![\"torchxrayvision\"](\"https://raw.githubusercontent.com/mlmed/torchxrayvision/master/docs/torchxrayvision-logo.png\"){kind=logo}
Hi everyone, I will update this dataset frequently.
\nhttps://www.kaggle.com/ammarali32/startingpointschestx
\nHope it will be useful )) good luck.
\nUPD_1: DenseNet121 added.
\nUPD_2: EfficientNet_B5 added.
\nUPD_3: SEResNet152D added
Hi everyone, I will update this dataset frequently.
\nhttps://www.kaggle.com/ammarali32/startingpointschestx
\nHope it will be useful )) good luck.
\nUPD_1: DenseNet121 added.
\nUPD_2: EfficientNet_B5 added.
\nUPD_3: SEResNet152D added
Thanks for sharing. Also, there is a pytorch module that contains on DenseNet models pretrained on different medical imaging dataset. torchxrayvision link
\n
Hope it helps you all.
", "rawMarkdown": "Thanks for sharing. Also, there is a pytorch module that contains on DenseNet models pretrained on different medical imaging dataset. [torchxrayvision link](https://github.com/mlmed/torchxrayvision)\n\n\n\nHope it helps you all.", "votes": 7 }, { "id": 1217620, "postDate": "2021-02-25T07:45:46.270Z", "content": "If someone had problems with uploading of the checkpoint. Here is a code example
\nstart_point = torch.load('/data/additional_data/startingpoints/tf_efficientnet_b5_ns_chestx.pth', map_location='cpu')\n\nnew_start_point = OrderedDict()\n\nfor k, v in start_point['model'].items():\n if k.startswith('classifier'):\n new_start_point[k] = v\n else:\n # Ignore `.module` at the begining \n new_start_point[k[6:]] = v\n\nmodel = timm.create_model('tf_efficientnet_b5_ns', pretrained=False, num_classes=11)\nmodel.load_state_dict(new_start_point)\nThanks for the dataset! By ChestX do you mean CheXpert?
", "rawMarkdown": "Thanks for the dataset! By ChestX do you mean CheXpert?", "votes": 3, "replies": [ { "id": 1190772, "postDate": "2021-02-08T02:07:26.317Z", "content": "I guess it can also mean ChestX-ray8 or ChestX-ray14, the external dataset:
\nhttps://arxiv.org/abs/1705.02315
Look forward to any clarification.
", "rawMarkdown": "I guess it can also mean ChestX-ray8 or ChestX-ray14, the external dataset:\nhttps://arxiv.org/abs/1705.02315\n\nLook forward to any clarification. ", "votes": 3 }, { "id": 1190781, "postDate": "2021-02-08T02:13:30.593Z", "content": "Hello, welcome you are right it is my bad. This is the link to the data https://www.kaggle.com/nih-chest-xrays/data
", "rawMarkdown": "Hello, welcome you are right it is my bad. This is the link to the data https://www.kaggle.com/nih-chest-xrays/data", "votes": 4 }, { "id": 1190791, "postDate": "2021-02-08T02:32:15.273Z", "content": "Actually I should be the one who should thank u, @ammarali32, for your sharing !
", "rawMarkdown": "Actually I should be the one who should thank u, @ammarali32, for your sharing !", "votes": 2 }, { "id": 1190941, "postDate": "2021-02-08T06:08:50.320Z", "content": "Hello,have your weights trained from the externel dataset?
", "rawMarkdown": "Hello,have your weights trained from the externel dataset?", "votes": 1 }, { "id": 1190946, "postDate": "2021-02-08T06:26:08.037Z", "content": "Hi, Yes the weights published on the dataset are only trained on this dataset https://www.kaggle.com/nih-chest-xrays/data.
", "rawMarkdown": "Hi, Yes the weights published on the dataset are only trained on this dataset https://www.kaggle.com/nih-chest-xrays/data.", "votes": 1 }, { "id": 1224315, "postDate": "2021-03-02T16:08:57.603Z", "content": "have you drop the duplication image between chest-x dataset and ranczer compete when you train startpoint model with chestx dataset @ammarali32
", "rawMarkdown": "have you drop the duplication image between chest-x dataset and ranczer compete when you train startpoint model with chestx dataset @ammarali32 " } ] }, { "id": 1216922, "postDate": "2021-02-24T16:03:16.407Z", "content": "What image resolution do you use ?
