{ "id": 67036, "title": "Transforming the data into images?", "url": "/competitions/quickdraw-doodle-recognition/discussion/67036", "author_name": "Daniel Möller", "post_date": "2018-09-27T22:39:53.412000", "votes": 12, "comment_count": 19, "views": 0, "content": "

Hello, everyone!

\n\n

Are there preset tools capable of creating pixel images from the given data?\nWhich other approaches are interesting?

", "messages": [ { "id": 395069, "postDate": "2018-09-27T22:39:53.413Z", "content": "

Hello, everyone!

\n\n

Are there preset tools capable of creating pixel images from the given data?\nWhich other approaches are interesting?

", "rawMarkdown": "Hello, everyone!\n\nAre there preset tools capable of creating pixel images from the given data?\nWhich other approaches are interesting?\n\n", "votes": 12 }, { "id": 396480, "postDate": "2018-09-30T20:55:08.300Z", "content": "

matplotlib might be inefficient. Alternative option is using PIL/pillow :

\n\n
import numpy as np\nfrom PIL import Image, ImageDraw\n\ndef draw_it(raw_strokes):\n    image = Image.new(\"P\", (255,255), color=255)\n    image_draw = ImageDraw.Draw(image)\n\n    for stroke in eval(raw_strokes):\n        for i in range(len(stroke[0])-1):\n\n            image_draw.line([stroke[0][i], \n                             stroke[1][i],\n                             stroke[0][i+1], \n                             stroke[1][i+1]],\n                            fill=0, width=6)\n    return np.array(image)\n
", "rawMarkdown": "matplotlib might be inefficient. Alternative option is using PIL/pillow :\n\n import numpy as np\n from PIL import Image, ImageDraw\n\n def draw_it(raw_strokes):\n image = Image.new(\"P\", (255,255), color=255)\n image_draw = ImageDraw.Draw(image)\n\n for stroke in eval(raw_strokes):\n for i in range(len(stroke[0])-1):\n\n image_draw.line([stroke[0][i], \n stroke[1][i],\n stroke[0][i+1], \n stroke[1][i+1]],\n fill=0, width=6)\n return np.array(image)", "votes": 8, "replies": [ { "id": 397749, "postDate": "2018-10-03T03:22:18.010Z", "content": "

Pillow was much faster for me than matplotlib and without the memory spikes - thanks!

", "rawMarkdown": "Pillow was much faster for me than matplotlib and without the memory spikes - thanks!", "votes": 1 }, { "id": 398030, "postDate": "2018-10-03T13:18:31.987Z", "content": "

What's cline ? Is it dataframe ?

", "rawMarkdown": "What's cline ? Is it dataframe ?", "votes": 1 }, { "id": 398038, "postDate": "2018-10-03T13:26:45.053Z", "content": "

It was a typo, this should be \"raw_strokes\", the stroke string in dataframe.

\n\n

It's fixed now. Thanks for pointing it out.

", "rawMarkdown": "It was a typo, this should be \"raw_strokes\", the stroke string in dataframe.\n\nIt's fixed now. Thanks for pointing it out." }, { "id": 400057, "postDate": "2018-10-07T13:31:44.560Z", "content": "

Your approach is really fast.
\nThere is a problem that the image size (255x255) maybe not enough to draw some images. Therefore, you can get white image then.

", "rawMarkdown": "Your approach is really fast. \nThere is a problem that the image size (255x255) maybe not enough to draw some images. Therefore, you can get white image then." }, { "id": 400799, "postDate": "2018-10-08T22:58:35.100Z", "rawMarkdown": "", "isDeleted": true }, { "id": 402208, "postDate": "2018-10-11T10:27:14.287Z", "content": "

Your approach is really good, thank you for the sharing). BTW, looks like json.loads should be faster than ast.literal_eval or eval (https://stackoverflow.com/questions/9949533/python-eval-vs-ast-literal-eval-vs-json-decode), so I think it could be even faster.

