minsaver , jsonstat

examples/Inception/inception-bn.py

@@ -35,7 +35,7 @@ class Model(ModelDesc):
 
     def _get_cost(self, input_vars, is_training):
         image, label = input_vars
-        image = image / 128.0 - 1
+        image = image / 128.0
 
         def inception(name, x, nr1x1, nr3x3r, nr3x3, nr233r, nr233, nrpool, pooltype):
             stride = 2 if nr1x1 == 0 else 1
@@ -46,7 +46,6 @@ class Model(ModelDesc):
                 x2 = Conv2D('conv3x3r', x, nr3x3r, 1)
                 outs.append(Conv2D('conv3x3', x2, nr3x3, 3, stride=stride))
 
-
                 x3 = Conv2D('conv233r', x, nr233r, 1)
                 x3 = Conv2D('conv233a', x3, nr233, 3)
                 outs.append(Conv2D('conv233b', x3, nr233, 3, stride=stride))
@@ -133,6 +132,8 @@ def get_data(train_or_test):
     if isTrain:
         augmentors = [
             imgaug.Resize((256, 256)),
+            imgaug.Brightness(30, False),
+            imgaug.Contrast((0.8,1.2), True),
             imgaug.MapImage(lambda x: x - pp_mean),
             imgaug.RandomCrop((224, 224)),
             imgaug.Flip(horiz=True),
@@ -172,9 +173,9 @@ def get_config():
                 ClassificationError('wrong-top5', 'val-top5-error')]),
             #HumanHyperParamSetter('learning_rate', 'hyper-googlenet.txt')
             ScheduledHyperParamSetter('learning_rate',
-                                      [(8, 0.03), (13, 0.02), (21, 5e-3),
-                                       (28, 3e-3), (33, 1e-3), (44, 5e-4),
-                                       (49, 1e-4), (59, 2e-5)])
+                                      [(8, 0.03), (13, 0.02), (16, 5e-3),
+                                       (18, 3e-3), (24, 1e-3), (26, 2e-4),
+                                       (28, 5e-5) ])
         ]),
         session_config=sess_config,
         model=Model(),

tensorpack/callbacks/common.py

@@ -3,7 +3,7 @@
 # Author: Yuxin Wu <ppwwyyxx@gmail.com>
 
 import tensorflow as tf
-import os
+import os, shutil
 import re
 
 from .base import Callback
@@ -56,6 +56,24 @@ class ModelSaver(Callback):
 class MinSaver(Callback):
     def __init__(self, monitor_stat):
         self.monitor_stat = monitor_stat
+        self.min = None
+
+    def _get_stat(self):
+        return self.trainer.stat_holder.get_stat_now(self.monitor_stat)
 
     def _trigger_epoch(self):
-        pass
+        if self.min is None or self._get_stat() < self.min:
+            self.min = self._get_stat()
+            self._save()
+
+    def _save(self):
+        ckpt = tf.train.get_checkpoint_state(logger.LOG_DIR)
+        if ckpt is None:
+            raise RuntimeError(
+                "Cannot find a checkpoint state. Do you forget to use ModelSaver before MinSaver?")
+        path = chpt.model_checkpoint_path
+        newname = os.path.join(logger.LOG_DIR, 'min_' + self.monitor_stat)
+        shutil.copy(path, newname)
+        logger.info("Model with minimum {} saved.".format(self.monitor_stat))
+
+

tensorpack/callbacks/summary.py

@@ -6,7 +6,7 @@ import tensorflow as tf
 import re
 import os
 import operator
-import pickle
+import json
 
 from .base import Callback
 from ..utils import *
@@ -25,11 +25,11 @@ class StatHolder(object):
         self.stat_now = {}
 
         self.log_dir = log_dir
-        self.filename = os.path.join(log_dir, 'stat.pkl')
+        self.filename = os.path.join(log_dir, 'stat.json')
         if os.path.isfile(self.filename):
             logger.info("Loading stats from {}...".format(self.filename))
             with open(self.filename) as f:
-                self.stat_history = pickle.load(f)
+                self.stat_history = json.load(f)
         else:
             self.stat_history = []
 
@@ -47,6 +47,12 @@ class StatHolder(object):
         """
         self.print_tag = None if print_tag is None else set(print_tag)
 
+    def get_stat_now(self, k):
+        """
+        Return the value of a stat in the current epoch.
+        """
+        return self.stat_now[k]
+
     def finalize(self):
         """
         Called after finishing adding stats. Will print and write stats to disk.
@@ -64,7 +70,7 @@ class StatHolder(object):
     def _write_stat(self):
         tmp_filename = self.filename + '.tmp'
         with open(tmp_filename, 'wb') as f:
-            pickle.dump(self.stat_history, f)
+            json.dump(self.stat_history, f)
         os.rename(tmp_filename, self.filename)
 
 class StatPrinter(Callback):

tensorpack/train/base.py

@@ -19,6 +19,13 @@ __all__ = ['Trainer']
 class Trainer(object):
     """
     Base class for a trainer.
+
+    Available Attritbutes:
+        stat_holder: a `StatHolder` instance
+        summary_writer: a `tf.SummaryWriter`
+        config: a `TrainConfig`
+        model: a `ModelDesc`
+        global_step: a `int`
     """
     __metaclass__ = ABCMeta