{ "id": 62740, "title": "External Data and Pre-Trained Model Disclosure Thread", "url": "/competitions/google-ai-open-images-visual-relationship-track/discussion/62740", "author_name": "", "post_date": "2018-08-06T16:05:55.290918700Z", "votes": 3, "comment_count": 19, "views": 0, "content": "

Please use this thread to post any external data and pre-trained models you use for your solution. Reminder, disclosure is required one week prior to the submission deadline.

", "messages": [ { "id": "366839", "postDate": "08/06/2018 16:05:55", "content": "

Please use this thread to post any external data and pre-trained models you use for your solution. Reminder, disclosure is required one week prior to the submission deadline.

", "rawMarkdown": "Please use this thread to post any external data and pre-trained models you use for your solution. Reminder, disclosure is required one week prior to the submission deadline.", "votes": null }, { "id": "367633", "postDate": "08/08/2018 06:57:28", "content": "

I am not planning on using external data (that is apart from the data from object detection track but not sure this would qualify as external) . I am still in the exploratory phase and not sure if I will have time to work more on this, but if I do, I will most likely use a pretrained model (or a couple of models) from either of the following sources:

\n\n
    \n
  1. Tensorflow detection model zoo / Tensorflow pretrained models
    2. \n
  2. Pretrained models that come with the fastai library (resnet18, resnet34, resnet50, resnet101, resnet152, vgg16, \n vgg19, resnetxt50, resnext101, resnext101_64, wrn, inceptionresnet_2, inception_4, dn121, dn161, dn169, dn201).
    3. \n
  3. Any of the pretrained models from here and here, in particular the darknet53.conv.74
    4. \n
  4. pretrained models in PyTorch from this repository
    5. \n
\n\n

Not sure if more information would be required - please let me know if that would be the case.

", "rawMarkdown": "I am not planning on using external data (that is apart from the data from object detection track but not sure this would qualify as external) . I am still in the exploratory phase and not sure if I will have time to work more on this, but if I do, I will most likely use a pretrained model (or a couple of models) from either of the following sources:\n\n 1. [Tensorflow detection model zoo][1] / [Tensorflow pretrained models][2]\n 2. Pretrained models that come with the [fastai][3] library (resnet18, resnet34, resnet50, resnet101, resnet152, vgg16, \n vgg19, resnetxt50, resnext101, resnext101_64, wrn, inceptionresnet_2, inception_4, dn121, dn161, dn169, dn201).\n 3. Any of the pretrained models from [here][4] and [here][5], in particular the [darknet53.conv.74][6]\n 4. pretrained models in PyTorch from this [repository][7]\n\nNot sure if more information would be required - please let me know if that would be the case.\n\n\n [1]: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md\n [2]: https://github.com/tensorflow/models/tree/master/research/slim\n [3]: https://github.com/fastai/fastai\n [4]: https://pjreddie.com/darknet/imagenet/\n [5]: https://pjreddie.com/darknet/yolo/\n [6]: https://pjreddie.com/media/files/darknet53.conv.74\n [7]: https://github.com/Cadene/pretrained-models.pytorch", "votes": null }, { "id": "371701", "postDate": "08/17/2018 13:22:29", "content": "

Keras resnet pretrained on imagenet and tf faster_rcnn_inception_resnet_v2_atrous_oid weights if I get that to work ;)

", "rawMarkdown": "Keras resnet pretrained on imagenet and tf faster_rcnn_inception_resnet_v2_atrous_oid weights if I get that to work ;)", "votes": null }, { "id": "374257", "postDate": "08/22/2018 17:22:23", "content": "

I will perhaps use one or more pre-trained model(s) from either of the following sources:

\n\n

Pre-trained model on Open Image dataset for detection/fine_tune_checkpoint: http://download.tensorflow.org/models/object_detection/faster_rcnn_inception_resnet_v2_atrous_oid_2018_01_28.tar.gz

\n\n

YOLO pre-trained model: https://pjreddie.com/darknet/yolo/

\n\n

Pre-trained model from ImageAI:

\n\n

https://github.com/OlafenwaMoses/ImageAI/releases/download/1.0/resnet50_coco_best_v2.0.1.h5

\n\n

https://github.com/OlafenwaMoses/ImageAI/releases/download/1.0/yolo.h5

\n\n

https://github.com/OlafenwaMoses/ImageAI/releases/download/1.0/tiny-yolo.h5

\n\n

Pre-trained models from https://pjreddie.com/darknet/imagenet/.

