I'm doubting whether tensorflow is correctly configured on my gpu box, since it's about 100x slower per iteration to train a simple linear regression model (batchsize = 32, 1500 input features, 150 output variables) on my fancy gpu machine than on my laptop.
I'm using a Titan X, with a modern cpu, etc. nvidia-smi says that I'm only at 10% gpu utilization, but I expect that's because of the small batchsizes. I'm not using a feed_dict to move data into the computation graph. Everything is coming via a tf.decode_csv and tf.train.shuffle_batch.
Does anyone have any recommendations for how to easily test whether my install is correct? Are there any simple speed benchmarks? The speed difference between my laptop and the gpu machine is so dramatic that I'm expecting that things aren't configured properly.
Try tensorflow/tensorflow/models/image/mnist/convolutional.py, that'll print per-step timing.
On Tesla K40c that should get about 16 ms per step, while about 120 ms for CPU-only on my 3 year old machine
Edit: This moved to the models repositories: https://github.com/tensorflow/models/blob/master/tutorials/image/mnist/convolutional.py.
The convolutional.py file is now at models/tutorials/image/mnist/convolutional.py
Extending Yaroslavs answer:
Here is how to do the entire testing process (CUDA and cudNN installed already)
git clone https://github.com/tensorflow/models.git
Create a Virtual Environment for tensorflow and install tensorflow
virtualenv --system-site-packages -p python3 tf-venv3
source tf-venv3/bin/activate
pip install --upgrade pip
pip install --upgrade tensorflow-gpu
Run the model within your Virtual Environment
python models/tutorials/image/mnist/convolutional.py
My GTX 1070 needs ~5ms per step
Note: On Geforce 1050 Ti it takes ~10ms per step
The answer given by Yaroslav and extended by Patrice may be simply extended once again to work with TensorFlow v2.
Just in line 35 of convolutional.py, instead of:
import tensorflow as tf
type in:
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
The file convolutional.py was removed from the tensorflow repo but it can still be found in the repo history.
Related
I tried
%tensorflow_version 1.15
I used this code a couple days ago, but it doesn't work anymore since today.
The outcomes are
ValueError Traceback (most recent call last)
<ipython-input-6-24c52e77c597> in <module>()
----> 1 get_ipython().magic('tensorflow_version 1.15')
2 frames
/usr/local/lib/python3.7/dist-packages/IPython/core/interactiveshell.py in magic(self, arg_s)
2158 magic_name, _, magic_arg_s = arg_s.partition(' ')
2159 magic_name = magic_name.lstrip(prefilter.ESC_MAGIC)
-> 2160 return self.run_line_magic(magic_name, magic_arg_s)
2161
2162 #-------------------------------------------------------------------------
/usr/local/lib/python3.7/dist-packages/IPython/core/interactiveshell.py in run_line_magic(self, magic_name, line)
2079 kwargs['local_ns'] = sys._getframe(stack_depth).f_locals
2080 with self.builtin_trap:
-> 2081 result = fn(*args,**kwargs)
2082 return result
2083
/usr/local/lib/python3.7/dist-packages/google/colab/_tensorflow_magics.py in _tensorflow_version(line)
39
40 Your notebook should be updated to use Tensorflow 2.
---> 41 See the guide at https://www.tensorflow.org/guide/migrate#migrate-from-tensorflow-1x-to-tensorflow-2."""
42 ))
43
ValueError: Tensorflow 1 is unsupported in Colab.
Your notebook should be updated to use Tensorflow 2.
See the guide at https://www.tensorflow.org/guide/migrate#migrate-from-tensorflow-1x-to-tensorflow-2.
Is there any method I can try to fix it or does it means that colab will not support tensorflow 1.x anymore?
Google Colab removed support for Tensorflow 1, and it is not possible to use %tensorflow_version 1.x magic anymore. You must remove this instruction from your code if you have it.
Also the default python version as I update this answer is python 3.8 which is not compatible with tensorflow 1.x.
