Official TensorFlow for Jetson Nano!

jmiralles · May 10, 2020, 9:57pm

My post was not focused on a comparison with a laptop. The focus is that a got better numbers with a tensorflow that is not loading the cuda libraries than with one that is properly loading the cuda libraries.
As this seems against the logic, I was wondering if I may have done anything wrong or if there was any problem with tensorflow for jetpack 4.4.

Alexander_1 · May 11, 2020, 9:07am

What version of protobuf?

For x86_64, OSX/Linux, python 3.7/3.8, I had to install specific version of proto 3.8.0. Searching the tensorflow github PRs and issues, is usually my first stop for these issues.

graf2ar6z · May 11, 2020, 9:23am

Yes, I am now trying to build a new protobuf. The steps described by Nvidia in this thread install PB 3.0.6 on my machine. That is probably too old.

graf2ar6z · May 11, 2020, 11:13am

Unsuccessful so far. Neither Protobuf 3.6, 3.8 did fix anything.

505106035 · May 12, 2020, 8:19am

what the hell？
python3 tensorflow2.1+20.4 jetpack4.4

import tensorflow as tf
2020-05-12 16:13:47.706899: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.2
2020-05-12 16:13:50.404420: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libnvinfer.so.7
2020-05-12 16:13:50.406626: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libnvinfer_plugin.so.7
print(tf.version)
Traceback (most recent call last):
File “”, line 1, in
File “/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/util/module_wrapper.py”, line 192, in getattr
attr = getattr(self._tfmw_wrapped_module, name)
AttributeError: module ‘tensorflow’ has no attribute ‘version’

steffen.aus.karlsruhe · May 12, 2020, 8:59am

I think that Tensorflow for JP 4.4 was compiled with different flags than Tensorflow for JP 4.3.

I now get this (JP 4.4, tf_gpu-2.1.0+nv20.4-py3)

$ python3
Python 3.6.9 (default, Apr 18 2020, 01:56:04) 
[GCC 8.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow
2020-05-12 10:51:03.048952: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.2
2020-05-12 10:51:07.319165: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libnvinfer.so.7
2020-05-12 10:51:07.322609: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libnvinfer_plugin.so.7
>>> tensorflow.version
<module 'tensorflow_core._api.v1.version' from '/usr/local/lib/python3.6/dist-packages/tensorflow_core/_api/v1/version/__init__.py'>
>>> tensorflow.__version__
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/util/module_wrapper.py", line 192, in __getattr__
    attr = getattr(self._tfmw_wrapped_module, name)
AttributeError: module 'tensorflow' has no attribute '__version__'

Furthermore, a deep learning script that worked flawlessly in JP 4.3 with tensorflow-2.1.0+nv20.3-cp36-cp36m-linux_aarch64.whl now fails:

Traceback (most recent call last):
  File "./mnist_dcgan_tf2.py", line 162, in <module>
    train(train_dataset, EPOCHS)
  File "./mnist_dcgan_tf2.py", line 148, in train
    for image_batch in dataset:
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/data/ops/dataset_ops.py", line 2443, in __iter__
    return iter(self._dataset)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/data/ops/dataset_ops.py", line 426, in __iter__
    raise RuntimeError("__iter__() is only supported inside of tf.function "
RuntimeError: __iter__() is only supported inside of tf.function or when eager execution is enabled.

It only works now under JP 4.4 when I make those changes in the script:

import tensorflow.compat.v2 as tf
import tensorflow.compat.v2.keras as keras
tf.enable_v2_behavior()

This looks like the current Tensorflow for JP 4.4 was compiled with --config=v1 flag., as V2 behaviour seems to be disabled in default.

Can the maintainers of the TF packages check if this was the case?

Alexander_1 · May 13, 2020, 8:43pm

I was getting similar errors from the wheel. I’m just going to build from source.

lycaass · May 13, 2020, 9:26pm

This works:

tf.version.VERSION
‘2.1.0’

mictiemix · May 26, 2020, 9:41am

Is there are chance to get an updated version of TF 2.1.1?
Version 2.1.1 ‘Fixes a versioning bug which causes Keras layers from TF 1.x to be used instead of those from TF 2.x’

I tried to build TF from source, but failed (it seems TF2.1.1 requires cuDNN7.6, but Jetpack 4.4 installed TF2.1 with cuDNN 8.0). Downgrading cuDNN 8.0 to 7.6 gives a ‘broken package’ …

(using Jetson Nano, Ubuntu 18.04, Jetpack4.4)

taro · May 27, 2020, 2:40am

Dear all.

