As stated in the Rules, the use of external data is not permitted. You may, though, use pre-trained models, as long as you use this thread to list the pre-trained models you are planning to use. Once a model has been posted once, it does not need to be posted again. In order for you to use a pre-trained model, it must be posted here no later than one week prior to the competition close.
", "messages": [ { "id": 364105, "postDate": "2018-07-30T18:47:14.750Z", "content": "As stated in the Rules, the use of external data is not permitted. You may, though, use pre-trained models, as long as you use this thread to list the pre-trained models you are planning to use. Once a model has been posted once, it does not need to be posted again. In order for you to use a pre-trained model, it must be posted here no later than one week prior to the competition close.
", "rawMarkdown": "As stated in the Rules, the use of external data is not permitted. You may, though, use pre-trained models, as long as you use this thread to list the pre-trained models you are planning to use. Once a model has been posted once, it does not need to be posted again. In order for you to use a pre-trained model, it must be posted here no later than one week prior to the competition close.", "votes": 7 }, { "id": 412518, "postDate": "2018-10-30T10:32:10.933Z", "content": "Locally modified version of Mask-Rcnn : https://github.com/matterport/Mask_RCNN with coco weight: https://github.com/matterport/Mask_RCNN/releases/download/v2.0/mask_rcnn_coco.h5 and also this https://github.com/fizyr/keras-maskrcnn/releases/download/0.2.0/resnet50_coco_v0.2.0.h5
", "rawMarkdown": "Locally modified version of Mask-Rcnn : https://github.com/matterport/Mask_RCNN with coco weight: https://github.com/matterport/Mask_RCNN/releases/download/v2.0/mask_rcnn_coco.h5 and also this https://github.com/fizyr/keras-maskrcnn/releases/download/0.2.0/resnet50_coco_v0.2.0.h5 ", "votes": 1 }, { "id": 372338, "postDate": "2018-08-19T02:08:03.383Z", "content": "Hi @inversion,
\n\nIs it allowed to manually annotate training data to assist model training?
", "rawMarkdown": "Hi @inversion,\n\nIs it allowed to manually annotate training data to assist model training?", "votes": 1, "replies": [ { "id": 372646, "postDate": "2018-08-20T01:31:59.107Z", "content": "In general, as is the case in this contest, you can do what you wish with the training images. You can't, though, manually label the test images.
", "rawMarkdown": "In general, as is the case in this contest, you can do what you wish with the training images. You can't, though, manually label the test images." }, { "id": 372664, "postDate": "2018-08-20T03:07:49.357Z", "content": "Thank you :)
", "rawMarkdown": "Thank you :) " } ] }, { "id": 414467, "postDate": "2018-11-02T20:24:39.657Z", "content": "Pretrained models from https://github.com/mapillary/inplace_abn repository
", "rawMarkdown": "Pretrained models from https://github.com/mapillary/inplace_abn repository " }, { "id": 372999, "postDate": "2018-08-20T21:29:31.160Z", "content": "@inversion I was just wondering: would not disclosure of the pre-trained models used by different users give additional information about the goodness/relevance of this model? For instance, if a kaggle master with position top 5 on LB publishes a pre-trained model more users will consider it as a good start than if a novice will publish it. It gives an additional info, puts kaggle leaders under more attention in this regard, and probably is not so good for a fair competition. It is NOT allowed, however, to have two kaggle accounts. Would it be a better idea to create a new kaggle account named airbus_models with password pretrainedmodels for everyone to use the same account to publish pre-trained models they use? Then everyone has an opportunity to publish models anonymously, and we remove the bias created by people profiles and position on LB. I can create such account myself, but do not want to get in trouble for making two kaggle accounts. Since you work at kaggle you may create/comment on this solution.
", "rawMarkdown": "@inversion I was just wondering: would not disclosure of the pre-trained models used by different users give additional information about the goodness/relevance of this model? For instance, if a kaggle master with position top 5 on LB publishes a pre-trained model more users will consider it as a good start than if a novice will publish it. It gives an additional info, puts kaggle leaders under more attention in this regard, and probably is not so good for a fair competition. It is NOT allowed, however, to have two kaggle accounts. Would it be a better idea to create a new kaggle account named airbus_models with password pretrainedmodels for everyone to use the same account to publish pre-trained models they use? Then everyone has an opportunity to publish models anonymously, and we remove the bias created by people profiles and position on LB. I can create such account myself, but do not want to get in trouble for making two kaggle accounts. Since you work at kaggle you may create/comment on this solution. ", "votes": -2 }, { "id": 420868, "postDate": "2018-11-14T09:06:13.737Z", "content": "I use pretrained pytorch models:\n'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',\n'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
", "rawMarkdown": "I use pretrained pytorch models:\n'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',\n'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth'," }, { "id": 420205, "postDate": "2018-11-13T09:05:36.723Z", "content": "Locally modified version of Deeplabv3+ : https://github.com/tensorflow/models/tree/master/research/deeplab with weight: http://download.tensorflow.org/models/xception_41_2018_05_09.tar.gz