", "rawMarkdown": " Your approach is really good, thank you for the sharing). BTW, looks like json.loads should be faster than ast.literal_eval or eval (https://stackoverflow.com/questions/9949533/python-eval-vs-ast-literal-eval-vs-json-decode), so I think it could be even faster.", "votes": 1 }, { "id": 417692, "postDate": "2018-11-08T16:38:22.277Z", "content": "

how do you determine the image size ? seems the 256*256 is not big enough, and different drawings has different image size

", "rawMarkdown": "how do you determine the image size ? seems the 256*256 is not big enough, and different drawings has different image size" } ] }, { "id": 395135, "postDate": "2018-09-28T02:37:18.837Z", "content": "

Just one update, just got a working code with help of one of the written kernels, it's basicaly what RDizzl3 said:

\n\n
def drawing_to_np(drawing, shape=(64, 64)):\n    drawing = eval(drawing)\n    fig, ax = plt.subplots()\n    for x,y in drawing:\n        ax.plot(x, y, marker='.')\n        ax.axis('off')\n    fig.canvas.draw()\n    # Convert images to numpy arrat\n    np_drawing = np.array(fig.canvas.renderer._renderer)\n\n    return cv2.resize(np_drawing, shape) # Resize array\n
\n\n

How you could apply:

\n\n
train['drawing_converted'] = train['drawing'].map(drawing_to_np)\n
\n\n

but maybe what you may want to do is use the function with batch loading, hope i could help.

", "rawMarkdown": "Just one update, just got a working code with help of one of the written kernels, it's basicaly what RDizzl3 said:\n\n def drawing_to_np(drawing, shape=(64, 64)):\n drawing = eval(drawing)\n fig, ax = plt.subplots()\n for x,y in drawing:\n ax.plot(x, y, marker='.')\n ax.axis('off')\n fig.canvas.draw()\n # Convert images to numpy arrat\n np_drawing = np.array(fig.canvas.renderer._renderer)\n\n return cv2.resize(np_drawing, shape) # Resize array\n\nHow you could apply:\n\n train['drawing_converted'] = train['drawing'].map(drawing_to_np)\n\nbut maybe what you may want to do is use the function with batch loading, hope i could help.", "votes": 3, "replies": [ { "id": 395187, "postDate": "2018-09-28T05:21:25.040Z", "content": "

That's great for starting :)

", "rawMarkdown": "That's great for starting :)", "votes": 1 }, { "id": 395359, "postDate": "2018-09-28T12:29:49.527Z", "content": "

In case anyone want a working code as a demonstration of one way you can do this, checkout this kernel.

", "rawMarkdown": "In case anyone want a working code as a demonstration of one way you can do this, [checkout this kernel.](https://www.kaggle.com/dimitreoliveira/quick-tips-converting-drawings-to-numpy-arrays)" }, { "id": 395526, "postDate": "2018-09-28T18:51:03.183Z", "content": "

I just wanted to make one more comment here. I ran this method now a few times and it was unable to finish every time. The issue? Leaving the figures open inside the for loop causes a MASSIVE memory leak! I order to fix this I had to add

\n\n
plt.close('all')\nplt.gcf()\n
\n\n

after every figure creation. Here is the reference.\nhttps://stackoverflow.com/questions/2364945/matplotlib-runs-out-of-memory-when-plotting-in-a-loop

", "rawMarkdown": "I just wanted to make one more comment here. I ran this method now a few times and it was unable to finish every time. The issue? Leaving the figures open inside the for loop causes a MASSIVE memory leak! I order to fix this I had to add \n\n plt.close('all')\n plt.gcf()\n\nafter every figure creation. Here is the reference.\nhttps://stackoverflow.com/questions/2364945/matplotlib-runs-out-of-memory-when-plotting-in-a-loop", "votes": 6 }, { "id": 398031, "postDate": "2018-10-03T13:19:11.743Z", "content": "

Even after adding these two snippet, kernel crashes everytime. Is there any more optimized way ?

", "rawMarkdown": "Even after adding these two snippet, kernel crashes everytime. Is there any more optimized way ?" }, { "id": 399445, "postDate": "2018-10-05T19:41:14.740Z", "content": "

great code, worked well for me in ipython, but I have some problems running in python: https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/67820 does anybody knows why?

", "rawMarkdown": "great code, worked well for me in ipython, but I have some problems running in python: https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/67820 does anybody knows why?" } ] }, { "id": 395080, "postDate": "2018-09-27T23:34:44.747Z", "content": "

Hi Daniel, currently i'm using the following snippet.

