Link:https://github.com/spytensor/kaggle_human_protein_baseline
", "messages": [ { "id": "428480", "postDate": "11/27/2018 10:58:00", "content": "Link:https://github.com/spytensor/kaggle_human_protein_baseline
", "rawMarkdown": "Link:https://github.com/spytensor/kaggle_human_protein_baseline", "votes": null }, { "id": "430213", "postDate": "11/30/2018 03:10:32", "content": "Hi spytensor, thank you very much for your baseline. I run the code by myself and get 0.460.\nI've got a question, which i cannot figure out:\nDuring training, f1 macro for training set and val set is about 0.26 or so. But f1 macro for test set is 0.460.\nwhy is that?
", "rawMarkdown": "Hi spytensor, thank you very much for your baseline. I run the code by myself and get 0.460.\nI've got a question, which i cannot figure out:\nDuring training, f1 macro for training set and val set is about 0.26 or so. But f1 macro for test set is 0.460.\nwhy is that?", "votes": null }, { "id": "430356", "postDate": "11/30/2018 08:59:41", "content": "I also found that question, I get 0.41 on val set,0.56 on train set and 0.461 on test set. One reason may be the f1-score method I used in my code.You can write yourself and if solved ,please tell me.
", "rawMarkdown": "I also found that question, I get 0.41 on val set,0.56 on train set and 0.461 on test set. One reason may be the f1-score method I used in my code.You can write yourself and if solved ,please tell me.", "votes": null }, { "id": "434873", "postDate": "12/07/2018 04:53:25", "content": "Thanks to your share, i am a noob in image classification.
\n\nYour opensource is quite helpful to me, thank you!
", "rawMarkdown": "Thanks to your share, i am a noob in image classification.\n\nYour opensource is quite helpful to me, thank you!", "votes": null }, { "id": "436348", "postDate": "12/10/2018 06:40:39", "content": "Hello ,where is the pretrained models,thanks!
", "rawMarkdown": "Hello ,where is the pretrained models,thanks!", "votes": null }, { "id": "436407", "postDate": "12/10/2018 08:41:07", "content": "Run the model ,and it will be downloaded automatically to path \"~/user/.torch/models/\".Or you can download yourself after you googled .
", "rawMarkdown": "Run the model ,and it will be downloaded automatically to path \"~/user/.torch/models/\".Or you can download yourself after you googled .", "votes": null }, { "id": "436408", "postDate": "12/10/2018 08:41:35", "content": "Enjoy the competition.
", "rawMarkdown": "Enjoy the competition.", "votes": null }, { "id": "436642", "postDate": "12/10/2018 17:32:25", "content": "I guess you need to calculate f1score for entire validation data not batch. Averaging each batch f1score gives you wrong score.
", "rawMarkdown": "I guess you need to calculate f1score for entire validation data not batch. Averaging each batch f1score gives you wrong score.", "votes": null }, { "id": "436849", "postDate": "12/11/2018 02:54:07", "content": "I‘ve realized that,thanks!\nThere is “from pretrainedmodels.models import bninception” in your model.py file, so the pretrainedmodels.models is need.
", "rawMarkdown": "I‘ve realized that,thanks!\nThere is “from pretrainedmodels.models import bninception” in your model.py file, so the pretrainedmodels.models is need.", "votes": null }, { "id": "437538", "postDate": "12/12/2018 05:22:22", "content": "The author caculated loss each batch, which i thinks is okay, but i don't think it is suitable for f1 score. While val, he caculated f1-score each batch, and use the average of them to replace the real f1-score,\nwhich i think is not right.
\n\nHere is my code: while val, get all the preditection and real_label and then caculate the f1-score.