", "rawMarkdown": "Locally modified version of Deeplabv3+ : https://github.com/tensorflow/models/tree/master/research/deeplab with weight: http://download.tensorflow.org/models/xception_41_2018_05_09.tar.gz" }, { "id": 420067, "postDate": "2018-11-13T02:22:09.313Z", "content": "I used Keras ResNet52 pretrained models with Imagenet weight
", "rawMarkdown": "I used Keras ResNet52 pretrained models with Imagenet weight " }, { "id": 417093, "postDate": "2018-11-07T18:24:15.397Z", "content": "i use pretrained model pytorch resnet34 or18
", "rawMarkdown": "i use pretrained model pytorch resnet34 or18" }, { "id": 417012, "postDate": "2018-11-07T16:02:51.920Z", "content": "I am using pretrained pytorch models:\nresnet34: \nhttps://download.pytorch.org/models/resnet34-333f7ec4.pth\nalso I'll include resnet50 if I try it in the future:\nhttps://download.pytorch.org/models/resnet50-19c8e357.pth
", "rawMarkdown": "I am using pretrained pytorch models:\nresnet34: \nhttps://download.pytorch.org/models/resnet34-333f7ec4.pth\nalso I'll include resnet50 if I try it in the future:\nhttps://download.pytorch.org/models/resnet50-19c8e357.pth" }, { "id": 416926, "postDate": "2018-11-07T12:57:29.317Z", "content": "fastai, pretrained-models & ternaus
", "rawMarkdown": "fastai, pretrained-models & ternaus" }, { "id": 416373, "postDate": "2018-11-06T15:46:48.040Z", "content": "Model from public kernel https://www.kaggle.com/iafoss/fine-tuning-resnet34-on-ship-detection/output
", "rawMarkdown": "Model from public kernel https://www.kaggle.com/iafoss/fine-tuning-resnet34-on-ship-detection/output" }, { "id": 416368, "postDate": "2018-11-06T15:43:36.600Z", "content": "our team\nPretrained pytorch models:\n'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',\n'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',\n‘seresnext50’：'http://data.lip6.fr/cadene/pretrainedmodels/se_resnext50_32x4d-a260b3a4.pth'
", "rawMarkdown": "our team\nPretrained pytorch models:\n'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',\n'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',\n‘seresnext50’：'http://data.lip6.fr/cadene/pretrainedmodels/se_resnext50_32x4d-a260b3a4.pth'" }, { "id": 416212, "postDate": "2018-11-06T12:06:04.387Z", "content": "Pretrained models from https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/model_zoo.md\nhttp://download.tensorflow.org/models/resnet_v1_50_2018_05_04.tar.gz
", "rawMarkdown": "Pretrained models from https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/model_zoo.md\nhttp://download.tensorflow.org/models/resnet_v1_50_2018_05_04.tar.gz" }, { "id": 416158, "postDate": "2018-11-06T09:47:20.843Z", "content": "I used Pretrained models from fast.ai\nhttp://files.fast.ai/models/
", "rawMarkdown": "I used Pretrained models from fast.ai\nhttp://files.fast.ai/models/" }, { "id": 416032, "postDate": "2018-11-06T04:06:56.467Z", "content": "Pytorch models with weights trained on imagenet: resnet18, resnet34, densenet121
", "rawMarkdown": "Pytorch models with weights trained on imagenet: resnet18, resnet34, densenet121" }, { "id": 415713, "postDate": "2018-11-05T14:40:58.067Z", "content": "Keras/keras-contribution, pytorch/cadene resnet pretrained models as encoders
", "rawMarkdown": "Keras/keras-contribution, pytorch/cadene resnet pretrained models as encoders" }, { "id": 415407, "postDate": "2018-11-05T03:54:55.290Z", "content": "Perhaps one or two of the followings: \nPretrained models from 1) Torchvision, 2) https://github.com/Cadene/pretrained-models.pytorch, 3) https://modelzoo.co/model/pytorch-cnn-finetune, 4) Pretrained-models in keras https://github.com/fchollet/deep-learning-models/releases, specifically used for this repo https://github.com/fizyr and 5) Pretrained models in fastai.
", "rawMarkdown": "Perhaps one or two of the followings: \nPretrained models from 1) Torchvision, 2) https://github.com/Cadene/pretrained-models.pytorch, 3) https://modelzoo.co/model/pytorch-cnn-finetune, 4) Pretrained-models in keras https://github.com/fchollet/deep-learning-models/releases, specifically used for this repo https://github.com/fizyr and 5) Pretrained models in fastai.\n" }, { "id": 415354, "postDate": "2018-11-05T01:11:03.033Z", "content": "Pre-trained on the coco dataset https://www.dropbox.com/s/dr4qh2xo1ksaj6r/coco.h5?dl=0\nand this: https://github.com/tensorflow/models/tree/master/research/slim
", "rawMarkdown": "Pre-trained on the coco dataset https://www.dropbox.com/s/dr4qh2xo1ksaj6r/coco.h5?dl=0\nand this: https://github.com/tensorflow/models/tree/master/research/slim" }, { "id": 415067, "postDate": "2018-11-04T09:59:55.933Z", "content": "Pretrained model from https://github.com/junfu1115/DANet repo
", "rawMarkdown": "Pretrained model from https://github.com/junfu1115/DANet repo" }, { "id": 414699, "postDate": "2018-11-03T12:18:43.163Z", "content": "https://github.com/qubvel/segmentation_models https://github.com/broadinstitute/keras-resnet
", "rawMarkdown": "https://github.com/qubvel/segmentation_models https://github.com/broadinstitute/keras-resnet" }, { "id": 413998, "postDate": "2018-11-01T23:33:34.227Z", "content": "yolo v3: https://pjreddie.com/media/files/yolov3.weights
", "rawMarkdown": "yolo v3: https://pjreddie.com/media/files/yolov3.weights" }, { "id": 413764, "postDate": "2018-11-01T13:45:08.057Z", "content": "I use pretrained pytorch models (https://pytorch.org/docs/stable/torchvision/models.html).
