Related
I have function to draw a confusion matrix
import itertools
import matplotlib.pyplot as plt
def plot_confusion_matrix(cm,classes,normalize=False,title='Confusion matrix',cmap=plt.cm.Blues):
plt.figure(figsize=(15,10))
plt.imshow(cm,interpolation='nearest',cmap=cmap)
plt.title(title)
plt.colorbar()
tick_marks=np.arange(len(classes))
plt.xticks(tick_marks,classes,rotation=45)
plt.yticks(tick_marks,classes)
if normalize:
cm=cm.astype('float')/cm.sum(axis=1)[:,np.newaxis]
cm=np.around(cm,decimals=2)
cm[np.isnan(cm)]=0.0
print('Normalized confusion matrix')
else:
print('Confusion matrix, without normalization')
thresh=cm.max()/2
for i, j in itertools.product(range(cm.shape[0]), range(cm.shape[1])):
plt.text(j, i, cm[i, j],
horizontalalignment="center",
color="white" if cm[i, j] > thresh else "black")
plt.tight_layout()
plt.ylabel('True label')
plt.xlabel('Predicted label')
Then, implementing the function
cm=confusion_matrix(validation_generator.classes,y_pred)
plot_confusion_matrix(cm,target_names,title='confusion matrix')
Outputs this picture
enter image description here
How can I increase the size of the numbers in the image?
To change the font size of a matplotlib plot, you can try this approach -
plt.rcParams.update({'font.size': 22})
Moreover you can also explore other formatting options here
I've plotted a confusion matrix with scikit-learn / matplotlib thanks to different code examples I found on the web, but I'm stuck at finding how to add space between the xticklabels and the main title. As you can see on the image below, the plot title and the xticklabels are overlapping (+ the ylabel 'True' is cut out).
Link to my confusion matrix image
Here is the function I use:
from sklearn.metrics import confusion_matrix
import matplotlib.pyplot as plt
PLOTS = '/plots/' # Output folder
def plt_confusion_matrix(y_test, y_pred, normalize=False, title="Confusion matrix"):
"""
Plots a nice confusion matrix.
:param y_test: list of predicted labels
:param y_pred: list of labels that should have been predicted.
:param normalize: boolean. If False, the plots shows the number of sentences predicted.
If True, shows the percentage of sentences predicted.
:param title: string. Title of the plot.
:return: Nothing but saves the plot as a PNG file and shows it.
"""
labels = list(set(y_pred))
cm = confusion_matrix(y_test, y_pred, labels)
fig = plt.figure()
ax = fig.add_subplot(111)
cax = ax.matshow(cm, cmap=plt.cm.binary, interpolation='nearest')
if normalize:
cm = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis]
fig.suptitle(title, fontsize=14, wrap=True)
fig.colorbar(cax)
ax.set_xticklabels([''] + labels, rotation=45)
ax.set_yticklabels([''] + labels)
plt.xlabel('Predicted')
plt.ylabel('True')
plt.subplots_adjust(hspace=0.6)
fmt = '.2f' if normalize else 'd'
thresh = cm.max() / 1.5 if normalize else cm.max() / 2
for i in range(cm.shape[0]):
for j in range(cm.shape[1]):
ax.text(j, i, format(cm[i, j], fmt),
ha="center", va="center",
color="white" if cm[i, j] > thresh else "black")
plt.savefig(PLOTS + title)
plt.show()
I had to rotate the xticklabels because they are too long and otherwise overlapping each other horizontally, and I had to wrap the title because it is also too long and could not otherwise be displayed entirely in the image.
I've seen in another post that xticklabels can also be placed underneath the figure (like in this stackoverflow post), so maybe it could be a solution, however I haven't understood how to make it.
How do I solve the problem?
either to add some space between the title and the xticklabels
(making them appear entirely btw);
or to make the ylabel 'True' visible
or to move the xticklabels under the figure.
Edit : I tried both of geekzeus solutions, without success...
