How to convert a dataset which has two items-image and label , where image is depicted with a list of image names such as '12_left',12_right' and so on, and labels such as 0,1,2,3,4? Tensorflow already has certain datasets, or the built-in input pipeline but what about datasets which are not present ?? What could be done to convert such a dataset into tensor, so as to avoid shape incompatibility value error and logits and labels error?
Categorical has been implemented using Keras, accuracy was changed to Accuracy/SparseCategoricalAccuracy, categorical was implemented and even SparseCategoricalCrossEntropy was also used as well. I am still having some errors and realized this conversion may be helpful. While troubleshooting, all I found tensorflow has datasets such as CIFAR-10, MNIST etc. already inbuilt and converted to tensors. But what about image datasets which are not present and not converted to tensors? How to convert them to tensors?
Please help!
I believe what you are referring to is image sequence classification, where instead of a typical MNIST like dataset which has a shape of (samples, width, height, channel_number) you have a dataset with a number of images for each sample. This way the shape should look like this: (samples, num_images, width, height, channel_number), where you concatenate the images of a single sample into the second dimension of the tensor, then you can easily use TimeDistributed layers to train models on such set. For further information i'd suggest you to check this article out: https://medium.com/smileinnovation/how-to-work-with-time-distributed-data-in-a-neural-network-b8b39aa4ce00
Abdolkarim Saeedi No no. Its not a sequence. Its a single image. Every image has a corresponding label. When I train my model, I encounter shape incompatibility error. I have troubleshooted the problem with replacement of the class_mode to categorical, metric to Accuracy/SparseCategoricalAccuracy and loss to SparseCategoricalCrossEntropy . This led to the removal of shape incompatibility to some extent but again I have an error ' logits and labels must have the same first dimension, got logits shape [250,5] and labels shape [50] '. How to solve this?
I also wanted to know how to convert the images of the dataset to tensors , which is not present in the data input pipeline of tensorflow?
N.B.: 12_left and 12_right are not sequence of a particular image. 12_left has label say 1 and 12_right has label 0. Again, 13_left has label 2 and 13_right has label 4. It is a categorical problem and not binary.
Dolly Das, So from what i am inspecting you mention that " Every image has a corresponding label", then the error is "logits shape [250,5] and labels shape [50]".
This means that the size of the labels from the last network layer (AKA classification layer), which might for example be a Dense(1, activation='softmax') are not the same with the labels your are providing when using model.fit. And you also provide that you are using "SparseCategoricalEntropy", this indicates that if your labels are between 0 to 5, they are one hot encoded before the label and logit compared to calculate the loss, this requires the labels to be like this:
1 -> [1, 0, 0, 0, 0]
2 -> [0, 1, 0, 0, 0]
and so on.
Based on that if the labels you are giving the model are not in this format then you cannot train the model. Again this applies to the second dimension of [250, 5] and [50, 1] not being the same and does not solve the 250 != 50.
My suggestion is to use the scikit-learn LabelBinarizer on the labels before training the model or even doing train and testing splits.
That way all of your labels will be the same shape, for instance if you have 300 samples. It will be [300, 5], this also requires you to use a dense layer with a softmax activation that has 5 neurons as the lastlayer, now because your labels are already on hot encoded you have to switch back to using "categorical_crossentropy" as the loss function.
Another thing i could think of is that the layers of the network are somehow not flattened before being fed to the classification layer.
Again if these does not help you, a good solution would be for you to post your model architecture of summary so i can take a closer look.
Abdolkarim Saeedi When it comes to train and test splitting, I have splitted the dataset into three sections using standard splitting method , which has train, valid and test, before training the model . Valid ratio is 0.15 and test-ratio is 0.5. I have computed the train, valid and test sections and with a batch size of 10. Now, each section has images ranging between class 0 to class 4 with different number of images. The dataset is highly imbalanced. Is my way of proceeding correct? Or should I proceed with a balanced dataset?
For the time being, I am learning how to train my model with the image dataset and then proceed towards an optimal result. However, I shall be grateful and eager to know all threads attached to the topic, which will lead to an optimal solution. I have very less idea on image data and associated techniques.
Dolly Das Yes it does seem correct and i do not think that having an imbalanced dataset is wrong, you can work on doing training and splitting in a way that the sample distribution remains the same throughout training and testing.
More information on that can be found here: https://stackoverflow.com/questions/29438265/stratified-train-test-split-in-scikit-learn.
You can read more on image preprocessing steps on:
https://keras.io/api/preprocessing/image/
There are also ways to load the data from a predefined directory schema which leaves you out of providing labels as an array.
As an unpopular opinion i think you should actually start on a more simple tutorial like the mnist one: https://keras.io/examples/vision/mnist_convnet/.
From my understanding your dataset is not much different shape wise. Both datasets present two dimensional images, possibly with different channel size as you have not mentioned, there are also categorical labels, So all i can suggest is reading more on how to use keras and trying to compare the datasets to truly understand the problem. For example you can create a dummy dataset using numpy functions that allow random number generation. More on that in:
https://numpy.org/doc/stable/reference/random/generated/numpy.random.rand.html
Once you build that foundation, training neural nets on image sets will be easy.
For any more issues, i suggest you posting your problem on "stackoverflow.com", which is a better place for reproducing and debugging code.
- Badges
- Science topic
- Similar topics
- Computer Science
- Images