", "rawMarkdown": "I will perhaps use one or more pre-trained model(s) from either of the following sources:\n\nPre-trained model on Open Image dataset for detection/fine_tune_checkpoint: http://download.tensorflow.org/models/object_detection/faster_rcnn_inception_resnet_v2_atrous_oid_2018_01_28.tar.gz\n\nYOLO pre-trained model: https://pjreddie.com/darknet/yolo/\n\nPre-trained model from ImageAI:\n\nhttps://github.com/OlafenwaMoses/ImageAI/releases/download/1.0/resnet50_coco_best_v2.0.1.h5\n\nhttps://github.com/OlafenwaMoses/ImageAI/releases/download/1.0/yolo.h5\n\nhttps://github.com/OlafenwaMoses/ImageAI/releases/download/1.0/tiny-yolo.h5\n\nPre-trained models from https://pjreddie.com/darknet/imagenet/.", "votes": null }, { "id": "374345", "postDate": "08/22/2018 21:08:24", "content": "

Pre-trained resnet_50 and vgg16 models from gluon model zoo. Additional training on materials in context dataset. Object detectors trained on data from object detection track.

", "rawMarkdown": "Pre-trained resnet_50 and vgg16 models from [gluon model zoo][1]. Additional training on [materials in context][2] dataset. Object detectors trained on data from object detection track.\n\n\n [1]: http://mxnet.apache.org/api/python/gluon/model_zoo.html\n [2]: http://opensurfaces.cs.cornell.edu/publications/minc/", "votes": null }, { "id": "374415", "postDate": "08/23/2018 02:06:47", "content": "

The pretrained model we used is as follows:

\n\n
    \n
  1. Keras Xception pretrained model on imagenet.
    2. \n
  2. Pytorch ResNet 50, Densenet 121 pretrained model on imagenet.
    3. \n
  3. RetianNet pretrained model on coco.
    4. \n
  4. Faster RCNN pretrained model on oid.
    5. \n
", "rawMarkdown": "The pretrained model we used is as follows:\n\n 1. Keras [Xception][1] pretrained model on imagenet.\n 2. Pytorch [ResNet 50][2], [Densenet 121][3] pretrained model on imagenet.\n 3. [RetianNet][4] pretrained model on coco.\n 4. [Faster RCNN][5] pretrained model on oid.\n\n\n [1]: https://github.com/fchollet/deep-learning-models/\n [2]: https://download.pytorch.org/models/resnet50-19c8e357.pth\n [3]: https://download.pytorch.org/models/densenet121-241335ed.pth\n [4]: https://github.com/yhenon/pytorch-retinanet\n [5]: https://github.com/tensorflow/models/blob/505f554c6417931c96b59516f14d1ad65df6dbc5/research/object_detection/g3doc/detection_model_zoo.md", "votes": null }, { "id": "374433", "postDate": "08/23/2018 03:17:11", "content": "

The pretrained models we used are:

\n\n
    \n
  1. ResNeXt-101-64x4d model pretrained on COCO from the pytorch version of mask-RCNN: https://github.com/roytseng-tw/Detectron.pytorch
    2. \n
  2. The same model is finetuned on the OpenImage object detection dataset.
    3. \n
  3. We may or may not use Visual Genome to pre-train our relationship detector. It is still under consideration.
    4. \n
", "rawMarkdown": "The pretrained models we used are:\n\n1. ResNeXt-101-64x4d model pretrained on COCO from the pytorch version of mask-RCNN: https://github.com/roytseng-tw/Detectron.pytorch\n2. The same model is finetuned on the OpenImage object detection dataset.\n3. We may or may not use [Visual Genome][1] to pre-train our relationship detector. It is still under consideration.\n\n [1]: https://cs.stanford.edu/~danfei/scene-graph/", "votes": null }, { "id": "374635", "postDate": "08/23/2018 12:50:30", "content": "

Tensorflow model zoo: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md Pytorch model zoo: https://pytorch.org/docs/stable/torchvision/models.html Keras model zoo: https://keras.io/applications/