", "rawMarkdown": "I use pretrained pytorch models (https://pytorch.org/docs/stable/torchvision/models.html). " }, { "id": 413199, "postDate": "2018-10-31T13:20:41.673Z", "content": "pre-trained pytorch models(https://github.com/pytorch/vision/tree/master/torchvision/models)
", "rawMarkdown": "pre-trained pytorch models(https://github.com/pytorch/vision/tree/master/torchvision/models)" }, { "id": 411400, "postDate": "2018-10-28T03:21:50.897Z", "content": "Pretrained models from torchvision: https://download.pytorch.org/models/densenet121-a639ec97.pth
", "rawMarkdown": "Pretrained models from torchvision: https://download.pytorch.org/models/densenet121-a639ec97.pth" }, { "id": 410198, "postDate": "2018-10-25T15:32:06.513Z", "content": "Pretrained models from: http://models.tensorpack.com/
", "rawMarkdown": "Pretrained models from: http://models.tensorpack.com/" }, { "id": 405867, "postDate": "2018-10-18T08:39:15.193Z", "content": "Squeezenet for Keras https://github.com/rcmalli/keras-squeezenet/
", "rawMarkdown": "Squeezenet for Keras https://github.com/rcmalli/keras-squeezenet/" }, { "id": 404863, "postDate": "2018-10-16T13:25:59.343Z", "content": "I suppose if you take public kernel you don't need to list it's pretrained models?
", "rawMarkdown": "I suppose if you take public kernel you don't need to list it's pretrained models?" }, { "id": 404487, "postDate": "2018-10-15T19:52:53.793Z", "content": "Pretrained pytorch models:\n'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',\n 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
", "rawMarkdown": "Pretrained pytorch models:\n'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',\n 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth'," }, { "id": 401278, "postDate": "2018-10-09T18:27:42.510Z", "content": "resnet trained on ImageNet
", "rawMarkdown": "resnet trained on ImageNet" }, { "id": 398806, "postDate": "2018-10-04T15:56:38.927Z", "content": "Sorry, May I know what is the defination of pre-trained Model? I have been training the full train set of ships by using YOLO tiny model. Is it considered pre-trained model?
", "rawMarkdown": "Sorry, May I know what is the defination of pre-trained Model? I have been training the full train set of ships by using YOLO tiny model. Is it considered pre-trained model?" }, { "id": 391371, "postDate": "2018-09-21T17:50:01.440Z", "content": "Pre-trained on the coco dataset https://www.dropbox.com/s/dr4qh2xo1ksaj6r/coco.h5?dl=0
", "rawMarkdown": "Pre-trained on the coco dataset https://www.dropbox.com/s/dr4qh2xo1ksaj6r/coco.h5?dl=0" }, { "id": 390701, "postDate": "2018-09-20T17:00:45.800Z", "content": "ResNeXt models. I uploaded them to kaggle https://www.kaggle.com/iafoss/pytorch-pretrained-models/home , but originally they are taken from http://files.fast.ai/models/weights.tgz.
", "rawMarkdown": "ResNeXt models. I uploaded them to kaggle https://www.kaggle.com/iafoss/pytorch-pretrained-models/home , but originally they are taken from http://files.fast.ai/models/weights.tgz." }, { "id": 389671, "postDate": "2018-09-19T03:44:57.423Z", "content": "Pre-trained model from DRBox https://github.com/liulei01/DRBox/tree/master/examples/rbox/deploy/ship
", "rawMarkdown": "Pre-trained model from DRBox https://github.com/liulei01/DRBox/tree/master/examples/rbox/deploy/ship" }, { "id": 387031, "postDate": "2018-09-14T06:57:40.887Z", "content": "Keras implementation of Deeplabv3+(https://github.com/bonlime/keras-deeplab-v3-plus)
", "rawMarkdown": "Keras implementation of Deeplabv3+(https://github.com/bonlime/keras-deeplab-v3-plus)" }, { "id": 386980, "postDate": "2018-09-14T04:10:41.040Z", "content": "chainer resnext models \nhttps://chainercv-models.preferred.jp/se_resnext50_imagenet_converted_2018_06_28.npz\nhttps://chainercv-models.preferred.jp/se_resnext101_imagenet_converted_2018_06_28.npz
", "rawMarkdown": "chainer resnext models \nhttps://chainercv-models.preferred.jp/se_resnext50_imagenet_converted_2018_06_28.npz\nhttps://chainercv-models.preferred.jp/se_resnext101_imagenet_converted_2018_06_28.npz" }, { "id": 383579, "postDate": "2018-09-09T02:22:42.970Z", "content": "resnet weight trained on imagenet and coco
", "rawMarkdown": "resnet weight trained on imagenet and coco" }, { "id": 380212, "postDate": "2018-09-02T04:09:36.677Z", "content": "Pretrained-models for keras https://github.com/fchollet/deep-learning-models/releases
", "rawMarkdown": "Pretrained-models for keras https://github.com/fchollet/deep-learning-models/releases" }, { "id": 380060, "postDate": "2018-09-01T15:21:15.653Z", "content": "keras all pre-trained model
\n\nhttps://keras.io/applications/
", "rawMarkdown": "keras all pre-trained model\n\nhttps://keras.io/applications/\n" }, { "id": 379401, "postDate": "2018-08-31T10:08:06.500Z", "content": "Hi, mask r cnn with coco weights
", "rawMarkdown": "Hi, mask r cnn with coco weights" }, { "id": 379212, "postDate": "2018-08-31T02:57:15.117Z", "content": "I use pre-trained pytorch models(https://github.com/pytorch/vision/tree/master/torchvision/models)
", "rawMarkdown": "I use pre-trained pytorch models(https://github.com/pytorch/vision/tree/master/torchvision/models)" }, { "id": 376865, "postDate": "2018-08-28T07:33:27.440Z", "content": "Pretrained models from fast.ai
\n\n", "rawMarkdown": "Pretrained models from fast.ai\n\nhttp://files.fast.ai/models/" }, { "id": 376435, "postDate": "2018-08-27T13:42:37.207Z", "content": "Keras ResNet Imagenet pretrained models - https://github.com/qubvel/classification_models
", "rawMarkdown": "Keras ResNet Imagenet pretrained models - https://github.com/qubvel/classification_models" }, { "id": 373726, "postDate": "2018-08-21T20:06:14.863Z", "content": "https://github.com/Cadene/pretrained-models.pytorch
", "rawMarkdown": "https://github.com/Cadene/pretrained-models.pytorch" }, { "id": 370547, "postDate": "2018-08-15T01:41:37.227Z", "content": "I would also like clarification on the use of pretrained models.
