I am using InceptionV3. I've trained the modified layers and fine tuned the inner layers with 60% of the training set and 2 epochs only, due to execution time. The results are very pour for now [LB 0.27]. Did any one get better results?
", "messages": [ { "id": "223530", "postDate": "09/22/2017 13:48:41", "content": "I am using InceptionV3. I've trained the modified layers and fine tuned the inner layers with 60% of the training set and 2 epochs only, due to execution time. The results are very pour for now [LB 0.27]. Did any one get better results?
", "rawMarkdown": "I am using InceptionV3. I've trained the modified layers and fine tuned the inner layers with 60% of the training set and 2 epochs only, due to execution time. The results are very pour for now [LB 0.27]. Did any one get better results?", "votes": null }, { "id": "223534", "postDate": "09/22/2017 13:59:30", "content": "I am using similar network now, still no luck.
", "rawMarkdown": "I am using similar network now, still no luck.", "votes": null }, { "id": "223539", "postDate": "09/22/2017 14:11:50", "content": "Have you trained the network with all training data?
", "rawMarkdown": "Have you trained the network with all training data?", "votes": null }, { "id": "223572", "postDate": "09/22/2017 16:02:28", "content": "The dataset is highly unbalanced. What approach did you use to select 60% of training set?
", "rawMarkdown": "The dataset is highly unbalanced. What approach did you use to select 60% of training set?", "votes": null }, { "id": "223623", "postDate": "09/22/2017 18:57:59", "content": "No extra care with unbalance. Normal random sampling with the approach of bson random access avaliable in kernels.
", "rawMarkdown": "No extra care with unbalance. Normal random sampling with the approach of bson random access avaliable in kernels.", "votes": null }, { "id": "223625", "postDate": "09/22/2017 19:08:19", "content": "I tried using the output of VGG16's block5_pool layer, saving it to disk, and training a classifier on top of that.
\n\nThe good news is that the training is very fast, and gives okay results. 0.45 on LB. Struggling to improve much beyond that though; retraining the convolutional layers makes things much slower.
\n\nThis is when using 102400 out of 7069896 for training + validation, a 10% validation split, and only targeting the most common 2048 categories (upper bound of 94% accuracy).
\n\nLet me know if anyone wants to team up, I'm eager to get back into the competitive scores :)
", "rawMarkdown": "I tried using the output of VGG16's block5_pool layer, saving it to disk, and training a classifier on top of that.\n\nThe good news is that the training is very fast, and gives okay results. 0.45 on LB. Struggling to improve much beyond that though; retraining the convolutional layers makes things much slower.\n\nThis is when using 102400 out of 7069896 for training + validation, a 10% validation split, and only targeting the most common 2048 categories (upper bound of 94% accuracy).\n\nLet me know if anyone wants to team up, I'm eager to get back into the competitive scores :)", "votes": null } ], "comments": [ { "id": 223534, "author_name": "woodywang", "author_url": "", "post_date": "09/22/2017 13:59:30", "content": "I am using similar network now, still no luck.
", "votes": null, "replies": [ { "id": 223539, "author_name": "aloisiodn", "author_url": "", "post_date": "09/22/2017 14:11:50", "content": "Have you trained the network with all training data?
", "votes": null, "replies": [] } ] }, { "id": 223572, "author_name": "nabinn", "author_url": "", "post_date": "09/22/2017 16:02:28", "content": "The dataset is highly unbalanced. What approach did you use to select 60% of training set?
", "votes": null, "replies": [ { "id": 223623, "author_name": "aloisiodn", "author_url": "", "post_date": "09/22/2017 18:57:59", "content": "No extra care with unbalance. Normal random sampling with the approach of bson random access avaliable in kernels.
", "votes": null, "replies": [] } ] }, { "id": 223625, "author_name": "ndahlquist", "author_url": "", "post_date": "09/22/2017 19:08:19", "content": "I tried using the output of VGG16's block5_pool layer, saving it to disk, and training a classifier on top of that.
\n\nThe good news is that the training is very fast, and gives okay results. 0.45 on LB. Struggling to improve much beyond that though; retraining the convolutional layers makes things much slower.
\n\nThis is when using 102400 out of 7069896 for training + validation, a 10% validation split, and only targeting the most common 2048 categories (upper bound of 94% accuracy).
\n\nLet me know if anyone wants to team up, I'm eager to get back into the competitive scores :)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "223530": "I am using InceptionV3. I've trained the modified layers and fine tuned the inner layers with 60% of the training set and 2 epochs only, due to execution time. The results are very pour for now [LB 0.27]. Did any one get better results?", "223534": "I am using similar network now, still no luck.", "223539": "Have you trained the network with all training data?", "223572": "The dataset is highly unbalanced. What approach did you use to select 60% of training set?", "223623": "No extra care with unbalance. Normal random sampling with the approach of bson random access avaliable in kernels.", "223625": "I tried using the output of VGG16's block5_pool layer, saving it to disk, and training a classifier on top of that.\n\nThe good news is that the training is very fast, and gives okay results. 0.45 on LB. Struggling to improve much beyond that though; retraining the convolutional layers makes things much slower.\n\nThis is when using 102400 out of 7069896 for training + validation, a 10% validation split, and only targeting the most common 2048 categories (upper bound of 94% accuracy).\n\nLet me know if anyone wants to team up, I'm eager to get back into the competitive scores :)" }, "source": "meta" }