", "rawMarkdown": "Tensorflow model zoo: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md Pytorch model zoo: https://pytorch.org/docs/stable/torchvision/models.html Keras model zoo: https://keras.io/applications/", "votes": null }, { "id": "374670", "postDate": "08/23/2018 14:42:12", "content": "

Mask_RCNN

", "rawMarkdown": "[Mask_RCNN][1] \n\n\n [1]: https://github.com/matterport/Mask_RCNN", "votes": null }, { "id": "374706", "postDate": "08/23/2018 16:05:57", "content": "

Pretrained models:

\n\n

yolov3: https://pjreddie.com/darknet/yolo/

\n\n

Pretrained models from keras: https://keras.io/applications/

\n\n

Data:

\n\n

Open Images Dataset V4: https://storage.googleapis.com/openimages/web/download.html

\n\n

Google AI Open Images - Object Detection Track: https://www.kaggle.com/c/google-ai-open-images-object-detection-track

", "rawMarkdown": "Pretrained models:\n\nyolov3: https://pjreddie.com/darknet/yolo/\n\nPretrained models from keras: https://keras.io/applications/\n\n\n\nData:\n\nOpen Images Dataset V4: https://storage.googleapis.com/openimages/web/download.html\n\nGoogle AI Open Images - Object Detection Track: https://www.kaggle.com/c/google-ai-open-images-object-detection-track", "votes": null }, { "id": "374746", "postDate": "08/23/2018 17:57:17", "content": "

If I manage to get my solution working in time for the contest deadline, it will be based on Detectron (https://github.com/facebookresearch/Detectron) which initializes its training runs with the pre-trained classification models listed in the .yaml files in the config directory (https://github.com/facebookresearch/Detectron/tree/master/configs), e.g. https://s3-us-west-2.amazonaws.com/detectron/ImageNetPretrained/47261647/R-50-GN.pkl used by https://github.com/facebookresearch/Detectron/blob/master/configs/04_2018_gn_baselines/e2e_mask_rcnn_R-50-FPN_2x_gn.yaml

", "rawMarkdown": "If I manage to get my solution working in time for the contest deadline, it will be based on Detectron (https://github.com/facebookresearch/Detectron) which initializes its training runs with the pre-trained classification models listed in the .yaml files in the config directory (https://github.com/facebookresearch/Detectron/tree/master/configs), e.g. https://s3-us-west-2.amazonaws.com/detectron/ImageNetPretrained/47261647/R-50-GN.pkl used by https://github.com/facebookresearch/Detectron/blob/master/configs/04_2018_gn_baselines/e2e_mask_rcnn_R-50-FPN_2x_gn.yaml", "votes": null }, { "id": "374770", "postDate": "08/23/2018 19:03:50", "content": "

I may try \nTensorflow model zoo: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md Pytorch model zoo: https://pytorch.org/docs/stable/torchvision/models.html Keras model zoo: https://keras.io/applications/\nand vgg16 based pretrianed model here

", "rawMarkdown": "I may try \nTensorflow model zoo: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md Pytorch model zoo: https://pytorch.org/docs/stable/torchvision/models.html Keras model zoo: https://keras.io/applications/\nand vgg16 based pretrianed model [here][1]\n [1]: http://cs.stanford.edu/people/jcjohns/densecap/densecap-pretrained-vgg16.t7.zip", "votes": null }, { "id": "374774", "postDate": "08/23/2018 19:15:19", "content": "

Keras model zoo: https://keras.io/applications/

", "rawMarkdown": "Keras model zoo: https://keras.io/applications/", "votes": null }, { "id": "374866", "postDate": "08/24/2018 03:03:19", "content": "

I used the pre-trained model in Tensorflow detection model zoo for fine tuning:\nhttp://download.tensorflow.org/models/object_detection/faster_rcnn_inception_resnet_v2_atrous_oid_2018_01_28.tar.gz

", "rawMarkdown": "I used the pre-trained model in Tensorflow detection model zoo for fine tuning:\nhttp://download.tensorflow.org/models/object_detection/faster_rcnn_inception_resnet_v2_atrous_oid_2018_01_28.tar.gz", "votes": null }, { "id": "374944", "postDate": "08/24/2018 08:35:16", "content": "

May or may not use pretrained models and datasets mentioned in:

\n\n
    \n

  1. TensorFlow Object Detection
    \n(Open Images dataset, COCO dataset, Oxford-IIIT Pets dataset, model zoo, etc.)
    \nhttps://github.com/tensorflow/models/tree/master/research/object_detection

    2. \n

  2. Darknet
    \n(ImageNet dataset, YOLO, etc.)
    \nhttps://pjreddie.com/darknet

    3. \n

  3. ImageAI
    \nhttps://github.com/OlafenwaMoses/ImageAI

    4. \n

  4. ChainerCV
    \nhttps://github.com/chainer/chainercv

    5. \n

  5. External Data and Pre-Trained Model Disclosure Thread (some kind of recurrent relation here ;) )
    \nhttps://www.kaggle.com/c/google-ai-open-images-visual-relationship-track/discussion/62740

    6. \n
", "rawMarkdown": "May or may not use pretrained models and datasets mentioned in: \n\n1. TensorFlow Object Detection \n(Open Images dataset, COCO dataset, Oxford-IIIT Pets dataset, model zoo, etc.) \n[https://github.com/tensorflow/models/tree/master/research/object_detection](https://github.com/tensorflow/models/tree/master/research/object_detection)\n\n2. Darknet \n(ImageNet dataset, YOLO, etc.) \n[https://pjreddie.com/darknet](https://pjreddie.com/darknet)\n\n3. ImageAI \n[https://github.com/OlafenwaMoses/ImageAI](https://github.com/OlafenwaMoses/ImageAI)\n\n4. ChainerCV \n[https://github.com/chainer/chainercv](https://github.com/chainer/chainercv)\n\n5. External Data and Pre-Trained Model Disclosure Thread (some kind of recurrent relation here ;) ) \n[https://www.kaggle.com/c/google-ai-open-images-visual-relationship-track/discussion/62740](https://www.kaggle.com/c/google-ai-open-images-visual-relationship-track/discussion/62740)", "votes": null }, { "id": "375297", "postDate": "08/24/2018 22:35:41", "content": "

I am considering the models from the model zoo:\nhttps://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md\nhttps://keras.io/applications/\nhttps://pytorch.org/docs/stable/torchvision/models.html

", "rawMarkdown": "I am considering the models from the model zoo:\nhttps://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md\nhttps://keras.io/applications/\nhttps://pytorch.org/docs/stable/torchvision/models.html", "votes": null }, { "id": "375446", "postDate": "08/25/2018 07:57:32", "content": "

Visual Relationship Detection with Deep Structural Ranking

\n\n

https://github.com/GriffinLiang/vrd-dsr

", "rawMarkdown": "Visual Relationship Detection with Deep Structural Ranking\n\nhttps://github.com/GriffinLiang/vrd-dsr", "votes": null }, { "id": "380736", "postDate": "09/03/2018 10:55:04", "content": "

I'm sorry to be late.\nI used pretrained ResNet in chainer(https://chainer.org/) and chainer-fpn(https://github.com/Hakuyume/chainer-fpn).

", "rawMarkdown": "I'm sorry to be late.\nI used pretrained ResNet in chainer(https://chainer.org/) and chainer-fpn(https://github.com/Hakuyume/chainer-fpn).", "votes": null }, { "id": "427382", "postDate": "11/25/2018 12:09:12", "content": "

@radek, Very useful info. Could you let me know how can I import \"4. pretrained models in PyTorch from this repository\" to Kaggle kernel?

", "rawMarkdown": "radek, Very useful info. Could you let me know how can I import \"4. pretrained models in PyTorch from this repository\" to Kaggle kernel?", "votes": null }, { "id": "624510", "postDate": "09/12/2019 06:39:35", "content": "

We'll use the pretrained faster rcnn in torchvison.models.detection, actually with only the resnet part of the convolution layers.