\n\nI am using detectron that has been trained on ipatch data which is publicly available.
\n\nCan I not use this model in conjunction with whatever algorithm I make?
\n\nCould I use the model for transfer learning?
\n\nTechnically whatever model I have that had been trained on external data can also be a pretrained model. Which we have stated in the rules that external data is not allowed, but pretrained models are. Do the rules mean publicly available pretrained models?
\n\nIt seems like the rules are contradictory and I would like to know the specifics of using a pretrained model without breaking the rules.
\n\nCan we use a model we have created with external data for fine-tuning, transfer learning, or in earlier portions of some novel algorithm when the model is not publicly available?
", "rawMarkdown": "I would also like clarification on the use of pretrained models.\n\nI am using detectron that has been trained on ipatch data which is publicly available. \n\nCan I not use this model in conjunction with whatever algorithm I make?\n\nCould I use the model for transfer learning?\n\nTechnically whatever model I have that had been trained on external data can also be a pretrained model. Which we have stated in the rules that external data is not allowed, but pretrained models are. Do the rules mean publicly available pretrained models?\n\nIt seems like the rules are contradictory and I would like to know the specifics of using a pretrained model without breaking the rules.\n\nCan we use a model we have created with external data for fine-tuning, transfer learning, or in earlier portions of some novel algorithm when the model is not publicly available?\n" }, { "id": 365807, "postDate": "2018-08-03T12:59:56.743Z", "content": "Pre-trained models from torchvision (https://github.com/pytorch/vision/tree/master/torchvision/models)
", "rawMarkdown": "Pre-trained models from torchvision (https://github.com/pytorch/vision/tree/master/torchvision/models)" }, { "id": 365231, "postDate": "2018-08-02T06:48:41.897Z", "content": "It is a fair point, but most of satellite data-sets I can imagine are publicly available anyway, so it is hard to gain unfair advantage in this way. Unless you have private fleet of satellites on Earths orbit.
", "rawMarkdown": "It is a fair point, but most of satellite data-sets I can imagine are publicly available anyway, so it is hard to gain unfair advantage in this way. Unless you have private fleet of satellites on Earths orbit." }, { "id": 365128, "postDate": "2018-08-02T00:08:41.793Z", "content": "At the risk of stating the obvious: ImageNet and COCO
", "rawMarkdown": "At the risk of stating the obvious: ImageNet and COCO", "replies": [ { "id": 365137, "postDate": "2018-08-02T00:35:56.923Z", "content": "My understanding is that you cannot use external datasets. So you MAY NOT take your own private model and train it on imagenet (or better yet, some remote sensing/satellite imagery dataset) and then do transfer learning to train it on the contest dataset. You can take a publicly available model (e.g. Resnet50 weights trained on imagenet) and then customize/fine-tune for this contest (as long as that model has been listed in this thread).
\n\nDo I have that right @inversion?
", "rawMarkdown": "My understanding is that you cannot use external datasets. So you MAY NOT take your own private model and train it on imagenet (or better yet, some remote sensing/satellite imagery dataset) and then do transfer learning to train it on the contest dataset. You can take a publicly available model (e.g. Resnet50 weights trained on imagenet) and then customize/fine-tune for this contest (as long as that model has been listed in this thread).\n\nDo I have that right @inversion?" }, { "id": 365149, "postDate": "2018-08-02T01:32:32.943Z", "content": "I'm not training my model on those datasets, I'm just initializing my models with weights from models which were trained on those two datasets. Which is exactly what you and inversion are describing. So should I list the model type instead?
", "rawMarkdown": "I'm not training my model on those datasets, I'm just initializing my models with weights from models which were trained on those two datasets. Which is exactly what you and inversion are describing. So should I list the model type instead?" }, { "id": 365204, "postDate": "2018-08-02T05:29:19.967Z", "content": "Paul, isn't tranfer learning imply existence of external dasta set, which was used to train Neural Net to extract general features?
\n\nIn that case how is using ResNet different from using algorithm trained on other satelite data?
", "rawMarkdown": "Paul, isn't tranfer learning imply existence of external dasta set, which was used to train Neural Net to extract general features? \n\nIn that case how is using ResNet different from using algorithm trained on other satelite data?", "votes": 1 }, { "id": 365212, "postDate": "2018-08-02T05:50:24.137Z", "content": "I don't claim to know the exact reason for this rule. Maybe the contest organizers can clarify.
\n\nHowever my best guess would be to make sure there is an even playing field. Models pretrained on common data sets are publicly available so anyone can grab the weights and start refining it for their own purpose. If you have access to some non-publicly available model trained or a large amount of labeled satellite imagery of the ocean, you might have a significant advantage over another contestant.
", "rawMarkdown": "I don't claim to know the exact reason for this rule. Maybe the contest organizers can clarify.\n\nHowever my best guess would be to make sure there is an even playing field. Models pretrained on common data sets are publicly available so anyone can grab the weights and start refining it for their own purpose. If you have access to some non-publicly available model trained or a large amount of labeled satellite imagery of the ocean, you might have a significant advantage over another contestant." }, { "id": 365435, "postDate": "2018-08-02T15:30:19.137Z", "content": "@Michał -
\n\nIt is a blurry line, but we do differentiate between external data and pre-trained models. The use of pre-trained models are so ubiquitous, it often doesn't make sense to exclude them. On the other hand, it is a very common request that, other than the use of pre-trained models starting point, only the competition data is used for further training.
