation)
-        Args:
-            inputs: the list of input values
-            outputs: list of output values after running this inputs
-            cost: the cost value after running this input
+        Could be useful to apply some tricks on parameters (clipping, low-rank, etc)
         """
 
     def trigger_epoch(self):

tensorpack/callbacks/common.py

@@ -56,3 +56,6 @@ class SummaryWriter(Callback):
                 logger.info('{}: {:.4f}'.format(val.tag, val.simple_value))
         self.writer.add_summary(summary, get_global_step())
 
+    def after_train(self):
+        self.writer.close()
+

tensorpack/callbacks/group.py

@@ -78,9 +78,13 @@ class TrainCallbacks(Callback):
             cb.before_train()
         self.writer = tf.get_collection(SUMMARY_WRITER_COLLECTION_KEY)[0]
 
-    def trigger_step(self, inputs, outputs, cost):
+    def after_train(self):
         for cb in self.cbs:
-            cb.trigger_step(inputs, outputs, cost)
+            cb.after_train()
+
+    def trigger_step(self):
+        for cb in self.cbs:
+            cb.trigger_step()
 
     def trigger_epoch(self):
         tm = CallbackTimeLogger()
@@ -111,6 +115,10 @@ class TestCallbacks(Callback):
             for cb in self.cbs:
                 cb.before_train()
 
+    def after_train(self):
+        for cb in self.cbs:
+            cb.after_train()
+
     def trigger_epoch(self):
         if not self.cbs:
             return
@@ -153,8 +161,12 @@ class Callbacks(Callback):
         self.train.before_train()
         self.test.before_train()
 
-    def trigger_step(self, inputs, outputs, cost):
-        self.train.trigger_step(inputs, outputs, cost)
+    def after_train(self):
+        self.train.after_train()
+        self.test.after_train()
+
+    def trigger_step(self):
+        self.train.trigger_step()
         # test callback don't have trigger_step
 
     def trigger_epoch(self):

tensorpack/dataflow/imgaug/imgproc.py

@@ -6,7 +6,7 @@
 from .base import ImageAugmentor
 import numpy as np
 
-__all__ = ['BrightnessAdd', 'Contrast', 'PerImageWhitening']
+__all__ = ['BrightnessAdd', 'Contrast', 'MeanVarianceNormalize']
 
 class BrightnessAdd(ImageAugmentor):
     """
@@ -35,7 +35,7 @@ class Contrast(ImageAugmentor):
         img.arr = (arr - mean) * r + mean
         img.arr = np.clip(img.arr, 0, 255)
 
-class PerImageWhitening(ImageAugmentor):
+class MeanVarianceNormalize(ImageAugmentor):
     """
     Linearly scales image to have zero mean and unit norm.
     x = (x - mean) / adjusted_stddev
@@ -43,7 +43,6 @@ class PerImageWhitening(ImageAugmentor):
     """
     def __init__(self, all_channel=True):
         self.all_channel = all_channel
-        pass
 
     def _augment(self, img):
         if self.all_channel:

tensorpack/train.py

@@ -10,7 +10,7 @@ import argparse
 
 import tqdm
 from utils import *
-from utils.concurrency import EnqueueThread,coordinator_guard
+from utils.concurrency import EnqueueThread
 from callbacks import *
 from utils.summary import summary_moving_average
 from utils.modelutils import describe_model
@@ -77,26 +77,9 @@ class TrainConfig(object):
 def average_gradients(tower_grads):
     average_grads = []
     for grad_and_vars in zip(*tower_grads):
-    # Note that each grad_and_vars looks like the following:
-    #   ((grad0_gpu0, var0_gpu0), ... , (grad0_gpuN, var0_gpuN))
-    grads = []
-    for g, _ in grad_and_vars:
-      # Add 0 dimension to the gradients to represent the tower.
-      expanded_g = tf.expand_dims(g, 0)
-
-      # Append on a 'tower' dimension which we will average over below.
-      grads.append(expanded_g)
-
-    # Average over the 'tower' dimension.
-    grad = tf.concat(0, grads)
-    grad = tf.reduce_mean(grad, 0)
-
-    # Keep in mind that the Variables are redundant because they are shared
-    # across towers. So .. we will just return the first tower's pointer to
-    # the Variable.
+        grad = tf.add_n([x[0] for x in grad_and_vars]) / float(len(tower_grads))
         v = grad_and_vars[0][1]
-    grad_and_var = (grad, v)
-    average_grads.append(grad_and_var)
+        average_grads.append((grad, v))
     return average_grads
 