I installed Jetpack 4.4 and Nvidia TensorFlow 2.1.0.
I have a question for tf.convert_to_tensor function.

In my Jetson Nano, following return without numpy object.
>>> import numpy as np
>>> import tensorflow as tf
>>> a = np.array([1,2])
>>> a
array([1, 2])
>>> t = tf.convert_to_tensor(a)
>>> t
<tf.Tensor ‘Const:0’ shape=(2,) dtype=int64>

But tensorflow-gpu 2.0.0 and other system’s tensorflow 2.1.0 includes numpy.
For example, my mac return following.

>>> import numpy as np
>>> import tensorflow as tf
>>> a = np.array([1,2])
>>> a
array([1, 2])
>>> t = tf.convert_to_tensor(a)
>>> t
<tf.Tensor: shape=(2,), dtype=int64, numpy=array([1, 2])>

I have a trouble with Nvidia’s TensorFlow 2.1.0 because I test TensorFlow Lite example but tf.lite.Interpreter.set_tensor raise ValueError.

github.com

tensorflow/tensorflow/blob/v2.1.0/tensorflow/lite/python/interpreter_wrapper/interpreter_wrapper.cc#L308


      
            TFLITE_PY_ENSURE_VALID_INTERPRETER();
            TFLITE_PY_TENSOR_BOUNDS_CHECK(i);
            const TfLiteTensor* tensor = interpreter_->tensor(i);
            return PyTupleFromQuantizationParam(tensor->params);
          }
          
          
PyObject* InterpreterWrapper::SetTensor(int i, PyObject* value) {
            TFLITE_PY_ENSURE_VALID_INTERPRETER();
            TFLITE_PY_TENSOR_BOUNDS_CHECK(i);
          
          
  std::unique_ptr<PyObject, PyDecrefDeleter> array_safe(
                PyArray_FromAny(value, nullptr, 0, 0, NPY_ARRAY_CARRAY, nullptr));
            if (!array_safe) {
              PyErr_SetString(PyExc_ValueError,
                              "Failed to convert value into readable tensor.");
              return nullptr;
            }
          
          
  PyArrayObject* array = reinterpret_cast<PyArrayObject*>(array_safe.get());
            TfLiteTensor* tensor = interpreter_->tensor(i);

mictiemix · May 30, 2020, 8:26am

BTW, I was finally able to compile (app 40hours!!!) and install TF2.1.1 from source and install on my Nano
Seems much more stable compared to 2.1

berykubik · June 1, 2020, 2:51pm

@AastaLLL would it be possible to release a new TF 2.1 wheel for JetPack 4.4 that would be correctly compiled? I’m also having trouble with running TF 2 on Jetson Nano with JetPack 4.4.

mictiemix · June 4, 2020, 9:35am

@berykubik:
in the meantime you can try my build: Tensorflow 2.x on Jetson nano - #19 by mictiemix

quangminhbui.dev · June 4, 2020, 1:48pm

I have some problems with tf 2.1.0 on AGX Xavier.
I installed tf 2.1.0 with “sudo pip3 install --pre --extra-index-url https://developer.download.nvidia.com/compute/redist/jp/v44 tensorflow”

But i feel a little bit strange.

I checked tf version with tensorflow.version but it doesnt have version.
I can’t use function .numpy() to convert tensor to numpy array which is possible in tf2.

mictiemix · June 4, 2020, 2:26pm

@quangminhbui.dev
See the bottom if this post: Official TensorFlow for Jetson Nano! - #127 by steffen.aus.karlsruhe

quangminhbui.dev · June 5, 2020, 3:03am

May it work on AGX Xavier?

mictiemix · June 5, 2020, 3:21am

Most likely not due to different Cuda architecture (5.3 vs 7.2)

pieterjan.kuyten · June 5, 2020, 1:56pm

I’m also running into this version problem with tensorflow 2.1.0 on my Jetson Xavier NX.

NVIDIA, please release a fixed version ASAP as this is really messing up my use of the module.

fabian.s.mack · June 23, 2020, 2:23pm

Also have the same problem for the Jetson NX.
The command tf.version.VERSION works but it seams to mess up other stuff.
Is there somthing new?

pieterjan.kuyten · June 23, 2020, 2:25pm

We are still waiting for a new Jetpack 4.4 compatible build.