Reproduce Python 2 PyQt4 QImage constructor behavior in Python 3

Question

I have written a small GUI using PyQt4 that displays an image and gets point coordinates that the user clicks on. I need to display a 2D numpy array as a grayscale, so I am creating a QImage from the array, then from that creating a QPixmap. In Python 2 it works fine.

When I moved to Python 3, however, it can't decide on a constructor for QImage - it gives me the following error:

TypeError: arguments did not match any overloaded call:
  QImage(): too many arguments
  QImage(QSize, QImage.Format): argument 1 has unexpected type 'numpy.ndarray'
  QImage(int, int, QImage.Format): argument 1 has unexpected type 'numpy.ndarray'
  QImage(str, int, int, QImage.Format): argument 1 has unexpected type 'numpy.ndarray'
  QImage(sip.voidptr, int, int, QImage.Format): argument 1 has unexpected type 'numpy.ndarray'
  QImage(str, int, int, int, QImage.Format): argument 1 has unexpected type 'numpy.ndarray'
  QImage(sip.voidptr, int, int, int, QImage.Format): argument 1 has unexpected type 'numpy.ndarray'
  QImage(list-of-str): argument 1 has unexpected type 'numpy.ndarray'
  QImage(str, str format=None): argument 1 has unexpected type 'numpy.ndarray'
  QImage(QImage): argument 1 has unexpected type 'numpy.ndarray'
  QImage(object): too many arguments

As far as I can tell, the QImage constructor I was calling previously was one of these:

• QImage(str, int, int, QImage.Format)
• QImage(sip.voidptr, int, int, QImage.Format)

I'm assuming that a numpy array fits one of the protocols necessary for one of these. I'm thinking it might have to do with an array versus a view, but all the variations I've tried either produce the above error or just make the GUI exit without doing anything. How can I reproduce the Python 2 behavior in Python 3?

The following is a small example, in which the same exact code works fine under Python 2 but not Python 3:

from __future__ import (print_function, division)

from PyQt4 import QtGui, QtCore
import numpy as np

class MouseView(QtGui.QGraphicsView):

    mouseclick = QtCore.pyqtSignal(tuple)

    def __init__(self, scene, parent=None):
        super(MouseView, self).__init__(scene, parent=parent)

    def mousePressEvent(self, event):
        self.mouseclick.emit((event.x(),
                              self.scene().sceneRect().height() - event.y()))


class SimplePicker(QtGui.QDialog):

    def __init__(self, data, parent=None):
        super(SimplePicker, self).__init__(parent)

        mind = data.min()
        maxd = data.max()
        bdata = ((data - mind) / (maxd - mind) * 255.).astype(np.uint8)

        wdims = data.shape
        wid = wdims[0]
        hgt = wdims[1]

        # This is the line of interest - it works fine under Python 2, but not Python 3
        img = QtGui.QImage(bdata.T, wid, hgt,
                           QtGui.QImage.Format_Indexed8)

        self.scene = QtGui.QGraphicsScene(0, 0, wid, hgt)
        self.px = self.scene.addPixmap(QtGui.QPixmap.fromImage(img))

        view = MouseView(self.scene)
        view.setHorizontalScrollBarPolicy(QtCore.Qt.ScrollBarAlwaysOff)
        view.setVerticalScrollBarPolicy(QtCore.Qt.ScrollBarAlwaysOff)
        view.setSizePolicy(QtGui.QSizePolicy.Fixed,
                           QtGui.QSizePolicy.Fixed)
        view.setMinimumSize(wid, hgt)
        view.mouseclick.connect(self.click_point)

        quitb = QtGui.QPushButton("Done")
        quitb.clicked.connect(self.quit)

        lay = QtGui.QVBoxLayout()
        lay.addWidget(view)
        lay.addWidget(quitb)

        self.setLayout(lay)

        self.points = []

    def click_point(self, xy):
        self.points.append(xy)

    def quit(self):
        self.accept()


def test_picker():

    x, y = np.mgrid[0:100, 0:100]
    img = x * y

    app = QtGui.QApplication.instance()
    if app is None:
        app = QtGui.QApplication(['python'])

    picker = SimplePicker(img)
    picker.show()
    app.exec_()

    print(picker.points)

if __name__ == "__main__":
    test_picker()

I am using an Anaconda installation on Windows 7 64-bit, Qt 4.8.7, PyQt 4.10.4, numpy 1.9.2.

Answer 1

In the PyQt constructor above, the following behavior is observed from a Numpy array called bdata:

• bdata works correctly for both Python 2 and Python 3
• bdata.T works for 2, not for 3 (constructor error)
• bdata.T.copy() works for both
• bdata[::-1,:] does not work for either 2 or 3 (the same error)
• bdata[::-1,:].copy() works for both
• bdata[::-1,:].base works for both, but loses the result of the reverse operation

As mentioned by @ekhumoro in the comments, you need something which supports the Python buffer protocol. The actual Qt constructor of interest here is this QImage constructor, or the const version of it:

QImage(uchar * data, int width, int height, Format format)

From the PyQt 4.10.4 documentation kept here, what PyQt expects for an unsigned char * is different in Python 2 and 3:

Python 2:

If Qt expects a char *, signed char * or an unsigned char * (or a const version) then PyQt4 will accept a unicode or QString that contains only ASCII characters, a str, a QByteArray, or a Python object that implements the buffer protocol.

Python 3:

If Qt expects a signed char * or an unsigned char * (or a const version) then PyQt4 will accept a bytes.

A Numpy array satisfies both of these, but apparently a Numpy view doesn't satisfy either. It's actually baffling that bdata.T works at all in Python 2, as it purportedly returns a view:

>>> a = np.ones((2,3))
>>> b = a.T
>>> b.base is a
True

The final answer: If you need to do transformations that result in a view, you can avoid errors by doing a copy() of the result to a new array for passing into the constructor. This may not be the best answer, but it will work.