This post is continuation of link.
\nIn order to visualize and log F2 score during validation of yolov5 follow the steps in this discussion, (thanks to @amiiiney) or below I am providing all the steps.
metrics.py filef2 = 5 * p * r / (4 * p + r + 1e-16) and below this change f1 to f2 everywhere in the ap_per_class function and return return tp, fp, p, r, f2, ap, unique_classes.astype(\"int32\").fitness function as below:def fitness(x):\n # Model fitness as a weighted combination of metrics\n w = [0.0, 0.0, 0.0, 0.0, 1.0] # weights for [P, R, mAP@0.5, mAP@0.5:0.95, F2@0.3:0.8]\n return (x[:, :5] * w).sum(1)\nval.py fileiouv= torch.from_numpy(np.arange(0.3, 0.85, 0.05)).to(device)f1 in this to f2 and same for this line and -- modify next 2 lines to ap50, ap, f2 = ap[:, 0], ap.mean(1), f2.mean(0) # AP@0.5, AP@0.5:0.95\nmp, mr, f2, map50, map = p.mean(), r.mean(), f2.mean(), ap50.mean(), ap.mean()\n-- change the following two lines in val.py here and here, we can have Yolo display the F2 on the command line while training. (thanks @cdeotte )
\n--and return (mp, mr, map50, map, f2, *(loss.cpu() / len(dataloader)).tolist()), maps, t
\"metrics/F2\" to this list and in a list self.best_keys in next line \"best/F2\" on_fit_epoch_end function modify this line with best_results = [epoch] + vals[3:8] After making these modifications you will be able to visualize and log F2 score on wandb(not sure about other loggers).
LAST UPDATED:
\nThis post is continuation of link.
\nIn order to visualize and log F2 score during validation of yolov5 follow the steps in this discussion, (thanks to @amiiiney) or below I am providing all the steps.
metrics.py filef2 = 5 * p * r / (4 * p + r + 1e-16) and below this change f1 to f2 everywhere in the ap_per_class function and return return tp, fp, p, r, f2, ap, unique_classes.astype(\"int32\").fitness function as below:def fitness(x):\n # Model fitness as a weighted combination of metrics\n w = [0.0, 0.0, 0.0, 0.0, 1.0] # weights for [P, R, mAP@0.5, mAP@0.5:0.95, F2@0.3:0.8]\n return (x[:, :5] * w).sum(1)\nval.py fileiouv= torch.from_numpy(np.arange(0.3, 0.85, 0.05)).to(device)f1 in this to f2 and same for this line and -- modify next 2 lines to ap50, ap, f2 = ap[:, 0], ap.mean(1), f2.mean(0) # AP@0.5, AP@0.5:0.95\nmp, mr, f2, map50, map = p.mean(), r.mean(), f2.mean(), ap50.mean(), ap.mean()\n-- change the following two lines in val.py here and here, we can have Yolo display the F2 on the command line while training. (thanks @cdeotte )
\n--and return (mp, mr, map50, map, f2, *(loss.cpu() / len(dataloader)).tolist()), maps, t
\"metrics/F2\" to this list and in a list self.best_keys in next line \"best/F2\" on_fit_epoch_end function modify this line with best_results = [epoch] + vals[3:8] After making these modifications you will be able to visualize and log F2 score on wandb(not sure about other loggers).
LAST UPDATED:
\nThanks @sanchitvj and @amiiiney, this works perfectly! Note the first file's name is metrics.py not matrix.py. Also, if we change the following two lines in val.py here and here, we can have Yolo display the F2 on the command line while training.
Thank you @cdeotte for valuable suggestion. I have made the changes :)
", "rawMarkdown": "Thank you @cdeotte for valuable suggestion. I have made the changes :)", "votes": 1 }, { "id": 1671048, "postDate": "2022-02-01T05:47:25.557Z", "content": "I have a question for you experts！！
\nI trained yolov5s6 with size3600 and then used the following notebook https://www.kaggle.com/kocha1/only-yolov5-tracking-lb-642 for inference. conf0.34 iou0.5, in the case where I only change the img-size of inference:
\nimg-size 3600 lb0.579
\nimg-size 9000 lb0.577
\nWhy does my self-trained model infer no change at large scale?
