Develop Reference

Implementing model from repo error: truth value of an array with more than one element is ambiguous

I'm trying to reproduce the results from the Graph-RCNN model here and am running into an error when I try to train it. I know about the error "truth value of an array with more than one element is ambiguous", but I don't know when you can have syntax like that without throwing the error. Is a solution downgrading my version of Numpy? Generally, is there a likely package incompatibility error I should check first? Are there older versions of Numpy or h5py that don't throw this error. I just followed the instructions on the environment setup within the repo and installed everything using Conda. This is all in python 3.6.2 in a Conda environment
Notable package versions:
cython 0.29.23 py36h2531618_0
h5py 2.8.0 py36h3010b51_1003 conda-forge
mkl 2020.4 h726a3e6_304 conda-forge
numpy 1.19.2 py36h54aff64_0
numpy-base 1.19.2 py36hfa32c7d_0
pycocotools 2.0.2 py36h8c4c3a4_1 conda-forge
pytorch 1.0.0 py3.6_cuda9.0.176_cudnn7.4.1_1 pytorch
scipy 1.4.1 py36h2d22cac_3 conda-forge
torchvision 0.2.1 py_2 pytorch

Tensorflow version mismatch on conda environments

I had initially installed tf-nightly by mistake and later uninstalled it. Now, I have installed two different versions of tensorflow on two different conda environments (tf1.14-gpu and tf2.0-gpu). When I execute the command
conda list -n tf1.14-gpu tensorflow it shows the following output
# Name Version Build Channel
tensorflow 1.14.0 gpu_py36h3fb9ad6_0
tensorflow-base 1.14.0 gpu_py36he45bfe2_0
tensorflow-estimator 1.14.0 py_0
tensorflow-gpu 1.14.0 h0d30ee6_0
When I execute the command conda list -n tf2.0-gpu tensorflow it shows the following output
# Name Version Build Channel
tensorflow 2.1.0 gpu_py36h2e5cdaa_0
tensorflow-base 2.1.0 gpu_py36h6c5654b_0
tensorflow-estimator 2.1.0 pyhd54b08b_0
tensorflow-gpu 2.1.0 h0d30ee6_0
But in both the environments when i import tensorflow and check for its version, it gives the same output as '2.2.0-dev20200218' which I assume is the version for tensorflow nightly build. I am not able to use this version for my existing models. I tried uninstalling anaconda and reinstalling the two environments with tensorflow 1.14 and tensorflow 2.0, but it tensorflow version still shows the same as '2.2.0-dev20200218'. Any idea how to overcome this ?

I ran to the same problem. Could it be possible that you installed tf-nightly using pip and not Conda? But when you run import tensorflow as tf; print(tf.__version__)it picks up the global pip version which is troublesome to get rid of?
p.s. Sorry that I'm posting instead of commenting. Don't have 50 reputation points yet.