", "rawMarkdown": "What image resolution do you use ?", "votes": 1, "replies": [ { "id": 1217055, "postDate": "2021-02-24T18:18:02.337Z", "content": "Hi, 640x640 used for training all models
", "rawMarkdown": "Hi, 640x640 used for training all models", "votes": 1 } ] }, { "id": 1190082, "postDate": "2021-02-07T13:15:56.077Z", "content": "Hi,did you train that in just one notebook?
", "rawMarkdown": "Hi,did you train that in just one notebook?", "votes": 1, "replies": [ { "id": 1190090, "postDate": "2021-02-07T13:29:09.037Z", "content": "Sorry, I didn't get the question. If you meant the starting point then yes.
", "rawMarkdown": "Sorry, I didn't get the question. If you meant the starting point then yes.", "votes": 2 }, { "id": 1190095, "postDate": "2021-02-07T13:35:21.383Z", "content": "Aha,I thought it is just a pretrained model,like imagenet, is it right ?
", "rawMarkdown": "Aha,I thought it is just a pretrained model,like imagenet, is it right ?", "votes": 1 }, { "id": 1190098, "postDate": "2021-02-07T13:38:04.343Z", "content": "Well, yes kinda
", "rawMarkdown": "Well, yes kinda", "votes": 2 }, { "id": 1190126, "postDate": "2021-02-07T14:02:29.557Z", "content": "Thank you！
", "rawMarkdown": "Thank you！", "votes": 1 }, { "id": 1221888, "postDate": "2021-03-01T11:29:07.613Z", "content": "Hi, did you pretrain your model fold by fold? I mean I can see the startpoint from ammara, but I don't know whether it's trained on a single fold or five-folds, If I use this starting point to train my model in 5 folds. I guess every fold might train the data in other folds(test folds)?
", "rawMarkdown": "Hi, did you pretrain your model fold by fold? I mean I can see the startpoint from ammara, but I don't know whether it's trained on a single fold or five-folds, If I use this starting point to train my model in 5 folds. I guess every fold might train the data in other folds(test folds)?" } ] }, { "id": 1215871, "postDate": "2021-02-24T02:32:22.240Z", "content": "Thanks for sharing. How can I load the weights? As you said, I defined the classifier with nn.Linear(n_features, 14). But, it failed to load.
", "rawMarkdown": "Thanks for sharing. How can I load the weights? As you said, I defined the classifier with nn.Linear(n_features, 14). But, it failed to load. ", "votes": 2, "replies": [ { "id": 1217058, "postDate": "2021-02-24T18:18:55.153Z", "content": "Hi please see the inference of my ResNet200D model it is exactly the same ))
", "rawMarkdown": "Hi please see the inference of my ResNet200D model it is exactly the same ))", "votes": 1 }, { "id": 1217126, "postDate": "2021-02-24T19:17:28.253Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 1204270, "postDate": "2021-02-16T02:04:13.070Z", "content": "Hello thanks for sharing.
\nHow did you get these starting point ? Have you trained them on the 14 labels of the chest x ray, and use the checkpoint as a starting point, or did you use something else ?
Hi, No labeling was used. My best model was a teacher model that transfers the ability for extracting features to the student using MSE loss between the features.
", "rawMarkdown": "Hi, No labeling was used. My best model was a teacher model that transfers the ability for extracting features to the student using MSE loss between the features.", "votes": 2 }, { "id": 1204547, "postDate": "2021-02-16T07:52:12.163Z", "content": "Very interesting @ammarali32 , can u please give an outline ,the steps for this process .
", "rawMarkdown": "Very interesting @ammarali32 , can u please give an outline ,the steps for this process .", "votes": 2 }, { "id": 1205700, "postDate": "2021-02-16T22:21:58.920Z", "content": "@ammarali32 Thanks for your answer, but how did you create the teacher model ?
