Suppose, I create a pipeline to load and resize an image in tensorflow like this:
\ndef decode_image(file_name, label=None, IMAGE_SIZE=IMAGE_DIMENSIONS):\n bits = tf.io.read_file(file_name)\n image = tf.image.decode_jpeg(bits, channels=IMAGE_CHANNELS)\n image = tf.cast(image, tf.float32)/255.0\n image = tf.image.resize(image, size=IMAGE_SIZE)\n if label is None:\n return image\n return image, label\n\ntrain_ds = (\n tf.data.Dataset\n .from_tensor_slices((train_df[\"StudyInstanceUID\"].values, train_df[labels].values))\n .map(decode_image, num_parallel_calls=AUTO)\n .map(data_augment, num_parallel_calls=AUTO)\n .shuffle(512)\n .repeat()\n .batch(BATCH_SIZE)\n .prefetch(AUTO)\n)\nHow can I apply functions like: cv2.cvtColor(image, cv2.COLOR_BGR2GRAY) on an image (which is a tensor) inside the decode_image function mentioned above? Please let me know if I should elaborate more on the details.
Note that tensorflow is not compatible with cv2 transformations/Albumentations when using TPU since it tensors cannot be executed eagerly when using TPU. However, for GPU you can wrap your transformations in a tf.numpy_function, refer to this
I implemented it but with help of imagedatagenerator and flow_from_dataframe or flow_from_directory there you can use custom processing function through setting processing_function property
", "rawMarkdown": "I implemented it but with help of imagedatagenerator and flow_from_dataframe or flow_from_directory there you can use custom processing function through setting processing_function property", "votes": 1, "replies": [ { "id": 1179640, "postDate": "2021-01-31T17:39:07.487Z", "content": "Thank you for the information @dmitryleshchinskiy. I will try this technique.
", "rawMarkdown": "Thank you for the information @dmitryleshchinskiy. I will try this technique.", "votes": 1 } ] }, { "id": 1178522, "postDate": "2021-01-31T00:09:51.487Z", "content": "Suppose, I create a pipeline to load and resize an image in tensorflow like this:
\ndef decode_image(file_name, label=None, IMAGE_SIZE=IMAGE_DIMENSIONS):\n bits = tf.io.read_file(file_name)\n image = tf.image.decode_jpeg(bits, channels=IMAGE_CHANNELS)\n image = tf.cast(image, tf.float32)/255.0\n image = tf.image.resize(image, size=IMAGE_SIZE)\n if label is None:\n return image\n return image, label\n\ntrain_ds = (\n tf.data.Dataset\n .from_tensor_slices((train_df[\"StudyInstanceUID\"].values, train_df[labels].values))\n .map(decode_image, num_parallel_calls=AUTO)\n .map(data_augment, num_parallel_calls=AUTO)\n .shuffle(512)\n .repeat()\n .batch(BATCH_SIZE)\n .prefetch(AUTO)\n)\nHow can I apply functions like: cv2.cvtColor(image, cv2.COLOR_BGR2GRAY) on an image (which is a tensor) inside the decode_image function mentioned above? Please let me know if I should elaborate more on the details.
we always preprocess images before converting them into tfrecords. If you are using this as an augmentation, you can use tf.image.rgb_to_grayscale in your augmentation function. But, the dataset already contains grayscale images, can you describe why you need to do this conversion?
@kingofarmy Thank you for the information. I am in the process of applying some opencv transformations to each image. Now, some of them are not available as a part of tensorflow e.g. contrast adaptive histogram equalization. So, to apply such functions via the pipeline I mentioned, I was searching for a solution. Hence, I gave a sample transformation to make my case more clear to the audience reading it.
", "rawMarkdown": "@kingofarmy Thank you for the information. I am in the process of applying some opencv transformations to each image. Now, some of them are not available as a part of tensorflow e.g. contrast adaptive histogram equalization. So, to apply such functions via the pipeline I mentioned, I was searching for a solution. Hence, I gave a sample transformation to make my case more clear to the audience reading it." }, { "id": 1181038, "postDate": "2021-02-01T16:07:54.147Z", "content": "Check this issue in github.
\nCheck this implementation of MCLAHE
\nI think these infos might help you. But, also keep in mind you can't use such methods if you gonna train in TPU. Refer the previous comment by Andy.