", "rawMarkdown": "@Michał -\n\nIt is a blurry line, but we do differentiate between external data and pre-trained models. The use of pre-trained models are so ubiquitous, it often doesn't make sense to exclude them. On the other hand, it is a very common request that, other than the use of pre-trained models starting point, only the competition data is used for further training." }, { "id": 367226, "postDate": "2018-08-07T10:53:54.630Z", "content": "Hi @inversion,
\nlet's suppose I would like to perform following procedure:
The last point is not very likely, but the previous can be. Is this procedure allowed, or not? On one hand I use pretrained network that is publicly available, so it is allowed, on the other hand I use external dataset, that is not allowed. Please kindly clarify that.
", "rawMarkdown": "Hi @inversion,This is a good point. In my mind, there isn't much difference between this and using an open source pre-trained model (other than the public availability, of course).
", "rawMarkdown": "This is a good point. In my mind, there isn't much difference between this and using an open source pre-trained model (other than the public availability, of course). " } ] }, { "id": 759303, "postDate": "2020-02-28T21:38:36.197Z", "rawMarkdown": "", "isDeleted": true }, { "id": 392285, "postDate": "2018-09-23T13:16:23.420Z", "rawMarkdown": "", "isDeleted": true }, { "id": 367001, "postDate": "2018-08-06T21:35:18.200Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 412518, "author_name": "Vaghawan ", "author_url": "", "post_date": "2018-10-30T10:32:10.933000", "content": "Locally modified version of Mask-Rcnn : https://github.com/matterport/Mask_RCNN with coco weight: https://github.com/matterport/Mask_RCNN/releases/download/v2.0/mask_rcnn_coco.h5 and also this https://github.com/fizyr/keras-maskrcnn/releases/download/0.2.0/resnet50_coco_v0.2.0.h5
", "votes": 1, "replies": [] }, { "id": 372338, "author_name": "Komaki", "author_url": "", "post_date": "2018-08-19T02:08:03.383000", "content": "Hi @inversion,
\n\nIs it allowed to manually annotate training data to assist model training?
", "votes": 1, "replies": [ { "id": 372646, "author_name": "inversion", "author_url": "", "post_date": "2018-08-20T01:31:59.107000", "content": "In general, as is the case in this contest, you can do what you wish with the training images. You can't, though, manually label the test images.
", "votes": 0, "replies": [] }, { "id": 372664, "author_name": "Komaki", "author_url": "", "post_date": "2018-08-20T03:07:49.357000", "content": "Thank you :)
", "votes": 0, "replies": [] } ] }, { "id": 414467, "author_name": "x0x0w1", "author_url": "", "post_date": "2018-11-02T20:24:39.657000", "content": "Pretrained models from https://github.com/mapillary/inplace_abn repository
", "votes": 0, "replies": [] }, { "id": 372999, "author_name": "Blonde", "author_url": "", "post_date": "2018-08-20T21:29:31.160000", "content": "@inversion I was just wondering: would not disclosure of the pre-trained models used by different users give additional information about the goodness/relevance of this model? For instance, if a kaggle master with position top 5 on LB publishes a pre-trained model more users will consider it as a good start than if a novice will publish it. It gives an additional info, puts kaggle leaders under more attention in this regard, and probably is not so good for a fair competition. It is NOT allowed, however, to have two kaggle accounts. Would it be a better idea to create a new kaggle account named airbus_models with password pretrainedmodels for everyone to use the same account to publish pre-trained models they use? Then everyone has an opportunity to publish models anonymously, and we remove the bias created by people profiles and position on LB. I can create such account myself, but do not want to get in trouble for making two kaggle accounts. Since you work at kaggle you may create/comment on this solution.
", "votes": -2, "replies": [] }, { "id": 420868, "author_name": "Oleg Z.", "author_url": "", "post_date": "2018-11-14T09:06:13.737000", "content": "I use pretrained pytorch models:\n'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',\n'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
", "votes": 0, "replies": [] }, { "id": 420205, "author_name": "guzi", "author_url": "", "post_date": "2018-11-13T09:05:36.723000", "content": "Locally modified version of Deeplabv3+ : https://github.com/tensorflow/models/tree/master/research/deeplab with weight: http://download.tensorflow.org/models/xception_41_2018_05_09.tar.gz
", "votes": 0, "replies": [] }, { "id": 420067, "author_name": "jison", "author_url": "", "post_date": "2018-11-13T02:22:09.313000", "content": "I used Keras ResNet52 pretrained models with Imagenet weight
", "votes": 0, "replies": [] }, { "id": 417093, "author_name": "Jaideep", "author_url": "", "post_date": "2018-11-07T18:24:15.397000", "content": "i use pretrained model pytorch resnet34 or18