Result with geekzeus' 1st solution : See confusion matrix
Result with geekzeus' 2nd solution : See confusion matrix
Do it like this
ax.set_xlabel('Predicted labels')
ax.set_ylabel('True labels')
ax.set_title('Confusion Matrix')
#xaxisticks
ax.xaxis.set_ticklabels(['A', 'B'])
#yaxis ticks
ax.yaxis.set_ticklabels(['B', 'A'])
OR
use seaborn with matplotlib,you can also directly provide list variable to ticks
import seaborn as sns
import matplotlib.pyplot as plt
cm = confusion_matrix(true_classes, predicted_classes)
ax= plt.subplot()
sns.heatmap(cm, annot=True, ax = ax); #annot=True to annotate cells
# labels, title and ticks
ax.set_xlabel('Predicted labels')
ax.set_ylabel('True labels')
ax.set_title('Confusion Matrix')
ax.xaxis.set_ticklabels(['A', 'B'])
ax.yaxis.set_ticklabels(['B', 'A'])
You can specify the location of the title using parameters x and y. If you tweak the values of y, the desired plot can be generated.
fig.suptitle(title, fontsize=14, wrap=True, y=1.2)
The following code is for generating the 3 subplots. And on all the 3 subplots scale is mentioned. I want to stack them in such a way that x-axis and y-axis scale appear once like this. Can I get this plot with plt.subplot() or fig.add_axes is compulsory for this? I actually want to do this with subplots because in fig.add_subplot I havve to specify the width and height of each plot that I don't want.
`fig,axes = plt.figure(nrow=3, ncolmn=1)
ax1 = fig.add_subplot(311)
ax2 = fig.add_subplot(312)
ax3 = fig.add_subplot(313)
ind1 =[1,2,3]
ind2 = [4,5,6]
for i in range(len(3)):
data1=np.load(..)
data2=np.load(..)
axes[i].plot(data1, data2)`
Here is one solution using subplots_adjust where you put the space between two plots to 0 using hspace. Also, use sharex=True to have a shared x-axis
fig, axes = plt.subplots(nrows=3, ncols=1,sharex=True)
x = np.linspace(0, 2*np.pi, 100)
y = np.sin(x)
for i, ax in enumerate(axes.ravel()): # or axes.flatten() or axes.flat
ax.plot(x, y, label='File %d' %i)
ax.legend()
fig.text(0.5, 0.01, 'X-label', ha='center')
fig.text(0.01, 0.5, 'Y-label', va='center', rotation='vertical')
plt.tight_layout() # To get a better spacing between the subplots
plt.subplots_adjust(hspace=.0)
I'm trying to create a plot using pyplot that has a discontinuous x-axis. The usual way this is drawn is that the axis will have something like this:
(values)----//----(later values)
where the // indicates that you're skipping everything between (values) and (later values).
I haven't been able to find any examples of this, so I'm wondering if it's even possible. I know you can join data over a discontinuity for, eg, financial data, but I'd like to make the jump in the axis more explicit. At the moment I'm just using subplots but I'd really like to have everything end up on the same graph in the end.
Paul's answer is a perfectly fine method of doing this.
However, if you don't want to make a custom transform, you can just use two subplots to create the same effect.
Rather than put together an example from scratch, there's an excellent example of this written by Paul Ivanov in the matplotlib examples (It's only in the current git tip, as it was only committed a few months ago. It's not on the webpage yet.).
This is just a simple modification of this example to have a discontinuous x-axis instead of the y-axis. (Which is why I'm making this post a CW)
Basically, you just do something like this:
import matplotlib.pylab as plt
import numpy as np
# If you're not familiar with np.r_, don't worry too much about this. It's just
# a series with points from 0 to 1 spaced at 0.1, and 9 to 10 with the same spacing.
x = np.r_[0:1:0.1, 9:10:0.1]
y = np.sin(x)
fig,(ax,ax2) = plt.subplots(1, 2, sharey=True)
# plot the same data on both axes
ax.plot(x, y, 'bo')
ax2.plot(x, y, 'bo')
# zoom-in / limit the view to different portions of the data
ax.set_xlim(0,1) # most of the data
ax2.set_xlim(9,10) # outliers only
# hide the spines between ax and ax2
ax.spines['right'].set_visible(False)
ax2.spines['left'].set_visible(False)
ax.yaxis.tick_left()
ax.tick_params(labeltop='off') # don't put tick labels at the top
ax2.yaxis.tick_right()
# Make the spacing between the two axes a bit smaller
plt.subplots_adjust(wspace=0.15)
plt.show()
To add the broken axis lines // effect, we can do this (again, modified from Paul Ivanov's example):
