{ "id": 240790, "title": "How can i implement Mixup In Tensorflow", "url": "/competitions/seti-breakthrough-listen/discussion/240790", "author_name": "Mithil Salunkhe", "post_date": "2021-05-21T11:45:08.662000", "votes": 3, "comment_count": 6, "views": 0, "content": "

All the notebook That implement Mixup are from Pytorch. Can someone Help me out In implementing Mixup in Tensor flow ? Any help would be greatly appreciated

", "messages": [ { "id": 1317442, "postDate": "2021-05-21T12:16:30.617Z", "content": "

Hello. It's so nice to meet you.
\nI'm not sure about TensorFlow, but we need to apply the steps below.

\n
    \n

  1. define mixup function with usual python language.
    \nx <- original inputs (have items as many as batch size)
    \nindex <- random permutation of x index
    \ny = x[index] <- permutation of x.
    \nz = alpha * x + (1-alpha) * y <- some input between (original inputs, permutation of x)
    \nnew_loss = alphaloss(predict_x) + (1-alpha)loss(predict_y)

    2. \n

  2. define tensorflow dataset with batchsize

    3. \n

  3. mapping dataset to new dataset using mixup function

    4. \n
", "rawMarkdown": "Hello. It's so nice to meet you.\nI'm not sure about TensorFlow, but we need to apply the steps below.\n\n1. define mixup function with usual python language.\nx <- original inputs (have items as many as batch size)\nindex <- random permutation of x index\ny = x[index] <- permutation of x.\nz = alpha * x + (1-alpha) * y <- some input between (original inputs, permutation of x)\nnew_loss = alpha*loss(predict_x) + (1-alpha)*loss(predict_y)\n\n2. define tensorflow dataset with batchsize\n\n3. mapping dataset to new dataset using mixup function", "votes": 3 }, { "id": 1317403, "postDate": "2021-05-21T11:45:08.663Z", "content": "

All the notebook That implement Mixup are from Pytorch. Can someone Help me out In implementing Mixup in Tensor flow ? Any help would be greatly appreciated

", "rawMarkdown": "All the notebook That implement Mixup are from Pytorch. Can someone Help me out In implementing Mixup in Tensor flow ? Any help would be greatly appreciated ", "votes": 3 }, { "id": 1317713, "postDate": "2021-05-21T16:21:44.503Z", "content": "

You can also look in MixUp augmentation for image classification in Keras documentation.

", "rawMarkdown": "You can also look in [MixUp augmentation for image classification](https://keras.io/examples/vision/mixup/) in Keras documentation.", "votes": 1 }, { "id": 1318150, "postDate": "2021-05-22T04:01:22.890Z", "content": "

For those who are using the notebook @micheomaano shared make sure to change your metrics from keras.metrics.AUC() to tf.keras.metrics.BinaryAccuracy

", "rawMarkdown": "For those who are using the notebook @micheomaano shared make sure to change your metrics from keras.metrics.AUC() to tf.keras.metrics.BinaryAccuracy\n", "votes": 2 }, { "id": 1317428, "postDate": "2021-05-21T12:06:25.950Z", "content": "

Here you go. https://www.kaggle.com/cdeotte/cutmix-and-mixup-on-gpu-tpu

", "rawMarkdown": "Here you go. [https://www.kaggle.com/cdeotte/cutmix-and-mixup-on-gpu-tpu](https://www.kaggle.com/cdeotte/cutmix-and-mixup-on-gpu-tpu)", "votes": 2, "replies": [ { "id": 1318149, "postDate": "2021-05-22T03:59:37.547Z", "content": "

Thank You very much

", "rawMarkdown": "Thank You very much\n" } ] }, { "id": 1318557, "postDate": "2021-05-22T12:02:34.137Z", "content": "

Here's a blog post that I have written: https://wandb.ai/authors/tfaugmentation/reports/Modern-Data-Augmentation-Techniques-for-Computer-Vision--VmlldzoxNzU3NTU

\n

Code snippet:

\n
def mixup(a, b):\n\n  alpha = [0.2]\n  beta = [0.2]\n\n  # unpack (image, label) pairs\n  (image1, label1), (image2, label2) = a, b\n\n  # define beta distribution\n  dist = tfd.Beta(alpha, beta)\n  # sample from this distribution\n  l = dist.sample(1)[0][0]\n\n  # mixup augmentation\n  img = l*image1+(1-l)*image2\n  lab = l*label1+(1-l)*label2\n\n  return img, lab\n\n\ntrainloader1 = tf.data.Dataset.from_tensor_slices((x_train, y_train)).shuffle(1024).map(preprocess_image, num_parallel_calls=AUTO)\ntrainloader2 = tf.data.Dataset.from_tensor_slices((x_train, y_train)).shuffle(1024).map(preprocess_image, num_parallel_calls=AUTO)\n\ntrainloader = tf.data.Dataset.zip((trainloader1, trainloader2))\ntrainloader = (\n    trainloader\n    .shuffle(1024)\n    .map(mixup, num_parallel_calls=AUTO)\n    .batch(BATCH_SIZE)\n    .prefetch(AUTO)\n)\n
", "rawMarkdown": "Here's a blog post that I have written: https://wandb.ai/authors/tfaugmentation/reports/Modern-Data-Augmentation-Techniques-for-Computer-Vision--VmlldzoxNzU3NTU\n\nCode snippet:\n\n```\ndef mixup(a, b):\n \n alpha = [0.2]\n beta = [0.2]\n \n # unpack (image, label) pairs\n (image1, label1), (image2, label2) = a, b\n\n # define beta distribution\n dist = tfd.Beta(alpha, beta)\n # sample from this distribution\n l = dist.sample(1)[0][0]\n \n # mixup augmentation\n img = l*image1+(1-l)*image2\n lab = l*label1+(1-l)*label2\n\n return img, lab\n\n\ntrainloader1 = tf.data.Dataset.from_tensor_slices((x_train, y_train)).shuffle(1024).map(preprocess_image, num_parallel_calls=AUTO)\ntrainloader2 = tf.data.Dataset.from_tensor_slices((x_train, y_train)).shuffle(1024).map(preprocess_image, num_parallel_calls=AUTO)\n\ntrainloader = tf.data.Dataset.zip((trainloader1, trainloader2))\ntrainloader = (\n trainloader\n .shuffle(1024)\n .map(mixup, num_parallel_calls=AUTO)\n .batch(BATCH_SIZE)\n .prefetch(AUTO)\n)\n```" } ], "comments": [ { "id": 1317442, "author_name": "WOOSUNG YOON", "author_url": "", "post_date": "2021-05-21T12:16:30.617000", "content": "

Hello. It's so nice to meet you.
\nI'm not sure about TensorFlow, but we need to apply the steps below.

\n
    \n

  1. define mixup function with usual python language.
    \nx <- original inputs (have items as many as batch size)
    \nindex <- random permutation of x index
    \ny = x[index] <- permutation of x.
    \nz = alpha * x + (1-alpha) * y <- some input between (original inputs, permutation of x)
    \nnew_loss = alphaloss(predict_x) + (1-alpha)loss(predict_y)

    2. \n

  2. define tensorflow dataset with batchsize

    3. \n

  3. mapping dataset to new dataset using mixup function

    4. \n
", "votes": 3, "replies": [] }, { "id": 1317713, "author_name": "Geir Drange", "author_url": "", "post_date": "2021-05-21T16:21:44.503000", "content": "

You can also look in MixUp augmentation for image classification in Keras documentation.

", "votes": 1, "replies": [] }, { "id": 1318150, "author_name": "Mithil Salunkhe", "author_url": "", "post_date": "2021-05-22T04:01:22.890000", "content": "

For those who are using the notebook @micheomaano shared make sure to change your metrics from keras.metrics.AUC() to tf.keras.metrics.BinaryAccuracy

", "votes": 2, "replies": [] }, { "id": 1317428, "author_name": "Salman", "author_url": "", "post_date": "2021-05-21T12:06:25.950000", "content": "

Here you go. https://www.kaggle.com/cdeotte/cutmix-and-mixup-on-gpu-tpu

", "votes": 2, "replies": [ { "id": 1318149, "author_name": "Mithil Salunkhe", "author_url": "", "post_date": "2021-05-22T03:59:37.547000", "content": "