\n\n
drawings = [ast.literal_eval(pts) for pts in train[:1]['drawing'].values]\n\nfor drawing in drawings:\n    for x,y in drawing:\n        plt.plot(x, y, marker='.')\n
\n\n

where train is the training set, and you get \"ast\" lib with \"import ast\"

", "rawMarkdown": "Hi Daniel, currently i'm using the following snippet.\n\n drawings = [ast.literal_eval(pts) for pts in train[:1]['drawing'].values]\n\n for drawing in drawings:\n for x,y in drawing:\n plt.plot(x, y, marker='.')\n\nwhere train is the training set, and you get \"ast\" lib with \"import ast\"", "votes": 1, "replies": [ { "id": 395084, "postDate": "2018-09-28T00:01:27.963Z", "content": "

Thanks for the tip :)

\n\n

But I meant not only to \"visualize\", but use as image data, a numpy image, for instance.

", "rawMarkdown": "Thanks for the tip :)\n\nBut I meant not only to \"visualize\", but use as image data, a numpy image, for instance.", "votes": 1 }, { "id": 395085, "postDate": "2018-09-28T00:11:38.480Z", "content": "

I am using something similar :) From this step you can also extract the matplotlib plot to a numpy array then you can use opencv to do some resizing and convert to grayscale if you wanted too. For now at least I feel like this would be much more efficient than loading the raw dataset.

\n\n

For context here is what the subsequent steps might look like:

\n\n
fig, ax = plt.subplots()\n\ncode from above ...\n\nfig.canvas.draw()\ndata = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')\ndata = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))\ndata = cv2.resize(data, (28, 28), interpolation = cv2.INTER_AREA)\ndata =  cv2.cvtColor(data, cv2.COLOR_BGR2GRAY)\n
", "rawMarkdown": "I am using something similar :) From this step you can also extract the matplotlib plot to a numpy array then you can use opencv to do some resizing and convert to grayscale if you wanted too. For now at least I feel like this would be much more efficient than loading the raw dataset. \n\nFor context here is what the subsequent steps might look like:\n\n fig, ax = plt.subplots()\n \n code from above ...\n \n fig.canvas.draw()\n data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')\n data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))\n data = cv2.resize(data, (28, 28), interpolation = cv2.INTER_AREA)\n data = cv2.cvtColor(data, cv2.COLOR_BGR2GRAY)", "votes": 3 }, { "id": 395114, "postDate": "2018-09-28T01:42:51.383Z", "content": "

Oh Daniel, now i got what you meant, i'm still writing my CNN code, but this may help you, it's a way to convert plot images into numpy arrays.

", "rawMarkdown": "Oh Daniel, now i got what you meant, i'm still writing my CNN code, but this may help you, it's a way to [convert plot images into numpy arrays](https://matplotlib.org/gallery/misc/agg_buffer_to_array.html).", "votes": 1 } ] }, { "id": 397681, "postDate": "2018-10-02T22:54:07.723Z", "content": "

There's also the direct conversion method like in Xander May's kernel. It's fast and memory efficient and should work well. I didn't get good results with classification though.

", "rawMarkdown": "There's also the direct conversion method like in [Xander May's kernel](https://www.kaggle.com/xmaayy/converting-vector-images-to-binary-images). It's fast and memory efficient and should work well. I didn't get good results with classification though." } ], "comments": [ { "id": 396480, "author_name": "Miha Skalic", "author_url": "", "post_date": "2018-09-30T20:55:08.300000", "content": "

matplotlib might be inefficient. Alternative option is using PIL/pillow :

\n\n
import numpy as np\nfrom PIL import Image, ImageDraw\n\ndef draw_it(raw_strokes):\n    image = Image.new(\"P\", (255,255), color=255)\n    image_draw = ImageDraw.Draw(image)\n\n    for stroke in eval(raw_strokes):\n        for i in range(len(stroke[0])-1):\n\n            image_draw.line([stroke[0][i], \n                             stroke[1][i],\n                             stroke[0][i+1], \n                             stroke[1][i+1]],\n                            fill=0, width=6)\n    return np.array(image)\n
", "votes": 8, "replies": [ { "id": 397749, "author_name": "JohnM", "author_url": "", "post_date": "2018-10-03T03:22:18.010000", "content": "

Pillow was much faster for me than matplotlib and without the memory spikes - thanks!

