misc fixes..

examples/HED/README.md

@@ -38,7 +38,6 @@ Models I trained can be downloaded [here](https://drive.google.com/drive/folders
 To view the loss curve:
 ```bash
 cat train_log/hed/stat.json | jq '.[] |
-[.xentropy1,.xentropy2,.xentropy3,.xentropy4,.xentropy5,.xentropy6] |
-map(tostring) | join("\t") | .' -r | \
+"\(.xentropy1)\t\(.xentropy2)\t\(.xentropy3)\t\(.xentropy4)\t\(.xentropy5)\t\(.xentropy6)"' -r | \
 				../../scripts/plot-point.py --legend 1,2,3,4,5,final --decay 0.8
 ```

examples/ResNet/imagenet-resnet.py

@@ -18,8 +18,8 @@ from tensorpack.tfutils.symbolic_functions import *
 from tensorpack.tfutils.summary import *
 
 """
-Training code of Pre-Activation version of ResNet on ImageNet. Work In Progress.
-Top1 error is now about 0.5% higher than fb.resnet.torch.
+Training code of Pre-Activation version of ResNet on ImageNet.
+Mainly follow the setup in fb.resnet.torch
 """
 
 

tensorpack/callbacks/param.py

@@ -207,8 +207,9 @@ class StatMonitorParamSetter(HyperParamSetter):
             ):
         """
         Change param by `new_value = value_func(old_value)`,
-        if `stat_name` wasn't decreasing > threshold  times in the lastest
-            last_k times of statistics update.
+        if :
+            min(stats) >= stats[0] - threshold, where
+            stats = [`stat_nam` in latest `last_k` epochs]
 
         For example, if error wasn't decreasing, anneal the learning rate:
             StatMonitorParamSetter('learning_rate', 'val-error', lambda x: x * 0.2)
@@ -235,11 +236,11 @@ class StatMonitorParamSetter(HyperParamSetter):
         hist_first = hist[0]
         if not self.reverse:
             hist_min = min(hist)
-            if hist_min <= hist_first - self.threshold: # small enough
+            if hist_min < hist_first - self.threshold: # small enough
                 return None
         else:
             hist_max = max(hist)
-            if hist_max >= hist_first + self.threshold: # large enough
+            if hist_max > hist_first + self.threshold: # large enough
                 return None
         self.last_changed_epoch = self.epoch_num
         return self.value_func(self.get_current_value())

tensorpack/dataflow/dftools.py

@@ -4,7 +4,7 @@
 
 import sys, os
 import cv2
-import multiprocessing
+import multiprocessing as mp
 
 from ..utils.concurrency import DIE
 from ..utils.fs import mkdir_p
@@ -44,8 +44,8 @@ def dataflow_to_process_queue(ds, size, nr_consumer):
     :returns: (queue, process). The process will take data from `ds` to fill
         the queue once you start it. Each element is (task_id, dp).
     """
-    q = multiprocessing.Queue(size)
-    class EnqueProc(multiprocessing.Process):
+    q = mp.Queue(size)
+    class EnqueProc(mp.Process):
         def __init__(self, ds, q, nr_consumer):
             super(EnqueProc, self).__init__()
             self.ds = ds

tensorpack/predict/dataset.py

@@ -101,7 +101,7 @@ class MultiProcessDatasetPredictor(DatasetPredictorBase):
         # setup all the procs
         self.inqueue_proc.start()
         for p, gpuid in zip(self.workers, gpus):
-            if gpuid == '':
+            if gpuid == '-1':
                 logger.info("Worker {} uses CPU".format(p.idx))
             else:
                 logger.info("Worker {} uses GPU {}".format(p.idx, gpuid))

tensorpack/tfutils/sessinit.py

@@ -62,7 +62,7 @@ class SaverRestore(SessionInit):
         if os.path.basename(model_path) == 'checkpoint':
             model_path = tf.train.latest_checkpoint(os.path.dirname(model_path))
             # to be consistent with either v1 or v2
-        assert os.path.isfile(model_path) or os.path.isfile(model_path + '.index')
+        assert os.path.isfile(model_path) or os.path.isfile(model_path + '.index'), model_path
         self.set_path(model_path)
         self.prefix = prefix