", "rawMarkdown": "@ammarali32 Thanks for your answer, but how did you create the teacher model ?", "votes": 1 }, { "id": 1205957, "postDate": "2021-02-17T04:17:53.193Z", "content": "Well, For ResNet200D I used this model https://www.kaggle.com/ammarali32/resnet200d-inference-single-model-lb-96-5 which trained using this approach https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/215910
\nFor other models also the same but they are not public ))
@kudzayiking Well, I guess it is clear if you read my discussions. My notebook has a bad code style if I had time I will go back to this competition and publish these notebooks after reorganizing. But for now, I have a master thesis. That should be written ))
", "rawMarkdown": "@kudzayiking Well, I guess it is clear if you read my discussions. My notebook has a bad code style if I had time I will go back to this competition and publish these notebooks after reorganizing. But for now, I have a master thesis. That should be written ))", "votes": 1 }, { "id": 1206003, "postDate": "2021-02-17T05:29:38.620Z", "content": "@ammarali32 thanks for the links, I'll check them out, good luck and all the best in your master thesis!
", "rawMarkdown": "@ammarali32 thanks for the links, I'll check them out, good luck and all the best in your master thesis!" }, { "id": 1211893, "postDate": "2021-02-20T16:31:08.110Z", "content": "@ammarali32 can you introduce what \"no labeling\" means in detail ? how do you train the ChestX dataset to get your pretrained weight? Thanks a lot !!
", "rawMarkdown": "@ammarali32 can you introduce what \"no labeling\" means in detail ? how do you train the ChestX dataset to get your pretrained weight? Thanks a lot !!", "votes": 1 } ] }, { "id": 1197436, "postDate": "2021-02-12T06:50:18.517Z", "content": "I tried pretraining on the https://www.kaggle.com/nih-chest-xrays/data dataset as well. Curious how you went about implementing it. I directly had the model train on the original targets from the dataset and then tried using that starting point to train for the ranzcr dataset. Performance did not really seem to improve though.
", "rawMarkdown": "I tried pretraining on the https://www.kaggle.com/nih-chest-xrays/data dataset as well. Curious how you went about implementing it. I directly had the model train on the original targets from the dataset and then tried using that starting point to train for the ranzcr dataset. Performance did not really seem to improve though. ", "votes": 2, "replies": [ { "id": 1197494, "postDate": "2021-02-12T07:23:38.507Z", "content": "Hi, I took my best model and train with no labels. The loss was the MSELoss between the features of the training model and my best model. The improvement on the CV from 0.962 to 0.967. on the LB was not so much for ResNet200D about 0.002 but for others was clear. Single model inception got 0.959 LB.
", "rawMarkdown": "Hi, I took my best model and train with no labels. The loss was the MSELoss between the features of the training model and my best model. The improvement on the CV from 0.962 to 0.967. on the LB was not so much for ResNet200D about 0.002 but for others was clear. Single model inception got 0.959 LB.", "votes": 2 }, { "id": 1197495, "postDate": "2021-02-12T07:25:42.483Z", "content": "Ah interesting. Lots of people have been using that idea of something sort of like a mean teacher or contrastive loss recently. WIll have to try that out for myself
", "rawMarkdown": "Ah interesting. Lots of people have been using that idea of something sort of like a mean teacher or contrastive loss recently. WIll have to try that out for myself", "votes": 2 }, { "id": 1197500, "postDate": "2021-02-12T07:28:18.703Z", "content": "Thanks )) and good luck with that.
", "rawMarkdown": "Thanks )) and good luck with that." } ] }, { "id": 1190174, "postDate": "2021-02-07T14:39:46.087Z", "content": "Thank you for sharing your nice dataset and notebooks! May I ask you a question? To generate your ChestX starting points, did you use the ImageNet weights? Or did you use randomly initialized ones?
", "rawMarkdown": "Thank you for sharing your nice dataset and notebooks! May I ask you a question? To generate your ChestX starting points, did you use the ImageNet weights? Or did you use randomly initialized ones?", "votes": 2, "replies": [ { "id": 1190219, "postDate": "2021-02-07T15:06:35.523Z", "content": "Welcome,)). Yes, I initialized with ImageNet weights.