", "rawMarkdown": "Check this [issue] (https://github.com/tensorflow/tensorflow/issues/24181) in github.\n\nCheck this implementation of [MCLAHE](https://github.com/VincentStimper/mclahe) \n\n\nI think these infos might help you. But, also keep in mind you can't use such methods if you gonna train in TPU. Refer the previous comment by Andy.", "votes": 1 } ] }, { "id": 1179548, "postDate": "2021-01-31T16:10:18.607Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1179777, "author_name": "Andy", "author_url": "", "post_date": "2021-01-31T20:08:14.230000", "content": "Note that tensorflow is not compatible with cv2 transformations/Albumentations when using TPU since it tensors cannot be executed eagerly when using TPU. However, for GPU you can wrap your transformations in a tf.numpy_function, refer to this
I implemented it but with help of imagedatagenerator and flow_from_dataframe or flow_from_directory there you can use custom processing function through setting processing_function property
", "votes": 1, "replies": [ { "id": 1179640, "author_name": "Aishwary Shukla", "author_url": "", "post_date": "2021-01-31T17:39:07.487000", "content": "Thank you for the information @dmitryleshchinskiy. I will try this technique.
", "votes": 1, "replies": [] } ] }, { "id": 1179608, "author_name": "TEnsorSAge", "author_url": "", "post_date": "2021-01-31T17:10:59.230000", "content": "we always preprocess images before converting them into tfrecords. If you are using this as an augmentation, you can use tf.image.rgb_to_grayscale in your augmentation function. But, the dataset already contains grayscale images, can you describe why you need to do this conversion?
@kingofarmy Thank you for the information. I am in the process of applying some opencv transformations to each image. Now, some of them are not available as a part of tensorflow e.g. contrast adaptive histogram equalization. So, to apply such functions via the pipeline I mentioned, I was searching for a solution. Hence, I gave a sample transformation to make my case more clear to the audience reading it.
", "votes": 0, "replies": [] }, { "id": 1181038, "author_name": "TEnsorSAge", "author_url": "", "post_date": "2021-02-01T16:07:54.147000", "content": "Check this issue in github.
\nCheck this implementation of MCLAHE
\nI think these infos might help you. But, also keep in mind you can't use such methods if you gonna train in TPU. Refer the previous comment by Andy.
", "votes": 1, "replies": [] } ] }, { "id": 1179548, "author_name": "", "author_url": "", "post_date": "2021-01-31T16:10:18.607000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1179777": "Note that tensorflow is not compatible with cv2 transformations/Albumentations when using TPU since it tensors cannot be executed eagerly when using TPU. However, for GPU you can wrap your transformations in a `tf.numpy_function`, refer to [this](https://www.tensorflow.org/api_docs/python/tf/numpy_function)", "1179245": "I implemented it but with help of imagedatagenerator and flow_from_dataframe or flow_from_directory there you can use custom processing function through setting processing_function property", "1178522": "Suppose, I create a pipeline to load and resize an image in tensorflow like this:\n\n```\ndef decode_image(file_name, label=None, IMAGE_SIZE=IMAGE_DIMENSIONS):\n bits = tf.io.read_file(file_name)\n image = tf.image.decode_jpeg(bits, channels=IMAGE_CHANNELS)\n image = tf.cast(image, tf.float32)/255.0\n image = tf.image.resize(image, size=IMAGE_SIZE)\n if label is None:\n return image\n return image, label\n\ntrain_ds = (\n tf.data.Dataset\n .from_tensor_slices((train_df[\"StudyInstanceUID\"].values, train_df[labels].values))\n .map(decode_image, num_parallel_calls=AUTO)\n .map(data_augment, num_parallel_calls=AUTO)\n .shuffle(512)\n .repeat()\n .batch(BATCH_SIZE)\n .prefetch(AUTO)\n)\n```\nHow can I apply functions like: `cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)` on an image (which is a tensor) inside the decode_image function mentioned above? Please let me know if I should elaborate more on the details.", "1179608": "we always preprocess images before converting them into tfrecords. If you are using this as an augmentation, you can use ```tf.image.rgb_to_grayscale``` in your augmentation function. But, the dataset already contains grayscale images, can you describe why you need to do this conversion?", "1179548": "" } }