\n\n

No extra data used besides official dataset, though we are struggling training on this data size with limited resource

", "rawMarkdown": "We'll use the pretrained faster rcnn in torchvison.models.detection, actually with only the resnet part of the convolution layers.\n\nNo extra data used besides official dataset, though we are struggling training on this data size with limited resource", "votes": null } ], "comments": [ { "id": 367633, "author_name": "radek1", "author_url": "", "post_date": "08/08/2018 06:57:28", "content": "

I am not planning on using external data (that is apart from the data from object detection track but not sure this would qualify as external) . I am still in the exploratory phase and not sure if I will have time to work more on this, but if I do, I will most likely use a pretrained model (or a couple of models) from either of the following sources:

\n\n
    \n
  1. Tensorflow detection model zoo / Tensorflow pretrained models
    2. \n
  2. Pretrained models that come with the fastai library (resnet18, resnet34, resnet50, resnet101, resnet152, vgg16, \n vgg19, resnetxt50, resnext101, resnext101_64, wrn, inceptionresnet_2, inception_4, dn121, dn161, dn169, dn201).
    3. \n
  3. Any of the pretrained models from here and here, in particular the darknet53.conv.74
    4. \n
  4. pretrained models in PyTorch from this repository
    5. \n
\n\n

Not sure if more information would be required - please let me know if that would be the case.

", "votes": null, "replies": [ { "id": 427382, "author_name": "lftuwujie", "author_url": "", "post_date": "11/25/2018 12:09:12", "content": "

@radek, Very useful info. Could you let me know how can I import \"4. pretrained models in PyTorch from this repository\" to Kaggle kernel?

", "votes": null, "replies": [] } ] }, { "id": 371701, "author_name": "mukeshmithrakumar", "author_url": "", "post_date": "08/17/2018 13:22:29", "content": "

Keras resnet pretrained on imagenet and tf faster_rcnn_inception_resnet_v2_atrous_oid weights if I get that to work ;)

", "votes": null, "replies": [] }, { "id": 374257, "author_name": "muhammedazamkhan", "author_url": "", "post_date": "08/22/2018 17:22:23", "content": "

I will perhaps use one or more pre-trained model(s) from either of the following sources:

\n\n

Pre-trained model on Open Image dataset for detection/fine_tune_checkpoint: http://download.tensorflow.org/models/object_detection/faster_rcnn_inception_resnet_v2_atrous_oid_2018_01_28.tar.gz

\n\n

YOLO pre-trained model: https://pjreddie.com/darknet/yolo/

\n\n

Pre-trained model from ImageAI:

\n\n

https://github.com/OlafenwaMoses/ImageAI/releases/download/1.0/resnet50_coco_best_v2.0.1.h5

\n\n

https://github.com/OlafenwaMoses/ImageAI/releases/download/1.0/yolo.h5

\n\n

https://github.com/OlafenwaMoses/ImageAI/releases/download/1.0/tiny-yolo.h5

\n\n

Pre-trained models from https://pjreddie.com/darknet/imagenet/.

", "votes": null, "replies": [] }, { "id": 374345, "author_name": "ardeshp2", "author_url": "", "post_date": "08/22/2018 21:08:24", "content": "

Pre-trained resnet_50 and vgg16 models from gluon model zoo. Additional training on materials in context dataset. Object detectors trained on data from object detection track.

", "votes": null, "replies": [] }, { "id": 374415, "author_name": "xiaochus", "author_url": "", "post_date": "08/23/2018 02:06:47", "content": "

The pretrained model we used is as follows:

\n\n
    \n
  1. Keras Xception pretrained model on imagenet.
    2. \n
  2. Pytorch ResNet 50, Densenet 121 pretrained model on imagenet.
    3. \n
  3. RetianNet pretrained model on coco.
    4. \n
  4. Faster RCNN pretrained model on oid.
    5. \n
", "votes": null, "replies": [] }, { "id": 374433, "author_name": "zhangjixyz", "author_url": "", "post_date": "08/23/2018 03:17:11", "content": "

The pretrained models we used are:

\n\n
    \n
  1. ResNeXt-101-64x4d model pretrained on COCO from the pytorch version of mask-RCNN: https://github.com/roytseng-tw/Detectron.pytorch
    2. \n
  2. The same model is finetuned on the OpenImage object detection dataset.
    3. \n
  3. We may or may not use Visual Genome to pre-train our relationship detector. It is still under consideration.
    4. \n
", "votes": null, "replies": [] }, { "id": 374635, "author_name": "kylelee", "author_url": "", "post_date": "08/23/2018 12:50:30", "content": "

Tensorflow model zoo: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md Pytorch model zoo: https://pytorch.org/docs/stable/torchvision/models.html Keras model zoo: https://keras.io/applications/