\n\n```
\n\ndef evaluate(val_loader,model,criterion,epoch,train_loss,best_results,start):\n # only meter loss and f1 score\n losses = AverageMeter()\n f1 = AverageMeter()\n # switch mode for evaluation\n model.cuda()\n model.eval()\n with torch.no_grad():\n for i, (images,target) in enumerate(val_loader):\n # 1、获取新的一个batch的结果\n images_var = images.cuda(non_blocking=True)\n target = torch.from_numpy(np.array(target)).float().cuda(non_blocking=True)\n #image_var = Variable(images).cuda()\n #target = Variable(torch.from_numpy(np.array(target)).long()).cuda()\n output = model(images_var)
\n\n # 2、把每一个batch的结果合并\n if i==0:\n total_output = output\n total_target = target\n else:\n total_output = torch.cat([total_output, output],0)\n total_target = torch.cat([total_target, target],0)\n\n # 3、计算整个验证集上的得分\n loss = criterion(total_output, total_target)\n losses.update(loss.item(),images_var.size(0))\n f1_batch = f1_score(total_target, total_output.sigmoid().cpu().data.numpy() > 0.15,average='macro')\n f1.update(f1_batch,images_var.size(0))\n print('\\r',end='',flush=True)\n message = '%s %5.1f %6.1f | %0.3f %0.3f | %0.3f %0.4f | %s %s | %s' % (\\\n \"val\", i/len(val_loader) + epoch, epoch, \n train_loss[0], train_loss[1], \n losses.avg, f1.avg,\n str(best_results[0])[:8],str(best_results[1])[:8],\n time_to_str((timer() - start),'min'))\n\n print(message, end='',flush=True)\n log.write(\"\\n\")\n #log.write(message)\n #log.write(\"\\n\")\n\nreturn [losses.avg,f1.avg]\n```
", "rawMarkdown": "The author caculated loss each batch, which i thinks is okay, but i don't think it is suitable for f1 score. While val, he caculated f1-score each batch, and use the average of them to replace the real f1-score,\nwhich i think is not right.\n\nHere is my code: while val, get all the preditection and real_label and then caculate the f1-score.\n\n```\n# 2. evaluate fuunction\ndef evaluate(val_loader,model,criterion,epoch,train_loss,best_results,start):\n # only meter loss and f1 score\n losses = AverageMeter()\n f1 = AverageMeter()\n # switch mode for evaluation\n model.cuda()\n model.eval()\n with torch.no_grad():\n for i, (images,target) in enumerate(val_loader):\n # 1、获取新的一个batch的结果\n images_var = images.cuda(non_blocking=True)\n target = torch.from_numpy(np.array(target)).float().cuda(non_blocking=True)\n #image_var = Variable(images).cuda()\n #target = Variable(torch.from_numpy(np.array(target)).long()).cuda()\n output = model(images_var)\n \n # 2、把每一个batch的结果合并\n if i==0:\n total_output = output\n total_target = target\n else:\n total_output = torch.cat([total_output, output],0)\n total_target = torch.cat([total_target, target],0)\n\n # 3、计算整个验证集上的得分\n loss = criterion(total_output, total_target)\n losses.update(loss.item(),images_var.size(0))\n f1_batch = f1_score(total_target, total_output.sigmoid().cpu().data.numpy() > 0.15,average='macro')\n f1.update(f1_batch,images_var.size(0))\n print('\\r',end='',flush=True)\n message = '%s %5.1f %6.1f | %0.3f %0.3f | %0.3f %0.4f | %s %s | %s' % (\\\n \"val\", i/len(val_loader) + epoch, epoch, \n train_loss[0], train_loss[1], \n losses.avg, f1.avg,\n str(best_results[0])[:8],str(best_results[1])[:8],\n time_to_str((timer() - start),'min'))\n\n print(message, end='',flush=True)\n log.write(\"\\n\")\n #log.write(message)\n #log.write(\"\\n\")\n \n return [losses.avg,f1.avg]\n```", "votes": null }, { "id": "439884", "postDate": "12/16/2018 15:22:57", "content": "Thanks for sharing a really good BSL.
