{
  "id": 132189,
  "title": "Does Increasing Image size helps?",
  "url": "/competitions/bengaliai-cv19/discussion/132189",
  "author_name": "",
  "post_date": "2020-02-24T19:27:22.802989600Z",
  "votes": null,
  "comment_count": 7,
  "views": 0,
  "content": "<p>Stuck at 0.93 with the below NN architecture. Wondering if anbody tried with greater image size and got better results. Currently my image size is (128,128,3).  Appreciate any other pointers that might help me to climb up the leaderboard</p>\n\n<p>```\nmodel.add(Conv2D(filters=32,kernel_size=(2,2),padding='same',activation='relu',input_shape=(size,size,3)))\nmodel.add(Conv2D(filters=32,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=32,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))</p>\n\n<p>model.add(Conv2D(filters=64,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=64,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=64,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))</p>\n\n<p>model.add(Conv2D(filters=128,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=128,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))</p>\n\n<p>model.add(Conv2D(filters=256,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=256,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))</p>\n\n<p>model.add(Conv2D(filters=512,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=512,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(5,5)))</p>\n\n<p>model.add(Flatten())</p>\n\n<p>model.add(Dense(units=512,activation='relu'))\nmodel.add(Dropout(0.3))\nmodel.add(Dense(units=256,activation='relu'))\nmodel.add(Dense(units=168,activation='softmax'))\n```</p>",
  "messages": [
    {
      "id": "755447",
      "postDate": "02/24/2020 19:27:22",
      "content": "<p>Stuck at 0.93 with the below NN architecture. Wondering if anbody tried with greater image size and got better results. Currently my image size is (128,128,3).  Appreciate any other pointers that might help me to climb up the leaderboard</p>\n\n<p>```\nmodel.add(Conv2D(filters=32,kernel_size=(2,2),padding='same',activation='relu',input_shape=(size,size,3)))\nmodel.add(Conv2D(filters=32,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=32,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))</p>\n\n<p>model.add(Conv2D(filters=64,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=64,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=64,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))</p>\n\n<p>model.add(Conv2D(filters=128,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=128,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))</p>\n\n<p>model.add(Conv2D(filters=256,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=256,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))</p>\n\n<p>model.add(Conv2D(filters=512,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=512,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(5,5)))</p>\n\n<p>model.add(Flatten())</p>\n\n<p>model.add(Dense(units=512,activation='relu'))\nmodel.add(Dropout(0.3))\nmodel.add(Dense(units=256,activation='relu'))\nmodel.add(Dense(units=168,activation='softmax'))\n```</p>",
      "rawMarkdown": "Stuck at 0.93 with the below NN architecture. Wondering if anbody tried with greater image size and got better results. Currently my image size is (128,128,3).  Appreciate any other pointers that might help me to climb up the leaderboard\n\n```\nmodel.add(Conv2D(filters=32,kernel_size=(2,2),padding='same',activation='relu',input_shape=(size,size,3)))\nmodel.add(Conv2D(filters=32,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=32,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))\n\nmodel.add(Conv2D(filters=64,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=64,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=64,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))\n\nmodel.add(Conv2D(filters=128,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=128,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))\n\nmodel.add(Conv2D(filters=256,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=256,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))\n\nmodel.add(Conv2D(filters=512,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=512,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(5,5)))\n\nmodel.add(Flatten())\n\nmodel.add(Dense(units=512,activation='relu'))\nmodel.add(Dropout(0.3))\nmodel.add(Dense(units=256,activation='relu'))\nmodel.add(Dense(units=168,activation='softmax'))\n```",
      "votes": null
    },
    {
      "id": "755455",
      "postDate": "02/24/2020 19:38:32",
      "content": "<p>128 is good enough, try pretrained models...</p>",
      "rawMarkdown": "128 is good enough, try pretrained models...",
      "votes": null
    },
    {
      "id": "755464",
      "postDate": "02/24/2020 19:49:56",
      "content": "<p>Thanks for the reply !! Do you advise to use pretrained weights?</p>",
      "rawMarkdown": "Thanks for the reply !! Do you advise to use pretrained weights?",
      "votes": null
    },
    {
      "id": "755468",
      "postDate": "02/24/2020 19:52:41",
      "content": "<p>Yes</p>",
      "rawMarkdown": "Yes",
      "votes": null
    },
    {
      "id": "755471",
      "postDate": "02/24/2020 19:55:12",
      "content": "<p>Thanks</p>",
      "rawMarkdown": "Thanks",
      "votes": null
    },
    {
      "id": "755478",
      "postDate": "02/24/2020 20:02:47",
      "content": "<p>128x128 is a good resolution, is it discussable if larger resolution lead to better results or not.\nLike <a href=\"/greatgamedota\">@greatgamedota</a>  mentioned, use pretrained models with pretrained weights, they do help\nAnother important aspects is to use serios data augmentation, I suggest to try affine transforms, cutoff, cutmix and mixup.\nAlso, be careful with the optimizer, scheduler and learning rate.\nFor more info you can check discussion from <a href=\"https://www.kaggle.com/c/bengaliai-cv19/discussion/130311\">https://www.kaggle.com/c/bengaliai-cv19/discussion/130311</a> or <a href=\"https://www.kaggle.com/c/bengaliai-cv19/discussion/131734\">https://www.kaggle.com/c/bengaliai-cv19/discussion/131734</a></p>",