\nHave you encountered this situation?
\n\nchange the following two lines in val.py here and here, we can have Yolo display the F2 on the command line while training.
\n
What should I change in these lines? i tried to add \"f2\" but it did not work. Need help.
", "rawMarkdown": "> change the following two lines in val.py [here](https://github.com/ultralytics/yolov5/blob/bd815d48df18a23e2bb08d88e430183bfb48eb78/val.py#L175) and [here](https://github.com/ultralytics/yolov5/blob/bd815d48df18a23e2bb08d88e430183bfb48eb78/val.py#L263), we can have Yolo display the F2 on the command line while training.\n\nWhat should I change in these lines? i tried to add \"f2\" but it did not work. Need help.\n", "votes": 3, "replies": [ { "id": 1675041, "postDate": "2022-02-04T01:31:22.747Z", "content": "Change the first line to
\ns = ('%20s' + '%11s' * 6) % ('Class', 'Images', 'Labels', 'P', 'R', 'F2', 'mAP@.5')
Change the second line to
\nLOGGER.info(pf % ('all', seen, nt.sum(), mp, mr, f2, map50))
(and do all the other changes in original discussion post above)
", "rawMarkdown": "Change the first line to \n`s = ('%20s' + '%11s' * 6) % ('Class', 'Images', 'Labels', 'P', 'R', 'F2', 'mAP@.5')`\n\nChange the second line to\n`LOGGER.info(pf % ('all', seen, nt.sum(), mp, mr, f2, map50))`\n\n(and do all the other changes in original discussion post above)", "votes": 5 }, { "id": 1675702, "postDate": "2022-02-04T12:22:32.880Z", "content": "Thank you, it is working now.
", "rawMarkdown": "Thank you, it is working now.", "votes": 1 } ] }, { "id": 1674065, "postDate": "2022-02-03T07:44:18.903Z", "content": "Why change the self.csv = False? If we keep this line in __init__.py , we can have the PLT chart result.
I think, The code should only modify w = [0.2, 0.8, 0.0, 0.0] in fitness function. Because, in F2 score P less important than R 4 times.
", "rawMarkdown": "I think, The code should only modify w = [0.2, 0.8, 0.0, 0.0] in fitness function. Because, in F2 score P less important than R 4 times.\n", "votes": 1, "replies": [ { "id": 1670282, "postDate": "2022-01-31T11:39:28.160Z", "content": "w = [0.0, 0.0, 0.0, 0.0, 1.0] here we are giving priority to F2 score. So by default P is less prioritized
Hi @sanchitvj,
\nDo you use the same hyperparameters as \"Sheep\" shared in this post: https://www.kaggle.com/steamedsheep/yolov5-high-resolution-training/data
Do you use albumentations and is the LB score that you reached with yolov5l6 or a smaller model?
\nThank you for the interesting discussions in this competition!
", "rawMarkdown": "Hi @sanchitvj,\nDo you use the same hyperparameters as \"Sheep\" shared in this post: [https://www.kaggle.com/steamedsheep/yolov5-high-resolution-training/data ](https://www.kaggle.com/steamedsheep/yolov5-high-resolution-training/data)\n\nDo you use albumentations and is the LB score that you reached with yolov5l6 or a smaller model? \n\nThank you for the interesting discussions in this competition!", "votes": 1, "replies": [ { "id": 1667073, "postDate": "2022-01-28T09:44:41.597Z", "content": "Hi, no hyperparams are different, no albumentations and model is s6 only. Happy to help :)
", "rawMarkdown": "Hi, no hyperparams are different, no albumentations and model is s6 only. Happy to help :) ", "votes": 3 }, { "id": 1667110, "postDate": "2022-01-28T10:22:57.100Z", "content": "Oh wow! TY @sanchitvj I'm now testing what the right img size is. What img size do you use for train and infer now?
