Good question. The starter code uses a TensorFlow dataset, API here. You can call the map function on your dataset and write routines to perform augmentation. However I think the map function gets called on each image individually so it wouldn't work for CutMix where you need to create new Images/Labels from 2 Images/Labels.
\n\nI'm reading about this now and will report if I figure an efficient way to do CutMix on TensorFlow dataset.
", "rawMarkdown": "Good question. The starter code uses a TensorFlow dataset, API [here][1]. You can call the map function on your dataset and write routines to perform augmentation. However I think the map function gets called on each image individually so it wouldn't work for CutMix where you need to create new Images/Labels from 2 Images/Labels.\n\nI'm reading about this now and will report if I figure an efficient way to do CutMix on TensorFlow dataset.\n\n[1]: https://www.tensorflow.org/api_docs/python/tf/data/Dataset", "votes": null }, { "id": "744308", "postDate": "02/12/2020 18:10:32", "content": "If you want other simple augmentation, it's easy. Just add TensorFlow calls inside the starter codes function data_augment which gets called with dataset.map() already.
The map function is mapped on either single or multiple elements if you have already batched the dataset.
", "rawMarkdown": "The map function is mapped on either single or multiple elements if you have already batched the dataset.", "votes": null }, { "id": "744344", "postDate": "02/12/2020 18:54:27", "content": "This is a great idea by the way! Let us know how it works.
", "rawMarkdown": "This is a great idea by the way! Let us know how it works.", "votes": null }, { "id": "744555", "postDate": "02/12/2020 23:40:31", "content": "Ah yes, i forgot that, thanks for pointing it out Martin. What Martin is saying is as follows. If you call map() before batch() then your map function receives a single image and if you call map() after batch() then your function receives a batch of images.
# EXAMPLE OF MAP BEFORE BATCH\ndataset = dataset.map(data_augment, num_parallel_calls=AUTO)\ndataset = dataset.batch(BATCH_SIZE)\nThen inside your augment function, you can check whether you received a single image or batch of images by checking the input's dimensions:
\n\n@tf.function\ndef transformation(data):\n if image.shape.__len__() ==4:\n # WE HAVE BATCH\n if image.shape.__len__() ==3:\n # WE HAVE SINGLE IMAGES\n return data\nYes exactly and if you want to mix multiple images into one you can do something like this:\n\ndataset = dataset.batch(2)\ndataset = dataset.map(my_mix) # your own mixing function, returns individual images again\ndataset = dataset.batch(BATCH_SIZE)\n
I am using your method.It can work.Suddenly it throw an exception tensorflow.python.framework.errors_impl.InvalidArgumentError: Index out of range using input dim 0; input has only 0 dims [Op:StridedSlice] name: strided_slice/ and distributed_function -> distributed_function\nI need 4 picture to merge 1 image.I think maybe i have 15 picture, when I getting the last 3 picture, program can't split.So How do you solve?
Good question. The starter code uses a TensorFlow dataset, API here. You can call the map function on your dataset and write routines to perform augmentation. However I think the map function gets called on each image individually so it wouldn't work for CutMix where you need to create new Images/Labels from 2 Images/Labels.
\n\nI'm reading about this now and will report if I figure an efficient way to do CutMix on TensorFlow dataset.
", "votes": null, "replies": [ { "id": 744308, "author_name": "cdeotte", "author_url": "", "post_date": "02/12/2020 18:10:32", "content": "If you want other simple augmentation, it's easy. Just add TensorFlow calls inside the starter codes function data_augment which gets called with dataset.map() already.
The map function is mapped on either single or multiple elements if you have already batched the dataset.
", "votes": null, "replies": [] }, { "id": 744555, "author_name": "cdeotte", "author_url": "", "post_date": "02/12/2020 23:40:31", "content": "Ah yes, i forgot that, thanks for pointing it out Martin. What Martin is saying is as follows. If you call map() before batch() then your map function receives a single image and if you call map() after batch() then your function receives a batch of images.
# EXAMPLE OF MAP BEFORE BATCH\ndataset = dataset.map(data_augment, num_parallel_calls=AUTO)\ndataset = dataset.batch(BATCH_SIZE)\nThen inside your augment function, you can check whether you received a single image or batch of images by checking the input's dimensions:
\n\n@tf.function\ndef transformation(data):\n if image.shape.__len__() ==4:\n # WE HAVE BATCH\n if image.shape.__len__() ==3:\n # WE HAVE SINGLE IMAGES\n return data\nYes exactly and if you want to mix multiple images into one you can do something like this:\n\ndataset = dataset.batch(2)\ndataset = dataset.map(my_mix) # your own mixing function, returns individual images again\ndataset = dataset.batch(BATCH_SIZE)\n
This is a great idea by the way! Let us know how it works.
", "votes": null, "replies": [] }, { "id": 861639, "author_name": "footrunist", "author_url": "", "post_date": "05/26/2020 06:22:12", "content": "I am using your method.It can work.Suddenly it throw an exception tensorflow.python.framework.errors_impl.InvalidArgumentError: Index out of range using input dim 0; input has only 0 dims [Op:StridedSlice] name: strided_slice/ and distributed_function -> distributed_function\nI need 4 picture to merge 1 image.I think maybe i have 15 picture, when I getting the last 3 picture, program can't split.So How do you solve?