I consider myself lucky since I only had one chance, to train only one model. I have a GTX1070 and each epoch took me about 2.5 days to train.
\n\nI used the pretrained InceptionResNetV2 from Keras, I trained it for 10 epochs using: SGD, lr=0.02, momentum=0.9. Then for one last epoch I decreased, while training, the learning rate from 0.02 to 0.002.
\n\nFinal private LB score: 0.75328
\n\nWhat's your best single model score?
", "messages": [ { "id": "257881", "postDate": "12/15/2017 02:50:15", "content": "I consider myself lucky since I only had one chance, to train only one model. I have a GTX1070 and each epoch took me about 2.5 days to train.
\n\nI used the pretrained InceptionResNetV2 from Keras, I trained it for 10 epochs using: SGD, lr=0.02, momentum=0.9. Then for one last epoch I decreased, while training, the learning rate from 0.02 to 0.002.
\n\nFinal private LB score: 0.75328
\n\nWhat's your best single model score?
", "rawMarkdown": "I consider myself lucky since I only had one chance, to train only one model. I have a GTX1070 and each epoch took me about 2.5 days to train.\n\nI used the pretrained InceptionResNetV2 from Keras, I trained it for 10 epochs using: SGD, lr=0.02, momentum=0.9. Then for one last epoch I decreased, while training, the learning rate from 0.02 to 0.002.\n\nFinal private LB score: 0.75328\n\nWhat's your best single model score?", "votes": null }, { "id": "257890", "postDate": "12/15/2017 03:02:16", "content": "My InceptionResNetV2 only achieved 0.735. I resized input to 224x224, set initial learning rate to 0.01 with a decay factor of 0.8. I trained 17 epoches with horizontal flip and random shift of 10%. May I know your augmentation settings and input sizes?
", "rawMarkdown": "My InceptionResNetV2 only achieved 0.735. I resized input to 224x224, set initial learning rate to 0.01 with a decay factor of 0.8. I trained 17 epoches with horizontal flip and random shift of 10%. May I know your augmentation settings and input sizes?", "votes": null }, { "id": "257891", "postDate": "12/15/2017 03:04:19", "content": "No augmentation, 180x180
", "rawMarkdown": "No augmentation, 180x180", "votes": null }, { "id": "257893", "postDate": "12/15/2017 03:06:32", "content": "Without augmentation I got slightly worse validation result. Maybe I need to recheck my code...
", "rawMarkdown": "Without augmentation I got slightly worse validation result. Maybe I need to recheck my code...", "votes": null }, { "id": "257972", "postDate": "12/15/2017 07:34:25", "content": "My InceptionResNetV2 only achieved about 0.68. I train the model from scratch for 55epochs. No augmentation, 196*196.
", "rawMarkdown": "My InceptionResNetV2 only achieved about 0.68. I train the model from scratch for 55epochs. No augmentation, 196*196.", "votes": null }, { "id": "257986", "postDate": "12/15/2017 08:00:16", "content": "Wow, that is great effort. I use inceptionresnetv2 too, but I just get 72.5 on LB because I started too late and didnt have time to train more epochs.
", "rawMarkdown": "Wow, that is great effort. I use inceptionresnetv2 too, but I just get 72.5 on LB because I started too late and didnt have time to train more epochs.", "votes": null }, { "id": "258003", "postDate": "12/15/2017 08:24:06", "content": "Very impressive performance with a 1070, vs the heavy lifters in 4x 1080Ti, congratulations.
", "rawMarkdown": "Very impressive performance with a 1070, vs the heavy lifters in 4x 1080Ti, congratulations.", "votes": null }, { "id": "258245", "postDate": "12/15/2017 19:57:27", "content": "Very impressive! I wonder how far could one go with such model trained longer and on more powerful hardware.
", "rawMarkdown": "Very impressive! I wonder how far could one go with such model trained longer and on more powerful hardware.", "votes": null }, { "id": "258522", "postDate": "12/16/2017 10:12:01", "content": "Very impressive indeed, I was just wondering if you could tell us whetehr you trained from scratch all the layers or some blocks ..?
", "rawMarkdown": "Very impressive indeed, I was just wondering if you could tell us whetehr you trained from scratch all the layers or some blocks ..?", "votes": null }, { "id": "258566", "postDate": "12/16/2017 12:56:06", "content": "I use keras imagenet pretrained InceptionResNetV2
\n\nTrain:
\n\nTest:
\n\nI was able to get 0.776 on private using combination of softmax and KNN on top of finetuned SE-ResNet-101. See https://www.kaggle.com/c/cdiscount-image-classification-challenge/discussion/45850
", "rawMarkdown": "I was able to get 0.776 on private using combination of softmax and KNN on top of finetuned SE-ResNet-101. See https://www.kaggle.com/c/cdiscount-image-classification-challenge/discussion/45850", "votes": null } ], "comments": [ { "id": 257890, "author_name": "wowfattie", "author_url": "", "post_date": "12/15/2017 03:02:16", "content": "My InceptionResNetV2 only achieved 0.735. I resized input to 224x224, set initial learning rate to 0.01 with a decay factor of 0.8. I trained 17 epoches with horizontal flip and random shift of 10%. May I know your augmentation settings and input sizes?