import matplotlib.pylab as plt
import numpy as np
# If you're not familiar with np.r_, don't worry too much about this. It's just
# a series with points from 0 to 1 spaced at 0.1, and 9 to 10 with the same spacing.
x = np.r_[0:1:0.1, 9:10:0.1]
y = np.sin(x)
fig,(ax,ax2) = plt.subplots(1, 2, sharey=True)
# plot the same data on both axes
ax.plot(x, y, 'bo')
ax2.plot(x, y, 'bo')
# zoom-in / limit the view to different portions of the data
ax.set_xlim(0,1) # most of the data
ax2.set_xlim(9,10) # outliers only
# hide the spines between ax and ax2
ax.spines['right'].set_visible(False)
ax2.spines['left'].set_visible(False)
ax.yaxis.tick_left()
ax.tick_params(labeltop='off') # don't put tick labels at the top
ax2.yaxis.tick_right()
# Make the spacing between the two axes a bit smaller
plt.subplots_adjust(wspace=0.15)
# This looks pretty good, and was fairly painless, but you can get that
# cut-out diagonal lines look with just a bit more work. The important
# thing to know here is that in axes coordinates, which are always
# between 0-1, spine endpoints are at these locations (0,0), (0,1),
# (1,0), and (1,1). Thus, we just need to put the diagonals in the
# appropriate corners of each of our axes, and so long as we use the
# right transform and disable clipping.
d = .015 # how big to make the diagonal lines in axes coordinates
# arguments to pass plot, just so we don't keep repeating them
kwargs = dict(transform=ax.transAxes, color='k', clip_on=False)
ax.plot((1-d,1+d),(-d,+d), **kwargs) # top-left diagonal
ax.plot((1-d,1+d),(1-d,1+d), **kwargs) # bottom-left diagonal
kwargs.update(transform=ax2.transAxes) # switch to the bottom axes
ax2.plot((-d,d),(-d,+d), **kwargs) # top-right diagonal
ax2.plot((-d,d),(1-d,1+d), **kwargs) # bottom-right diagonal
# What's cool about this is that now if we vary the distance between
# ax and ax2 via f.subplots_adjust(hspace=...) or plt.subplot_tool(),
# the diagonal lines will move accordingly, and stay right at the tips
# of the spines they are 'breaking'
plt.show()
I see many suggestions for this feature but no indication that it's been implemented. Here is a workable solution for the time-being. It applies a step-function transform to the x-axis. It's a lot of code, but it's fairly simple since most of it is boilerplate custom scale stuff. I have not added any graphics to indicate the location of the break, since that is a matter of style. Good luck finishing the job.
from matplotlib import pyplot as plt
from matplotlib import scale as mscale
from matplotlib import transforms as mtransforms
import numpy as np
def CustomScaleFactory(l, u):
class CustomScale(mscale.ScaleBase):
name = 'custom'
def __init__(self, axis, **kwargs):
mscale.ScaleBase.__init__(self)
self.thresh = None #thresh
def get_transform(self):
return self.CustomTransform(self.thresh)
def set_default_locators_and_formatters(self, axis):
pass
class CustomTransform(mtransforms.Transform):
input_dims = 1
output_dims = 1
is_separable = True
lower = l
upper = u
def __init__(self, thresh):
mtransforms.Transform.__init__(self)
self.thresh = thresh
def transform(self, a):
aa = a.copy()
aa[a>self.lower] = a[a>self.lower]-(self.upper-self.lower)
aa[(a>self.lower)&(a<self.upper)] = self.lower
return aa
def inverted(self):
return CustomScale.InvertedCustomTransform(self.thresh)
class InvertedCustomTransform(mtransforms.Transform):
input_dims = 1
output_dims = 1
is_separable = True
lower = l
upper = u
def __init__(self, thresh):
mtransforms.Transform.__init__(self)
self.thresh = thresh
def transform(self, a):
aa = a.copy()
aa[a>self.lower] = a[a>self.lower]+(self.upper-self.lower)
return aa
def inverted(self):
return CustomScale.CustomTransform(self.thresh)
return CustomScale
mscale.register_scale(CustomScaleFactory(1.12, 8.88))
x = np.concatenate((np.linspace(0,1,10), np.linspace(9,10,10)))
xticks = np.concatenate((np.linspace(0,1,6), np.linspace(9,10,6)))
y = np.sin(x)
plt.plot(x, y, '.')
ax = plt.gca()
ax.set_xscale('custom')
ax.set_xticks(xticks)
plt.show()
Check the brokenaxes package:
import matplotlib.pyplot as plt
from brokenaxes import brokenaxes
import numpy as np
fig = plt.figure(figsize=(5,2))
bax = brokenaxes(
xlims=((0, .1), (.4, .7)),
ylims=((-1, .7), (.79, 1)),
hspace=.05
)
x = np.linspace(0, 1, 100)
bax.plot(x, np.sin(10 * x), label='sin')
bax.plot(x, np.cos(10 * x), label='cos')
bax.legend(loc=3)
bax.set_xlabel('time')
bax.set_ylabel('value')
A very simple hack is to
scatter plot rectangles over the axes' spines and
draw the "//" as text at that position.