 def summary_grads(grads):
@@ -139,8 +122,8 @@ def start_train(config):
         kept_summaries = {}
         grads = []
         for i in range(config.nr_tower):
-            with tf.device('/gpu:{}'.format(i)):
-                with tf.name_scope('tower{}'.format(i)) as scope:
+            with tf.device('/gpu:{}'.format(i)), \
+                    tf.name_scope('tower{}'.format(i)) as scope:
                 model_inputs = get_model_inputs()
                 output_vars, cost_var = config.get_model_func(model_inputs, is_training=True)
                 grads.append(
@@ -172,14 +155,14 @@ def start_train(config):
     # start training:
     coord = tf.train.Coordinator()
     # a thread that keeps filling the queue
-    input_th = EnqueueThread(sess, coord, enqueue_op, config.dataset, input_queue)
     model_th = tf.train.start_queue_runners(
         sess=sess, coord=coord, daemon=True, start=True)
+    input_th = EnqueueThread(sess, coord, enqueue_op, config.dataset, input_queue)
     input_th.start()
 
-    with sess.as_default(), \
-            coordinator_guard(sess, coord):
-        logger.info("Start with global_step={}".format(get_global_step()))
+    with sess.as_default():
+        try:
+            logger.info("Start training with global_step={}".format(get_global_step()))
             callbacks.before_train()
             for epoch in xrange(1, config.max_epoch):
                 with timed_operation('epoch {}'.format(epoch)):
@@ -187,14 +170,16 @@ def start_train(config):
                             config.step_per_epoch, leave=True, mininterval=0.2):
                         if coord.should_stop():
                             return
-                    # TODO if no one uses trigger_step, train_op can be
-                    # faster, see: https://github.com/soumith/convnet-benchmarks/pull/67/files
-                    fetches = [train_op, cost_var] + output_vars + model_inputs
-                    results = sess.run(fetches)
-                    cost = results[1]
-                    outputs = results[2:2 + len(output_vars)]
-                    inputs = results[-len(model_inputs):]
-                    callbacks.trigger_step(inputs, outputs, cost)
+                        sess.run([train_op])    # faster since train_op return None
+                        callbacks.trigger_step()
 
                     # note that summary_op will take a data from the queue.
                     callbacks.trigger_epoch()
+        except (KeyboardInterrupt, Exception):
+            raise
+        finally:
+            coord.request_stop()
+            queue.close(cancel_pending_enqueues=True)
+            callbacks.after_train()
+            sess.close()
+

tensorpack/utils/__init__.py

@@ -37,8 +37,7 @@ def get_default_sess_config():
     Tensorflow default session config consume too much resources
     """
     conf = tf.ConfigProto()
-    conf.device_count['GPU'] = 1
-    conf.gpu_options.per_process_gpu_memory_fraction = 0.8
+    conf.gpu_options.per_process_gpu_memory_fraction = 0.6
     conf.gpu_options.allocator_type = 'BFC'
     conf.allow_soft_placement = True
     return conf

tensorpack/utils/concurrency.py

@@ -51,13 +51,3 @@ class EnqueueThread(threading.Thread):
             logger.exception("Exception in EnqueueThread:")
             self.queue.close(cancel_pending_enqueues=True)
             self.coord.request_stop()
-
-@contextmanager
-def coordinator_guard(sess, coord):
-    try:
-        yield
-    except (KeyboardInterrupt, Exception) as e:
-        raise
-    finally:
-        coord.request_stop()
-        sess.close()