Hi everyone! I cropped the images into 224*224 blocks randomly, then I used a pretrained ResNet 50 model in keras and fine-tuned the fc layers. I got both ~90% accuracy in training data and validation data. However, what puzzles me is that, I can only get about 0.40 LB scocre. I'm sure that there is no class info leakage when training my model, and the training data are corretly splitted. So I'm really confused about the situation. Could you give me some suggestions?
", "messages": [ { "id": "275944", "postDate": "01/30/2018 09:42:44", "content": "Hi everyone! I cropped the images into 224*224 blocks randomly, then I used a pretrained ResNet 50 model in keras and fine-tuned the fc layers. I got both ~90% accuracy in training data and validation data. However, what puzzles me is that, I can only get about 0.40 LB scocre. I'm sure that there is no class info leakage when training my model, and the training data are corretly splitted. So I'm really confused about the situation. Could you give me some suggestions?
", "rawMarkdown": "Hi everyone! I cropped the images into 224*224 blocks randomly, then I used a pretrained ResNet 50 model in keras and fine-tuned the fc layers. I got both ~90% accuracy in training data and validation data. However, what puzzles me is that, I can only get about 0.40 LB scocre. I'm sure that there is no class info leakage when training my model, and the training data are corretly splitted. So I'm really confused about the situation. Could you give me some suggestions?", "votes": null }, { "id": "275952", "postDate": "01/30/2018 10:13:27", "content": "I guess that the training and validation datasets are cropped from training fold. They have similar contents. But the testing dataset is totally different, whose content is not similar.\nYour model learned more about content than camera model.
", "rawMarkdown": "I guess that the training and validation datasets are cropped from training fold. They have similar contents. But the testing dataset is totally different, whose content is not similar.\nYour model learned more about content than camera model.", "votes": null }, { "id": "275974", "postDate": "01/30/2018 11:16:09", "content": "Augment your training and validation data.
", "rawMarkdown": "Augment your training and validation data.", "votes": null }, { "id": "275980", "postDate": "01/30/2018 11:37:06", "content": "Thanks young. How did augment your data? By using ImageDataGenerator in Keras? If so, what's your parameter setting?
", "rawMarkdown": "Thanks young. How did augment your data? By using ImageDataGenerator in Keras? If so, what's your parameter setting?", "votes": null }, { "id": "276011", "postDate": "01/30/2018 13:29:44", "content": "Using the eight possible processing operations that were performed are as follows:\nJPEG compression with quality factor = 70\nJPEG compression with quality factor = 90\nresizing (via bicubic interpolation) by a factor of 0.5\nresizing (via bicubic interpolation) by a factor of 0.8\nresizing (via bicubic interpolation) by a factor of 1.5\nresizing (via bicubic interpolation) by a factor of 2.0\ngamma correction using gamma = 0.8\ngamma correction using gamma = 1.2
", "rawMarkdown": "Using the eight possible processing operations that were performed are as follows:\nJPEG compression with quality factor = 70\nJPEG compression with quality factor = 90\nresizing (via bicubic interpolation) by a factor of 0.5\nresizing (via bicubic interpolation) by a factor of 0.8\nresizing (via bicubic interpolation) by a factor of 1.5\nresizing (via bicubic interpolation) by a factor of 2.0\ngamma correction using gamma = 0.8\ngamma correction using gamma = 1.2", "votes": null }, { "id": "276119", "postDate": "01/30/2018 18:34:44", "content": "Hi Young, is your single best model using 224x224 crop or 512x512?
", "rawMarkdown": "Hi Young, is your single best model using 224x224 crop or 512x512?", "votes": null }, { "id": "276278", "postDate": "01/31/2018 04:47:20", "content": "I don't try to use 512 crop.
", "rawMarkdown": "I don't try to use 512 crop.", "votes": null } ], "comments": [ { "id": 275952, "author_name": "yaozhenjie", "author_url": "", "post_date": "01/30/2018 10:13:27", "content": "I guess that the training and validation datasets are cropped from training fold. They have similar contents. But the testing dataset is totally different, whose content is not similar.\nYour model learned more about content than camera model.
", "votes": null, "replies": [] }, { "id": 275974, "author_name": "youngkl", "author_url": "", "post_date": "01/30/2018 11:16:09", "content": "Augment your training and validation data.
", "votes": null, "replies": [ { "id": 275980, "author_name": "hzywish", "author_url": "", "post_date": "01/30/2018 11:37:06", "content": "Thanks young. How did augment your data? By using ImageDataGenerator in Keras? If so, what's your parameter setting?
", "votes": null, "replies": [] }, { "id": 276011, "author_name": "youngkl", "author_url": "", "post_date": "01/30/2018 13:29:44", "content": "Using the eight possible processing operations that were performed are as follows:\nJPEG compression with quality factor = 70\nJPEG compression with quality factor = 90\nresizing (via bicubic interpolation) by a factor of 0.5\nresizing (via bicubic interpolation) by a factor of 0.8\nresizing (via bicubic interpolation) by a factor of 1.5\nresizing (via bicubic interpolation) by a factor of 2.0\ngamma correction using gamma = 0.8\ngamma correction using gamma = 1.2
", "votes": null, "replies": [] }, { "id": 276119, "author_name": "shunjiading", "author_url": "", "post_date": "01/30/2018 18:34:44", "content": "Hi Young, is your single best model using 224x224 crop or 512x512?
", "votes": null, "replies": [] }, { "id": 276278, "author_name": "youngkl", "author_url": "", "post_date": "01/31/2018 04:47:20", "content": "I don't try to use 512 crop.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "275944": "Hi everyone! I cropped the images into 224*224 blocks randomly, then I used a pretrained ResNet 50 model in keras and fine-tuned the fc layers. I got both ~90% accuracy in training data and validation data. However, what puzzles me is that, I can only get about 0.40 LB scocre. I'm sure that there is no class info leakage when training my model, and the training data are corretly splitted. So I'm really confused about the situation. Could you give me some suggestions?", "275952": "I guess that the training and validation datasets are cropped from training fold. They have similar contents. But the testing dataset is totally different, whose content is not similar.\nYour model learned more about content than camera model.", "275974": "Augment your training and validation data.", "275980": "Thanks young. How did augment your data? By using ImageDataGenerator in Keras? If so, what's your parameter setting?", "276011": "Using the eight possible processing operations that were performed are as follows:\nJPEG compression with quality factor = 70\nJPEG compression with quality factor = 90\nresizing (via bicubic interpolation) by a factor of 0.5\nresizing (via bicubic interpolation) by a factor of 0.8\nresizing (via bicubic interpolation) by a factor of 1.5\nresizing (via bicubic interpolation) by a factor of 2.0\ngamma correction using gamma = 0.8\ngamma correction using gamma = 1.2", "276119": "Hi Young, is your single best model using 224x224 crop or 512x512?", "276278": "I don't try to use 512 crop." }, "source": "meta" }