Compiling scikit-image on Windows to run tests

I'm stumped. I'm developing some enhancements to scikit-image which are failing the automated build tests, probably due to rounding errors. I therefore need to get the automated tests running on my Windows system so that I can debug and work out what's wrong. I've so far tried two approaches, neither of which are working:
In my Anaconda Python 3.6 environment, when I try to run the automated tests, I am getting the following error:
RuntimeError: module compiled against API version 0xc but this version of numpy is 0xb
...which I have found reference to in other contexts, but have not been able to eliminate.
Since the automated test do run (but fail) on a Python 3.5-based system, I thought things might work if I tried a local Python 3.5 environment. Here, I am running into the issue that, despite being installed, the environment cannot find the MS C++ compiler cl.exe. It is installed in C:\Program Files (x86)\Microsoft Visual Studio\2017\BuildTools\VC\Tools\MSVC\14.15.26726\bin\HostX86\x64\ and is found and executed by my Python 3.6 environment, but my Python 3.5 environment doesn't find it despite me adding that directory to my PATH. I should add that my Python 3.6 environment finds it without the directory being added to the PATH. I understand that both Python 3.5 and 3.6 use MSVC 14.0.
I would prefer to fix the problem in my Python 3.6 environment if possible. Any assistance much appreciated.
Update
I have made a box-fresh Python 3.6 conda environment as follows:
conda create --name sk36 python=3.6
conda activate sk36
conda install scikit-image --only-deps
conda install cython
git clone https://github.com/scikit-image/scikit-image.git
cd scikit-image
pip install -e .
pytest skimage/feature
The specific error I am getting is as follows:
..\Anaconda3\lib\site-packages\py\_path\local.py:662: in pyimport
__import__(modname)
skimage\__init__.py:135: in <module>
from .data import data_dir
skimage\data\__init__.py:13: in <module>
from ..io import imread, use_plugin
skimage\io\__init__.py:7: in <module>
from .manage_plugins import *
skimage\io\manage_plugins.py:24: in <module>
from .collection import imread_collection_wrapper
skimage\io\collection.py:12: in <module>
from ..external.tifffile import TiffFile
skimage\external\tifffile\__init__.py:1: in <module>
from .tifffile import imsave, imread, imshow, TiffFile, TiffWriter, TiffSequence
skimage\external\tifffile\tifffile.py:292: in <module>
from . import _tifffile
E RuntimeError: module compiled against API version 0xc but this version of numpy is 0xb
...which appears to have something to do with tifffile. Since this package wasn't originally explicitly installed in my new environment, I tried installing various versions of it, including some which downgraded numpy and scipy. Still the same error as above.
Having done some more research it would appear that something is seeing numpy 1.13.x when in fact version 1.15.4 is installed. Here is the full output from conda list:
# Name Version Build Channel
blas 1.0 mkl anaconda
ca-certificates 2018.03.07 0 anaconda
certifi 2018.10.15 py36_0 anaconda
cloudpickle 0.6.1 py36_0 anaconda
cycler 0.10.0 py36h009560c_0 anaconda
cython 0.29 py36ha925a31_0 anaconda
dask-core 0.20.0 py36_0 anaconda
decorator 4.3.0 py36_0 anaconda
freetype 2.9.1 ha9979f8_1 anaconda
icc_rt 2017.0.4 h97af966_0 anaconda
icu 58.2 ha66f8fd_1 anaconda
imageio 2.4.1 py36_0 anaconda
intel-openmp 2019.0 118 anaconda
jpeg 9b hb83a4c4_2 anaconda
kiwisolver 1.0.1 py36h6538335_0 anaconda
libpng 1.6.35 h2a8f88b_0 anaconda
libtiff 4.0.9 h36446d0_2 anaconda
matplotlib 3.0.1 py36hc8f65d3_0 anaconda
mkl 2019.0 118 anaconda
mkl_fft 1.0.6 py36hdbbee80_0 anaconda
mkl_random 1.0.1 py36h77b88f5_1 anaconda
networkx 2.2 py36_1 anaconda
numpy 1.15.4 py36ha559c80_0 anaconda
numpy-base 1.15.4 py36h8128ebf_0 anaconda
olefile 0.46 py36_0 anaconda
openssl 1.0.2p hfa6e2cd_0 anaconda
package_has_been_revoked 1.0 0 enable_revoked
pillow 5.3.0 py36hdc69c19_0 anaconda
pip 18.1 py36_0 anaconda
pyparsing 2.3.0 py36_0 anaconda
pyqt 5.9.2 py36h6538335_2 anaconda
python 3.6.7 h33f27b4_1 anaconda
python-dateutil 2.7.5 py36_0 anaconda
pytz 2018.7 py36_0 anaconda
pywavelets 1.0.1 py36h8c2d366_0 anaconda
qt 5.9.6 vc14h1e9a669_2 anaconda
scikit-image 0.15.dev0 <pip>
scipy 1.1.0 py36h4f6bf74_1 anaconda
setuptools 40.5.0 py36_0 anaconda
sip 4.19.8 py36h6538335_0 anaconda
six 1.11.0 py36_1 anaconda
sqlite 3.25.2 hfa6e2cd_0 anaconda
tifffile 0.15.1 py36h452e1ab_1001 conda-forge
tk 8.6.8 hfa6e2cd_0 anaconda
toolz 0.9.0 py36_0 anaconda
tornado 5.1.1 py36hfa6e2cd_0 anaconda
vc 14.1 h21ff451_3 anaconda
vs2015_runtime 15.5.2 3 anaconda
wheel 0.32.2 py36_0 anaconda
wincertstore 0.2 py36h7fe50ca_0 anaconda
zlib 1.2.11 h8395fce_2 anaconda
Update 2
I've solved the problem for Python 3.6, and I think there's enough information above for the astute to be able to work out what was wrong. I'll put the solution in an answer below.
A cleanly built Python 3.5 environment can't find the compiler, so that issue still remains.