", "rawMarkdown": "Thanks for sharing a really good BSL.", "votes": null }, { "id": "440196", "postDate": "12/17/2018 07:42:31", "content": "I am wondering why SGD optimizer is used in your code. As far as I know, Adam performs better as it combines RMSprop and Momentum.\nI tried Adam but turned to get a terrible 0.33LB.\nCan anyone tell me the reason.
", "rawMarkdown": "I am wondering why SGD optimizer is used in your code. As far as I know, Adam performs better as it combines RMSprop and Momentum.\nI tried Adam but turned to get a terrible 0.33LB.\nCan anyone tell me the reason.", "votes": null }, { "id": "442363", "postDate": "12/19/2018 21:29:21", "content": "Hi EeAary,\nDid you also modify the learning rate that is set to 0.03. Such a learning rate is usable for the SGD optimizer...but for Adam you'd better start with the default of 0.001 or other minor variation on it.\nSee if that works for you.
", "rawMarkdown": "Hi EeAary,\nDid you also modify the learning rate that is set to 0.03. Such a learning rate is usable for the SGD optimizer...but for Adam you'd better start with the default of 0.001 or other minor variation on it.\nSee if that works for you.", "votes": null }, { "id": "442447", "postDate": "12/20/2018 01:18:08", "content": "Thanks for ur sharing
", "rawMarkdown": "Thanks for ur sharing", "votes": null }, { "id": "442521", "postDate": "12/20/2018 04:01:27", "content": "Definely I have modified the learning rate to 0.001 using Adam, and I set a StepLR with gamma 0.1 every 5 epoch. Anyway I failed to reach the loss as low as SGD did.\nI've obtain a better score but still interested in this problem. Maybe it just likes playing Hearthstone, I don't know what cards I'm going to get.
", "rawMarkdown": "Definely I have modified the learning rate to 0.001 using Adam, and I set a StepLR with gamma 0.1 every 5 epoch. Anyway I failed to reach the loss as low as SGD did.\nI've obtain a better score but still interested in this problem. Maybe it just likes playing Hearthstone, I don't know what cards I'm going to get.", "votes": null }, { "id": "442685", "postDate": "12/20/2018 10:29:57", "content": "Give it a try with a stepsize of 8 or 10 epochs. Very likely it is to early to lower the learningrate. Alternativly ... Increase gamma to say something like 0.5. That should likely increase your results. Good luck.
", "rawMarkdown": "Give it a try with a stepsize of 8 or 10 epochs. Very likely it is to early to lower the learningrate. Alternativly ... Increase gamma to say something like 0.5. That should likely increase your results. Good luck.", "votes": null }, { "id": "443660", "postDate": "12/22/2018 04:01:11", "content": "Thanks so much, it works!
", "rawMarkdown": "Thanks so much, it works!", "votes": null }, { "id": "447148", "postDate": "12/29/2018 08:15:20", "content": "Hi @Spytensor, i have some difficulty understanding your code. If you can help me, it would be great. Thanks in advance.\nIn your HumanDataset class:
\n\ny = str(self.images_df.iloc[index].Id.absolute()) # what is this absolute() function? \n@Spytensor could you please explain how can i feed 512X512 images to resnext architecture. Please.
", "rawMarkdown": "Spytensor could you please explain how can i feed 512X512 images to resnext architecture. Please.", "votes": null }, { "id": "447550", "postDate": "12/30/2018 03:49:58", "content": "Change the model's avgpool to adaptiveavgpool like this:\npython\nmodel = resnext 50()\nmodel.avg_pool = nn.AdaptiveAvgPool2d(1)\n
Thanks
", "rawMarkdown": "Thanks", "votes": null }, { "id": "447966", "postDate": "12/30/2018 23:48:12", "content": "The elements in Id column are objects from pathlib, the method gets you the absolute path to the image file.