Worked like a charm for me:
# FAKE BROKEN AXES
# plot a white rectangle on the x-axis-spine to "break" it
xpos = 10 # x position of the "break"
ypos = plt.gca().get_ylim()[0] # y position of the "break"
plt.scatter(xpos, ypos, color='white', marker='s', s=80, clip_on=False, zorder=100)
# draw "//" on the same place as text
plt.text(xpos, ymin-0.125, r'//', fontsize=label_size, zorder=101, horizontalalignment='center', verticalalignment='center')
Example Plot:
For those interested, I've expanded upon #Paul's answer and added it to the matplotlib wrapper proplot. It can do axis "jumps", "speedups", and "slowdowns".
There is no way currently to add "crosses" that indicate the discrete jump like in Joe's answer, but I plan to add this in the future. I also plan to add a default "tick locator" that sets sensible default tick locations depending on the CutoffScale arguments.
Adressing Frederick Nord's question how to enable parallel orientation of the diagonal "breaking" lines when using a gridspec with ratios unequal 1:1, the following changes based on the proposals of Paul Ivanov and Joe Kingtons may be helpful. Width ratio can be varied using variables n and m.
import matplotlib.pylab as plt
import numpy as np
import matplotlib.gridspec as gridspec
x = np.r_[0:1:0.1, 9:10:0.1]
y = np.sin(x)
n = 5; m = 1;
gs = gridspec.GridSpec(1,2, width_ratios = [n,m])
plt.figure(figsize=(10,8))
ax = plt.subplot(gs[0,0])
ax2 = plt.subplot(gs[0,1], sharey = ax)
plt.setp(ax2.get_yticklabels(), visible=False)
plt.subplots_adjust(wspace = 0.1)
ax.plot(x, y, 'bo')
ax2.plot(x, y, 'bo')
ax.set_xlim(0,1)
ax2.set_xlim(10,8)
# hide the spines between ax and ax2
ax.spines['right'].set_visible(False)
ax2.spines['left'].set_visible(False)
ax.yaxis.tick_left()
ax.tick_params(labeltop='off') # don't put tick labels at the top
ax2.yaxis.tick_right()
d = .015 # how big to make the diagonal lines in axes coordinates
# arguments to pass plot, just so we don't keep repeating them
kwargs = dict(transform=ax.transAxes, color='k', clip_on=False)
on = (n+m)/n; om = (n+m)/m;
ax.plot((1-d*on,1+d*on),(-d,d), **kwargs) # bottom-left diagonal
ax.plot((1-d*on,1+d*on),(1-d,1+d), **kwargs) # top-left diagonal
kwargs.update(transform=ax2.transAxes) # switch to the bottom axes
ax2.plot((-d*om,d*om),(-d,d), **kwargs) # bottom-right diagonal
ax2.plot((-d*om,d*om),(1-d,1+d), **kwargs) # top-right diagonal
plt.show()
This is a hacky but pretty solution for x-axis breaks.
The solution is based on https://matplotlib.org/stable/gallery/subplots_axes_and_figures/broken_axis.html, which gets rid of the problem with positioning the break above the spine, solved by How can I plot points so they appear over top of the spines with matplotlib?