One approach you could try is to upgrade your numpy with
pip install numpy --upgrade
as described here: RuntimeError: module compiled against API version a but this version of numpy is 9
Otherwise (if for some reason you cannot upgrade numpy) I would suggest going with a virtual environment for scikit-image project. I just tried it on Windows 10 and was able to successfully execute tests. My steps (from cmd, inside the project folder):
conda uninstall scikit-image to remove any previously built/installed versions
conda -n scikit-image python=3.6 to create a virtual environment for this project (I used python 3.6, but you can change it to 3.5)
activate scikit-image activated the new virtual env
pip install -r requirements.txt -- installed dependencies (without this step I wasn't getting the dependencies for tests installed)
pip install -e .
pytest

It turns out that pytest wasn't actually installed in the correct environment, it was being invoked from base which did indeed have numpy 1.13.3 installed. Installing it in the cleanly built Python 3.6 environment solved the problem for Python 3.6 at least.

Tensorflow on 32-bit Windows 10: is not a supported wheel on this platform; package missing in current win-32 channels, no matching distribution

I have been following the installation guide from http://machinelearningmastery.com/setup-python-environment-machine-learning-deep-learning-anaconda/
I got & am using:
conda 4.3.22
Python 3.5.3 :: Anaconda 4.4.0 (32-bit)
scipy: 0.19.0
numpy: 1.12.1
matplotlib: 2.0.2
pandas: 0.20.1
statsmodels: 0.8.0
sklearn: 0.18.2
I successfully installed theano & keras. HOWEVER, I FAIL at installing tensorflow. Please HELP.
I created a conda ‘tensorflow’ environment with python 3.5. With command
『pip install –ignore-installed –upgrade https://storage.googleapis.com/tensorflow/windows/cpu/tensorflow-1.2.1-cp35-cp35m-win_amd64.whl』
I got ERROR saying
『tensorflow-1.2.1-cp35-cp35m-win_amd64.whl is not a supported wheel on this platform』
So i changed to version 1.0.1 and same error.
Version 1.1.0 also same error.
So i deactivated the environment, and type command
『conda install -c conda-forge tensorflow』
I got ERROR
『PackageNotFoundError: Package missing in current win-32 channels』
Instead it says the close match found is “xtensor” which i know is a C++ library that I'm not looking for.
Is it because I’m using a 32-bit Windows 10?
So I also tried running the following :
『python -m pip install –upgrade tensorflow』
and got ERROR of
『Could not find a version that satisfies the requirement tensorflow (from versions: ) No matching distribution found for tensorflow』
What more requirements do i need for this?
I tried 『pip3 install tensorflow』 but somehow it could not recognized ‘pip3’. So i type 『where pip3』 and it could not find files for the given pattern. So i type『where python』. It ouput the directory of my python. Then checked if it’s already put under the path inside the environmental variable. And it has. But i still couldn't use pip3 command.
I have tried all the solutions provided from people having similar problem with me and none of them work.

This question has been answered here.
In short, yes, TensorFlow does not support 32-bit platforms. Although if you only plan on writing abstract high-level Keras code then Theano will do just fine.