", "rawMarkdown": "The elements in Id column are objects from pathlib, the method gets you the absolute path to the image file.", "votes": null }, { "id": "448781", "postDate": "01/02/2019 04:42:51", "content": "@pascal1129\nThanks for sharing,does the train code need to be changed?I changed it in your way, but I got it wrong.\nIf you don’t change how to judge the gap between train and val.
", "rawMarkdown": "pascal1129\nThanks for sharing,does the train code need to be changed?I changed it in your way, but I got it wrong.\nIf you don’t change how to judge the gap between train and val.", "votes": null }, { "id": "448921", "postDate": "01/02/2019 11:06:58", "content": "Indeed, I changed lots, but the details are difficult to explain here.
", "rawMarkdown": "Indeed, I changed lots, but the details are difficult to explain here.", "votes": null }, { "id": "448944", "postDate": "01/02/2019 12:00:28", "content": "ok,thanks
", "rawMarkdown": "ok,thanks", "votes": null }, { "id": "1496033", "postDate": "08/30/2021 03:26:22", "content": "Thanks for sharing. This has really made me learn a lot. I also found that change to (better) model or let the image size larger could get better results.
", "rawMarkdown": "Thanks for sharing. This has really made me learn a lot. I also found that change to (better) model or let the image size larger could get better results.", "votes": null } ], "comments": [ { "id": 1496033, "author_name": "nanamiyvonne", "author_url": "", "post_date": "08/30/2021 03:26:22", "content": "Thanks for sharing. This has really made me learn a lot. I also found that change to (better) model or let the image size larger could get better results.
", "votes": null, "replies": [] }, { "id": 430213, "author_name": "irvingzhang0512", "author_url": "", "post_date": "11/30/2018 03:10:32", "content": "Hi spytensor, thank you very much for your baseline. I run the code by myself and get 0.460.\nI've got a question, which i cannot figure out:\nDuring training, f1 macro for training set and val set is about 0.26 or so. But f1 macro for test set is 0.460.\nwhy is that?
", "votes": null, "replies": [ { "id": 437538, "author_name": "pascal1129", "author_url": "", "post_date": "12/12/2018 05:22:22", "content": "The author caculated loss each batch, which i thinks is okay, but i don't think it is suitable for f1 score. While val, he caculated f1-score each batch, and use the average of them to replace the real f1-score,\nwhich i think is not right.
\n\nHere is my code: while val, get all the preditection and real_label and then caculate the f1-score.
\n\n```
\n\ndef evaluate(val_loader,model,criterion,epoch,train_loss,best_results,start):\n # only meter loss and f1 score\n losses = AverageMeter()\n f1 = AverageMeter()\n # switch mode for evaluation\n model.cuda()\n model.eval()\n with torch.no_grad():\n for i, (images,target) in enumerate(val_loader):\n # 1、获取新的一个batch的结果\n images_var = images.cuda(non_blocking=True)\n target = torch.from_numpy(np.array(target)).float().cuda(non_blocking=True)\n #image_var = Variable(images).cuda()\n #target = Variable(torch.from_numpy(np.array(target)).long()).cuda()\n output = model(images_var)
\n\n # 2、把每一个batch的结果合并\n if i==0:\n total_output = output\n total_target = target\n else:\n total_output = torch.cat([total_output, output],0)\n total_target = torch.cat([total_target, target],0)\n\n # 3、计算整个验证集上的得分\n loss = criterion(total_output, total_target)\n losses.update(loss.item(),images_var.size(0))\n f1_batch = f1_score(total_target, total_output.sigmoid().cpu().data.numpy() > 0.15,average='macro')\n f1.update(f1_batch,images_var.size(0))\n print('\\r',end='',flush=True)\n message = '%s %5.1f %6.1f | %0.3f %0.3f | %0.3f %0.4f | %s %s | %s' % (\\\n \"val\", i/len(val_loader) + epoch, epoch, \n train_loss[0], train_loss[1], \n losses.avg, f1.avg,\n str(best_results[0])[:8],str(best_results[1])[:8],\n time_to_str((timer() - start),'min'))\n\n print(message, end='',flush=True)\n log.write(\"\\n\")\n #log.write(message)\n #log.write(\"\\n\")\n\nreturn [losses.avg,f1.avg]\n```
", "votes": null, "replies": [] }, { "id": 448781, "author_name": "wwchaoo", "author_url": "", "post_date": "01/02/2019 04:42:51", "content": "@pascal1129\nThanks for sharing,does the train code need to be changed?I changed it in your way, but I got it wrong.\nIf you don’t change how to judge the gap between train and val.