from matplotlib.patches import Rectangle
import matplotlib.pyplot as plt
def axis_break(axis, xpos=[0.1, 0.125], slant=1.5):
d = slant # proportion of vertical to horizontal extent of the slanted line
anchor = (xpos[0], -1)
w = xpos[1] - xpos[0]
h = 1
kwargs = dict(marker=[(-1, -d), (1, d)], markersize=12, zorder=3,
linestyle="none", color='k', mec='k', mew=1, clip_on=False)
axis.add_patch(Rectangle(
anchor, w, h, fill=True, color="white",
transform=axis.transAxes, clip_on=False, zorder=3)
)
axis.plot(xpos, [0, 0], transform=axis.transAxes, **kwargs)
fig, ax = plt.subplots(1,1)
plt.plot(np.arange(10))
axis_break(ax, xpos=[0.1, 0.12], slant=1.5)
axis_break(ax, xpos=[0.3, 0.31], slant=-10)
if you want to replace an axis label, this would do the trick:
from matplotlib import ticker
def replace_pos_with_label(fig, pos, label, axis):
fig.canvas.draw() # this is needed to set up the x-ticks
labs = axis.get_xticklabels()
labels = []
locs = []
for text in labs:
x = text._x
lab = text._text
if x == pos:
lab = label
labels.append(lab)
locs.append(x)
axis.xaxis.set_major_locator(ticker.FixedLocator(locs))
axis.set_xticklabels(labels)
fig, ax = plt.subplots(1,1)
plt.plot(np.arange(10))
replace_pos_with_label(fig, 0, "-10", axis=ax)
replace_pos_with_label(fig, 6, "$10^{4}$", axis=ax)
axis_break(ax, xpos=[0.1, 0.12], slant=2)
I am trying to make some figures for a scientific article, so I want my figures to have a specific size. I also see that Matplotlib by default adds a lot of padding on the border of the figures, which I don't need (since the figures will be on a white background anyway).
To set a specific figure size I simply use plt.figure(figsize = [w, h]), and I add the argument tight_layout = {'pad': 0} to remove the padding. This works perfectly, and even works if I add a title, y/x-labels etc. Example:
fig = plt.figure(
figsize = [3,2],
tight_layout = {'pad': 0}
)
ax = fig.add_subplot(111)
plt.title('title')
ax.set_ylabel('y label')
ax.set_xlabel('x label')
plt.savefig('figure01.pdf')
This creates a pdf file with exact size 3x2 (inches).
The issue I have is that when I for example add a text box outside the axis (typically a legend box), Matplotlib does not make room for the text box like it does when adding titles/axis labels. Typically the text box is cut off, or does not show in the saved figure at all. Example:
plt.close('all')
fig = plt.figure(
figsize = [3,2],
tight_layout = {'pad': 0}
)
ax = fig.add_subplot(111)
plt.title('title')
ax.set_ylabel('y label')
ax.set_xlabel('x label')
t = ax.text(0.7, 1.1, 'my text here', bbox = dict(boxstyle = 'round'))
plt.savefig('figure02.pdf')
A solution I found elsewhere on SO was to add the argument bbox_inches = 'tight' to the savefig command. The text box is now included like I wanted, but the pdf is now the wrong size. It seems like Matplotlib just makes the figure bigger, instead of reducing the size of the axes like it does when adding titles and x/y-labels.
Example:
plt.close('all')
fig = plt.figure(
figsize = [3,2],
tight_layout = {'pad': 0}
)
ax = fig.add_subplot(111)
plt.title('title')
ax.set_ylabel('y label')
ax.set_xlabel('x label')
t = ax.text(0.7, 1.1, 'my text here', bbox = dict(boxstyle = 'round'))
plt.savefig('figure03.pdf', bbox_inches = 'tight')
(This figure is 3.307x2.248)
Is there any solution to this that covers most cases with a legend just outside the axes?
So the requirements are:
Having a fixed, predefined figure size
Adding a text label or legend outside the axes
Axes and text cannot overlap
The axes, together with the title and axis labels, sits tightly agains the figure border.
So tight_layout with pad = 0, solves 1. and 4. but contradicts 2.
One could think on setting pad to a larger value. This would solve 2. However, since it's is symmetric in all directions, it would contradict 4.
Using bbox_inches = 'tight' changes the figure size. Contradicts 1.
So I think there is no generic solution to this problem.
Something I can come up with is the following: It sets the text in figure coordinates and then resizes the axes either in horizontal or in vertical direction such that there is no overlap between the axes and the text.