", "votes": null, "replies": [ { "id": 257891, "author_name": "radustoicescu", "author_url": "", "post_date": "12/15/2017 03:04:19", "content": "No augmentation, 180x180
", "votes": null, "replies": [] }, { "id": 257893, "author_name": "wowfattie", "author_url": "", "post_date": "12/15/2017 03:06:32", "content": "Without augmentation I got slightly worse validation result. Maybe I need to recheck my code...
", "votes": null, "replies": [] }, { "id": 258522, "author_name": "", "author_url": "", "post_date": "12/16/2017 10:12:01", "content": "Very impressive indeed, I was just wondering if you could tell us whetehr you trained from scratch all the layers or some blocks ..?
", "votes": null, "replies": [] } ] }, { "id": 257972, "author_name": "youngkl", "author_url": "", "post_date": "12/15/2017 07:34:25", "content": "My InceptionResNetV2 only achieved about 0.68. I train the model from scratch for 55epochs. No augmentation, 196*196.
", "votes": null, "replies": [] }, { "id": 257986, "author_name": "sondaoduy", "author_url": "", "post_date": "12/15/2017 08:00:16", "content": "Wow, that is great effort. I use inceptionresnetv2 too, but I just get 72.5 on LB because I started too late and didnt have time to train more epochs.
", "votes": null, "replies": [] }, { "id": 258003, "author_name": "ericperbos", "author_url": "", "post_date": "12/15/2017 08:24:06", "content": "Very impressive performance with a 1070, vs the heavy lifters in 4x 1080Ti, congratulations.
", "votes": null, "replies": [] }, { "id": 258245, "author_name": "ceperaang", "author_url": "", "post_date": "12/15/2017 19:57:27", "content": "Very impressive! I wonder how far could one go with such model trained longer and on more powerful hardware.
", "votes": null, "replies": [] }, { "id": 258566, "author_name": "outrunner", "author_url": "", "post_date": "12/16/2017 12:56:06", "content": "I use keras imagenet pretrained InceptionResNetV2
\n\nTrain:
\n\nTest:
\n\nI was able to get 0.776 on private using combination of softmax and KNN on top of finetuned SE-ResNet-101. See https://www.kaggle.com/c/cdiscount-image-classification-challenge/discussion/45850
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "257881": "I consider myself lucky since I only had one chance, to train only one model. I have a GTX1070 and each epoch took me about 2.5 days to train.\n\nI used the pretrained InceptionResNetV2 from Keras, I trained it for 10 epochs using: SGD, lr=0.02, momentum=0.9. Then for one last epoch I decreased, while training, the learning rate from 0.02 to 0.002.\n\nFinal private LB score: 0.75328\n\nWhat's your best single model score?", "257890": "My InceptionResNetV2 only achieved 0.735. I resized input to 224x224, set initial learning rate to 0.01 with a decay factor of 0.8. I trained 17 epoches with horizontal flip and random shift of 10%. May I know your augmentation settings and input sizes?", "257891": "No augmentation, 180x180", "257893": "Without augmentation I got slightly worse validation result. Maybe I need to recheck my code...", "257972": "My InceptionResNetV2 only achieved about 0.68. I train the model from scratch for 55epochs. No augmentation, 196*196.", "257986": "Wow, that is great effort. I use inceptionresnetv2 too, but I just get 72.5 on LB because I started too late and didnt have time to train more epochs.", "258003": "Very impressive performance with a 1070, vs the heavy lifters in 4x 1080Ti, congratulations.", "258245": "Very impressive! I wonder how far could one go with such model trained longer and on more powerful hardware.", "258522": "Very impressive indeed, I was just wondering if you could tell us whetehr you trained from scratch all the layers or some blocks ..?", "258566": "I use keras imagenet pretrained InceptionResNetV2\n\nTrain:\n\n - 18 epochs, SGD without momentum, train the top layer 0.1 epoch, then train all.\n - lr=0.1 (10 epochs), 0.03 (5), 0.01 (2), 0.003 (1)\n - batch size: 70\n - random pick 1 image per product when training for efficiency (due to the duplicate image in data set)\n - random resize between 171 to 203 then random crop 171, horizontal flip, color adjustment.\n - about 1 day per epoch on GTX 1080 (7M product)\n\nTest:\n\n - 12x TTA(the same as training stage) per product\n - geometric mean on mean of image's probability (acc +0.002 vs. arithmetic mean)\n - Private LB : 0.76035", "258600": "I was able to get 0.776 on private using combination of softmax and KNN on top of finetuned SE-ResNet-101. See https://www.kaggle.com/c/cdiscount-image-classification-challenge/discussion/45850" }, "source": "meta" }