", "votes": null, "replies": [] }, { "id": 448921, "author_name": "pascal1129", "author_url": "", "post_date": "01/02/2019 11:06:58", "content": "Indeed, I changed lots, but the details are difficult to explain here.
", "votes": null, "replies": [] }, { "id": 448944, "author_name": "wwchaoo", "author_url": "", "post_date": "01/02/2019 12:00:28", "content": "ok,thanks
", "votes": null, "replies": [] } ] }, { "id": 430356, "author_name": "spytensor", "author_url": "", "post_date": "11/30/2018 08:59:41", "content": "I also found that question, I get 0.41 on val set,0.56 on train set and 0.461 on test set. One reason may be the f1-score method I used in my code.You can write yourself and if solved ,please tell me.
", "votes": null, "replies": [ { "id": 436642, "author_name": "appian", "author_url": "", "post_date": "12/10/2018 17:32:25", "content": "I guess you need to calculate f1score for entire validation data not batch. Averaging each batch f1score gives you wrong score.
", "votes": null, "replies": [] } ] }, { "id": 434873, "author_name": "pascal1129", "author_url": "", "post_date": "12/07/2018 04:53:25", "content": "Thanks to your share, i am a noob in image classification.
\n\nYour opensource is quite helpful to me, thank you!
", "votes": null, "replies": [ { "id": 436408, "author_name": "spytensor", "author_url": "", "post_date": "12/10/2018 08:41:35", "content": "Enjoy the competition.
", "votes": null, "replies": [] } ] }, { "id": 436348, "author_name": "ispectre", "author_url": "", "post_date": "12/10/2018 06:40:39", "content": "Hello ,where is the pretrained models,thanks!
", "votes": null, "replies": [ { "id": 436407, "author_name": "spytensor", "author_url": "", "post_date": "12/10/2018 08:41:07", "content": "Run the model ,and it will be downloaded automatically to path \"~/user/.torch/models/\".Or you can download yourself after you googled .
", "votes": null, "replies": [] }, { "id": 436849, "author_name": "ispectre", "author_url": "", "post_date": "12/11/2018 02:54:07", "content": "I‘ve realized that,thanks!\nThere is “from pretrainedmodels.models import bninception” in your model.py file, so the pretrainedmodels.models is need.
", "votes": null, "replies": [] } ] }, { "id": 439884, "author_name": "kentchun33333", "author_url": "", "post_date": "12/16/2018 15:22:57", "content": "Thanks for sharing a really good BSL.
", "votes": null, "replies": [] }, { "id": 440196, "author_name": "eeaary", "author_url": "", "post_date": "12/17/2018 07:42:31", "content": "I am wondering why SGD optimizer is used in your code. As far as I know, Adam performs better as it combines RMSprop and Momentum.\nI tried Adam but turned to get a terrible 0.33LB.\nCan anyone tell me the reason.
", "votes": null, "replies": [ { "id": 442363, "author_name": "rsmits", "author_url": "", "post_date": "12/19/2018 21:29:21", "content": "Hi EeAary,\nDid you also modify the learning rate that is set to 0.03. Such a learning rate is usable for the SGD optimizer...but for Adam you'd better start with the default of 0.001 or other minor variation on it.\nSee if that works for you.