", "rawMarkdown": "Oh wow! TY @sanchitvj I'm now testing what the right img size is. What img size do you use for train and infer now?" } ] }, { "id": 1659734, "postDate": "2022-01-22T04:44:56.327Z", "content": "Did you check your LB when you infer with original size? (size when you training) Thanks
", "rawMarkdown": "Did you check your LB when you infer with original size? (size when you training) Thanks", "votes": 1, "replies": [ { "id": 1659747, "postDate": "2022-01-22T05:10:59.820Z", "content": "Not yet. But planning to do this once done with hyperparameter optimization.
", "rawMarkdown": "Not yet. But planning to do this once done with hyperparameter optimization.", "votes": 1 }, { "id": 1659891, "postDate": "2022-01-22T07:20:53.420Z", "content": "I understood. So your LB is upscale inference?
", "rawMarkdown": "I understood. So your LB is upscale inference?" }, { "id": 1660061, "postDate": "2022-01-22T11:03:20.040Z", "content": "yes….1.6, 2x, 2.2x
", "rawMarkdown": "yes....1.6, 2x, 2.2x\n", "votes": 1 }, { "id": 1660267, "postDate": "2022-01-22T14:40:24.237Z", "content": "@sanchitvj is is best conf selected from F2_curves ?
", "rawMarkdown": "@sanchitvj is is best conf selected from F2_curves ? " }, { "id": 1660334, "postDate": "2022-01-22T15:45:23.557Z", "content": "We are still working on that.
", "rawMarkdown": "We are still working on that. ", "votes": 1 } ] }, { "id": 1659140, "postDate": "2022-01-21T15:15:23.347Z", "content": "Hi @sanchitvj
\nthis line seem to compute F1 AP50? so you are calculating F2 at AP50 and not average 0.3:0.8:0.05 per class (this case is 1). Correct me if I wrong. Thanks
In step2 I am taking the average torch.from_numpy(np.arange(0.3, 0.85, 0.05)).to(device)
Traceback (most recent call last):
\n File \"train.py\", line 637, in
\n main(opt)
\n File \"train.py\", line 534, in main
\n train(opt.hyp, opt, device, callbacks)
\n File \"train.py\", line 376, in train
\n compute_loss=compute_loss)
\n File \"/home/ubuntu/anaconda3/envs/yolo5/lib/python3.7/site-packages/torch/autograd/grad_mode.py\", line 28, in decorate_context
\n return func(*args, **kwargs)
\n File \"/home/ubuntu/Kaggle/yolov5/val.py\", line 255, in run
\n ap50, ap, f2 = ap[:, 0], ap.mean(1), f2.mean(1) # AP@0.5, AP@0.5:0.95
\n File \"/home/ubuntu/anaconda3/envs/yolo5/lib/python3.7/site-packages/numpy/core/_methods.py\", line 167, in _mean
\n rcount = _count_reduce_items(arr, axis, keepdims=keepdims, where=where)
\n File \"/home/ubuntu/anaconda3/envs/yolo5/lib/python3.7/site-packages/numpy/core/_methods.py\", line 76, in _count_reduce_items
\n items *= arr.shape[mu.normalize_axis_index(ax, arr.ndim)]
\nnumpy.AxisError: axis 1 is out of bounds for array of dimension 1
GOT this error
", "rawMarkdown": "Traceback (most recent call last):\n File \"train.py\", line 637, in@sanchitvj can u help in this?
