noviscl / design2code Goto Github PK

improve API design of the eval code so users can either provide one single filename, one single directory, or multiple directories for running eval. also need to benchmark eval speed for people's reference.

visual_eval_v3_multi(input_list, debug=False) right now requires screenshots of the generated webpages as input, which shouldn't be necessary.

Fatal Python error: init_sys_streams: can't initialize sys standard streams
Python runtime state: core initialized
OSError: [Errno 9] Bad file descriptor

Looked through some examples and it seems that the scoring can still be improved, esp. for cases where certain elements are entirely missing from the generation.

• Include image matching (size, positions, etc.). Also consider missing / extra image files.
• Consider visual similarity of bounding box elements: colors, fonts (for text), etc.
• Generate a more explainable multi-dimensional report apart from the final aggregate score. E.g., text element score & image element score / content score & layout score. We can ask humans to judge along the same dimensions.

Need to fix the overlapping and merging issue of text blocks (the right one below):

Need to check what would be an appropriate license for our test data. Also, are we ok with releasing them as part of the repo directly (e.g., should we somehow avoid them from being part of the future GPT training data)?

@NoviScl
Can we remove something like

@import url("http://fonts.googleapis.com/css?family=Open+Sans");

during preprocessing? I find it can sometimes lead to render failure while taking screenshot:

after taking screenshot from (python3 data_utils/screenshot.py ).
How to input that screenshot to the model and gets generated code?

There are some duplicates right now (e.g., screenshot code is also in the metrics modules; image rescaling code is copied over in the GPT-4V module). Would be nice to refactor the code to avoid these.

Example 11625.png in gpt4v_visual_revision_prompting, error below:

Traceback (most recent call last):
 File “eval.py”, line 46, in <module>
  matched, final_score, multi_score = visual_eval_v3(os.path.join(predictions_dir, filename.replace(“.html”, “.png”)), os.path.join(reference_dir, filename.replace(“.html”, “.png”)))
 File “/Users/clsi/Desktop/Pix2Code/Pix2Code/metrics/visual_score.py”, line 963, in visual_eval_v3
  blocks1 = get_blocks_ocr_free(gpt_img)
 File “/Users/clsi/Desktop/Pix2Code/Pix2Code/metrics/ocr_free_utils.py”, line 229, in get_blocks_ocr_free
  different_pixels = find_different_pixels(p_png, p_png_1)
 File “/Users/clsi/Desktop/Pix2Code/Pix2Code/metrics/ocr_free_utils.py”, line 74, in find_different_pixels
  raise ValueError(“Images are not the same size”)
ValueError: Images are not the same size

@StevenyzZhang

The answer is html code, what is the question when train the model?

Bad example (looks like some sort of dating website).

In the Websight dataset, the HTML code contains an image URL. Should the link be replaced?

Add argparse to individual modules like the visual_score or GPT-4V calling so that they can be used in a standalone way.

Add a demo where users give a web link, and we generate webpages with different models and methods and compare all of them side by side.

^

6941.html didn't get a score

Because Playwright is not supported on it.

Right now the final_score from visual_score is not within the range [0,1], would be nice to normalize it.

Where is Pix2Code/metrics/visual_score.py? Did I miss something in the requirements.txt?

The closest reference I found was https://github.com/tonybeltramelli/pix2code/tree/master, which doesn't include any metrics.

For low-level metric, are blocks only contains the text elements? Do the size and position of the image elements also need to be added to the block for comparison?

1. We need high resolutions AND flexible aspect ratio.
   1. playwright: flexible aspect ratio (full page option) but low res(?)
   2. html2image: can set to high res, but need to set aspect ratio.
2. Default background colors can lead to unrecognizable words by OCR.
   1. --default-background-color flag in browser settings
   2. Color differences pose challenges to OCR (the dark color foot note in 1390.png)