", "rawMarkdown": "Welcome,)). Yes, I initialized with ImageNet weights.", "votes": 3 }, { "id": 1190223, "postDate": "2021-02-07T15:09:59.897Z", "content": "Thank you for your reply!
", "rawMarkdown": "Thank you for your reply!" } ] }, { "id": 1189990, "postDate": "2021-02-07T11:50:24.057Z", "content": "Thank you for sharing, although I don't know how to use pytorch. It's a pity that I can't ensemble your 0.967 model. I ensemble your 0.965 model can get 0.968 score, but I identify with you it is good you will not share the trained models. Good luck!!
", "rawMarkdown": "Thank you for sharing, although I don't know how to use pytorch. It's a pity that I can't ensemble your 0.967 model. I ensemble your 0.965 model can get 0.968 score, but I identify with you it is good you will not share the trained models. Good luck!!", "votes": 2, "replies": [ { "id": 1189993, "postDate": "2021-02-07T11:57:52.200Z", "content": "Hi thanks for your feedback. if had enough time i will generate starting points for tensorflow.
", "rawMarkdown": "Hi thanks for your feedback. if had enough time i will generate starting points for tensorflow.", "votes": 2 }, { "id": 1190054, "postDate": "2021-02-07T13:00:25.933Z", "content": "Hi,How did u blend it ?
\nDifferent weights?
Yes, just different weights. My single model does not perform well, but combining with other models can improve performance.
", "rawMarkdown": "Yes, just different weights. My single model does not perform well, but combining with other models can improve performance. " } ] }, { "id": 1241305, "postDate": "2021-03-17T01:51:16.610Z", "content": "@ammarali32 Thank you very much for the weights because we could progress so high without your initial weight . i was extremely useful to make our first confidence for achieving this rank
\nAs @projdev asked we would like to experiment and see how you actually made these weights
@ammarali32 Thank you very much for your datasets. It is extremely helpful in my model training and achieving this rank. Is it possible for you to publish a script regarding how you perform pretraining? I want to try it using a customized model
", "rawMarkdown": "@ammarali32 Thank you very much for your datasets. It is extremely helpful in my model training and achieving this rank. Is it possible for you to publish a script regarding how you perform pretraining? I want to try it using a customized model" }, { "id": 1221822, "postDate": "2021-03-01T10:24:05.280Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1215661, "author_name": "Dr. Amritpal Singh", "author_url": "", "post_date": "2021-02-23T20:54:44.377000", "content": "Thanks for sharing. Also, there is a pytorch module that contains on DenseNet models pretrained on different medical imaging dataset. torchxrayvision link
\n
Hope it helps you all.
", "votes": 7, "replies": [] }, { "id": 1217620, "author_name": "Volodymyr", "author_url": "", "post_date": "2021-02-25T07:45:46.270000", "content": "If someone had problems with uploading of the checkpoint. Here is a code example
\nstart_point = torch.load('/data/additional_data/startingpoints/tf_efficientnet_b5_ns_chestx.pth', map_location='cpu')\n\nnew_start_point = OrderedDict()\n\nfor k, v in start_point['model'].items():\n if k.startswith('classifier'):\n new_start_point[k] = v\n else:\n # Ignore `.module` at the begining \n new_start_point[k[6:]] = v\n\nmodel = timm.create_model('tf_efficientnet_b5_ns', pretrained=False, num_classes=11)\nmodel.load_state_dict(new_start_point)\nThanks for the dataset! By ChestX do you mean CheXpert?
", "votes": 3, "replies": [ { "id": 1190772, "author_name": "FP", "author_url": "", "post_date": "2021-02-08T02:07:26.317000", "content": "I guess it can also mean ChestX-ray8 or ChestX-ray14, the external dataset:
\nhttps://arxiv.org/abs/1705.02315
Look forward to any clarification.
", "votes": 3, "replies": [] }, { "id": 1190781, "author_name": "ammarali32", "author_url": "", "post_date": "2021-02-08T02:13:30.593000", "content": "Hello, welcome you are right it is my bad. This is the link to the data https://www.kaggle.com/nih-chest-xrays/data
", "votes": 4, "replies": [] }, { "id": 1190791, "author_name": "FP", "author_url": "", "post_date": "2021-02-08T02:32:15.273000", "content": "Actually I should be the one who should thank u, @ammarali32, for your sharing !