Managing packages: PyCharm vs conda vs pip

I'm new to Python and recently installed PyCharm 2016.3 on Windows 10. I'm also using Anaconda 3.
I don't know much about package management and would like to understand it better. Normally I just use conda update --all but I noticed (by checking the package list of my local PyCharm Interpreter) that this doesn't upgrade all packages to the latest version.
One such package is Pillow of which there's a version 4.0.0 but conda (4.3.11) won't update it past 3.4.2. I tried conda install pillow: 4.0.0 and got:
UnsatisfiableError: The following specifications were found to be in conflict:
- pillow 4.0.0*
- python 3.5*
- spyder-app
Use "conda info <package>" to see the dependencies for each package.
Later I found out that Pillow is also available on conda-forge so I tried conda install -c conda-forge pillow=4.0.0 and got:
The following NEW packages will be INSTALLED:
libiconv: 1.14-vc14_4 conda-forge [vc14]
libxml2: 2.9.3-vc14_9 conda-forge [vc14]
olefile: 0.44-py35_0 conda-forge
vc: 14-0 conda-forge
The following packages will be UPDATED:
freetype: 2.5.5-vc14_2 [vc14] --> 2.7-vc14_0 conda-forge [vc14]
jpeg: 8d-vc14_2 [vc14] --> 9b-vc14_0 conda-forge [vc14]
libtiff: 4.0.6-vc14_2 [vc14] --> 4.0.6-vc14_7 conda-forge [vc14]
pillow: 3.4.2-py35_0 --> 4.0.0-py35_2 conda-forge
The following packages will be SUPERCEDED by a higher-priority channel:
conda: 4.3.11-py35_0 --> 4.2.13-py35_0 conda-forge
conda-env: 2.6.0-0 --> 2.6.0-0 conda-forge
qt: 4.8.7-vc14_9 [vc14] --> 4.8.7-vc14_6 conda-forge [vc14]
I decided not to proceed and instead tried pip install pillow. Since this command doesn't ask for confirmation the package was simply installed. Now when I type conda list I get:
Pillow 4.0.0 <pip>
pillow 3.4.2 py35_0
The package list of the PyCharm Interpreter now shows Pillow as being version 4.0.0 but conda update pillow still returns:
# All requested packages already installed.
pillow 3.4.2 py35_0
My questions are:
1) What should I rely on to keep all my packages up to date, without compatibility issues?
2) Why did conda install pillow: 4.0.0 return an error but conda install -c conda-forge pillow=4.0.0 didn't?
3) What do the * next to pillow 4.0.0 and python 3.5 in the list of dependencies mean?
4) Since now I have both Pillow 3.4.2 (in /anaconda3/pkgs) and Pillow 4.0.0 (in /anaconda3/lib/site-packages) which one would be used if I imported Pillow?
5) Does the superseding conda: 4.3.11-py35_0 --> 4.2.13-py35_0 conda-forge mean conda is getting downgraded?
6) What is the difference between the tags pip, py35_0, py35_4, np111py35_2, etc?
7) PyCharm tells me there's a version 2.9.5 of package Jinja2 but both normal conda and conda-forge only find 2.9.4. From which channel is PyCharm getting this information?

Ok, I can't answer all of your questions but here goes:
1) Conda defers to the "pain up front" approach for handling dependency/conflict resolution. You'll have to get all of your packages to play nicely together in the repo's/channels that you have available to even make a package or keep them in an environment together. You can try running it with --force or --no-deps to try getting it in but ..... that can cause issues for you in the future (I don't know if that would even work with the later versions of conda, it changes a lot). Simply keeping packages up to date, and up to latest, I would just use pip. Its come a long way in the last few years (https://glyph.twistedmatrix.com/2016/08/python-packaging.html)
2) I am not completely sure, I believe it would have something to do with providing an explicit non-url channel for conda to look at. Typically you pass it the URL to the conda-forge repo (I think, again we don't use conda-forge internally).
3) The * means you are ignoring the patch/build 4.0.0 == Major.Minor.Build. Likewise, 3.5* == any version of 3.5
4) I would import pillow in a terminal, and then print out the module to see where its getting pulled from, why guess?
5) pass (although I think so)
6)
pip : means you installed that package via pip. It will not be picked up if you do conda list --explicit
py35_0 : has a requirement / only available to envs / packages that use python 3.5
py35_4 : not sure (always forget that one)
np111py35_2 : requires python3.5 and also numpy 1.11 (I think *)
7) I tend to steer clear of pycharm, I believe that you can inspect the python interpreter that pycharm is pointing at to see what environment its in. Based on the root environment, you can do a conda info and get a list of all of the channels you are pointing to.
Note: if you are going to use conda, you may just want to add conda-forge to your channels list instead of passing the -c (but seeing how the other channels are organized should help you see how you should pass the -c flag)