Merge pull request #12 from psteinb/try-tf-1.6

Try tf 1.6

Author: Peter Steinbach

models/__init__.py

@@ -121,7 +121,7 @@ def versions(self):
 
         value = ""
 
-        if "keras" in self.backend.lower():
+        if self.backend.lower().startswith("keras"):
 
             import keras
             from keras import backend as K
@@ -143,7 +143,7 @@ def versions(self):
 
         else:
 
-            if "tensorflow" in self.backend.lower():
+            if self.backend.lower() == "tensorflow" or self.backend.lower() == "tf":
                 import tensorflow as tf
                 value = "tensorflow:{ver}".format(ver=tf.__version__)
             else:

models/resnet.py

@@ -33,7 +33,7 @@ def can_train():
         warnings.simplefilter(action='ignore', category=FutureWarning)
 
         from tensorflow import __version__ as tfv
-        required = "1.7.0"
+        required = "1.6.0"
 
         #only require major and minor release number as the patch number may contain 'rc' etc
         if versiontuple(tfv,2) >= versiontuple(required,2):
@@ -111,4 +111,9 @@ def train(train, test, datafraction, opts):
     logging.info('handing over \n >> %s \n >>  %s',flags,opts)
     history, timings = run_loop.resnet_main(flags, cfmain.cifar10_model_fn, cfmain.input_fn, opts)
 
+    if not opts['checkpoint_epochs']:
+        logging.info("unable to ensure pure no-checkpoint behavior with resnet in pure tensorflow, removing result directory")
+        import shutil
+        shutil.rmtree(model_dir)
+
     return history, timings, { 'num_weights' : None }

models/tf_details/resnet_details.py

@@ -327,6 +327,7 @@ def resnet_main(flags, model_function, input_function, opts = None):
     logging.warning("batch sizes differ in model %i %s", flags.batch_size, opts["batch_size"])
 
   if ngpus > 1:
+    steps_per_epoch -= 1
     validate_batch_size_for_multi_gpu(flags.batch_size)
     # There are two steps required if using multi-GPU: (1) wrap the model_fn,
     # and (2) wrap the optimizer. The first happens here, and (2) happens
@@ -407,12 +408,6 @@ def input_fn_eval():
     validation_results = classifier.evaluate(input_fn=input_fn_eval,
                                        steps=flags.max_train_steps)
 
-
-    # for (k,v) in train_hooks["CaptureTensorsHook"].captured.items():
-    #   print(">> ",k,v[:5],v[-2:])
-
-    #epoch_times.extend(train_hooks["TimePerEpochHook"].epoch_durations)
-
     for k in validation_results.keys():
       if "global_step" in k:
         continue

models/tf_details/resnet_run_loop.py

@@ -50,7 +50,7 @@ def add(self,other):
 class TimePerEpochHook(tf.train.SessionRunHook):
   def __init__(self,
                every_n_steps,
-               warm_steps=0):
+               warm_steps=-1):
 
     self.every_n_steps = every_n_steps
     logging.info("TimePerEpochHook triggering every %i steps",every_n_steps)
@@ -112,8 +112,8 @@ def after_run(self, run_context, run_values):  # pylint: disable=unused-argument
     global_step = run_values.results
     sess = run_context.session
 
-
-    if self._timer.should_trigger_for_step(global_step) and global_step > self._warm_steps:
+    #if self._timer.should_trigger_for_step(global_step) and global_step > self._warm_steps:
+    if self._step % self.every_n_steps == 0:
       elapsed_time, elapsed_steps = self._timer.update_last_triggered_step(
           global_step)
       if elapsed_time is not None:
@@ -124,6 +124,8 @@ def after_run(self, run_context, run_values):  # pylint: disable=unused-argument
         tf.logging.info('Epoch [%g steps]: %g (%s)', self._total_steps,self._epoch_train_time,str(self.epoch_durations))
 
         self._epoch_train_time = 0
+      else:
+        logging.warning("step %i, elapsed_time is None!", global_step)