To make everything work you first have to downgrade python. Python 3.6 should work. As suggested by #s-abbaasi here's a guide on how to do so:
%%bash
MINICONDA_INSTALLER_SCRIPT=Miniconda3-4.5.4-Linux-x86_64.sh
MINICONDA_PREFIX=/usr/local
wget https://repo.continuum.io/miniconda/$MINICONDA_INSTALLER_SCRIPT
chmod +x $MINICONDA_INSTALLER_SCRIPT
./$MINICONDA_INSTALLER_SCRIPT -b -f -p $MINICONDA_PREFIX
Then add to path:
import sys
_ = (sys.path.append("/usr/local/lib/python3.6/site-packages"))
At this point you can manually uninstall and re-install tensorflow through pip:
!pip uninstall tensorflow
!pip install tensorflow-gpu==1.15
Doing just so I sometimes encounter some errors due to the Cuda version. If this happens to you, you can execute the following:
!apt install --allow-change-held-packages libcudnn7=7.4.1.5-1+cuda10.0
The most appropriate version of cuda and libcudnn to use with the tensorflow version you want to install can be found here.
The versions available of libcudnn can be found with the following command:
!apt list -a libcudnn7
This will list all libcudnn7 versions available.
I was having the same problems while trying to use StyleGAN2-ADA, which only supports TensorFlow 1.
I found out that unfortunately Google Colab removed support for TensorFlow 1 in their latest release of 2022/8/11.
'Removed support for TensorFlow 1'
You can find more information in their notebook Release-Notes: https://colab.research.google.com/notebooks/relnotes.ipynb
There are some answers to this question in a Python environment, but the solutions did not work for my RStudio environment. Here is my code:
library(keras)
library(tensorflow)
use_condaenv("tf")
train_dir = "C:/training_images/"
train_datagen <- image_data_generator(rescale = 1/255)
validation_datagen <- image_data_generator(rescale = 1/255)
train_generator <- flow_images_from_directory(
train_dir,
train_datagen,
target_size = c(150, 150),
batch_size = 20,
class_mode = "binary"
)
batch <- generator_next(train_generator)
The code works until the last "batch" line where it explodes like this:
Error in py_iter_next(it, completed) :
ImportError: Could not import PIL.Image. The use of `load_img` requires PIL.
Detailed traceback:
File "C:\Users\mory3\ANACON~1\envs\tf\lib\site-packages\keras_preprocessing\image\iterator.py", line 104, in __next__
return self.next(*args, **kwargs)
File "C:\Users\mory3\ANACON~1\envs\tf\lib\site-packages\keras_preprocessing\image\iterator.py", line 116, in next
return self._get_batches_of_transformed_samples(index_array)
File "C:\Users\mory3\ANACON~1\envs\tf\lib\site-packages\keras_preprocessing\image\iterator.py", line 230, in _get_batches_of_transformed_samples
interpolation=self.interpolation)
File "C:\Users\mory3\ANACON~1\envs\tf\lib\site-packages\keras_preprocessing\image\utils.py", line 108, in load_img
raise ImportError('Could not import PIL.Image. '
R version 3.6.1
Conda version 4.7
Python version 3.7
I had this same problem
After a few hours of looking, I came up with a solution that worked for me.
I used this code for solving the PIL problem. I tried using anaconda prompt but this code worked in r for me...
reticulate::py_install("pillow",env=tf)
I came up with this error next...
loaded runtime CuDNN library: 7.4.2 but source was compiled with: 7.6.0.
Make sure you have the correct cudnn version installed. For me it was CUDA 10 with 7.6.0 cudnn with 10. The output of the error will tell you which one to use.
Make sure you have cleaned any extra path variables that are in your environmental variables from installing previous versions.
I'm using windows 10
gpu = GeForce GTX 1060 with Max-Q Design
R - 3.6.1
tensorflow = 1.13
python = 3.7
anaconda = Anaconda3–2019.03-Windows-x86_64.exe
I ended up uninstalling Anaconda altogether, which made troubleshooting the remaining errors in the Python connection to R much simpler.
I had same problem with "Deep Learning with R" CNN example on Win7. I solved it like this:
I added Anaconda3 paths to PATH. In my case it was Windows so paths were like that:
C:\Anaconda3\Scripts;C:\Anaconda3\Library\bin By default there were no paths to conda in $PATH.
installed pillow (it contains PIL) to python with:
pip install pillow
configured r-reticulate.
This answer Could not import PIL.Image even if Pillow already installed? helped me. I had pillow already but conda environment wasn't configured properly so pillow wasn't visible.
Also install Nvidia CUDA if you don't have it - you need it too for tensorflow.
