\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1708071%2F20e89c79f294d6abd021778084095b68%2Flog_scale.png?generation=1680208071099507&alt=media\")
An example training chart where both the x-axis and y-axis have logarithmic scales: '039d3e82ebaf.jpg' The image is below with an added green rectangle to highlight a point plus coordinate axes added around the image for pixel numbers (created with OpenCV/matplotlib).
\n
Consider the point with the green rectangle added around it. That point clearly has the fifth largest y-value. Looking at the annotation file that means the ground truth for the point is x=114, y=23. Looking at a numpy array for the image it can be checked that the center of the green point has pixel coordinates of pixel_x=281, pixel_y=198. Also, in the image the tick marks on the y-axis tell us y=100 is pixel_y=42, y=10 is pixel_y=224, and y=1 is pixel_y=406. The tick marks and using a log-scale give us a formula of pixel_y = 406 - 182log_10(y). That means the green point (or any point with y=23) should have pixel_y = 406 - 182log_10(23) = 158.2.
\nWhat seems to be happening in the ground truth is the use of a linear scale. Looking between 10 and 100 the linear pixel_y formula would be pixel_y = 224 - 182/90 * (y-10). That means for a point with y=23 we get pixel_y = 224-182/90 * (23-10) = 197.7. The lowest point (x=132, y=3) also seems to be linear, but with a different formula extrapolated from the y-values of 1 and 10.
", "messages": [ { "id": "2203474", "postDate": "03/30/2023 20:44:08", "content": "An example training chart where both the x-axis and y-axis have logarithmic scales: '039d3e82ebaf.jpg' The image is below with an added green rectangle to highlight a point plus coordinate axes added around the image for pixel numbers (created with OpenCV/matplotlib).
\n
Consider the point with the green rectangle added around it. That point clearly has the fifth largest y-value. Looking at the annotation file that means the ground truth for the point is x=114, y=23. Looking at a numpy array for the image it can be checked that the center of the green point has pixel coordinates of pixel_x=281, pixel_y=198. Also, in the image the tick marks on the y-axis tell us y=100 is pixel_y=42, y=10 is pixel_y=224, and y=1 is pixel_y=406. The tick marks and using a log-scale give us a formula of pixel_y = 406 - 182log_10(y). That means the green point (or any point with y=23) should have pixel_y = 406 - 182log_10(23) = 158.2.
\nWhat seems to be happening in the ground truth is the use of a linear scale. Looking between 10 and 100 the linear pixel_y formula would be pixel_y = 224 - 182/90 * (y-10). That means for a point with y=23 we get pixel_y = 224-182/90 * (23-10) = 197.7. The lowest point (x=132, y=3) also seems to be linear, but with a different formula extrapolated from the y-values of 1 and 10.
", "rawMarkdown": "An example training chart where both the x-axis and y-axis have logarithmic scales: '039d3e82ebaf.jpg' The image is below with an added green rectangle to highlight a point plus coordinate axes added around the image for pixel numbers (created with OpenCV/matplotlib).\n\n\nConsider the point with the green rectangle added around it. That point clearly has the fifth largest y-value. Looking at the annotation file that means the ground truth for the point is x=114, y=23. Looking at a numpy array for the image it can be checked that the center of the green point has pixel coordinates of pixel_x=281, pixel_y=198. Also, in the image the tick marks on the y-axis tell us y=100 is pixel_y=42, y=10 is pixel_y=224, and y=1 is pixel_y=406. The tick marks and using a log-scale give us a formula of pixel_y = 406 - 182*log_10(y). That means the green point (or any point with y=23) should have pixel_y = 406 - 182*log_10(23) = 158.2.\n\nWhat seems to be happening in the ground truth is the use of a linear scale. Looking between 10 and 100 the linear pixel_y formula would be pixel_y = 224 - 182/90 * (y-10). That means for a point with y=23 we get pixel_y = 224-182/90 * (23-10) = 197.7. The lowest point (x=132, y=3) also seems to be linear, but with a different formula extrapolated from the y-values of 1 and 10.", "votes": null }, { "id": "2203503", "postDate": "03/30/2023 21:23:46", "content": "good workππππ
", "rawMarkdown": "good workππππ", "votes": null }, { "id": "2225799", "postDate": "04/18/2023 12:52:34", "content": "So in otherwords there's a bug in the scoring!?!
", "rawMarkdown": "So in otherwords there's a bug in the scoring!?!", "votes": null }, { "id": "2225842", "postDate": "04/18/2023 13:24:55", "content": "Hi @pabyrnes, Thanks for the heads up. The inclusion of a log-scale plot in was a mistake. We reviewed the data however, and it appears this is the only instance of one -- there are no log-scale plots in the test set. You can ignore the 039d3e82ebaf image.
good workππππ
", "votes": null, "replies": [] }, { "id": 2225799, "author_name": "nigelhenry", "author_url": "", "post_date": "04/18/2023 12:52:34", "content": "So in otherwords there's a bug in the scoring!?!
", "votes": null, "replies": [] }, { "id": 2225842, "author_name": "ryanholbrook", "author_url": "", "post_date": "04/18/2023 13:24:55", "content": "Hi @pabyrnes, Thanks for the heads up. The inclusion of a log-scale plot in was a mistake. We reviewed the data however, and it appears this is the only instance of one -- there are no log-scale plots in the test set. You can ignore the 039d3e82ebaf image.