", "votes": null, "replies": [] }, { "id": 442521, "author_name": "eeaary", "author_url": "", "post_date": "12/20/2018 04:01:27", "content": "Definely I have modified the learning rate to 0.001 using Adam, and I set a StepLR with gamma 0.1 every 5 epoch. Anyway I failed to reach the loss as low as SGD did.\nI've obtain a better score but still interested in this problem. Maybe it just likes playing Hearthstone, I don't know what cards I'm going to get.
", "votes": null, "replies": [] }, { "id": 442685, "author_name": "rsmits", "author_url": "", "post_date": "12/20/2018 10:29:57", "content": "Give it a try with a stepsize of 8 or 10 epochs. Very likely it is to early to lower the learningrate. Alternativly ... Increase gamma to say something like 0.5. That should likely increase your results. Good luck.
", "votes": null, "replies": [] }, { "id": 443660, "author_name": "eeaary", "author_url": "", "post_date": "12/22/2018 04:01:11", "content": "Thanks so much, it works!
", "votes": null, "replies": [] } ] }, { "id": 442447, "author_name": "desperadod", "author_url": "", "post_date": "12/20/2018 01:18:08", "content": "Thanks for ur sharing
", "votes": null, "replies": [] }, { "id": 447148, "author_name": "adilurrahim", "author_url": "", "post_date": "12/29/2018 08:15:20", "content": "Hi @Spytensor, i have some difficulty understanding your code. If you can help me, it would be great. Thanks in advance.\nIn your HumanDataset class:
\n\ny = str(self.images_df.iloc[index].Id.absolute()) # what is this absolute() function? \nThe elements in Id column are objects from pathlib, the method gets you the absolute path to the image file.
", "votes": null, "replies": [] } ] }, { "id": 447422, "author_name": "adilurrahim", "author_url": "", "post_date": "12/29/2018 19:39:01", "content": "@Spytensor could you please explain how can i feed 512X512 images to resnext architecture. Please.
", "votes": null, "replies": [ { "id": 447550, "author_name": "spytensor", "author_url": "", "post_date": "12/30/2018 03:49:58", "content": "Change the model's avgpool to adaptiveavgpool like this:\npython\nmodel = resnext 50()\nmodel.avg_pool = nn.AdaptiveAvgPool2d(1)\n
Thanks
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "428480": "Link:https://github.com/spytensor/kaggle_human_protein_baseline", "430213": "Hi spytensor, thank you very much for your baseline. I run the code by myself and get 0.460.\nI've got a question, which i cannot figure out:\nDuring training, f1 macro for training set and val set is about 0.26 or so. But f1 macro for test set is 0.460.\nwhy is that?", "430356": "I also found that question, I get 0.41 on val set,0.56 on train set and 0.461 on test set. One reason may be the f1-score method I used in my code.You can write yourself and if solved ,please tell me.", "434873": "Thanks to your share, i am a noob in image classification.\n\nYour opensource is quite helpful to me, thank you!", "436348": "Hello ,where is the pretrained models,thanks!", "436407": "Run the model ,and it will be downloaded automatically to path \"~/user/.torch/models/\".Or you can download yourself after you googled .", "436408": "Enjoy the competition.", "436642": "I guess you need to calculate f1score for entire validation data not batch. Averaging each batch f1score gives you wrong score.", "436849": "I‘ve realized that,thanks!\nThere is “from pretrainedmodels.models import bninception” in your model.py file, so the pretrainedmodels.models is need.", "437538": "The author caculated loss each batch, which i thinks is okay, but i don't think it is suitable for f1 score. While val, he caculated f1-score each batch, and use the average of them to replace the real f1-score,\nwhich i think is not right.\n\nHere is my code: while val, get all the preditection and real_label and then caculate the f1-score.