", "votes": 2, "replies": [] }, { "id": 1190941, "author_name": "Zekun", "author_url": "", "post_date": "2021-02-08T06:08:50.320000", "content": "Hello,have your weights trained from the externel dataset?
", "votes": 1, "replies": [] }, { "id": 1190946, "author_name": "ammarali32", "author_url": "", "post_date": "2021-02-08T06:26:08.037000", "content": "Hi, Yes the weights published on the dataset are only trained on this dataset https://www.kaggle.com/nih-chest-xrays/data.
", "votes": 1, "replies": [] }, { "id": 1224315, "author_name": "cswwp", "author_url": "", "post_date": "2021-03-02T16:08:57.603000", "content": "have you drop the duplication image between chest-x dataset and ranczer compete when you train startpoint model with chestx dataset @ammarali32
", "votes": 0, "replies": [] } ] }, { "id": 1216922, "author_name": "Volodymyr", "author_url": "", "post_date": "2021-02-24T16:03:16.407000", "content": "What image resolution do you use ?
", "votes": 1, "replies": [ { "id": 1217055, "author_name": "ammarali32", "author_url": "", "post_date": "2021-02-24T18:18:02.337000", "content": "Hi, 640x640 used for training all models
", "votes": 1, "replies": [] } ] }, { "id": 1190082, "author_name": "Zekun", "author_url": "", "post_date": "2021-02-07T13:15:56.077000", "content": "Hi,did you train that in just one notebook?
", "votes": 1, "replies": [ { "id": 1190090, "author_name": "ammarali32", "author_url": "", "post_date": "2021-02-07T13:29:09.037000", "content": "Sorry, I didn't get the question. If you meant the starting point then yes.
", "votes": 2, "replies": [] }, { "id": 1190095, "author_name": "Zekun", "author_url": "", "post_date": "2021-02-07T13:35:21.383000", "content": "Aha,I thought it is just a pretrained model,like imagenet, is it right ?
", "votes": 1, "replies": [] }, { "id": 1190098, "author_name": "ammarali32", "author_url": "", "post_date": "2021-02-07T13:38:04.343000", "content": "Well, yes kinda
", "votes": 2, "replies": [] }, { "id": 1190126, "author_name": "Zekun", "author_url": "", "post_date": "2021-02-07T14:02:29.557000", "content": "Thank you！
", "votes": 1, "replies": [] }, { "id": 1221888, "author_name": "Looking for Luck", "author_url": "", "post_date": "2021-03-01T11:29:07.613000", "content": "Hi, did you pretrain your model fold by fold? I mean I can see the startpoint from ammara, but I don't know whether it's trained on a single fold or five-folds, If I use this starting point to train my model in 5 folds. I guess every fold might train the data in other folds(test folds)?
", "votes": 0, "replies": [] } ] }, { "id": 1215871, "author_name": "LeoF", "author_url": "", "post_date": "2021-02-24T02:32:22.240000", "content": "Thanks for sharing. How can I load the weights? As you said, I defined the classifier with nn.Linear(n_features, 14). But, it failed to load.
", "votes": 2, "replies": [ { "id": 1217058, "author_name": "ammarali32", "author_url": "", "post_date": "2021-02-24T18:18:55.153000", "content": "Hi please see the inference of my ResNet200D model it is exactly the same ))
", "votes": 1, "replies": [] }, { "id": 1217126, "author_name": "", "author_url": "", "post_date": "2021-02-24T19:17:28.253000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 1204270, "author_name": "Shiro", "author_url": "", "post_date": "2021-02-16T02:04:13.070000", "content": "Hello thanks for sharing.
\nHow did you get these starting point ? Have you trained them on the 14 labels of the chest x ray, and use the checkpoint as a starting point, or did you use something else ?
Hi, No labeling was used. My best model was a teacher model that transfers the ability for extracting features to the student using MSE loss between the features.
