I am using TensorFlow and I am trying to train a model from the segment data to learn which vertebrae is present in the picture. I have the dicom image data as an EagorTensor of shape (512, 512, 1) and I used CategoricalEncoding to create a multi-hot encoded labels, which end up being EagorTensors of shape (8,). However, when I try to call model.fit the code fails. I tried to write the training loop from scratch using the Tensorflow documentation https://www.tensorflow.org/guide/keras/writing_a_training_loop_from_scratch#using_the_gradienttape_a_first_end-to-end_example and the code fails when I try to pass my data to the model with model(train_ds). I feel like the shape in my model might be off, but I am not quite sure. I looked at a couple different tutorials and I copied their code and debugged it to try and get the same shapes as they did with their data. But I have still had no luck. I looked at https://towardsdatascience.com/multi-label-image-classification-in-tensorflow-2-0-7d4cf8a4bc72 and https://medium.com/deep-learning-with-keras/how-to-solve-multi-label-classification-problems-in-deep-learning-with-tensorflow-keras-7fb933243595. My shapes seem to be the same with the image at, (batch_size, height, width, channels) / (75, 512, 512, 1) and the labels at (batch_size, num_labels)/ (75, 8). Maybe the problem is in my model and I need my last layer to be of shape (batch_size, num_labels). Not quite sure. Any help would be of greatly appreciated. Here is a link to my notebook on my Github.
\nhttps://github.com/cbarb15/RSNACervicalSpineFractureDetection
**The code in the notebook has a last dense layer with 7, but I have tried it with 8 and get the same results as listed above.
", "messages": [ { "id": 1957505, "postDate": "2022-09-27T02:53:12.843Z", "content": "I am using TensorFlow and I am trying to train a model from the segment data to learn which vertebrae is present in the picture. I have the dicom image data as an EagorTensor of shape (512, 512, 1) and I used CategoricalEncoding to create a multi-hot encoded labels, which end up being EagorTensors of shape (8,). However, when I try to call model.fit the code fails. I tried to write the training loop from scratch using the Tensorflow documentation https://www.tensorflow.org/guide/keras/writing_a_training_loop_from_scratch#using_the_gradienttape_a_first_end-to-end_example and the code fails when I try to pass my data to the model with model(train_ds). I feel like the shape in my model might be off, but I am not quite sure. I looked at a couple different tutorials and I copied their code and debugged it to try and get the same shapes as they did with their data. But I have still had no luck. I looked at https://towardsdatascience.com/multi-label-image-classification-in-tensorflow-2-0-7d4cf8a4bc72 and https://medium.com/deep-learning-with-keras/how-to-solve-multi-label-classification-problems-in-deep-learning-with-tensorflow-keras-7fb933243595. My shapes seem to be the same with the image at, (batch_size, height, width, channels) / (75, 512, 512, 1) and the labels at (batch_size, num_labels)/ (75, 8). Maybe the problem is in my model and I need my last layer to be of shape (batch_size, num_labels). Not quite sure. Any help would be of greatly appreciated. Here is a link to my notebook on my Github.
\nhttps://github.com/cbarb15/RSNACervicalSpineFractureDetection
**The code in the notebook has a last dense layer with 7, but I have tried it with 8 and get the same results as listed above.
", "rawMarkdown": "I am using TensorFlow and I am trying to train a model from the segment data to learn which vertebrae is present in the picture. I have the dicom image data as an EagorTensor of shape (512, 512, 1) and I used CategoricalEncoding to create a multi-hot encoded labels, which end up being EagorTensors of shape (8,). However, when I try to call model.fit the code fails. I tried to write the training loop from scratch using the Tensorflow documentation [https://www.tensorflow.org/guide/keras/writing_a_training_loop_from_scratch#using_the_gradienttape_a_first_end-to-end_example](url) and the code fails when I try to pass my data to the model with model(train_ds). I feel like the shape in my model might be off, but I am not quite sure. I looked at a couple different tutorials and I copied their code and debugged it to try and get the same shapes as they did with their data. But I have still had no luck. I looked at https://towardsdatascience.com/multi-label-image-classification-in-tensorflow-2-0-7d4cf8a4bc72 and https://medium.com/deep-learning-with-keras/how-to-solve-multi-label-classification-problems-in-deep-learning-with-tensorflow-keras-7fb933243595. My shapes seem to be the same with the image at, (batch_size, height, width, channels) / (75, 512, 512, 1) and the labels at (batch_size, num_labels)/ (75, 8). Maybe the problem is in my model and I need my last layer to be of shape (batch_size, num_labels). Not quite sure. Any help would be of greatly appreciated. Here is a link to my notebook on my Github. \nhttps://github.com/cbarb15/RSNACervicalSpineFractureDetection \n\n **The code in the notebook has a last dense layer with 7, but I have tried it with 8 and get the same results as listed above. ", "votes": 2 } ], "comments": [], "raw_markdown_by_id": { "1957505": "I am using TensorFlow and I am trying to train a model from the segment data to learn which vertebrae is present in the picture. I have the dicom image data as an EagorTensor of shape (512, 512, 1) and I used CategoricalEncoding to create a multi-hot encoded labels, which end up being EagorTensors of shape (8,). However, when I try to call model.fit the code fails. I tried to write the training loop from scratch using the Tensorflow documentation [https://www.tensorflow.org/guide/keras/writing_a_training_loop_from_scratch#using_the_gradienttape_a_first_end-to-end_example](url) and the code fails when I try to pass my data to the model with model(train_ds). I feel like the shape in my model might be off, but I am not quite sure. I looked at a couple different tutorials and I copied their code and debugged it to try and get the same shapes as they did with their data. But I have still had no luck. I looked at https://towardsdatascience.com/multi-label-image-classification-in-tensorflow-2-0-7d4cf8a4bc72 and https://medium.com/deep-learning-with-keras/how-to-solve-multi-label-classification-problems-in-deep-learning-with-tensorflow-keras-7fb933243595. My shapes seem to be the same with the image at, (batch_size, height, width, channels) / (75, 512, 512, 1) and the labels at (batch_size, num_labels)/ (75, 8). Maybe the problem is in my model and I need my last layer to be of shape (batch_size, num_labels). Not quite sure. Any help would be of greatly appreciated. Here is a link to my notebook on my Github. \nhttps://github.com/cbarb15/RSNACervicalSpineFractureDetection \n\n **The code in the notebook has a last dense layer with 7, but I have tried it with 8 and get the same results as listed above. " } }