", "votes": 0, "replies": [] }, { "id": 417012, "author_name": "Cozy Doomer", "author_url": "", "post_date": "2018-11-07T16:02:51.920000", "content": "I am using pretrained pytorch models:\nresnet34: \nhttps://download.pytorch.org/models/resnet34-333f7ec4.pth\nalso I'll include resnet50 if I try it in the future:\nhttps://download.pytorch.org/models/resnet50-19c8e357.pth
", "votes": 0, "replies": [] }, { "id": 416926, "author_name": "Max", "author_url": "", "post_date": "2018-11-07T12:57:29.317000", "content": "fastai, pretrained-models & ternaus
", "votes": 0, "replies": [] }, { "id": 416373, "author_name": "sheep", "author_url": "", "post_date": "2018-11-06T15:46:48.040000", "content": "Model from public kernel https://www.kaggle.com/iafoss/fine-tuning-resnet34-on-ship-detection/output
", "votes": 0, "replies": [] }, { "id": 416368, "author_name": "SeuTao", "author_url": "", "post_date": "2018-11-06T15:43:36.600000", "content": "our team\nPretrained pytorch models:\n'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',\n'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',\n‘seresnext50’：'http://data.lip6.fr/cadene/pretrainedmodels/se_resnext50_32x4d-a260b3a4.pth'
", "votes": 0, "replies": [] }, { "id": 416212, "author_name": "guzi", "author_url": "", "post_date": "2018-11-06T12:06:04.387000", "content": "Pretrained models from https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/model_zoo.md\nhttp://download.tensorflow.org/models/resnet_v1_50_2018_05_04.tar.gz
", "votes": 0, "replies": [] }, { "id": 416158, "author_name": "e-toppo", "author_url": "", "post_date": "2018-11-06T09:47:20.843000", "content": "I used Pretrained models from fast.ai\nhttp://files.fast.ai/models/
", "votes": 0, "replies": [] }, { "id": 416032, "author_name": "phun", "author_url": "", "post_date": "2018-11-06T04:06:56.467000", "content": "Pytorch models with weights trained on imagenet: resnet18, resnet34, densenet121
", "votes": 0, "replies": [] }, { "id": 415713, "author_name": "Igor Praznik", "author_url": "", "post_date": "2018-11-05T14:40:58.067000", "content": "Keras/keras-contribution, pytorch/cadene resnet pretrained models as encoders
", "votes": 0, "replies": [] }, { "id": 415407, "author_name": "Mohammad Azam Khan", "author_url": "", "post_date": "2018-11-05T03:54:55.290000", "content": "Perhaps one or two of the followings: \nPretrained models from 1) Torchvision, 2) https://github.com/Cadene/pretrained-models.pytorch, 3) https://modelzoo.co/model/pytorch-cnn-finetune, 4) Pretrained-models in keras https://github.com/fchollet/deep-learning-models/releases, specifically used for this repo https://github.com/fizyr and 5) Pretrained models in fastai.
", "votes": 0, "replies": [] }, { "id": 415354, "author_name": "nan", "author_url": "", "post_date": "2018-11-05T01:11:03.033000", "content": "Pre-trained on the coco dataset https://www.dropbox.com/s/dr4qh2xo1ksaj6r/coco.h5?dl=0\nand this: https://github.com/tensorflow/models/tree/master/research/slim
", "votes": 0, "replies": [] }, { "id": 415067, "author_name": "Byung Soo Ko", "author_url": "", "post_date": "2018-11-04T09:59:55.933000", "content": "Pretrained model from https://github.com/junfu1115/DANet repo
", "votes": 0, "replies": [] }, { "id": 414699, "author_name": "Kostiantyn Maksymov", "author_url": "", "post_date": "2018-11-03T12:18:43.163000", "content": "https://github.com/qubvel/segmentation_models https://github.com/broadinstitute/keras-resnet
", "votes": 0, "replies": [] }, { "id": 413998, "author_name": "Michael Schwendeman", "author_url": "", "post_date": "2018-11-01T23:33:34.227000", "content": "yolo v3: https://pjreddie.com/media/files/yolov3.weights
", "votes": 0, "replies": [] }, { "id": 413764, "author_name": "toshi_k", "author_url": "", "post_date": "2018-11-01T13:45:08.057000", "content": "I use pretrained pytorch models (https://pytorch.org/docs/stable/torchvision/models.html).
", "votes": 0, "replies": [] }, { "id": 413199, "author_name": "ZhuoZheng", "author_url": "", "post_date": "2018-10-31T13:20:41.673000", "content": "pre-trained pytorch models(https://github.com/pytorch/vision/tree/master/torchvision/models)
", "votes": 0, "replies": [] }, { "id": 411400, "author_name": "XudongMao", "author_url": "", "post_date": "2018-10-28T03:21:50.897000", "content": "Pretrained models from torchvision: https://download.pytorch.org/models/densenet121-a639ec97.pth
", "votes": 0, "replies": [] }, { "id": 410198, "author_name": "YaG320", "author_url": "", "post_date": "2018-10-25T15:32:06.513000", "content": "Pretrained models from: http://models.tensorpack.com/
", "votes": 0, "replies": [] }, { "id": 405867, "author_name": "Borys Tymchenko", "author_url": "", "post_date": "2018-10-18T08:39:15.193000", "content": "Squeezenet for Keras https://github.com/rcmalli/keras-squeezenet/
", "votes": 0, "replies": [] }, { "id": 404863, "author_name": "Sergei Tsimbalist", "author_url": "", "post_date": "2018-10-16T13:25:59.343000", "content": "I suppose if you take public kernel you don't need to list it's pretrained models?