import matplotlib.pyplot as plt
import matplotlib.transforms
fig = plt.figure(figsize = [3,2])
ax = fig.add_subplot(111)
plt.title('title')
ax.set_ylabel('y label')
ax.set_xlabel('x label')
def text_legend(ax, x0, y0, text, direction = "v", padpoints = 3, margin=1.,**kwargs):
ha = kwargs.pop("ha", "right")
va = kwargs.pop("va", "top")
t = ax.figure.text(x0, y0, text, ha=ha, va=va, **kwargs)
otrans = ax.figure.transFigure
plt.tight_layout(pad=0)
ax.figure.canvas.draw()
plt.tight_layout(pad=0)
offs = t._bbox_patch.get_boxstyle().pad * t.get_size() + margin # adding 1pt
trans = otrans + \
matplotlib.transforms.ScaledTranslation(-offs/72.,-offs/72.,fig.dpi_scale_trans)
t.set_transform(trans)
ax.figure.canvas.draw()
ppar = [0,-padpoints/72.] if direction == "v" else [-padpoints/72.,0]
trans2 = matplotlib.transforms.ScaledTranslation(ppar[0],ppar[1],fig.dpi_scale_trans) + \
ax.figure.transFigure.inverted()
tbox = trans2.transform(t._bbox_patch.get_window_extent())
bbox = ax.get_position()
if direction=="v":
ax.set_position([bbox.x0, bbox.y0,bbox.width, tbox[0][1]-bbox.y0])
else:
ax.set_position([bbox.x0, bbox.y0,tbox[0][0]-bbox.x0, bbox.height])
# case 1: place text label at top right corner of figure (1,1). Adjust axes height.
#text_legend(ax, 1,1, 'my text here', bbox = dict(boxstyle = 'round'), )
# case 2: place text left of axes, (1, y), direction=="v"
text_legend(ax, 1., 0.8, 'my text here', margin=2., direction="h", bbox = dict(boxstyle = 'round') )
plt.savefig(__file__+'.pdf')
plt.show()
case 1 (left) and case 2 (right):
Doin the same with a legend is slightly easier, because we can directly use the bbox_to_anchor argument and don't need to control the fancy box around the legend.
import matplotlib.pyplot as plt
import matplotlib.transforms
fig = plt.figure(figsize = [3.5,2])
ax = fig.add_subplot(111)
ax.set_title('title')
ax.set_ylabel('y label')
ax.set_xlabel('x label')
ax.plot([1,2,3], marker="o", label="quantity 1")
ax.plot([2,1.7,1.2], marker="s", label="quantity 2")
def legend(ax, x0=1,y0=1, direction = "v", padpoints = 3,**kwargs):
otrans = ax.figure.transFigure
t = ax.legend(bbox_to_anchor=(x0,y0), loc=1, bbox_transform=otrans,**kwargs)
plt.tight_layout(pad=0)
ax.figure.canvas.draw()
plt.tight_layout(pad=0)
ppar = [0,-padpoints/72.] if direction == "v" else [-padpoints/72.,0]
trans2=matplotlib.transforms.ScaledTranslation(ppar[0],ppar[1],fig.dpi_scale_trans)+\
ax.figure.transFigure.inverted()
tbox = t.get_window_extent().transformed(trans2 )
bbox = ax.get_position()
if direction=="v":
ax.set_position([bbox.x0, bbox.y0,bbox.width, tbox.y0-bbox.y0])
else:
ax.set_position([bbox.x0, bbox.y0,tbox.x0-bbox.x0, bbox.height])
# case 1: place text label at top right corner of figure (1,1). Adjust axes height.
#legend(ax, borderaxespad=0)
# case 2: place text left of axes, (1, y), direction=="h"
legend(ax,y0=0.8, direction="h", borderaxespad=0.2)
plt.savefig(__file__+'.pdf')
plt.show()
Why 72? The 72 is the number of points per inch (ppi). This is a fixed typographic unit e.g. fontsizes are always given in points (like 12pt). Because matplotlib defines the padding of the text box in units relative to fontsize, which is points, we need to use 72 to transform back to inches (and then to display coordinates). The default dots per inch (dpi) is not touched here, but is accounted for in fig.dpi_scale_trans. If you want to change dpi you need to make sure the figure dpi is set when creating the figure as well as when saving it (use dpi=.. in the call to plt.figure() as well as plt.savefig()).
As of matplotlib==3.1.3, you can use constrained_layout=True to achieve the desired result. This is currently experimental, but see the docs for a very helpful guide (and the section specifically on legends). Note that the legend will steal space from the plot, but this is unavoidable. I've found that as long as the legend does not take up too much space relative to the size of the plot, then the figure gets saved without cropping anything.
import matplotlib.pyplot as plt
fig, ax = plt.subplots(figsize=(3, 2), constrained_layout=True)
ax.set_title('title')
ax.set_ylabel('y label')
ax.set_xlabel('x label')
ax.plot([0,1], [0,1], label='my text here')
ax.legend(loc='center left', bbox_to_anchor=(1.1, 0.5))
fig.savefig('figure03.pdf')