      "rawMarkdown": "128x128 is a good resolution, is it discussable if larger resolution lead to better results or not.\nLike @greatgamedota  mentioned, use pretrained models with pretrained weights, they do help\nAnother important aspects is to use serios data augmentation, I suggest to try affine transforms, cutoff, cutmix and mixup.\nAlso, be careful with the optimizer, scheduler and learning rate.\nFor more info you can check discussion from https://www.kaggle.com/c/bengaliai-cv19/discussion/130311 or https://www.kaggle.com/c/bengaliai-cv19/discussion/131734",
      "votes": null
    },
    {
      "id": "756125",
      "postDate": "02/25/2020 12:41:52",
      "content": "<p><a href=\"/vladvdv\">@vladvdv</a> You say, \"use pretrained models with pretrained weights\". Isn't pretrained model, a model with pretrained weights. Am I missing something?</p>",
      "rawMarkdown": "vladvdv You say, \"use pretrained models with pretrained weights\". Isn't pretrained model, a model with pretrained weights. Am I missing something?",
      "votes": null
    },
    {
      "id": "756161",
      "postDate": "02/25/2020 13:12:48",
      "content": "<p>I did not express myself good.\nWhat I mean to say was \"already implemented model arhitectures(resnet, etc) and use pretrained weights (like imagenet) \"</p>",
      "rawMarkdown": "I did not express myself good.\nWhat I mean to say was \"already implemented model arhitectures(resnet, etc) and use pretrained weights (like imagenet) \"",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 755455,
      "author_name": "greatgamedota",
      "author_url": "",
      "post_date": "02/24/2020 19:38:32",
      "content": "<p>128 is good enough, try pretrained models...</p>",
      "votes": null,
      "replies": [
        {
          "id": 755464,
          "author_name": "avranil",
          "author_url": "",
          "post_date": "02/24/2020 19:49:56",
          "content": "<p>Thanks for the reply !! Do you advise to use pretrained weights?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 755468,
          "author_name": "greatgamedota",
          "author_url": "",
          "post_date": "02/24/2020 19:52:41",
          "content": "<p>Yes</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 755471,
          "author_name": "avranil",
          "author_url": "",
          "post_date": "02/24/2020 19:55:12",
          "content": "<p>Thanks</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 755478,
      "author_name": "vladvdv",
      "author_url": "",
      "post_date": "02/24/2020 20:02:47",
      "content": "<p>128x128 is a good resolution, is it discussable if larger resolution lead to better results or not.\nLike <a href=\"/greatgamedota\">@greatgamedota</a>  mentioned, use pretrained models with pretrained weights, they do help\nAnother important aspects is to use serios data augmentation, I suggest to try affine transforms, cutoff, cutmix and mixup.\nAlso, be careful with the optimizer, scheduler and learning rate.\nFor more info you can check discussion from <a href=\"https://www.kaggle.com/c/bengaliai-cv19/discussion/130311\">https://www.kaggle.com/c/bengaliai-cv19/discussion/130311</a> or <a href=\"https://www.kaggle.com/c/bengaliai-cv19/discussion/131734\">https://www.kaggle.com/c/bengaliai-cv19/discussion/131734</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 756125,
          "author_name": "santosh16k",
          "author_url": "",
          "post_date": "02/25/2020 12:41:52",
          "content": "<p><a href=\"/vladvdv\">@vladvdv</a> You say, \"use pretrained models with pretrained weights\". Isn't pretrained model, a model with pretrained weights. Am I missing something?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 756161,
          "author_name": "vladvdv",
          "author_url": "",
          "post_date": "02/25/2020 13:12:48",
          "content": "<p>I did not express myself good.\nWhat I mean to say was \"already implemented model arhitectures(resnet, etc) and use pretrained weights (like imagenet) \"</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "755447": "Stuck at 0.93 with the below NN architecture. Wondering if anbody tried with greater image size and got better results. Currently my image size is (128,128,3).  Appreciate any other pointers that might help me to climb up the leaderboard\n\n```\nmodel.add(Conv2D(filters=32,kernel_size=(2,2),padding='same',activation='relu',input_shape=(size,size,3)))\nmodel.add(Conv2D(filters=32,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=32,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))\n\nmodel.add(Conv2D(filters=64,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=64,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=64,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))\n\nmodel.add(Conv2D(filters=128,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=128,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))\n\nmodel.add(Conv2D(filters=256,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=256,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(2,2)))\n\nmodel.add(Conv2D(filters=512,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(Conv2D(filters=512,kernel_size=(2,2),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(MaxPool2D(pool_size=(5,5)))\n\nmodel.add(Flatten())\n\nmodel.add(Dense(units=512,activation='relu'))\nmodel.add(Dropout(0.3))\nmodel.add(Dense(units=256,activation='relu'))\nmodel.add(Dense(units=168,activation='softmax'))\n```",
    "755455": "128 is good enough, try pretrained models...",
    "755464": "Thanks for the reply !! Do you advise to use pretrained weights?",
    "755468": "Yes",
    "755471": "Thanks",
    "755478": "128x128 is a good resolution, is it discussable if larger resolution lead to better results or not.\nLike @greatgamedota  mentioned, use pretrained models with pretrained weights, they do help\nAnother important aspects is to use serios data augmentation, I suggest to try affine transforms, cutoff, cutmix and mixup.\nAlso, be careful with the optimizer, scheduler and learning rate.\nFor more info you can check discussion from https://www.kaggle.com/c/bengaliai-cv19/discussion/130311 or https://www.kaggle.com/c/bengaliai-cv19/discussion/131734",
    "756125": "vladvdv You say, \"use pretrained models with pretrained weights\". Isn't pretrained model, a model with pretrained weights. Am I missing something?",
    "756161": "I did not express myself good.\nWhat I mean to say was \"already implemented model arhitectures(resnet, etc) and use pretrained weights (like imagenet) \""
  },
  "source": "meta"
}