", "votes": 1, "replies": [] }, { "id": 398030, "author_name": "Prajjwal", "author_url": "", "post_date": "2018-10-03T13:18:31.987000", "content": "

What's cline ? Is it dataframe ?

", "votes": 1, "replies": [] }, { "id": 398038, "author_name": "Miha Skalic", "author_url": "", "post_date": "2018-10-03T13:26:45.053000", "content": "

It was a typo, this should be \"raw_strokes\", the stroke string in dataframe.

\n\n

It's fixed now. Thanks for pointing it out.

", "votes": 0, "replies": [] }, { "id": 400057, "author_name": "cab", "author_url": "", "post_date": "2018-10-07T13:31:44.560000", "content": "

Your approach is really fast.
\nThere is a problem that the image size (255x255) maybe not enough to draw some images. Therefore, you can get white image then.

", "votes": 0, "replies": [] }, { "id": 400799, "author_name": "", "author_url": "", "post_date": "2018-10-08T22:58:35.100000", "content": "", "votes": 0, "replies": [] }, { "id": 402208, "author_name": "Danylo Honcharov", "author_url": "", "post_date": "2018-10-11T10:27:14.287000", "content": "

Your approach is really good, thank you for the sharing). BTW, looks like json.loads should be faster than ast.literal_eval or eval (https://stackoverflow.com/questions/9949533/python-eval-vs-ast-literal-eval-vs-json-decode), so I think it could be even faster.

", "votes": 1, "replies": [] }, { "id": 417692, "author_name": "yyqing", "author_url": "", "post_date": "2018-11-08T16:38:22.277000", "content": "

how do you determine the image size ? seems the 256*256 is not big enough, and different drawings has different image size

", "votes": 0, "replies": [] } ] }, { "id": 395135, "author_name": "DimitreOliveira", "author_url": "", "post_date": "2018-09-28T02:37:18.837000", "content": "

Just one update, just got a working code with help of one of the written kernels, it's basicaly what RDizzl3 said:

\n\n
def drawing_to_np(drawing, shape=(64, 64)):\n    drawing = eval(drawing)\n    fig, ax = plt.subplots()\n    for x,y in drawing:\n        ax.plot(x, y, marker='.')\n        ax.axis('off')\n    fig.canvas.draw()\n    # Convert images to numpy arrat\n    np_drawing = np.array(fig.canvas.renderer._renderer)\n\n    return cv2.resize(np_drawing, shape) # Resize array\n
\n\n

How you could apply:

\n\n
train['drawing_converted'] = train['drawing'].map(drawing_to_np)\n
\n\n

but maybe what you may want to do is use the function with batch loading, hope i could help.

", "votes": 3, "replies": [ { "id": 395187, "author_name": "Daniel Möller", "author_url": "", "post_date": "2018-09-28T05:21:25.040000", "content": "

That's great for starting :)

", "votes": 1, "replies": [] }, { "id": 395359, "author_name": "DimitreOliveira", "author_url": "", "post_date": "2018-09-28T12:29:49.527000", "content": "

In case anyone want a working code as a demonstration of one way you can do this, checkout this kernel.

", "votes": 0, "replies": [] }, { "id": 395526, "author_name": "RDizzl3", "author_url": "", "post_date": "2018-09-28T18:51:03.183000", "content": "

I just wanted to make one more comment here. I ran this method now a few times and it was unable to finish every time. The issue? Leaving the figures open inside the for loop causes a MASSIVE memory leak! I order to fix this I had to add

\n\n
plt.close('all')\nplt.gcf()\n
\n\n

after every figure creation. Here is the reference.\nhttps://stackoverflow.com/questions/2364945/matplotlib-runs-out-of-memory-when-plotting-in-a-loop

", "votes": 6, "replies": [] }, { "id": 398031, "author_name": "Prajjwal", "author_url": "", "post_date": "2018-10-03T13:19:11.743000", "content": "

Even after adding these two snippet, kernel crashes everytime. Is there any more optimized way ?

", "votes": 0, "replies": [] }, { "id": 399445, "author_name": "Vitaliy", "author_url": "", "post_date": "2018-10-05T19:41:14.740000", "content": "

great code, worked well for me in ipython, but I have some problems running in python: https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/67820 does anybody knows why?

