Hi All,
\nI am having a hard time getting my submissions to have a decent score. I essentially forked the code from another notebook and submitted my score, but while the original notebook scored >0.80, mine scored 0.8, but after making a submission it goes down significantly…
\nI thought maybe this just a bad fluke and re-trained it using 5-fold CV and averaged their predictions in my inference notebook, and still got very awful results. Would anyone be able to quickly glance at my notebook and give me a pointer in to what is going wrong here? I really appreciate any help.
\nMy prediction notebook:
\nhttps://www.kaggle.com/funky15/tpus-with-5-fold
My inference notebook:
\nhttps://www.kaggle.com/funky15/5-fold-tpu-inference-plus1
Hi All,
\nI am having a hard time getting my submissions to have a decent score. I essentially forked the code from another notebook and submitted my score, but while the original notebook scored >0.80, mine scored 0.8, but after making a submission it goes down significantly…
\nI thought maybe this just a bad fluke and re-trained it using 5-fold CV and averaged their predictions in my inference notebook, and still got very awful results. Would anyone be able to quickly glance at my notebook and give me a pointer in to what is going wrong here? I really appreciate any help.
\nMy prediction notebook:
\nhttps://www.kaggle.com/funky15/tpus-with-5-fold
My inference notebook:
\nhttps://www.kaggle.com/funky15/5-fold-tpu-inference-plus1
It could be because you are normalizing pixel values in training but not during inference.
\nIn your TPU inference notebook add code that normalizes pixels.
\nfor i in glob.glob(image_dir+\"/*.jpg\"):\n image = Image.open(i)\n image = image.resize((IMAGE_SIZE, IMAGE_SIZE))\n image = np.expand_dims(image, axis = 0)\n image = image/255 #ADD THIS LINE\n preds.append(np.argmax(model_obj.predict(image))+1)\n img_ids.append(i.replace(image_dir+\"/\", \"\"))\nLet me know if this works!
", "rawMarkdown": "It could be because you are normalizing pixel values in training but not during inference.\n\nIn your TPU inference notebook add code that normalizes pixels.\n\n```\nfor i in glob.glob(image_dir+\"/*.jpg\"):\n image = Image.open(i)\n image = image.resize((IMAGE_SIZE, IMAGE_SIZE))\n image = np.expand_dims(image, axis = 0)\n image = image/255 #ADD THIS LINE\n preds.append(np.argmax(model_obj.predict(image))+1)\n img_ids.append(i.replace(image_dir+\"/\", \"\"))\n```\n\nLet me know if this works!", "votes": null }, { "id": "1145038", "postDate": "01/08/2021 20:28:38", "content": "@brendanartley Thank you so much! That looks like it is a problem.
", "rawMarkdown": "brendanartley Thank you so much! That looks like it is a problem.", "votes": null } ], "comments": [ { "id": 1145023, "author_name": "brendanartley", "author_url": "", "post_date": "01/08/2021 20:14:16", "content": "It could be because you are normalizing pixel values in training but not during inference.
\nIn your TPU inference notebook add code that normalizes pixels.
\nfor i in glob.glob(image_dir+\"/*.jpg\"):\n image = Image.open(i)\n image = image.resize((IMAGE_SIZE, IMAGE_SIZE))\n image = np.expand_dims(image, axis = 0)\n image = image/255 #ADD THIS LINE\n preds.append(np.argmax(model_obj.predict(image))+1)\n img_ids.append(i.replace(image_dir+\"/\", \"\"))\nLet me know if this works!
", "votes": null, "replies": [] }, { "id": 1145038, "author_name": "funky15", "author_url": "", "post_date": "01/08/2021 20:28:38", "content": "@brendanartley Thank you so much! That looks like it is a problem.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1145007": "Hi All,\n\nI am having a hard time getting my submissions to have a decent score. I essentially forked the code from [another notebook](https://www.kaggle.com/wuliaokaola/getting-started-tpus-new-tfrecords) and submitted my score, but while the original notebook scored >0.80, mine scored <0.60. No idea why. My validation accuracy is always >0.8, but after making a submission it goes down significantly...\n\nI thought maybe this just a bad fluke and re-trained it using 5-fold CV and averaged their predictions in my inference notebook, and still got very awful results. **Would anyone be able to quickly glance at my notebook and give me a pointer in to what is going wrong here?** I really appreciate any help.\n\nMy prediction notebook:\n[https://www.kaggle.com/funky15/tpus-with-5-fold](https://www.kaggle.com/funky15/tpus-with-5-fold)\n\nMy inference notebook:\n[https://www.kaggle.com/funky15/5-fold-tpu-inference-plus1](https://www.kaggle.com/funky15/5-fold-tpu-inference-plus1)", "1145023": "It could be because you are normalizing pixel values in training but not during inference.\n\nIn your TPU inference notebook add code that normalizes pixels.\n\n```\nfor i in glob.glob(image_dir+\"/*.jpg\"):\n image = Image.open(i)\n image = image.resize((IMAGE_SIZE, IMAGE_SIZE))\n image = np.expand_dims(image, axis = 0)\n image = image/255 #ADD THIS LINE\n preds.append(np.argmax(model_obj.predict(image))+1)\n img_ids.append(i.replace(image_dir+\"/\", \"\"))\n```\n\nLet me know if this works!", "1145038": "brendanartley Thank you so much! That looks like it is a problem." }, "source": "meta" }