Anyone know if Tensorflow Lite has GPU support for Python? I've seen guides for Android and iOS, but I haven't come across anything about Python. If tensorflow-gpu is installed and tensorflow.lite.python.interpreter is imported, will GPU be used automatically?
According to this thread, it is not.
one solution is to convert tflite to onnx and use onnxruntime-gpu
convert to onnx with https://github.com/onnx/tensorflow-onnx:
pip install tf2onnx
python3 -m tf2onnx.convert --opset 11 --tflite path/to/model.tflite --output path/to/model.onnx
then pip install onnxruntime-gpu
and run like:
session = onnxruntime.InferenceSession(('/path/to/model.onnx'))
raw_output = self.detection_session.run(['output_name'], {'input_name': img})
you can get the input and output names by:
for i in range(len(session.get_inputs)):
print(session.get_inputs()[i].name)
and the same but replace 'get_inputs' with 'get_outputs'
You can force the computation to take place on a GPU:
import tensorflow as tf
with tf.device('/gpu:0'):
for i in range(10):
t = np.random.randint(len(x_test) )
...
Hope this helps.
I'm new to Tensorflow.
I am using a 64 bit version of Windows 10 and I would like to install Tensorflow for the CPU.
I don't remember the exact steps that I followed to install it, however when I checked for the installation using:
import tensorflow as tf
hello = tf.constant('Hello, TensorFlow!')
sess = tf.Session()
print(sess.run(hello))
I have the following output:
2017-10-18 09:56:21.656601: W C:\tf_jenkins\home\workspace\rel-win\M\windows\PY\36\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2017-10-18 09:56:21.656984: W C:\tf_jenkins\home\workspace\rel-win\M\windows\PY\36\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.
b'Hello, TensorFlow!'
I am running python in Sublime Text 3 using the package SublimeREPL.
I tried to search these errors and found out that it means that the tensorflow is built without these instructions which could improve performances for the CPU. I also found the code to hide these warnings, but I actually I want to use these instructions.
The code that I found that enables this is:
bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-msse4.2 --copt=-msse4.1 --copt=-msse3 --copt=-mfma -k //tensorflow/tools/pip_package:build_pip_package
but I got this output:
ERROR: Skipping '//tensorflow/tools/pip_package:build_pip_package': no such package 'tensorflow/tools/pip_package': BUILD file not found on package path.
WARNING: Target pattern parsing failed. Continuing anyway.
INFO: Found 0 targets...
ERROR: command succeeded, but there were errors parsing the target pattern.
INFO: Elapsed time: 8,147s, Critical Path: 0,02s
How can I solve this problem?
Lastly, I don't understand what pip, wheel and bazel are so I need a step by step instructions.
Thank you a lot!
if you want to download TensorFlow source, compile+install, use this link. If you want to download binaries, then use this link.
Is there a version of TensorFlow for 32-bit Linux? I only see the 64-bit wheel available, and didn't find anything about it on the site.
We have only tested the TensorFlow distribution on 64-bit Linux and Mac OS X, and distribute binary packages for those platforms only. Try following the source installation instructions to build a version for your platform.
EDIT: One user has published instructions for running TensorFlow on a 32-bit ARM processor, which is promising for other 32-bit architectures. These instructions may have useful pointers for getting TensorFlow and Bazel to work in a 32-bit environment.
I've built a CPU-only version of TensorFlow on 32-bit Ubuntu (16.04.1 Xubuntu). It went a lot more smoothly than anticipated, for such a complex library that doesn't support 32-bit architectures officially.
It can be done by following a subset of the intersection of these two guides:
November 2015 walkthrough about Jetson TK1.
November 2016 walkthrough about Jetson TX1.
If I haven't forgotten anything, here are the steps I've taken:
Install Oracle Java 8 JDK:
$ sudo apt-get remove icedtea-8-plugin #This is just in case
$ sudo add-apt-repository ppa:webupd8team/java
$ sudo apt-get update
$ sudo apt-get install oracle-java8-installer
(This is all you need in a pristine Xubuntu install, but google the above keywords otherwise, to read about selecting a default JRE and javac.)
Dependencies:
sudo apt-get update
sudo apt-get install git zip unzip swig python-numpy python-dev python-pip python-wheel
pip install --upgrade pip
Following the instructions that come with Bazel, download a Bazel source zip (I got bazel-0.4.3-dist.zip), make a directory like ~/tf/bazel/ and unzip it there.