", "votes": 2, "replies": [] }, { "id": 1204547, "author_name": "Kudzayi Matinyarare", "author_url": "", "post_date": "2021-02-16T07:52:12.163000", "content": "Very interesting @ammarali32 , can u please give an outline ,the steps for this process .
", "votes": 2, "replies": [] }, { "id": 1205700, "author_name": "Shiro", "author_url": "", "post_date": "2021-02-16T22:21:58.920000", "content": "@ammarali32 Thanks for your answer, but how did you create the teacher model ?
", "votes": 1, "replies": [] }, { "id": 1205957, "author_name": "ammarali32", "author_url": "", "post_date": "2021-02-17T04:17:53.193000", "content": "Well, For ResNet200D I used this model https://www.kaggle.com/ammarali32/resnet200d-inference-single-model-lb-96-5 which trained using this approach https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/215910
\nFor other models also the same but they are not public ))
@kudzayiking Well, I guess it is clear if you read my discussions. My notebook has a bad code style if I had time I will go back to this competition and publish these notebooks after reorganizing. But for now, I have a master thesis. That should be written ))
", "votes": 1, "replies": [] }, { "id": 1206003, "author_name": "Kudzayi Matinyarare", "author_url": "", "post_date": "2021-02-17T05:29:38.620000", "content": "@ammarali32 thanks for the links, I'll check them out, good luck and all the best in your master thesis!
", "votes": 0, "replies": [] }, { "id": 1211893, "author_name": "Yineng Xiong", "author_url": "", "post_date": "2021-02-20T16:31:08.110000", "content": "@ammarali32 can you introduce what \"no labeling\" means in detail ? how do you train the ChestX dataset to get your pretrained weight? Thanks a lot !!
", "votes": 1, "replies": [] } ] }, { "id": 1197436, "author_name": "ryches", "author_url": "", "post_date": "2021-02-12T06:50:18.517000", "content": "I tried pretraining on the https://www.kaggle.com/nih-chest-xrays/data dataset as well. Curious how you went about implementing it. I directly had the model train on the original targets from the dataset and then tried using that starting point to train for the ranzcr dataset. Performance did not really seem to improve though.
", "votes": 2, "replies": [ { "id": 1197494, "author_name": "ammarali32", "author_url": "", "post_date": "2021-02-12T07:23:38.507000", "content": "Hi, I took my best model and train with no labels. The loss was the MSELoss between the features of the training model and my best model. The improvement on the CV from 0.962 to 0.967. on the LB was not so much for ResNet200D about 0.002 but for others was clear. Single model inception got 0.959 LB.
", "votes": 2, "replies": [] }, { "id": 1197495, "author_name": "ryches", "author_url": "", "post_date": "2021-02-12T07:25:42.483000", "content": "Ah interesting. Lots of people have been using that idea of something sort of like a mean teacher or contrastive loss recently. WIll have to try that out for myself
", "votes": 2, "replies": [] }, { "id": 1197500, "author_name": "ammarali32", "author_url": "", "post_date": "2021-02-12T07:28:18.703000", "content": "Thanks )) and good luck with that.
", "votes": 0, "replies": [] } ] }, { "id": 1190174, "author_name": "rikein12", "author_url": "", "post_date": "2021-02-07T14:39:46.087000", "content": "Thank you for sharing your nice dataset and notebooks! May I ask you a question? To generate your ChestX starting points, did you use the ImageNet weights? Or did you use randomly initialized ones?
", "votes": 2, "replies": [ { "id": 1190219, "author_name": "ammarali32", "author_url": "", "post_date": "2021-02-07T15:06:35.523000", "content": "Welcome,)). Yes, I initialized with ImageNet weights.
", "votes": 3, "replies": [] }, { "id": 1190223, "author_name": "rikein12", "author_url": "", "post_date": "2021-02-07T15:09:59.897000", "content": "Thank you for your reply!
", "votes": 0, "replies": [] } ] }, { "id": 1189990, "author_name": "Alien", "author_url": "", "post_date": "2021-02-07T11:50:24.057000", "content": "Thank you for sharing, although I don't know how to use pytorch. It's a pity that I can't ensemble your 0.967 model. I ensemble your 0.965 model can get 0.968 score, but I identify with you it is good you will not share the trained models. Good luck!!
