Does anyone have any idea why pseudo labeling on effdet is not working very well as with yolov5? my base effdet-d5 with TTA can reach ~0.74, but pseudo labeling can only leverage it by 0.01, which comparing the rise on yolov5 is small. And is there anything I can test out to make it better? Appreciate any suggestion or idea!
", "messages": [ { "id": "949859", "postDate": "07/29/2020 02:48:36", "content": "Does anyone have any idea why pseudo labeling on effdet is not working very well as with yolov5? my base effdet-d5 with TTA can reach ~0.74, but pseudo labeling can only leverage it by 0.01, which comparing the rise on yolov5 is small. And is there anything I can test out to make it better? Appreciate any suggestion or idea!
", "rawMarkdown": "Does anyone have any idea why pseudo labeling on effdet is not working very well as with yolov5? my base effdet-d5 with TTA can reach ~0.74, but pseudo labeling can only leverage it by 0.01, which comparing the rise on yolov5 is small. And is there anything I can test out to make it better? Appreciate any suggestion or idea!", "votes": null }, { "id": "949875", "postDate": "07/29/2020 03:09:35", "content": "I have the same question.
", "rawMarkdown": "I have the same question.", "votes": null }, { "id": "950333", "postDate": "07/29/2020 10:44:04", "content": "A tin foil hat explanation is that this particular dataset was pre-labeled with Yolov5 and then fine tuned by humans, yadda yadda leaking something something.
", "rawMarkdown": "A tin foil hat explanation is that this particular dataset was pre-labeled with Yolov5 and then fine tuned by humans, yadda yadda leaking something something.", "votes": null }, { "id": "951337", "postDate": "07/30/2020 04:37:46", "content": "In may case PL down my score
", "rawMarkdown": "In may case PL down my score", "votes": null }, { "id": "951658", "postDate": "07/30/2020 09:50:31", "content": "I dont agree it worked well for me. My base model 0.7407 boosted to 0.7558
", "rawMarkdown": "I dont agree it worked well for me. My base model 0.7407 boosted to 0.7558", "votes": null }, { "id": "951824", "postDate": "07/30/2020 12:40:20", "content": "Nice job! Would you mind to share how did you set up the PL, like how many epoch and the lr?
", "rawMarkdown": "Nice job! Would you mind to share how did you set up the PL, like how many epoch and the lr?", "votes": null }, { "id": "951974", "postDate": "07/30/2020 14:28:42", "content": "i use this kernel \nhttps://www.kaggle.com/gofarther/efficientdet-framework-tta-pl-etc\nI only change it to apex and my image size is 1024
", "rawMarkdown": "i use this kernel \nhttps://www.kaggle.com/gofarther/efficientdet-framework-tta-pl-etc\nI only change it to apex and my image size is 1024", "votes": null }, { "id": "952306", "postDate": "07/30/2020 19:54:49", "content": "Hi, did you change the image size in inference?
", "rawMarkdown": "Hi, did you change the image size in inference?", "votes": null }, { "id": "952553", "postDate": "07/31/2020 03:36:43", "content": "Hi , can you solve this?
\n\nloss, _, _ = self.model(images, boxes, labels)\nTypeError: forward() takes 3 positional arguments but 4 were given
", "rawMarkdown": "Hi , can you solve this?\n\nloss, _, _ = self.model(images, boxes, labels)\nTypeError: forward() takes 3 positional arguments but 4 were given", "votes": null }, { "id": "952575", "postDate": "07/31/2020 04:00:36", "content": "Yes I change the image size in inference to 1024
", "rawMarkdown": "Yes I change the image size in inference to 1024", "votes": null }, { "id": "952587", "postDate": "07/31/2020 04:15:43", "content": "Here is my train_one_epoch function where I add the gradient accumulation:
\n\n`
\n\n images = torch.stack(images)\n images = images.to(self.device).float()\n batch_size = images.shape[0]\n boxes = [target['boxes'].to(self.device).float() for target in targets]\n labels = [target['labels'].to(self.device).float() for target in targets]\n\n loss, _, _ = self.model(images, boxes, labels)\n if self.mixed_precision:\n with amp.scale_loss(loss, self.optimizer) as scaled_loss:\n scaled_loss.backward()\n else:\n loss.backward()\n\n if step % self.accumulate == 0:\n self.optimizer.step()\n self.optimizer.zero_grad()\n\n summary_loss.update(loss.detach().item(), batch_size)\n\n return summary_loss`\nThanks
", "rawMarkdown": "Thanks", "votes": null }, { "id": "952598", "postDate": "07/31/2020 04:33:37", "content": "What is the corresponding validation loss that gives a LB score 0.74
", "rawMarkdown": "What is the corresponding validation loss that gives a LB score 0.74", "votes": null }, { "id": "952668", "postDate": "07/31/2020 06:04:54", "content": "@zekunn this comment will help you
", "rawMarkdown": "zekunn this [comment ](https://www.kaggle.com/shonenkov/training-efficientdet/comments#889181)will help you", "votes": null }, { "id": "952682", "postDate": "07/31/2020 06:29:41", "content": "Thank you!