", "votes": 0, "replies": [] }, { "id": 404487, "author_name": "Tom57", "author_url": "", "post_date": "2018-10-15T19:52:53.793000", "content": "Pretrained pytorch models:\n'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',\n 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
", "votes": 0, "replies": [] }, { "id": 401278, "author_name": "kinetical", "author_url": "", "post_date": "2018-10-09T18:27:42.510000", "content": "resnet trained on ImageNet
", "votes": 0, "replies": [] }, { "id": 398806, "author_name": "", "author_url": "", "post_date": "2018-10-04T15:56:38.927000", "content": "", "votes": 0, "replies": [] }, { "id": 391371, "author_name": "", "author_url": "", "post_date": "2018-09-21T17:50:01.440000", "content": "", "votes": 0, "replies": [] }, { "id": 390701, "author_name": "", "author_url": "", "post_date": "2018-09-20T17:00:45.800000", "content": "", "votes": 0, "replies": [] }, { "id": 389671, "author_name": "", "author_url": "", "post_date": "2018-09-19T03:44:57.423000", "content": "", "votes": 0, "replies": [] }, { "id": 387031, "author_name": "", "author_url": "", "post_date": "2018-09-14T06:57:40.887000", "content": "", "votes": 0, "replies": [] }, { "id": 386980, "author_name": "", "author_url": "", "post_date": "2018-09-14T04:10:41.040000", "content": "", "votes": 0, "replies": [] }, { "id": 383579, "author_name": "", "author_url": "", "post_date": "2018-09-09T02:22:42.970000", "content": "", "votes": 0, "replies": [] }, { "id": 380212, "author_name": "", "author_url": "", "post_date": "2018-09-02T04:09:36.677000", "content": "", "votes": 0, "replies": [] }, { "id": 380060, "author_name": "", "author_url": "", "post_date": "2018-09-01T15:21:15.653000", "content": "", "votes": 0, "replies": [] }, { "id": 379401, "author_name": "", "author_url": "", "post_date": "2018-08-31T10:08:06.500000", "content": "", "votes": 0, "replies": [] }, { "id": 379212, "author_name": "", "author_url": "", "post_date": "2018-08-31T02:57:15.117000", "content": "", "votes": 0, "replies": [] }, { "id": 376865, "author_name": "", "author_url": "", "post_date": "2018-08-28T07:33:27.440000", "content": "", "votes": 0, "replies": [] }, { "id": 376435, "author_name": "", "author_url": "", "post_date": "2018-08-27T13:42:37.207000", "content": "", "votes": 0, "replies": [] }, { "id": 373726, "author_name": "", "author_url": "", "post_date": "2018-08-21T20:06:14.863000", "content": "", "votes": 0, "replies": [] }, { "id": 370547, "author_name": "", "author_url": "", "post_date": "2018-08-15T01:41:37.227000", "content": "", "votes": 0, "replies": [] }, { "id": 365807, "author_name": "", "author_url": "", "post_date": "2018-08-03T12:59:56.743000", "content": "", "votes": 0, "replies": [] }, { "id": 365231, "author_name": "", "author_url": "", "post_date": "2018-08-02T06:48:41.897000", "content": "", "votes": 0, "replies": [] }, { "id": 365128, "author_name": "", "author_url": "", "post_date": "2018-08-02T00:08:41.793000", "content": "", "votes": 0, "replies": [ { "id": 365137, "author_name": "", "author_url": "", "post_date": "2018-08-02T00:35:56.923000", "content": "", "votes": 0, "replies": [] }, { "id": 365149, "author_name": "", "author_url": "", "post_date": "2018-08-02T01:32:32.943000", "content": "", "votes": 0, "replies": [] }, { "id": 365204, "author_name": "", "author_url": "", "post_date": "2018-08-02T05:29:19.967000", "content": "", "votes": 1, "replies": [] }, { "id": 365212, "author_name": "", "author_url": "", "post_date": "2018-08-02T05:50:24.137000", "content": "", "votes": 0, "replies": [] }, { "id": 365435, "author_name": "", "author_url": "", "post_date": "2018-08-02T15:30:19.137000", "content": "", "votes": 0, "replies": [] }, { "id": 367226, "author_name": "", "author_url": "", "post_date": "2018-08-07T10:53:54.630000", "content": "", "votes": 8, "replies": [] }, { "id": 367946, "author_name": "", "author_url": "", "post_date": "2018-08-08T22:08:49.937000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 759303, "author_name": "", "author_url": "", "post_date": "2020-02-28T21:38:36.197000", "content": "", "votes": 0, "replies": [] }, { "id": 392285, "author_name": "", "author_url": "", "post_date": "2018-09-23T13:16:23.420000", "content": "", "votes": 0, "replies": [] }, { "id": 367001, "author_name": "", "author_url": "", "post_date": "2018-08-06T21:35:18.200000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "364105": "As stated in the Rules, the use of external data is not permitted. You may, though, use pre-trained models, as long as you use this thread to list the pre-trained models you are planning to use. Once a model has been posted once, it does not need to be posted again. In order for you to use a pre-trained model, it must be posted here no later than one week prior to the competition close.", "412518": "Locally modified version of Mask-Rcnn : https://github.com/matterport/Mask_RCNN with coco weight: https://github.com/matterport/Mask_RCNN/releases/download/v2.0/mask_rcnn_coco.h5 and also this https://github.com/fizyr/keras-maskrcnn/releases/download/0.2.0/resnet50_coco_v0.2.0.h5 ", "372338": "Hi @inversion,\n\nIs it allowed to manually annotate training data to assist model training?", "414467": "Pretrained models from https://github.com/mapillary/inplace_abn repository ", "372999": "@inversion I was just wondering: would not disclosure of the pre-trained models used by different users give additional information about the goodness/relevance of this model? For instance, if a kaggle master with position top 5 on LB publishes a pre-trained model more users will consider it as a good start than if a novice will publish it. It gives an additional info, puts kaggle leaders under more attention in this regard, and probably is not so good for a fair competition. It is NOT allowed, however, to have two kaggle accounts. Would it be a better idea to create a new kaggle account named airbus_models with password pretrainedmodels for everyone to use the same account to publish pre-trained models they use? Then everyone has an opportunity to publish models anonymously, and we remove the bias created by people profiles and position on LB. I can create