I was getting an OutOfMemoryError during the following build, but this fix took care of it (i.e. adding the -J-Xmx512m for the bootstrap build).
Call bash ./compile.sh, and wait for a long time (overnight for me, but see the remarks at the end).
$ git clone -b r0.12 https://github.com/tensorflow/tensorflow
This seems like the only change to the source code that was necessary!
$ cd tensorflow
$ grep -Rl "lib64"| xargs sed -i 's/lib64/lib/g'
Then $ ./configure and say no to everything. (Accept defaults where relevant.)
The following took quite a few hours with my setup:
$ bazel build -c opt --jobs 1 --local_resources 1024,0.5,1.0 --verbose_failures //tensorflow/tools/pip_package:build_pip_package
$ bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
$ pip install --user /tmp/tensorflow_pkg/ten<Press TAB here>
To see that it's installed, see if it works on the TensorFlow Beginners tutorial. I use jupyter qtconsole (i.e. the new name of IPython). Run the code in the mnist_softmax.py. It should take little time even on very limited machines.
For some reason, TensorFlow's guide to building from source doesn't suggest running the unit tests:
$ bazel test //tensorflow/...
(Yes, type in the ellipses.)
Though I couldn't run them — it spent 19 hours trying to link libtensorflow_cc.so, and then something killed the linker. This was with half a core and 1536 MB memory limit. Maybe someone else, with a larger machine, can report on how the unit tests go.
Why didn't we need to do the other things mentioned in those two walkthroughs? Firstly, most of that work is about taking care of GPU interfacing. Secondly, both Bazel and TensorFlow have become more self-contained since the first of those walkthroughs was written.
Note that the above settings provided to Bazel for the build are very conservative (1024 MB RAM, half a core, one job at a time), because I'm running this through VirtualBox using a single core of a $200 netbook of the type that Intel makes for disadvantaged kids in Venezuela, Pakistan and Nigeria. (By the way, if you do this, make sure the virtual HDD is 20 GB at the very least — trying to build the unit tests above took about 5 GB of space.) The build of the wheel took almost 20 hours and the modest deep CNN from the second tutorial, which is quoted to take up to half an hour to run on modern desktop CPUs, takes about 80 hours under this setup. One might wonder why I don't get a desktop, but the truth is that actual training with TensorFlow only makes sense on a high-end GPU (or a bunch thereof), and when we can hire an AWS spot instance with such a GPU for about 10 cents an hour without commitment and on a workable ad-hoc basis, it doesn't make a lot of sense to be training elsewhere. The 480000% speed-up is really noticeable. On the other hand, the convenience of having a local installation is well worth going through a process such as above.
It appears that Google does not yet support tensorflow on 32-bit machines.
On a 32-bit machine running Centos 6.5,the following error is received after the "import tensorflow as tf" command:
ImportError: tensorflow/python/_pywrap_tensorflow.so: wrong ELF class: ELFCLASS64
Until Google distributes a 32-bit version of tensorflow, I also recommend building tensorflow from source as specified here.
I have used the information from the responses to this question and generated a detailed instructions list to compile and install tensorflow in a 32 bits linux system.
The latest version of the instructions is available in github at: tensorflow-32-bits-linux
Instructions to install Tensorflow in a 32 bits linux system
I used the following steps to install tensorflow in a old Asus Eee-Pc 1000H. Granted, it has been upgraded from the original 1 GB of RAM and an 80 GB HDD, to 2 GB of RAM and to 480 GB of SSD storage.
I tested the these instructions with the following OS versions and worked without problems:
* Xubuntu 16.04.6 Xenial Xerus 32 bits.
* Xubuntu 18.04.3 Bionic Beaver 32 bits.
* Debian 9.11 Stretch 32 bits.
Choose a convenient linux system
I have tested both the Ubuntu 16.04 (Xenial) and Debian 9.11 (Stretch) systems with 2 GB of RAM.
I set up the system to have 4 GB of SWAP space. With only 1 GB of SWAP, some compilations failed.
It's critical that the distribution has the version 8 of the Java SDK: openjdk-8-jdk
Install the Java 8 SDK and build tools
sudo apt-get update
sudo apt-get install openjdk-8-jdk
sudo apt-get install git zip unzip autoconf automake libtool curl zlib1g-dev swig build-essential
Install Python libraries
Next, we install python 3 development libraries and the keras module that will be required by tensorflow.
sudo apt-get install python3-dev python3-pip python3-wheel
sudo python3 -m pip install --upgrade pip
python3 -m pip install --user keras
You can use eithr python 3 or python 2 and compile tensorflow for that version.
