https://www.kaggle.com/tanmaymaloo/final-bengali\nIn the notebook above I used the 3 different learned models for prediction of each feature of the image. But I think because I am converting all parquet files at once into the image and then using it for prediction is too memory consuming...\nTherefore, I used another method (Kind of jugaad) which is also clearly not working.\nhttps://www.kaggle.com/tanmaymaloo/fork-of-final-bengali-fe31e3\nIn[24]: which is commented right now is giving the correct result and also memory efficient but taking more the 2 hrs which I think because I am converting each array in parquet file into an image -> using open_image -> prediction.\nNow, at last, I think to eliminate the step of saving each image and using open_image instead i used.\n*img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB);\n img_fastai = Image(pil2tensor(img, dtype=np.uint8))*\nAfter all preprocessing to convert np.array format of image into fastai.Image format (On visualizing it is showing the exact same image as we got from *open_image*) but the prediction from these img is bizzare.\nAny help is appreciated
", "messages": [ { "id": "771732", "postDate": "03/14/2020 14:54:40", "content": "https://www.kaggle.com/tanmaymaloo/final-bengali\nIn the notebook above I used the 3 different learned models for prediction of each feature of the image. But I think because I am converting all parquet files at once into the image and then using it for prediction is too memory consuming...\nTherefore, I used another method (Kind of jugaad) which is also clearly not working.\nhttps://www.kaggle.com/tanmaymaloo/fork-of-final-bengali-fe31e3\nIn[24]: which is commented right now is giving the correct result and also memory efficient but taking more the 2 hrs which I think because I am converting each array in parquet file into an image -> using open_image -> prediction.\nNow, at last, I think to eliminate the step of saving each image and using open_image instead i used.\n*img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB);\n img_fastai = Image(pil2tensor(img, dtype=np.uint8))*\nAfter all preprocessing to convert np.array format of image into fastai.Image format (On visualizing it is showing the exact same image as we got from *open_image*) but the prediction from these img is bizzare.\nAny help is appreciated
", "rawMarkdown": "*https://www.kaggle.com/tanmaymaloo/final-bengali*\nIn the notebook above I used the 3 different learned models for prediction of each feature of the image. But I think because I am converting all parquet files at once into the image and then using it for prediction is too memory consuming...\nTherefore, I used another method (Kind of jugaad) which is also clearly not working.\n*https://www.kaggle.com/tanmaymaloo/fork-of-final-bengali-fe31e3*\nIn[24]: which is commented right now is giving the correct result and also memory efficient but taking more the 2 hrs which I think because I am converting each array in parquet file into an image -> using open_image -> prediction.\nNow, at last, I think to eliminate the step of saving each image and using open_image instead i used.\n*img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB);\n img_fastai = Image(pil2tensor(img, dtype=np.uint8))*\nAfter all preprocessing to convert np.array format of image into fastai.Image format (On visualizing it is showing the exact same image as we got from *open_image*) but the prediction from these *img* is bizzare.\n**Any help is appreciated**", "votes": null }, { "id": "771825", "postDate": "03/14/2020 17:11:41", "content": "Use the below way with three different models
\n\nhttps://www.kaggle.com/ianmoone0617/grapheme?scriptVersionId=29759128
\n\nor if you still face difficulty \nshare me your inference kernel
", "rawMarkdown": "Use the below way with three different models\n\nhttps://www.kaggle.com/ianmoone0617/grapheme?scriptVersionId=29759128\n\nor if you still face difficulty \nshare me your inference kernel", "votes": null }, { "id": "772521", "postDate": "03/15/2020 15:32:02", "content": "@ianmoone0617 it's working now man... thanks for replying.
", "rawMarkdown": "ianmoone0617 it's working now man... thanks for replying.", "votes": null } ], "comments": [ { "id": 771825, "author_name": "", "author_url": "", "post_date": "03/14/2020 17:11:41", "content": "Use the below way with three different models
\n\nhttps://www.kaggle.com/ianmoone0617/grapheme?scriptVersionId=29759128
\n\nor if you still face difficulty \nshare me your inference kernel
", "votes": null, "replies": [] }, { "id": 772521, "author_name": "tanmaymaloo", "author_url": "", "post_date": "03/15/2020 15:32:02", "content": "@ianmoone0617 it's working now man... thanks for replying.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "771732": "*https://www.kaggle.com/tanmaymaloo/final-bengali*\nIn the notebook above I used the 3 different learned models for prediction of each feature of the image. But I think because I am converting all parquet files at once into the image and then using it for prediction is too memory consuming...\nTherefore, I used another method (Kind of jugaad) which is also clearly not working.\n*https://www.kaggle.com/tanmaymaloo/fork-of-final-bengali-fe31e3*\nIn[24]: which is commented right now is giving the correct result and also memory efficient but taking more the 2 hrs which I think because I am converting each array in parquet file into an image -> using open_image -> prediction.\nNow, at last, I think to eliminate the step of saving each image and using open_image instead i used.\n*img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB);\n img_fastai = Image(pil2tensor(img, dtype=np.uint8))*\nAfter all preprocessing to convert np.array format of image into fastai.Image format (On visualizing it is showing the exact same image as we got from *open_image*) but the prediction from these *img* is bizzare.\n**Any help is appreciated**", "771825": "Use the below way with three different models\n\nhttps://www.kaggle.com/ianmoone0617/grapheme?scriptVersionId=29759128\n\nor if you still face difficulty \nshare me your inference kernel", "772521": "ianmoone0617 it's working now man... thanks for replying." }, "source": "meta" }