", "rawMarkdown": "Thank you!", "votes": null }, { "id": "952751", "postDate": "07/31/2020 07:29:42", "content": "I can confirm a boost of about 0.01 LB (0.74x to 0.75x) using the same TTA and PL configuration as this notebook for EfficientDet, using image size 512, not 1024.
", "rawMarkdown": "I can confirm a boost of about 0.01 LB (0.74x to 0.75x) using the same TTA and PL configuration as this notebook for EfficientDet, using image size 512, not 1024.", "votes": null }, { "id": "952782", "postDate": "07/31/2020 08:12:45", "content": "How much time did it take you to run 7 epochs? Once we submit. we do not need to keep GPU on in the original notebook, right? I submitted this notebook as is and it's been running for 3 hours. Cannot figure out what I am doing wrong.
", "rawMarkdown": "How much time did it take you to run 7 epochs? Once we submit. we do not need to keep GPU on in the original notebook, right? I submitted this notebook as is and it's been running for 3 hours. Cannot figure out what I am doing wrong.", "votes": null }, { "id": "952806", "postDate": "07/31/2020 08:30:01", "content": "More than 3 hours
", "rawMarkdown": "More than 3 hours", "votes": null }, { "id": "953053", "postDate": "07/31/2020 13:34:11", "content": "@ahmmadrashid Great job!\nYou mean you use 1024 size in both training and inferencing?
", "rawMarkdown": "ahmmadrashid Great job!\nYou mean you use 1024 size in both training and inferencing?", "votes": null }, { "id": "953837", "postDate": "08/01/2020 07:00:15", "content": "Notebook Exceeded Allowed Compute\nI always meet this....😧
", "rawMarkdown": "Notebook Exceeded Allowed Compute\nI always meet this....😧", "votes": null }, { "id": "954470", "postDate": "08/01/2020 18:51:22", "content": "@ahmmadrashid What is best loss did you get ,I can reach up to 328 but not sure why I don't get corresponding score
", "rawMarkdown": "ahmmadrashid What is best loss did you get ,I can reach up to 328 but not sure why I don't get corresponding score", "votes": null }, { "id": "954812", "postDate": "08/02/2020 05:50:00", "content": "@ahmmadrashid hi there, is your current score based on efficientdet as well?
", "rawMarkdown": "ahmmadrashid hi there, is your current score based on efficientdet as well?", "votes": null }, { "id": "955662", "postDate": "08/02/2020 19:50:25", "content": "Yes I used 1024 both training and inference
", "rawMarkdown": "Yes I used 1024 both training and inference", "votes": null } ], "comments": [ { "id": 949875, "author_name": "zekunn", "author_url": "", "post_date": "07/29/2020 03:09:35", "content": "I have the same question.
", "votes": null, "replies": [] }, { "id": 950333, "author_name": "suicaokhoailang", "author_url": "", "post_date": "07/29/2020 10:44:04", "content": "A tin foil hat explanation is that this particular dataset was pre-labeled with Yolov5 and then fine tuned by humans, yadda yadda leaking something something.
