To improve prediction accuracy, it may be useful to use data extensions within the pre-processing function. If this idea is correct, what would be the variety of images that the model should address in a more appropriate way?
\nI assume a random Since it is a graph, inversions and rotations may not have much significance.
\nAs a test, we modified the preprocessing function in 🍩 donut-infer (LB 0.44) [benetech] as follows
():\n pixel_values = []\n\n \n transform = transforms.Compose([\n transforms.RandomHorizontalFlip(),\n transforms.RandomRotation(),\n transforms.RandomResizedCrop((, ), scale=(, )),\n transforms.ColorJitter(brightness=, contrast=, saturation=, hue=),\n ])\n\n sample examples[]:\n arr = np.array(sample)\n\n \n \n (arr.shape) == :\n ()\n (arr.shape)\n arr = np.stack([arr]*, axis=-)\n\n \n arr = Image.fromarray(arr)\n\n \n arr = transform(arr)\n\n \n arr = np.array(arr)\n\n pixel_values.append(processor(arr, random_padding=).pixel_values)\n\n {\n : torch.tensor(np.vstack(pixel_values)),\n }\nHowever, although we tried and saw this correction with data expansion, the scores deteriorated rapidly. Is it wrong that we did it with trained data?Is it not a good idea to begin with?
\nI would be grateful for your wisdom
If you crop the image, the x or y-axis values could be gone.
\nIn real life, we don't have any rotated charts.
\nAlso, it's rare to see a flipped plot in the test images.
\nI believe the y-axis would be on the left side of the plot.
\nThat's just my opinion.
I appreciate your comments. It is certainly pointless for general data expansion.
", "rawMarkdown": "I appreciate your comments. It is certainly pointless for general data expansion." } ] }, { "id": 2252063, "postDate": "2023-05-09T21:16:08.763Z", "content": "To improve prediction accuracy, it may be useful to use data extensions within the pre-processing function. If this idea is correct, what would be the variety of images that the model should address in a more appropriate way?
\nI assume a random Since it is a graph, inversions and rotations may not have much significance.
\nAs a test, we modified the preprocessing function in 🍩 donut-infer (LB 0.44) [benetech] as follows
():\n pixel_values = []\n\n \n transform = transforms.Compose([\n transforms.RandomHorizontalFlip(),\n transforms.RandomRotation(),\n transforms.RandomResizedCrop((, ), scale=(, )),\n transforms.ColorJitter(brightness=, contrast=, saturation=, hue=),\n ])\n\n sample examples[]:\n arr = np.array(sample)\n\n \n \n (arr.shape) == :\n ()\n (arr.shape)\n arr = np.stack([arr]*, axis=-)\n\n \n arr = Image.fromarray(arr)\n\n \n arr = transform(arr)\n\n \n arr = np.array(arr)\n\n pixel_values.append(processor(arr, random_padding=).pixel_values)\n\n {\n : torch.tensor(np.vstack(pixel_values)),\n }\nHowever, although we tried and saw this correction with data expansion, the scores deteriorated rapidly. Is it wrong that we did it with trained data?Is it not a good idea to begin with?
\nI would be grateful for your wisdom
In the train set, generated and extracted images could be very different. This has already been discussed somewhere.
\nFor example, the scatter plots, when you look at maximum and minimum x-axis values of data ('data-series'), the ~100 extracted images have a very different distribution than ~11k generated images.
\nMaybe it would be good to generate more scatter images such that their x-values more resemble the x-values of the ~100 extracted images. I guess this can be done by simply duplicating some images or adding noise to them.
We appreciate your comments. We will read the other discussions carefully. Thank you also for your excellent ideas on data extensions.
", "rawMarkdown": "We appreciate your comments. We will read the other discussions carefully. Thank you also for your excellent ideas on data extensions." } ] } ], "comments": [ { "id": 2252106, "author_name": "june", "author_url": "", "post_date": "2023-05-09T23:11:01.847000", "content": "If you crop the image, the x or y-axis values could be gone.
\nIn real life, we don't have any rotated charts.
\nAlso, it's rare to see a flipped plot in the test images.
\nI believe the y-axis would be on the left side of the plot.
\nThat's just my opinion.
I appreciate your comments. It is certainly pointless for general data expansion.
", "votes": 0, "replies": [] } ] }, { "id": 2254259, "author_name": "", "author_url": "", "post_date": "2023-05-10T19:04:52.800000", "content": "In the train set, generated and extracted images could be very different. This has already been discussed somewhere.
\nFor example, the scatter plots, when you look at maximum and minimum x-axis values of data ('data-series'), the ~100 extracted images have a very different distribution than ~11k generated images.
\nMaybe it would be good to generate more scatter images such that their x-values more resemble the x-values of the ~100 extracted images. I guess this can be done by simply duplicating some images or adding noise to them.
We appreciate your comments. We will read the other discussions carefully. Thank you also for your excellent ideas on data extensions.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "2252106": "If you crop the image, the x or y-axis values could be gone.\nIn real life, we don't have any rotated charts.\nAlso, it's rare to see a flipped plot in the test images.\nI believe the y-axis would be on the left side of the plot. \nThat's just my opinion.", "2252063": "To improve prediction accuracy, it may be useful to use data extensions within the pre-processing function. If this idea is correct, what would be the variety of images that the model should address in a more appropriate way?\nI assume a random Since it is a graph, inversions and rotations may not have much significance.\nAs a test, we modified the preprocessing function in [🍩 donut-infer (LB 0.44) [benetech]](https://www.kaggle.com/code/nbroad/donut-infer-lb-0-44-benetech) as follows\n\n```python\ndef preprocess(examples, processor):\n pixel_values = []\n\n # Data augmentation\n transform = transforms.Compose([\n transforms.RandomHorizontalFlip(),\n transforms.RandomRotation(10),\n transforms.RandomResizedCrop((224, 224), scale=(0.8, 1.0)),\n transforms.ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1, hue=0.1),\n ])\n\n for sample in examples[\"image_path\"]:\n arr = np.array(sample)\n \n # There are some grayscale images that were making this fail\n # This prevents that.\n if len(arr.shape) == 2:\n print(\"Changing grayscale to 3 channel format\")\n print(arr.shape)\n arr = np.stack([arr]*3, axis=-1)\n \n # Convert numpy array to PIL Image\n arr = Image.fromarray(arr)\n \n # Apply data augmentation\n arr = transform(arr)\n \n # Convert PIL Image back to numpy array\n arr = np.array(arr)\n \n pixel_values.append(processor(arr, random_padding=True).pixel_values)\n \n return {\n \"pixel_values\": torch.tensor(np.vstack(pixel_values)),\n }\n```\nHowever, although we tried and saw this correction with data expansion, the scores deteriorated rapidly. Is it wrong that we did it with trained data?Is it not a good idea to begin with?\nI would be grateful for your wisdom", "2254259": "In the train set, generated and extracted images could be very different. This has already been discussed somewhere. \n\nFor example, the scatter plots, when you look at maximum and minimum x-axis values of data ('data-series'), the ~100 extracted images have a very different distribution than ~11k generated images.\nMaybe it would be good to generate more scatter images such that their x-values more resemble the x-values of the ~100 extracted images. I guess this can be done by simply duplicating some images or adding noise to them." } }