update about NVML and environment

docs/tutorial/philosophy/dataflow.md

@@ -160,7 +160,8 @@ or when you need to filter your data on the fly.
    but inefficient for generic data type or numpy arrays.
    Also, its implementation [does not always clean up the subprocesses correctly](https://github.com/pytorch/pytorch/issues/16608).
 
-PyTorch starts to improve on these bad assumptions (e.g., with [IterableDataset](https://github.com/pytorch/pytorch/pull/19228)).
+PyTorch starts to improve on bad assumptions 1-3, (e.g., with IterableDataset).
+But the interface still bears the history of these assumptions.
 On the other hand, DataFlow:
 
 1. Is an iterator, not necessarily has a length or can be indexed. This is more generic.

examples/A3C-Gym/train-atari.py

@@ -173,7 +173,7 @@ class MySimulatorMaster(SimulatorMaster, Callback):
             try:
                 distrib, value = outputs.result()
             except CancelledError:
-                logger.info("Client {} cancelled.".format(client.ident))
+                logger.info("Client {} cancelled.".format(client.ident.decode('utf-8')))
                 return
             assert np.all(np.isfinite(distrib)), distrib
             action = np.random.choice(len(distrib), p=distrib)

examples/FasterRCNN/README.md

@@ -116,7 +116,7 @@ Performance in [Detectron](https://github.com/facebookresearch/Detectron/) can b
  We compare models that have identical training & inference cost between the two implementations.
  Their numbers can be different due to small implementation details.
 
- <a id="ft2">2</a>: Our mAP is __7 point__ better than the official model in
+ <a id="ft2">2</a>: This model has __7 point__ better mAP than the official model in
  [matterport/Mask_RCNN](https://github.com/matterport/Mask_RCNN/releases/tag/v2.0) which has the same architecture.
  Our implementation is also [5x faster](https://github.com/tensorpack/benchmarks/tree/master/MaskRCNN).
 

tensorpack/callbacks/prof.py

@@ -124,10 +124,10 @@ class GPUUtilizationTracker(Callback):
         Args:
             devices (list[int])
         """
+        try:
             with NVMLContext() as ctx:
                 devices = [ctx.device(i) for i in devices]
                 while True:
-                try:
                     evt.wait()  # start epoch
                     evt.clear()
                     if stop_evt.is_set():   # or on exit

tensorpack/tfutils/common.py

@@ -10,6 +10,7 @@ import psutil
 import tensorflow as tf
 import numpy as np
 
+import tensorpack
 from ..compat import tfv1
 from ..utils.argtools import graph_memoized
 from ..utils.utils import find_library_full_path as find_library
@@ -172,10 +173,10 @@ def collect_env_info():
     data = []
     data.append(("sys.platform", sys.platform))
     data.append(("Python", sys.version.replace("\n", "")))
-    data.append(("Tensorpack", __git_version__))
+    data.append(("Tensorpack", __git_version__ + " @" + os.path.dirname(tensorpack.__file__)))
     data.append(("Numpy", np.__version__))
 
-    data.append(("TensorFlow", tfv1.VERSION + "/" + tfv1.GIT_VERSION))
+    data.append(("TensorFlow", tfv1.VERSION + "/" + tfv1.GIT_VERSION + " @" + os.path.dirname(tf.__file__)))
     data.append(("TF Compiler Version", tfv1.COMPILER_VERSION))
     has_cuda = tf.test.is_built_with_cuda()
     data.append(("TF CUDA support", has_cuda))
@@ -221,7 +222,7 @@ def collect_env_info():
     # Other important dependencies:
     try:
         import horovod
-        data.append(("Horovod", horovod.__version__))
+        data.append(("Horovod", horovod.__version__ + " @" + os.path.dirname(horovod.__file__)))
     except ImportError:
         pass
 

tensorpack/utils/gpu.py

@@ -43,19 +43,29 @@ def get_num_gpu():
             logger.warn(message + "But TensorFlow was not built with CUDA support and could not use GPUs!")
         return ret
 
+    try:
+        # Use NVML to query device properties
+        with NVMLContext() as ctx:
+            nvml_num_dev = ctx.num_devices()
+    except Exception:
+        nvml_num_dev = None
+
     env = os.environ.get('CUDA_VISIBLE_DEVICES', None)
     if env:
-        return warn_return(len(env.split(',')), "Found non-empty CUDA_VISIBLE_DEVICES. ")
+        num_dev = len(env.split(','))
+        assert num_dev <= nvml_num_dev, \
+            "Only {} GPU(s) available, but CUDA_VISIBLE_DEVICES is set to {}".format(nvml_num_dev, env)
+        return warn_return(num_dev, "Found non-empty CUDA_VISIBLE_DEVICES. ")
+
     output, code = subproc_call("nvidia-smi -L", timeout=5)
     if code == 0:
         output = output.decode('utf-8')
         return warn_return(len(output.strip().split('\n')), "Found nvidia-smi. ")
-    try:
-        # Use NVML to query device properties
-        with NVMLContext() as ctx:
-            return warn_return(ctx.num_devices(), "NVML found nvidia devices. ")
-    except Exception:
-        # Fallback
+
+    if nvml_num_dev is not None:
+        return warn_return(nvml_num_dev, "NVML found nvidia devices. ")
+
+    # Fallback to TF
     logger.info("Loading local devices by TensorFlow ...")
 
     try:

tensorpack/utils/nvml.py

@@ -191,6 +191,8 @@ class NVMLContext(object):
         Returns:
             NvidiaDevice: single GPU device
         """
+        num_dev = self.num_devices()
+        assert idx < num_dev, "Cannot obtain device {}: NVML only found {} devices.".format(idx, num_dev)
 
         class GpuDevice(Structure):
             pass