\n\n```\n# 2. evaluate fuunction\ndef evaluate(val_loader,model,criterion,epoch,train_loss,best_results,start):\n # only meter loss and f1 score\n losses = AverageMeter()\n f1 = AverageMeter()\n # switch mode for evaluation\n model.cuda()\n model.eval()\n with torch.no_grad():\n for i, (images,target) in enumerate(val_loader):\n # 1、获取新的一个batch的结果\n images_var = images.cuda(non_blocking=True)\n target = torch.from_numpy(np.array(target)).float().cuda(non_blocking=True)\n #image_var = Variable(images).cuda()\n #target = Variable(torch.from_numpy(np.array(target)).long()).cuda()\n output = model(images_var)\n \n # 2、把每一个batch的结果合并\n if i==0:\n total_output = output\n total_target = target\n else:\n total_output = torch.cat([total_output, output],0)\n total_target = torch.cat([total_target, target],0)\n\n # 3、计算整个验证集上的得分\n loss = criterion(total_output, total_target)\n losses.update(loss.item(),images_var.size(0))\n f1_batch = f1_score(total_target, total_output.sigmoid().cpu().data.numpy() > 0.15,average='macro')\n f1.update(f1_batch,images_var.size(0))\n print('\\r',end='',flush=True)\n message = '%s %5.1f %6.1f | %0.3f %0.3f | %0.3f %0.4f | %s %s | %s' % (\\\n \"val\", i/len(val_loader) + epoch, epoch, \n train_loss[0], train_loss[1], \n losses.avg, f1.avg,\n str(best_results[0])[:8],str(best_results[1])[:8],\n time_to_str((timer() - start),'min'))\n\n print(message, end='',flush=True)\n log.write(\"\\n\")\n #log.write(message)\n #log.write(\"\\n\")\n \n return [losses.avg,f1.avg]\n```", "439884": "Thanks for sharing a really good BSL.", "440196": "I am wondering why SGD optimizer is used in your code. As far as I know, Adam performs better as it combines RMSprop and Momentum.\nI tried Adam but turned to get a terrible 0.33LB.\nCan anyone tell me the reason.", "442363": "Hi EeAary,\nDid you also modify the learning rate that is set to 0.03. Such a learning rate is usable for the SGD optimizer...but for Adam you'd better start with the default of 0.001 or other minor variation on it.\nSee if that works for you.", "442447": "Thanks for ur sharing", "442521": "Definely I have modified the learning rate to 0.001 using Adam, and I set a StepLR with gamma 0.1 every 5 epoch. Anyway I failed to reach the loss as low as SGD did.\nI've obtain a better score but still interested in this problem. Maybe it just likes playing Hearthstone, I don't know what cards I'm going to get.", "442685": "Give it a try with a stepsize of 8 or 10 epochs. Very likely it is to early to lower the learningrate. Alternativly ... Increase gamma to say something like 0.5. That should likely increase your results. Good luck.", "443660": "Thanks so much, it works!", "447148": "Hi @Spytensor, i have some difficulty understanding your code. If you can help me, it would be great. Thanks in advance.\nIn your HumanDataset class:\n\n y = str(self.images_df.iloc[index].Id.absolute()) # what is this absolute() function?", "447422": "Spytensor could you please explain how can i feed 512X512 images to resnext architecture. Please.", "447550": "Change the model's avgpool to adaptiveavgpool like this:\n```python\nmodel = resnext 50()\nmodel.avg_pool = nn.AdaptiveAvgPool2d(1)\n```", "447892": "Thanks", "447966": "The elements in Id column are objects from pathlib, the method gets you the absolute path to the image file.", "448781": "pascal1129\nThanks for sharing,does the train code need to be changed?I changed it in your way, but I got it wrong.\nIf you don’t change how to judge the gap between train and val.", "448921": "Indeed, I changed lots, but the details are difficult to explain here.", "448944": "ok,thanks", "1496033": "Thanks for sharing. This has really made me learn a lot. I also found that change to (better) model or let the image size larger could get better results." }, "source": "meta" }