Install and compile Bazel from sources
We need the source code bazel 0.19.2 distribution. We can obtain it and install in a new folder.
cd $HOME
wget https://github.com/bazelbuild/bazel/releases/download/0.19.2/bazel-0.19.2-dist.zip
mkdir Bazel-0-19.2
cd Bazel-0-19.2
unzip ../bazel-0.19.2-dist.zip
Before compiling, we need to remove line 30 of ./src/tools/singlejar/mapped_file_posix.inc file (#error This code for 64 bit Unix.) that throws an error if we are not in a 64 bit machine. This bazel version works ok in 32 bits.
vi ./src/tools/singlejar/mapped_file_posix.inc
Also we need to increase the java memory available to Bazel and start compiling it.
export BAZEL_JAVAC_OPTS="-J-Xmx1g"
./compile.sh
When it finishes (It can take several hours), we move the bazel compiled executable to some location in the current user's path
sudo cp output/bazel /usr/local/bin
Compile Tensorflow from sources
Create a folder and clone tensorflow's 1.13.2 version to it. Starting from version 1.14, tensorflow uses the Intel MKL DNN optimization library that it only works in 64 bits systems. So 1.13.2 is the last version that runs in 32 bits.
cd $HOME
mkdir Tensorflow-1.13.2
cd Tensorflow-1.13.2
git clone -b v1.13.2 --depth=1 https://github.com/tensorflow/tensorflow .
Before compiling, we replace the references to 64 bit libraries to the 32 bit ones.
grep -Rl "lib64"| xargs sed -i 's/lib64/lib/g'
We start the tensorflow configuration. We need to explicity disable the use of several optional libraries that are not available or not supported on 32 bit systems.
export TF_NEED_CUDA=0
export TF_NEED_AWS=0
./configure
We have to take the following considerations:
* When asked to specify the location of python. [Default is /usr/bin/python]: We should respond /usr/bin/python3 to use python 3.
* When asked to input the desired Python library path to use. Default is [/usr/local/lib/python3.5/dist-packages] we just hit Enter
* We should respond N to all the Y/N questions.
* When asked to specify optimization flags to use during compilation when bazel option "--config=opt" is specified [Default is -march=native -Wno-sign-compare]: Just hit Enter
Now we start compiling tensorflow disabling optional components like aws, kafka, etc.
bazel build --config=noaws --config=nohdfs --config=nokafka --config=noignite --config=nonccl -c opt --verbose_failures //tensorflow/tools/pip_package:build_pip_package
If everything went ok, now we generate the pip package.
bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
And we install the pip package
python3 -m pip install --user /tmp/tensorflow_pkg/tensorflow-1.13.2-cp35-cp35m-linux_i686.whl
Test tensorflow
Now we run a small test to check that it works. We create a test.py file with the following contents:
import tensorflow as tf
mnist = tf.keras.datasets.mnist
(x_train, y_train),(x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0
model = tf.keras.models.Sequential([
tf.keras.layers.Flatten(input_shape=(28, 28)),
tf.keras.layers.Dense(512, activation=tf.nn.relu),
tf.keras.layers.Dropout(0.2),
tf.keras.layers.Dense(10, activation=tf.nn.softmax)
])
model.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
model.fit(x_train, y_train, epochs=5)
model.evaluate(x_test, y_test)
And we run the test
python3 test.py
Here is the output
Epoch 1/5
60000/60000 [==============================] - 87s 1ms/sample - loss: 0.2202 - acc: 0.9348
Epoch 2/5
60000/60000 [==============================] - 131s 2ms/sample - loss: 0.0963 - acc: 0.9703
Epoch 3/5
60000/60000 [==============================] - 135s 2ms/sample - loss: 0.0685 - acc: 0.9785
Epoch 4/5
60000/60000 [==============================] - 128s 2ms/sample - loss: 0.0526 - acc: 0.9828
Epoch 5/5
60000/60000 [==============================] - 128s 2ms/sample - loss: 0.0436 - acc: 0.9863
10000/10000 [==============================] - 3s 273us/sample - loss: 0.0666 - acc: 0.9800
Enjoy your new Tensorflow library !!