", "votes": null, "replies": [] }, { "id": 951337, "author_name": "doanquanvietnamca", "author_url": "", "post_date": "07/30/2020 04:37:46", "content": "In may case PL down my score
", "votes": null, "replies": [] }, { "id": 951658, "author_name": "ahmmadrashid", "author_url": "", "post_date": "07/30/2020 09:50:31", "content": "I dont agree it worked well for me. My base model 0.7407 boosted to 0.7558
", "votes": null, "replies": [ { "id": 951824, "author_name": "creatrol", "author_url": "", "post_date": "07/30/2020 12:40:20", "content": "Nice job! Would you mind to share how did you set up the PL, like how many epoch and the lr?
", "votes": null, "replies": [] }, { "id": 951974, "author_name": "ahmmadrashid", "author_url": "", "post_date": "07/30/2020 14:28:42", "content": "i use this kernel \nhttps://www.kaggle.com/gofarther/efficientdet-framework-tta-pl-etc\nI only change it to apex and my image size is 1024
", "votes": null, "replies": [] }, { "id": 952306, "author_name": "chrisstan", "author_url": "", "post_date": "07/30/2020 19:54:49", "content": "Hi, did you change the image size in inference?
", "votes": null, "replies": [] }, { "id": 952553, "author_name": "zekunn", "author_url": "", "post_date": "07/31/2020 03:36:43", "content": "Hi , can you solve this?
\n\nloss, _, _ = self.model(images, boxes, labels)\nTypeError: forward() takes 3 positional arguments but 4 were given
", "votes": null, "replies": [] }, { "id": 952575, "author_name": "ahmmadrashid", "author_url": "", "post_date": "07/31/2020 04:00:36", "content": "Yes I change the image size in inference to 1024
", "votes": null, "replies": [] }, { "id": 952587, "author_name": "ahmmadrashid", "author_url": "", "post_date": "07/31/2020 04:15:43", "content": "Here is my train_one_epoch function where I add the gradient accumulation:
\n\n`
\n\n images = torch.stack(images)\n images = images.to(self.device).float()\n batch_size = images.shape[0]\n boxes = [target['boxes'].to(self.device).float() for target in targets]\n labels = [target['labels'].to(self.device).float() for target in targets]\n\n loss, _, _ = self.model(images, boxes, labels)\n if self.mixed_precision:\n with amp.scale_loss(loss, self.optimizer) as scaled_loss:\n scaled_loss.backward()\n else:\n loss.backward()\n\n if step % self.accumulate == 0:\n self.optimizer.step()\n self.optimizer.zero_grad()\n\n summary_loss.update(loss.detach().item(), batch_size)\n\n return summary_loss`\nThanks
", "votes": null, "replies": [] }, { "id": 952598, "author_name": "sebastianji", "author_url": "", "post_date": "07/31/2020 04:33:37", "content": "What is the corresponding validation loss that gives a LB score 0.74
", "votes": null, "replies": [] }, { "id": 952668, "author_name": "arcticyen", "author_url": "", "post_date": "07/31/2020 06:04:54", "content": "@zekunn this comment will help you
", "votes": null, "replies": [] }, { "id": 952682, "author_name": "zekunn", "author_url": "", "post_date": "07/31/2020 06:29:41", "content": "Thank you!
", "votes": null, "replies": [] }, { "id": 952751, "author_name": "khoongweihao", "author_url": "", "post_date": "07/31/2020 07:29:42", "content": "I can confirm a boost of about 0.01 LB (0.74x to 0.75x) using the same TTA and PL configuration as this notebook for EfficientDet, using image size 512, not 1024.
", "votes": null, "replies": [] }, { "id": 953053, "author_name": "chrisstan", "author_url": "", "post_date": "07/31/2020 13:34:11", "content": "@ahmmadrashid Great job!\nYou mean you use 1024 size in both training and inferencing?
", "votes": null, "replies": [] }, { "id": 954470, "author_name": "jaideepvalani", "author_url": "", "post_date": "08/01/2020 18:51:22", "content": "@ahmmadrashid What is best loss did you get ,I can reach up to 328 but not sure why I don't get corresponding score
", "votes": null, "replies": [] }, { "id": 954812, "author_name": "reighns", "author_url": "", "post_date": "08/02/2020 05:50:00", "content": "@ahmmadrashid hi there, is your current score based on efficientdet as well?