", "votes": 2, "replies": [ { "id": 1189993, "author_name": "ammarali32", "author_url": "", "post_date": "2021-02-07T11:57:52.200000", "content": "Hi thanks for your feedback. if had enough time i will generate starting points for tensorflow.
", "votes": 2, "replies": [] }, { "id": 1190054, "author_name": "HEEEEE", "author_url": "", "post_date": "2021-02-07T13:00:25.933000", "content": "Hi,How did u blend it ?
\nDifferent weights?
Yes, just different weights. My single model does not perform well, but combining with other models can improve performance.
", "votes": 0, "replies": [] } ] }, { "id": 1241305, "author_name": "Mohammed Rizin V K", "author_url": "", "post_date": "2021-03-17T01:51:16.610000", "content": "@ammarali32 Thank you very much for the weights because we could progress so high without your initial weight . i was extremely useful to make our first confidence for achieving this rank
\nAs @projdev asked we would like to experiment and see how you actually made these weights
@ammarali32 Thank you very much for your datasets. It is extremely helpful in my model training and achieving this rank. Is it possible for you to publish a script regarding how you perform pretraining? I want to try it using a customized model
", "votes": 0, "replies": [] }, { "id": 1221822, "author_name": "", "author_url": "", "post_date": "2021-03-01T10:24:05.280000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1189876": "Hi everyone, I will update this dataset frequently.\nhttps://www.kaggle.com/ammarali32/startingpointschestx\nHope it will be useful )) good luck.\nUPD_1: DenseNet121 added.\nUPD_2: EfficientNet_B5 added.\nUPD_3: SEResNet152D added", "1215661": "Thanks for sharing. Also, there is a pytorch module that contains on DenseNet models pretrained on different medical imaging dataset. [torchxrayvision link](https://github.com/mlmed/torchxrayvision)\n\n\n\nHope it helps you all.", "1217620": "If someone had problems with uploading of the checkpoint. Here is a code example\n```\nstart_point = torch.load('/data/additional_data/startingpoints/tf_efficientnet_b5_ns_chestx.pth', map_location='cpu')\n\nnew_start_point = OrderedDict()\n\nfor k, v in start_point['model'].items():\n if k.startswith('classifier'):\n new_start_point[k] = v\n else:\n # Ignore `.module` at the begining \n new_start_point[k[6:]] = v\n \nmodel = timm.create_model('tf_efficientnet_b5_ns', pretrained=False, num_classes=11)\nmodel.load_state_dict(new_start_point)\n```", "1190706": "Thanks for the dataset! By ChestX do you mean CheXpert?", "1216922": "What image resolution do you use ?", "1190082": "Hi,did you train that in just one notebook?", "1215871": "Thanks for sharing. How can I load the weights? As you said, I defined the classifier with nn.Linear(n_features, 14). But, it failed to load. ", "1204270": "Hello thanks for sharing.\nHow did you get these starting point ? Have you trained them on the 14 labels of the chest x ray, and use the checkpoint as a starting point, or did you use something else ?", "1197436": "I tried pretraining on the https://www.kaggle.com/nih-chest-xrays/data dataset as well. Curious how you went about implementing it. I directly had the model train on the original targets from the dataset and then tried using that starting point to train for the ranzcr dataset. Performance did not really seem to improve though. ", "1190174": "Thank you for sharing your nice dataset and notebooks! May I ask you a question? To generate your ChestX starting points, did you use the ImageNet weights? Or did you use randomly initialized ones?", "1189990": "Thank you for sharing, although I don't know how to use pytorch. It's a pity that I can't ensemble your 0.967 model. I ensemble your 0.965 model can get 0.968 score, but I identify with you it is good you will not share the trained models. Good luck!!", "1241305": "@ammarali32 Thank you very much for the weights because we could progress so high without your initial weight . i was extremely useful to make our first confidence for achieving this rank\nAs @projdev asked we would like to experiment and see how you actually made these weights", "1241291": "@ammarali32 Thank you very much for your datasets. It is extremely helpful in my model training and achieving this rank. Is it possible for you to publish a script regarding how you perform pretraining? I want to try it using a customized model", "1221822": "" } }