", "votes": null, "replies": [] }, { "id": 374670, "author_name": "nandanam", "author_url": "", "post_date": "08/23/2018 14:42:12", "content": "

Mask_RCNN

", "votes": null, "replies": [] }, { "id": 374706, "author_name": "its7171", "author_url": "", "post_date": "08/23/2018 16:05:57", "content": "

Pretrained models:

\n\n

yolov3: https://pjreddie.com/darknet/yolo/

\n\n

Pretrained models from keras: https://keras.io/applications/

\n\n

Data:

\n\n

Open Images Dataset V4: https://storage.googleapis.com/openimages/web/download.html

\n\n

Google AI Open Images - Object Detection Track: https://www.kaggle.com/c/google-ai-open-images-object-detection-track

", "votes": null, "replies": [] }, { "id": 374746, "author_name": "particlebbq", "author_url": "", "post_date": "08/23/2018 17:57:17", "content": "

If I manage to get my solution working in time for the contest deadline, it will be based on Detectron (https://github.com/facebookresearch/Detectron) which initializes its training runs with the pre-trained classification models listed in the .yaml files in the config directory (https://github.com/facebookresearch/Detectron/tree/master/configs), e.g. https://s3-us-west-2.amazonaws.com/detectron/ImageNetPretrained/47261647/R-50-GN.pkl used by https://github.com/facebookresearch/Detectron/blob/master/configs/04_2018_gn_baselines/e2e_mask_rcnn_R-50-FPN_2x_gn.yaml

", "votes": null, "replies": [] }, { "id": 374770, "author_name": "reachkishore", "author_url": "", "post_date": "08/23/2018 19:03:50", "content": "

I may try \nTensorflow model zoo: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md Pytorch model zoo: https://pytorch.org/docs/stable/torchvision/models.html Keras model zoo: https://keras.io/applications/\nand vgg16 based pretrianed model here

", "votes": null, "replies": [] }, { "id": 374774, "author_name": "cttsai", "author_url": "", "post_date": "08/23/2018 19:15:19", "content": "

Keras model zoo: https://keras.io/applications/

", "votes": null, "replies": [] }, { "id": 374866, "author_name": "tonyyy", "author_url": "", "post_date": "08/24/2018 03:03:19", "content": "

I used the pre-trained model in Tensorflow detection model zoo for fine tuning:\nhttp://download.tensorflow.org/models/object_detection/faster_rcnn_inception_resnet_v2_atrous_oid_2018_01_28.tar.gz

", "votes": null, "replies": [] }, { "id": 374944, "author_name": "vecxoz", "author_url": "", "post_date": "08/24/2018 08:35:16", "content": "

May or may not use pretrained models and datasets mentioned in:

\n\n
    \n

  1. TensorFlow Object Detection
    \n(Open Images dataset, COCO dataset, Oxford-IIIT Pets dataset, model zoo, etc.)
    \nhttps://github.com/tensorflow/models/tree/master/research/object_detection

    2. \n

  2. Darknet
    \n(ImageNet dataset, YOLO, etc.)
    \nhttps://pjreddie.com/darknet

    3. \n

  3. ImageAI
    \nhttps://github.com/OlafenwaMoses/ImageAI

    4. \n

  4. ChainerCV
    \nhttps://github.com/chainer/chainercv

    5. \n

  5. External Data and Pre-Trained Model Disclosure Thread (some kind of recurrent relation here ;) )
    \nhttps://www.kaggle.com/c/google-ai-open-images-visual-relationship-track/discussion/62740

    6. \n
", "votes": null, "replies": [] }, { "id": 375297, "author_name": "novxin", "author_url": "", "post_date": "08/24/2018 22:35:41", "content": "

I am considering the models from the model zoo:\nhttps://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md\nhttps://keras.io/applications/\nhttps://pytorch.org/docs/stable/torchvision/models.html

", "votes": null, "replies": [] }, { "id": 375446, "author_name": "alexanderliao", "author_url": "", "post_date": "08/25/2018 07:57:32", "content": "

Visual Relationship Detection with Deep Structural Ranking

\n\n

https://github.com/GriffinLiang/vrd-dsr

", "votes": null, "replies": [] }, { "id": 380736, "author_name": "keitaotani", "author_url": "", "post_date": "09/03/2018 10:55:04", "content": "

I'm sorry to be late.\nI used pretrained ResNet in chainer(https://chainer.org/) and chainer-fpn(https://github.com/Hakuyume/chainer-fpn).