such account myself, but do not want to get in trouble for making two kaggle accounts. Since you work at kaggle you may create/comment on this solution. ", "420868": "I use pretrained pytorch models:\n'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',\n'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',", "420205": "Locally modified version of Deeplabv3+ : https://github.com/tensorflow/models/tree/master/research/deeplab with weight: http://download.tensorflow.org/models/xception_41_2018_05_09.tar.gz", "420067": "I used Keras ResNet52 pretrained models with Imagenet weight ", "417093": "i use pretrained model pytorch resnet34 or18", "417012": "I am using pretrained pytorch models:\nresnet34: \nhttps://download.pytorch.org/models/resnet34-333f7ec4.pth\nalso I'll include resnet50 if I try it in the future:\nhttps://download.pytorch.org/models/resnet50-19c8e357.pth", "416926": "fastai, pretrained-models & ternaus", "416373": "Model from public kernel https://www.kaggle.com/iafoss/fine-tuning-resnet34-on-ship-detection/output", "416368": "our team\nPretrained pytorch models:\n'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',\n'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',\n‘seresnext50’：'http://data.lip6.fr/cadene/pretrainedmodels/se_resnext50_32x4d-a260b3a4.pth'", "416212": "Pretrained models from https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/model_zoo.md\nhttp://download.tensorflow.org/models/resnet_v1_50_2018_05_04.tar.gz", "416158": "I used Pretrained models from fast.ai\nhttp://files.fast.ai/models/", "416032": "Pytorch models with weights trained on imagenet: resnet18, resnet34, densenet121", "415713": "Keras/keras-contribution, pytorch/cadene resnet pretrained models as encoders", "415407": "Perhaps one or two of the followings: \nPretrained models from 1) Torchvision, 2) https://github.com/Cadene/pretrained-models.pytorch, 3) https://modelzoo.co/model/pytorch-cnn-finetune, 4) Pretrained-models in keras https://github.com/fchollet/deep-learning-models/releases, specifically used for this repo https://github.com/fizyr and 5) Pretrained models in fastai.\n", "415354": "Pre-trained on the coco dataset https://www.dropbox.com/s/dr4qh2xo1ksaj6r/coco.h5?dl=0\nand this: https://github.com/tensorflow/models/tree/master/research/slim", "415067": "Pretrained model from https://github.com/junfu1115/DANet repo", "414699": "https://github.com/qubvel/segmentation_models https://github.com/broadinstitute/keras-resnet", "413998": "yolo v3: https://pjreddie.com/media/files/yolov3.weights", "413764": "I use pretrained pytorch models (https://pytorch.org/docs/stable/torchvision/models.html). ", "413199": "pre-trained pytorch models(https://github.com/pytorch/vision/tree/master/torchvision/models)", "411400": "Pretrained models from torchvision: https://download.pytorch.org/models/densenet121-a639ec97.pth", "410198": "Pretrained models from: http://models.tensorpack.com/", "405867": "Squeezenet for Keras https://github.com/rcmalli/keras-squeezenet/", "404863": "I suppose if you take public kernel you don't need to list it's pretrained models?", "404487": "Pretrained pytorch models:\n'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',\n 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',", "401278": "resnet trained on ImageNet", "398806": "Sorry, May I know what is the defination of pre-trained Model? I have been training the full train set of ships by using YOLO tiny model. Is it considered pre-trained model?", "391371": "Pre-trained on the coco dataset https://www.dropbox.com/s/dr4qh2xo1ksaj6r/coco.h5?dl=0", "390701": "ResNeXt models. I uploaded them to kaggle https://www.kaggle.com/iafoss/pytorch-pretrained-models/home , but originally they are taken from http://files.fast.ai/models/weights.tgz.", "389671": "Pre-trained model from DRBox https://github.com/liulei01/DRBox/tree/master/examples/rbox/deploy/ship", "387031": "Keras implementation of Deeplabv3+(https://github.com/bonlime/keras-deeplab-v3-plus)", "386980": "chainer resnext models \nhttps://chainercv-models.preferred.jp/se_resnext50_imagenet_converted_2018_06_28.npz\nhttps://chainercv-models.preferred.jp/se_resnext101_imagenet_converted_2018_06_28.npz", "383579": "resnet weight trained on imagenet and coco", "380212": "Pretrained-models for keras https://github.com/fchollet/deep-learning-models/releases", "380060": "keras all pre-trained model\n\nhttps://keras.io/applications/\n", "379401": "Hi, mask r cnn with coco weights", "379212": "I use pre-trained pytorch models(https://github.com/pytorch/vision/tree/master/torchvision/models)", "376865": "Pretrained models from fast.ai\n\nhttp://files.fast.ai/models/", "376435": "Keras ResNet Imagenet pretrained models - https://github.com/qubvel/classification_models", "373726": "https://github.com/Cadene/pretrained-models.pytorch", "370547": "I would also like clarification on the use of pretrained models.\n\nI am using detectron that has been trained on ipatch data which is publicly available. \n\nCan I not use this model in conjunction with whatever algorithm I make?\n\nCould I use the model for transfer learning?\n\nTechnically whatever model I have that had been trained on external data can also be a pretrained model. Which we have stated in the rules that external data is not allowed, but pretrained models are. Do the rules mean publicly available pretrained models?\n\nIt seems like the rules are contradictory and I would like to know the specifics of using a pretrained model without breaking the rules.\n\nCan we use a model we have created with external data for fine-tuning, transfer learning, or in earlier portions of some novel algorithm when the model is not publicly available?\n", "365807": "Pre-trained models from torchvision (https://github.com/pytorch/vision/tree/master/torchvision/models)", "365231": "It is a fair point, but most of satellite data-sets I can imagine are publicly available anyway, so it is hard to gain unfair advantage in this way. Unless you have private fleet of satellites on Earths orbit.", "365128": "At the risk of stating the obvious: ImageNet and COCO", "759303": "", "392285": "", "367001": "" } }