", "votes": 0, "replies": [] } ] }, { "id": 395080, "author_name": "DimitreOliveira", "author_url": "", "post_date": "2018-09-27T23:34:44.747000", "content": "

Hi Daniel, currently i'm using the following snippet.

\n\n
drawings = [ast.literal_eval(pts) for pts in train[:1]['drawing'].values]\n\nfor drawing in drawings:\n    for x,y in drawing:\n        plt.plot(x, y, marker='.')\n
\n\n

where train is the training set, and you get \"ast\" lib with \"import ast\"

", "votes": 1, "replies": [ { "id": 395084, "author_name": "Daniel Möller", "author_url": "", "post_date": "2018-09-28T00:01:27.963000", "content": "

Thanks for the tip :)

\n\n

But I meant not only to \"visualize\", but use as image data, a numpy image, for instance.

", "votes": 1, "replies": [] }, { "id": 395085, "author_name": "RDizzl3", "author_url": "", "post_date": "2018-09-28T00:11:38.480000", "content": "

I am using something similar :) From this step you can also extract the matplotlib plot to a numpy array then you can use opencv to do some resizing and convert to grayscale if you wanted too. For now at least I feel like this would be much more efficient than loading the raw dataset.

\n\n

For context here is what the subsequent steps might look like:

\n\n
fig, ax = plt.subplots()\n\ncode from above ...\n\nfig.canvas.draw()\ndata = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')\ndata = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))\ndata = cv2.resize(data, (28, 28), interpolation = cv2.INTER_AREA)\ndata =  cv2.cvtColor(data, cv2.COLOR_BGR2GRAY)\n
", "votes": 3, "replies": [] }, { "id": 395114, "author_name": "DimitreOliveira", "author_url": "", "post_date": "2018-09-28T01:42:51.383000", "content": "

Oh Daniel, now i got what you meant, i'm still writing my CNN code, but this may help you, it's a way to convert plot images into numpy arrays.

", "votes": 1, "replies": [] } ] }, { "id": 397681, "author_name": "JohnM", "author_url": "", "post_date": "2018-10-02T22:54:07.723000", "content": "

There's also the direct conversion method like in Xander May's kernel. It's fast and memory efficient and should work well. I didn't get good results with classification though.

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "395069": "Hello, everyone!\n\nAre there preset tools capable of creating pixel images from the given data?\nWhich other approaches are interesting?\n\n", "396480": "matplotlib might be inefficient. Alternative option is using PIL/pillow :\n\n import numpy as np\n from PIL import Image, ImageDraw\n\n def draw_it(raw_strokes):\n image = Image.new(\"P\", (255,255), color=255)\n image_draw = ImageDraw.Draw(image)\n\n for stroke in eval(raw_strokes):\n for i in range(len(stroke[0])-1):\n\n image_draw.line([stroke[0][i], \n stroke[1][i],\n stroke[0][i+1], \n stroke[1][i+1]],\n fill=0, width=6)\n return np.array(image)", "395135": "Just one update, just got a working code with help of one of the written kernels, it's basicaly what RDizzl3 said:\n\n def drawing_to_np(drawing, shape=(64, 64)):\n drawing = eval(drawing)\n fig, ax = plt.subplots()\n for x,y in drawing:\n ax.plot(x, y, marker='.')\n ax.axis('off')\n fig.canvas.draw()\n # Convert images to numpy arrat\n np_drawing = np.array(fig.canvas.renderer._renderer)\n\n return cv2.resize(np_drawing, shape) # Resize array\n\nHow you could apply:\n\n train['drawing_converted'] = train['drawing'].map(drawing_to_np)\n\nbut maybe what you may want to do is use the function with batch loading, hope i could help.", "395080": "Hi Daniel, currently i'm using the following snippet.\n\n drawings = [ast.literal_eval(pts) for pts in train[:1]['drawing'].values]\n\n for drawing in drawings:\n for x,y in drawing:\n plt.plot(x, y, marker='.')\n\nwhere train is the training set, and you get \"ast\" lib with \"import ast\"", "397681": "There's also the direct conversion method like in [Xander May's kernel](https://www.kaggle.com/xmaayy/converting-vector-images-to-binary-images). It's fast and memory efficient and should work well. I didn't get good results with classification though." } }