", "votes": null, "replies": [] }, { "id": 955662, "author_name": "ahmmadrashid", "author_url": "", "post_date": "08/02/2020 19:50:25", "content": "Yes I used 1024 both training and inference
", "votes": null, "replies": [] } ] }, { "id": 952782, "author_name": "sebastianji", "author_url": "", "post_date": "07/31/2020 08:12:45", "content": "How much time did it take you to run 7 epochs? Once we submit. we do not need to keep GPU on in the original notebook, right? I submitted this notebook as is and it's been running for 3 hours. Cannot figure out what I am doing wrong.
", "votes": null, "replies": [ { "id": 952806, "author_name": "guizengyou", "author_url": "", "post_date": "07/31/2020 08:30:01", "content": "More than 3 hours
", "votes": null, "replies": [] }, { "id": 953837, "author_name": "zekunn", "author_url": "", "post_date": "08/01/2020 07:00:15", "content": "Notebook Exceeded Allowed Compute\nI always meet this....😧
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "949859": "Does anyone have any idea why pseudo labeling on effdet is not working very well as with yolov5? my base effdet-d5 with TTA can reach ~0.74, but pseudo labeling can only leverage it by 0.01, which comparing the rise on yolov5 is small. And is there anything I can test out to make it better? Appreciate any suggestion or idea!", "949875": "I have the same question.", "950333": "A tin foil hat explanation is that this particular dataset was pre-labeled with Yolov5 and then fine tuned by humans, yadda yadda leaking something something.", "951337": "In may case PL down my score", "951658": "I dont agree it worked well for me. My base model 0.7407 boosted to 0.7558", "951824": "Nice job! Would you mind to share how did you set up the PL, like how many epoch and the lr?", "951974": "i use this kernel \nhttps://www.kaggle.com/gofarther/efficientdet-framework-tta-pl-etc\nI only change it to apex and my image size is 1024", "952306": "Hi, did you change the image size in inference?", "952553": "Hi , can you solve this?\n\nloss, _, _ = self.model(images, boxes, labels)\nTypeError: forward() takes 3 positional arguments but 4 were given", "952575": "Yes I change the image size in inference to 1024", "952587": "Here is my train_one_epoch function where I add the gradient accumulation:\n\n`\n \n images = torch.stack(images)\n images = images.to(self.device).float()\n batch_size = images.shape[0]\n boxes = [target['boxes'].to(self.device).float() for target in targets]\n labels = [target['labels'].to(self.device).float() for target in targets]\n\n loss, _, _ = self.model(images, boxes, labels)\n if self.mixed_precision:\n with amp.scale_loss(loss, self.optimizer) as scaled_loss:\n scaled_loss.backward()\n else:\n loss.backward()\n\n if step % self.accumulate == 0:\n self.optimizer.step()\n self.optimizer.zero_grad()\n\n summary_loss.update(loss.detach().item(), batch_size)\n\n return summary_loss`", "952592": "Thanks", "952598": "What is the corresponding validation loss that gives a LB score 0.74", "952668": "zekunn this [comment ](https://www.kaggle.com/shonenkov/training-efficientdet/comments#889181)will help you", "952682": "Thank you!", "952751": "I can confirm a boost of about 0.01 LB (0.74x to 0.75x) using the same TTA and PL configuration as this notebook for EfficientDet, using image size 512, not 1024.", "952782": "How much time did it take you to run 7 epochs? Once we submit. we do not need to keep GPU on in the original notebook, right? I submitted this notebook as is and it's been running for 3 hours. Cannot figure out what I am doing wrong.", "952806": "More than 3 hours", "953053": "ahmmadrashid Great job!\nYou mean you use 1024 size in both training and inferencing?", "953837": "Notebook Exceeded Allowed Compute\nI always meet this....😧", "954470": "ahmmadrashid What is best loss did you get ,I can reach up to 328 but not sure why I don't get corresponding score", "954812": "ahmmadrashid hi there, is your current score based on efficientdet as well?", "955662": "Yes I used 1024 both training and inference" }, "source": "meta" }