How to install COCO PythonAPI in python3

问题:

It seems the COCO PythonAPI only support python2. But peoples do use it in python3 environment.

I tried possible methods to install it, like

python3 setup.py build_ext --inplace
python3 setup.py install

But python3 setup.py install will fail due to coco.py and cocoeval.py containning python2 print function.

Update: solved by updating the COCO PythonAPI project. Leave this question for people facing the same issue.

回答1:

Try the following steps:

1. Use git clone to clone the folder into your drive. In this case, it should be git clone https://github.com/cocodataset/cocoapi.git.
2. Use terminal to enter the directory, or open a terminal inside the directory.
3. Type in 2to3 . -w. Note that you might have to install a package to get 2to3. It is an elegant tool to convert code from Python2 to Python3. This code converts all .py files from Python2-compatible to Python3-compatible.
4. Enter the folder for setup in terminal, or open a terminal inside the setup directory.
5. Type in python3 setup.py install.

This should help you install COCO or any package intended for Python2, and run the package using Python3. Cheers!

回答2:

Install

1. Instead of the official version (which has issues with python 3) use an alternative one. Install it on your local machine, globally (i.e., outside any virtual environment). You can do this by:

   pip install git+https://github.com/philferriere/cocoapi.git#subdirectory=PythonAPI

2. Check if it is installed globally:

   pip freeze | grep "pycocotools"

You should see something like pycocotools==2.0.0 in your output.

3. Now, inside your virtual-env (conda or whatever), first install numpy and cython (and maybe setuptools if it's not installed) using pip, and then:

   pip install pycocotools

Verify

Inside your project, import (for example) from pycocotools import mask as mask and then print(mask.__author__). This should print out the author's name, which is tsungyi.

Where Is It?

The installed package, like any other packages that are locally installed inside a virtual-env using pip, will go to External Libraries of your project, under site-packages. That means it is now part of your virtual-env and not part of your project. So, other users who may want to use your code, must repeat this installation on their virtual-env as well.

---

Troubleshooting:

The main source of confusion is that either you did not install the required packages before installing cocoapi, or you did install the required packages but for a different python version. And when you want to check if something is installed, you may check with, for instance, python3.6 and see that it exists, but you are actually running all your commands with python3.7. So suppose you are using python3.7. You need to make sure that:

1. python -V gives you python3.7 and NOT other version, and pip -V gives you pip 19.2.3 from /home/<USER>/.local/lib/python3.7/site-packages/pip (python3.7), that actually matches with your default python version. If this is not the case, you can change your default python using sudo update-alternatives --config python, and following the one-step instruction.

2. All the required packages are installed using the right python or pip version. You can check this using pip and pip3 to stop any differences that may cause an issue: pip freeze | grep "<SUBSTRING-NAME-OF-PACKAGE>" or pip show <PACKAGE-NAME> for more recent versions of pip.

3. To install the required packages, after you made sure about (1), you need to run: sudo apt install python-setuptools python3.7-dev python3-wheel build-essential and pip install numpy cython matplotlib

---

Environment: The above steps were tested on Ubuntu 18.4, python 3.6.8, pip 19.0.3.

回答3:

There are alternative versions of the cocoapi that you can download and use too (I'm using python 3.5). Here's a solution that you might want to try out:
How to download and use object detection datasets (e.g. coco or pascal)

回答4:

I have completed it with a simple step

pip install "git+https://github.com/philferriere/cocoapi.git#egg=pycocotools&subdirectory=PythonAPI"

** before that you need to install Visual C++ 2015 build tools on your path