", "rawMarkdown": "@sanchitvj can u help in this?" }, { "id": 1659369, "postDate": "2022-01-21T18:30:57.590Z", "content": "\n", "rawMarkdown": "> ```\n> ap50, ap, f2 = ap[:, 0], ap.mean(1), f2.mean(0) # AP@0.5, AP@0.5:0.95\n> mp, mr, f2, map50, map = p.mean(), r.mean(), f2.mean(), ap50.mean(), ap.mean()\n> ```\nSorry for the mistake changed to `f2.mean(0)` in first line \n" }, { "id": 1659722, "postDate": "2022-01-22T04:27:34.520Z", "content": "\nap50, ap, f2 = ap[:, 0], ap.mean(1), f2.mean(0) # AP@0.5, AP@0.5:0.95\nmp, mr, f2, map50, map = p.mean(), r.mean(), f2.mean(), ap50.mean(), ap.mean()\nSorry for the mistake changed to
\nf2.mean(0)in first line
@sanchitvj thanks bro its working now
", "rawMarkdown": "@sanchitvj thanks bro its working now", "votes": 1 } ] }, { "id": 1659927, "postDate": "2022-01-22T08:17:01.590Z", "content": "For wandb logging, need to add this to add 'metrics/f2'.
", "rawMarkdown": "For wandb logging, need to add [this](https://github.com/ultralytics/yolov5/blob/bd815d48df18a23e2bb08d88e430183bfb48eb78/utils/loggers/__init__.py#L47) to add 'metrics/f2'.", "votes": 2, "replies": [ { "id": 1659940, "postDate": "2022-01-22T08:30:27.573Z", "content": "Thank you @ichimarugin…. I have updated :)
", "rawMarkdown": "Thank you @ichimarugin.... I have updated :)" } ] }, { "id": 1679104, "postDate": "2022-02-07T01:30:19.510Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1674619, "postDate": "2022-02-03T16:37:03.663Z", "content": "Thank you for sharing.
", "rawMarkdown": "Thank you for sharing.", "votes": 1 } ], "comments": [ { "id": 1670785, "author_name": "Chris Deotte", "author_url": "", "post_date": "2022-01-31T21:35:56.273000", "content": "Thanks @sanchitvj and @amiiiney, this works perfectly! Note the first file's name is metrics.py not matrix.py. Also, if we change the following two lines in val.py here and here, we can have Yolo display the F2 on the command line while training.
Thank you @cdeotte for valuable suggestion. I have made the changes :)
", "votes": 1, "replies": [] }, { "id": 1671048, "author_name": "zhiqiang he", "author_url": "", "post_date": "2022-02-01T05:47:25.557000", "content": "I have a question for you experts！！
\nI trained yolov5s6 with size3600 and then used the following notebook https://www.kaggle.com/kocha1/only-yolov5-tracking-lb-642 for inference. conf0.34 iou0.5, in the case where I only change the img-size of inference:
\nimg-size 3600 lb0.579
\nimg-size 9000 lb0.577
\nWhy does my self-trained model infer no change at large scale?
\nHave you encountered this situation?
\n\nchange the following two lines in val.py here and here, we can have Yolo display the F2 on the command line while training.
\n
What should I change in these lines? i tried to add \"f2\" but it did not work. Need help.
", "votes": 3, "replies": [ { "id": 1675041, "author_name": "Chris Deotte", "author_url": "", "post_date": "2022-02-04T01:31:22.747000", "content": "Change the first line to
\ns = ('%20s' + '%11s' * 6) % ('Class', 'Images', 'Labels', 'P', 'R', 'F2', 'mAP@.5')
Change the second line to
\nLOGGER.info(pf % ('all', seen, nt.sum(), mp, mr, f2, map50))
(and do all the other changes in original discussion post above)
", "votes": 5, "replies": [] }, { "id": 1675702, "author_name": "somuSan", "author_url": "", "post_date": "2022-02-04T12:22:32.880000", "content": "Thank you, it is working now.
", "votes": 1, "replies": [] } ] }, { "id": 1674065, "author_name": "grhuang", "author_url": "", "post_date": "2022-02-03T07:44:18.903000", "content": "Why change the self.csv = False? If we keep this line in __init__.py , we can have the PLT chart result.
I think, The code should only modify w = [0.2, 0.8, 0.0, 0.0] in fitness function. Because, in F2 score P less important than R 4 times.