", "votes": null, "replies": [] }, { "id": 624510, "author_name": "raynardj", "author_url": "", "post_date": "09/12/2019 06:39:35", "content": "

We'll use the pretrained faster rcnn in torchvison.models.detection, actually with only the resnet part of the convolution layers.

\n\n

No extra data used besides official dataset, though we are struggling training on this data size with limited resource

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "366839": "Please use this thread to post any external data and pre-trained models you use for your solution. Reminder, disclosure is required one week prior to the submission deadline.", "367633": "I am not planning on using external data (that is apart from the data from object detection track but not sure this would qualify as external) . I am still in the exploratory phase and not sure if I will have time to work more on this, but if I do, I will most likely use a pretrained model (or a couple of models) from either of the following sources:\n\n 1. [Tensorflow detection model zoo][1] / [Tensorflow pretrained models][2]\n 2. Pretrained models that come with the [fastai][3] library (resnet18, resnet34, resnet50, resnet101, resnet152, vgg16, \n vgg19, resnetxt50, resnext101, resnext101_64, wrn, inceptionresnet_2, inception_4, dn121, dn161, dn169, dn201).\n 3. Any of the pretrained models from [here][4] and [here][5], in particular the [darknet53.conv.74][6]\n 4. pretrained models in PyTorch from this [repository][7]\n\nNot sure if more information would be required - please let me know if that would be the case.\n\n\n [1]: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md\n [2]: https://github.com/tensorflow/models/tree/master/research/slim\n [3]: https://github.com/fastai/fastai\n [4]: https://pjreddie.com/darknet/imagenet/\n [5]: https://pjreddie.com/darknet/yolo/\n [6]: https://pjreddie.com/media/files/darknet53.conv.74\n [7]: https://github.com/Cadene/pretrained-models.pytorch", "371701": "Keras resnet pretrained on imagenet and tf faster_rcnn_inception_resnet_v2_atrous_oid weights if I get that to work ;)", "374257": "I will perhaps use one or more pre-trained model(s) from either of the following sources:\n\nPre-trained model on Open Image dataset for detection/fine_tune_checkpoint: http://download.tensorflow.org/models/object_detection/faster_rcnn_inception_resnet_v2_atrous_oid_2018_01_28.tar.gz\n\nYOLO pre-trained model: https://pjreddie.com/darknet/yolo/\n\nPre-trained model from ImageAI:\n\nhttps://github.com/OlafenwaMoses/ImageAI/releases/download/1.0/resnet50_coco_best_v2.0.1.h5\n\nhttps://github.com/OlafenwaMoses/ImageAI/releases/download/1.0/yolo.h5\n\nhttps://github.com/OlafenwaMoses/ImageAI/releases/download/1.0/tiny-yolo.h5\n\nPre-trained models from https://pjreddie.com/darknet/imagenet/.", "374345": "Pre-trained resnet_50 and vgg16 models from [gluon model zoo][1]. Additional training on [materials in context][2] dataset. Object detectors trained on data from object detection track.\n\n\n [1]: http://mxnet.apache.org/api/python/gluon/model_zoo.html\n [2]: http://opensurfaces.cs.cornell.edu/publications/minc/", "374415": "The pretrained model we used is as follows:\n\n 1. Keras [Xception][1] pretrained model on imagenet.\n 2. Pytorch [ResNet 50][2], [Densenet 121][3] pretrained model on imagenet.\n 3. [RetianNet][4] pretrained model on coco.\n 4. [Faster RCNN][5] pretrained model on oid.\n\n\n [1]: https://github.com/fchollet/deep-learning-models/\n [2]: https://download.pytorch.org/models/resnet50-19c8e357.pth\n [3]: https://download.pytorch.org/models/densenet121-241335ed.pth\n [4]: https://github.com/yhenon/pytorch-retinanet\n [5]: https://github.com/tensorflow/models/blob/505f554c6417931c96b59516f14d1ad65df6dbc5/research/object_detection/g3doc/detection_model_zoo.md", "374433": "The pretrained models we used are:\n\n1. ResNeXt-101-64x4d model pretrained on COCO from the pytorch version of mask-RCNN: https://github.com/roytseng-tw/Detectron.pytorch\n2. The same model is finetuned on the OpenImage object detection dataset.