amueller
commented
6 years ago Thanks for the analysis. Can you run lineprof to show the lines that actually take the time?
Sent from phone. Please excuse spelling and brevity.
On Thu, Apr 26, 2018, 00:02 Paula15 notifications@github.com wrote:
Thanks for developing such an amazing library.
I encountered some performance issues. I'm using it for poster generation (1500 x 2400). It takes about 30 seconds to generate a poster with mask. After setting scale = 12 (the maximum scale that yields acceptable result), it renders in 4 seconds, but still too slow for a web application for (probably impatient) users.
I profiled my code (shown below) with python -m cProfile --sort cumulative test.py >profile to show the most time-consuming functions and their callers:
import time from PIL import Imageimport jiebaimport numpy as npfrom wordcloud import WordCloud
inputs = open('words.txt').read().decode('utf-8') words = '\n'.join([word for word in jieba.cut(inputs)]) scale = 12 mask = Image.open('mask.png') width, height = mask.size mask = np.array(mask.resize((int(mask.size[0] / float(scale)), int(mask.size[1] / float(scale))))) wordcloud = WordCloud(scale=scale, background_color='white', width=int(width/float(scale)), height=int(height/float(scale)), margin=2, font_path='msyh.ttc', mask=mask).generate(words) wordcloud.to_file('out.png')
... And got:
ncalls tottime percall cumtime percall filename:lineno(function) 1 0.013 0.013 4.087 4.087 test.py:1(
) 271 0.002 0.000 1.453 0.005 ImageFont.py:216(truetype) 271 0.002 0.000 1.451 0.005 ImageFont.py:119(init) 271 1.448 0.005 1.448 0.005 {PIL._imagingft.getfont} 1 0.000 0.000 1.415 1.415 wordcloud.py:556(generate) 1 0.000 0.000 1.415 1.415 wordcloud.py:535(generate_from_text) 2/1 0.058 0.029 1.392 1.392 wordcloud.py:331(generate_from_frequencies) I was surprised that font construction was so time-consuming. Checking out the source code for wordcloud.py:generate_from_frequencies, it reveals that in this for-loop:
while True:
try to find a position
font = ImageFont.truetype(self.font_path, font_size) # transpose font optionally transposed_font = ImageFont.TransposedFont( font, orientation=orientation) # get size of resulting text box_size = draw.textsize(word, font=transposed_font) # find possible places using integral image: result = occupancy.sample_position(box_size[1] + self.margin, box_size[0] + self.margin, random_state)It first constructs a font object, then use it to get font size, then judge if there is a possible location. The problem is that even if the attempted location is not valid, we still have to construct a font object for text size computing (as required by the draw.textsize function). Run time of fonts for different languages tend to be similar.
So, I'm wondering if is there a better way to get text size in this looping procedure. I believe this to be a worthy optimization.
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/amueller/word_cloud/issues/366, or mute the thread https://github.com/notifications/unsubscribe-auth/AAbcFrARnIkZPr_BAsyTxFew7qRdQOzVks5tsUa9gaJpZM4TkfU0 .
zhengpingzhou
mikesmith1611
Thanks for developing such an amazing library.
I encountered some performance issues. I'm using it for poster generation (1500 x 2400). It takes about 30 seconds to generate a poster with mask. After setting
scale = 12(the maximum scale that yields acceptable result), it renders in 4 seconds, but still too slow for a web application for (probably impatient) users.I profiled my code (shown below) with
python -m cProfile --sort cumulative test.py >profileto show the most time-consuming functions and their callers:... And got:
I was surprised that font construction was so time-consuming. Checking out the source code for
wordcloud.py:generate_from_frequencies, it reveals that in this while-loop:It first constructs a font object, then uses it to get font size, then judges if there is a possible location. The problem is that even if the attempted location is not valid, we still have to construct a font object for text size computing (as required by the
draw.textsizefunction). Run time of fonts for different languages tend to be similar.So, I'm wondering if is there a better way to get text size in this looping procedure. I believe this to be a worthy optimization.