", "votes": 1, "replies": [ { "id": 1670282, "author_name": "Sanchit Vijay", "author_url": "", "post_date": "2022-01-31T11:39:28.160000", "content": "w = [0.0, 0.0, 0.0, 0.0, 1.0] here we are giving priority to F2 score. So by default P is less prioritized
Hi @sanchitvj,
\nDo you use the same hyperparameters as \"Sheep\" shared in this post: https://www.kaggle.com/steamedsheep/yolov5-high-resolution-training/data
Do you use albumentations and is the LB score that you reached with yolov5l6 or a smaller model?
\nThank you for the interesting discussions in this competition!
", "votes": 1, "replies": [ { "id": 1667073, "author_name": "Sanchit Vijay", "author_url": "", "post_date": "2022-01-28T09:44:41.597000", "content": "Hi, no hyperparams are different, no albumentations and model is s6 only. Happy to help :)
", "votes": 3, "replies": [] }, { "id": 1667110, "author_name": "Meesz9", "author_url": "", "post_date": "2022-01-28T10:22:57.100000", "content": "Oh wow! TY @sanchitvj I'm now testing what the right img size is. What img size do you use for train and infer now?
", "votes": 0, "replies": [] } ] }, { "id": 1659734, "author_name": "Hiếu Phạm Trần Minh", "author_url": "", "post_date": "2022-01-22T04:44:56.327000", "content": "Did you check your LB when you infer with original size? (size when you training) Thanks
", "votes": 1, "replies": [ { "id": 1659747, "author_name": "Sanchit Vijay", "author_url": "", "post_date": "2022-01-22T05:10:59.820000", "content": "Not yet. But planning to do this once done with hyperparameter optimization.
", "votes": 1, "replies": [] }, { "id": 1659891, "author_name": "Hiếu Phạm Trần Minh", "author_url": "", "post_date": "2022-01-22T07:20:53.420000", "content": "I understood. So your LB is upscale inference?
", "votes": 0, "replies": [] }, { "id": 1660061, "author_name": "Sanchit Vijay", "author_url": "", "post_date": "2022-01-22T11:03:20.040000", "content": "yes….1.6, 2x, 2.2x
", "votes": 1, "replies": [] }, { "id": 1660267, "author_name": "SeshuRaju 🧘♂️", "author_url": "", "post_date": "2022-01-22T14:40:24.237000", "content": "@sanchitvj is is best conf selected from F2_curves ?
", "votes": 0, "replies": [] }, { "id": 1660334, "author_name": "Sanchit Vijay", "author_url": "", "post_date": "2022-01-22T15:45:23.557000", "content": "We are still working on that.
", "votes": 1, "replies": [] } ] }, { "id": 1659140, "author_name": "Ơ con lừa!", "author_url": "", "post_date": "2022-01-21T15:15:23.347000", "content": "Hi @sanchitvj
\nthis line seem to compute F1 AP50? so you are calculating F2 at AP50 and not average 0.3:0.8:0.05 per class (this case is 1). Correct me if I wrong. Thanks
In step2 I am taking the average torch.from_numpy(np.arange(0.3, 0.85, 0.05)).to(device)
Traceback (most recent call last):
\n File \"train.py\", line 637, in
\n main(opt)
\n File \"train.py\", line 534, in main
\n train(opt.hyp, opt, device, callbacks)
\n File \"train.py\", line 376, in train
\n compute_loss=compute_loss)
\n File \"/home/ubuntu/anaconda3/envs/yolo5/lib/python3.7/site-packages/torch/autograd/grad_mode.py\", line 28, in decorate_context
\n return func(*args, **kwargs)
\n File \"/home/ubuntu/Kaggle/yolov5/val.py\", line 255, in run
\n ap50, ap, f2 = ap[:, 0], ap.mean(1), f2.mean(1) # AP@0.5, AP@0.5:0.95
\n File \"/home/ubuntu/anaconda3/envs/yolo5/lib/python3.7/site-packages/numpy/core/_methods.py\", line 167, in _mean
\n rcount = _count_reduce_items(arr, axis, keepdims=keepdims, where=where)
\n File \"/home/ubuntu/anaconda3/envs/yolo5/lib/python3.7/site-packages/numpy/core/_methods.py\", line 76, in _count_reduce_items
\n items *= arr.shape[mu.normalize_axis_index(ax, arr.ndim)]
\nnumpy.AxisError: axis 1 is out of bounds for array of dimension 1
GOT this error
", "votes": 1, "replies": [ { "id": 1659139, "author_name": "Anshul Khadse", "author_url": "", "post_date": "2022-01-21T15:15:18.307000", "content": "@sanchitvj can u help in this?