Thank You very much

", "votes": 0, "replies": [] } ] }, { "id": 1318557, "author_name": "Ayush Thakur", "author_url": "", "post_date": "2021-05-22T12:02:34.137000", "content": "

Here's a blog post that I have written: https://wandb.ai/authors/tfaugmentation/reports/Modern-Data-Augmentation-Techniques-for-Computer-Vision--VmlldzoxNzU3NTU

\n

Code snippet:

\n
def mixup(a, b):\n\n  alpha = [0.2]\n  beta = [0.2]\n\n  # unpack (image, label) pairs\n  (image1, label1), (image2, label2) = a, b\n\n  # define beta distribution\n  dist = tfd.Beta(alpha, beta)\n  # sample from this distribution\n  l = dist.sample(1)[0][0]\n\n  # mixup augmentation\n  img = l*image1+(1-l)*image2\n  lab = l*label1+(1-l)*label2\n\n  return img, lab\n\n\ntrainloader1 = tf.data.Dataset.from_tensor_slices((x_train, y_train)).shuffle(1024).map(preprocess_image, num_parallel_calls=AUTO)\ntrainloader2 = tf.data.Dataset.from_tensor_slices((x_train, y_train)).shuffle(1024).map(preprocess_image, num_parallel_calls=AUTO)\n\ntrainloader = tf.data.Dataset.zip((trainloader1, trainloader2))\ntrainloader = (\n    trainloader\n    .shuffle(1024)\n    .map(mixup, num_parallel_calls=AUTO)\n    .batch(BATCH_SIZE)\n    .prefetch(AUTO)\n)\n
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1317442": "Hello. It's so nice to meet you.\nI'm not sure about TensorFlow, but we need to apply the steps below.\n\n1. define mixup function with usual python language.\nx <- original inputs (have items as many as batch size)\nindex <- random permutation of x index\ny = x[index] <- permutation of x.\nz = alpha * x + (1-alpha) * y <- some input between (original inputs, permutation of x)\nnew_loss = alpha*loss(predict_x) + (1-alpha)*loss(predict_y)\n\n2. define tensorflow dataset with batchsize\n\n3. mapping dataset to new dataset using mixup function", "1317403": "All the notebook That implement Mixup are from Pytorch. Can someone Help me out In implementing Mixup in Tensor flow ? Any help would be greatly appreciated ", "1317713": "You can also look in [MixUp augmentation for image classification](https://keras.io/examples/vision/mixup/) in Keras documentation.", "1318150": "For those who are using the notebook @micheomaano shared make sure to change your metrics from keras.metrics.AUC() to tf.keras.metrics.BinaryAccuracy\n", "1317428": "Here you go. [https://www.kaggle.com/cdeotte/cutmix-and-mixup-on-gpu-tpu](https://www.kaggle.com/cdeotte/cutmix-and-mixup-on-gpu-tpu)", "1318557": "Here's a blog post that I have written: https://wandb.ai/authors/tfaugmentation/reports/Modern-Data-Augmentation-Techniques-for-Computer-Vision--VmlldzoxNzU3NTU\n\nCode snippet:\n\n```\ndef mixup(a, b):\n \n alpha = [0.2]\n beta = [0.2]\n \n # unpack (image, label) pairs\n (image1, label1), (image2, label2) = a, b\n\n # define beta distribution\n dist = tfd.Beta(alpha, beta)\n # sample from this distribution\n l = dist.sample(1)[0][0]\n \n # mixup augmentation\n img = l*image1+(1-l)*image2\n lab = l*label1+(1-l)*label2\n\n return img, lab\n\n\ntrainloader1 = tf.data.Dataset.from_tensor_slices((x_train, y_train)).shuffle(1024).map(preprocess_image, num_parallel_calls=AUTO)\ntrainloader2 = tf.data.Dataset.from_tensor_slices((x_train, y_train)).shuffle(1024).map(preprocess_image, num_parallel_calls=AUTO)\n\ntrainloader = tf.data.Dataset.zip((trainloader1, trainloader2))\ntrainloader = (\n trainloader\n .shuffle(1024)\n .map(mixup, num_parallel_calls=AUTO)\n .batch(BATCH_SIZE)\n .prefetch(AUTO)\n)\n```" } }