\n3. We may or may not use [Visual Genome][1] to pre-train our relationship detector. It is still under consideration.\n\n [1]: https://cs.stanford.edu/~danfei/scene-graph/", "374635": "Tensorflow model zoo: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md Pytorch model zoo: https://pytorch.org/docs/stable/torchvision/models.html Keras model zoo: https://keras.io/applications/", "374670": "[Mask_RCNN][1] \n\n\n [1]: https://github.com/matterport/Mask_RCNN", "374706": "Pretrained models:\n\nyolov3: https://pjreddie.com/darknet/yolo/\n\nPretrained models from keras: https://keras.io/applications/\n\n\n\nData:\n\nOpen Images Dataset V4: https://storage.googleapis.com/openimages/web/download.html\n\nGoogle AI Open Images - Object Detection Track: https://www.kaggle.com/c/google-ai-open-images-object-detection-track", "374746": "If I manage to get my solution working in time for the contest deadline, it will be based on Detectron (https://github.com/facebookresearch/Detectron) which initializes its training runs with the pre-trained classification models listed in the .yaml files in the config directory (https://github.com/facebookresearch/Detectron/tree/master/configs), e.g. https://s3-us-west-2.amazonaws.com/detectron/ImageNetPretrained/47261647/R-50-GN.pkl used by https://github.com/facebookresearch/Detectron/blob/master/configs/04_2018_gn_baselines/e2e_mask_rcnn_R-50-FPN_2x_gn.yaml", "374770": "I may try \nTensorflow model zoo: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md Pytorch model zoo: https://pytorch.org/docs/stable/torchvision/models.html Keras model zoo: https://keras.io/applications/\nand vgg16 based pretrianed model [here][1]\n [1]: http://cs.stanford.edu/people/jcjohns/densecap/densecap-pretrained-vgg16.t7.zip", "374774": "Keras model zoo: https://keras.io/applications/", "374866": "I used the pre-trained model in Tensorflow detection model zoo for fine tuning:\nhttp://download.tensorflow.org/models/object_detection/faster_rcnn_inception_resnet_v2_atrous_oid_2018_01_28.tar.gz", "374944": "May or may not use pretrained models and datasets mentioned in: \n\n1. TensorFlow Object Detection \n(Open Images dataset, COCO dataset, Oxford-IIIT Pets dataset, model zoo, etc.) \n[https://github.com/tensorflow/models/tree/master/research/object_detection](https://github.com/tensorflow/models/tree/master/research/object_detection)\n\n2. Darknet \n(ImageNet dataset, YOLO, etc.) \n[https://pjreddie.com/darknet](https://pjreddie.com/darknet)\n\n3. ImageAI \n[https://github.com/OlafenwaMoses/ImageAI](https://github.com/OlafenwaMoses/ImageAI)\n\n4. ChainerCV \n[https://github.com/chainer/chainercv](https://github.com/chainer/chainercv)\n\n5. External Data and Pre-Trained Model Disclosure Thread (some kind of recurrent relation here ;) ) \n[https://www.kaggle.com/c/google-ai-open-images-visual-relationship-track/discussion/62740](https://www.kaggle.com/c/google-ai-open-images-visual-relationship-track/discussion/62740)", "375297": "I am considering the models from the model zoo:\nhttps://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md\nhttps://keras.io/applications/\nhttps://pytorch.org/docs/stable/torchvision/models.html", "375446": "Visual Relationship Detection with Deep Structural Ranking\n\nhttps://github.com/GriffinLiang/vrd-dsr", "380736": "I'm sorry to be late.\nI used pretrained ResNet in chainer(https://chainer.org/) and chainer-fpn(https://github.com/Hakuyume/chainer-fpn).", "427382": "radek, Very useful info. Could you let me know how can I import \"4. pretrained models in PyTorch from this repository\" to Kaggle kernel?", "624510": "We'll use the pretrained faster rcnn in torchvison.models.detection, actually with only the resnet part of the convolution layers.\n\nNo extra data used besides official dataset, though we are struggling training on this data size with limited resource" }, "source": "meta" }