", "votes": 0, "replies": [] }, { "id": 1659369, "author_name": "Sanchit Vijay", "author_url": "", "post_date": "2022-01-21T18:30:57.590000", "content": "\n", "votes": 0, "replies": [] }, { "id": 1659722, "author_name": "Anshul Khadse", "author_url": "", "post_date": "2022-01-22T04:27:34.520000", "content": "\nap50, ap, f2 = ap[:, 0], ap.mean(1), f2.mean(0) # AP@0.5, AP@0.5:0.95\nmp, mr, f2, map50, map = p.mean(), r.mean(), f2.mean(), ap50.mean(), ap.mean()\nSorry for the mistake changed to
\nf2.mean(0)in first line
@sanchitvj thanks bro its working now
", "votes": 1, "replies": [] } ] }, { "id": 1659927, "author_name": "Ichimaru Gin", "author_url": "", "post_date": "2022-01-22T08:17:01.590000", "content": "For wandb logging, need to add this to add 'metrics/f2'.
", "votes": 2, "replies": [ { "id": 1659940, "author_name": "Sanchit Vijay", "author_url": "", "post_date": "2022-01-22T08:30:27.573000", "content": "Thank you @ichimarugin…. I have updated :)
", "votes": 0, "replies": [] } ] }, { "id": 1679104, "author_name": "", "author_url": "", "post_date": "2022-02-07T01:30:19.510000", "content": "", "votes": 0, "replies": [] }, { "id": 1674619, "author_name": "DongYuan", "author_url": "", "post_date": "2022-02-03T16:37:03.663000", "content": "Thank you for sharing.
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "1659077": "This post is continuation of [link](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/300638#1658343).\nIn order to visualize and log F2 score during validation of [yolov5](https://github.com/ultralytics/yolov5) follow the steps in [this](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/300638#1658343) discussion, (thanks to @amiiiney) or below I am providing all the steps.\n\n- In `metrics.py` file\n-- change [this](https://github.com/ultralytics/yolov5/blob/bd815d48df18a23e2bb08d88e430183bfb48eb78/utils/metrics.py#L73) line to `f2 = 5 * p * r / (4 * p + r + 1e-16)` and below this change `f1` to `f2` everywhere in the `ap_per_class` function and return `return tp, fp, p, r, f2, ap, unique_classes.astype(\"int32\")`.\n-- next modify `fitness` function as below:\n```\ndef fitness(x):\n # Model fitness as a weighted combination of metrics\n w = [0.0, 0.0, 0.0, 0.0, 1.0] # weights for [P, R, mAP@0.5, mAP@0.5:0.95, F2@0.3:0.8]\n return (x[:, :5] * w).sum(1)\n```\n\n- In `val.py` file\n-- modify [this](https://github.com/ultralytics/yolov5/blob/bd815d48df18a23e2bb08d88e430183bfb48eb78/val.py#L160) line with `iouv= torch.from_numpy(np.arange(0.3, 0.85, 0.05)).to(device)`\n-- change `f1` in [this](https://github.com/ultralytics/yolov5/blob/bd815d48df18a23e2bb08d88e430183bfb48eb78/val.py#L176) to `f2` and same for [this](https://github.com/ultralytics/yolov5/blob/bd815d48df18a23e2bb08d88e430183bfb48eb78/val.py#L254) line and -- modify next 2 lines to \n```\nap50, ap, f2 = ap[:, 0], ap.mean(1), f2.mean(0) # AP@0.5, AP@0.5:0.95\nmp, mr, f2, map50, map = p.mean(), r.mean(), f2.mean(), ap50.mean(), ap.mean()\n```\n-- change the following two lines in val.py [here](https://github.com/ultralytics/yolov5/blob/bd815d48df18a23e2bb08d88e430183bfb48eb78/val.py#L175) and [here](https://github.com/ultralytics/yolov5/blob/bd815d48df18a23e2bb08d88e430183bfb48eb78/val.py#L263), we can have Yolo display the F2 on the command line while training. (thanks @cdeotte )\n--and `return (mp, mr, map50, map, f2, *(loss.cpu() / len(dataloader)).tolist()), maps, t`\n\n- In [this](https://github.com/ultralytics/yolov5/blob/master/utils/loggers/__init__.py) file\n-- add `\"metrics/F2\"` to [this](https://github.com/ultralytics/yolov5/blob/bd815d48df18a23e2bb08d88e430183bfb48eb78/utils/loggers/__init__.py#L47) list and in a list `self.best_keys` in next line `\"best/F2\"` \n-- In `on_fit_epoch_end` function modify [this](https://github.com/ultralytics/yolov5/blob/bd815d48df18a23e2bb08d88e430183bfb48eb78/utils/loggers/__init__.py#L130) line with `best_results = [epoch] + vals[3:8]` \n\n---\n\nAfter making these modifications you will be able to visualize and log F2 score on `wandb`(not sure about other loggers).\n- Thanks to @amiiiney for providing solution.\n- Thanks to @cdeotte for suggestion.\nSugggestions or modifications are welcome :)\n\nLAST UPDATED:\n- 22 Jan, 2 PM (IST)\n- 1 Feb, 11 AM (IST)", "1670785": "Thanks @sanchitvj and @amiiiney, this works perfectly! Note the first file's name is `metrics.py` not `matrix.py`. Also, if we change the following two lines in `val.py` [here][1] and [here][2], we can have Yolo display the F2 on the command line while training.\n\n[1]: https://github.com/ultralytics/yolov5/blob/bd815d48df18a23e2bb08d88e430183bfb48eb78/val.py#L175\n[2]: https://github.com/ultralytics/yolov5/blob/bd815d48df18a23e2bb08d88e430183bfb48eb78/val.py#L263", "1674965": "> change the following two lines in val.py [here](https://github.com/ultralytics/yolov5/blob/bd815d48df18a23e2bb08d88e430183bfb48eb78/val.py#L175) and [here](https://github.com/ultralytics/yolov5/blob/bd815d48df18a23e2bb08d88e430183bfb48eb78/val.py#L263), we can have Yolo display the F2 on the command line while training.\n\nWhat should I change in these lines? i tried to add \"f2\" but it did not work. Need help.\n", "1674065": "Why change the `self.csv = False`? If we keep this line in `__init__.py` , we can have the PLT chart result.", "1668295": "I think, The code should only modify w = [0.2, 0.8, 0.0, 0.0] in fitness function. Because, in F2 score P less important than R 4 times.\n", "1667045": "Hi @sanchitvj,\nDo you use the same hyperparameters as \"Sheep\" shared in this post: [https://www.kaggle.com/steamedsheep/yolov5-high-resolution-training/data ](https://www.kaggle.com/steamedsheep/yolov5-high-resolution-training/data)\n\nDo you use albumentations and is the LB score that you reached with yolov5l6 or a smaller model? \n\nThank you for the interesting discussions in this competition!", "1659734": "Did you check your LB when you infer with original size? (size when you training) Thanks", "1659140": "Hi @sanchitvj \nthis line seem to compute F1 AP50? so you are calculating F2 at AP50 and not average 0.3:0.8:0.05 per class (this case is 1). Correct me if I wrong. Thanks ", "1659138": "Traceback (most recent call last):\n File \"train.py\", line 637, in