cnns | Convolutional Neural Networks in Go | Machine Learning library

I encountered many hardships when trying to fit a CNN (U-Net) to my tif training images in Python.

I have the following structure to my data:

• X
• • 0
• • • [Images] (tif, 3-band, 128x128, values ∈ [0, 255])
• X_val
• • 0
• • • [Images] (tif, 3-band, 128x128, values ∈ [0, 255])
• y
• • 0
• • • [Images] (tif, 1-band, 128x128, values ∈ [0, 255])
• y_val
• • 0
• • • [Images] (tif, 1-band, 128x128, values ∈ [0, 255])

Starting with this data, I defined ImageDataGenerators:

I'm trying to implement a 3D facial recognition algorithm using CNNs with multiple classes. I have an image generator for rgb images, and an image generator for depth images (grayscale). As I have two distinct inputs, I made two different CNN models, one with shape=(height, width, 3) and another with shape=(height, width, 1). Independently I can fit the models with its respective image generator, but after concatenating the two branches and merging both image generators, I got this warning and error:

WARNING:tensorflow:Model was constructed with shape (None, 400, 400, 1) for input KerasTensor(type_spec=TensorSpec(shape=(None, 400, 400, 1), dtype=tf.float32, name='Depth_Input_input'), name='Depth_Input_input', description="created by layer 'Depth_Input_input'"), but it was called on an input with incompatible shape (None, None)

"ValueError: Input 0 of layer Depth_Input is incompatible with the layer: : expected min_ndim=4, found ndim=2. Full shape received: (None, None)"

What can i do to solve this? Thanks

Here is my code:

I am building a CNN that can detect numbers and the addition and subtraction symbols.

I was following DeepLizards Tutorial on CNNs.

And I wanted to use my own test images but I keep getting this error when I make predictions:

I am trying to understand the training process of a object deetaction deeplearng algorithm and I am having some problems understanding how the backbone network (the network that performs feature extraction) is trained.

I understand that it is common to use CNNs like AlexNet, VGGNet, and ResNet but I don't understand if these networks are pre-trained or not. If they are not trained what does the training consist of?

I have a trained model (Which I trained myself), which is a .h5 format, It works fine by itself, but I need to convert it to .onnx format for deploying it inside Unity Engine, I searched how to convert .h5 models to .onnx format and stumbled upon keras2onnx library, and following some tutorials I got this:

I'm working on an astronomical images classification project and I'm currently using keras to build CNNs.

I'm trying to build a preprocessing pipeline to augment my dataset with keras/tensorflow layers. To keep things simple I would like to implement random transformations of the dihedral group (i.e., for square images, 90-degrees rotations and flips), but it seems that tf.keras.preprocessing.image.random_rotation only allows a random degree over a continuous range of choice following a uniform distribution.

I was wondering whether there is a way to instead choose from a list of specified degrees, in my case [0, 90, 180, 270].

When I read some classical papers about CNNs, like Inception family, ResNet, VGGnet and so on, I notice the terminology concatenation, summation and aggregation, which makes me confused(summation is easy to understand for me). Could someone tell me what the differences are among them? Maybe in a more sepcific way, like using examples to illustrate the dimensionality and representation ability differences.

I've been reading up a bit on different CNNs for object detection, and have found that most of the models I'm looking at are fully convolutional networks, like the latest YOLO versions and retinanet.

What are the benefits of FCNs over conventional CNNs with pooling, apart from FCNs having less different layers? I've read https://arxiv.org/pdf/1412.6806.pdf and as I read it the main interest of that paper was to simplify the networks structure. Is this the sole reason that modern detection/classification networks don't use pooling, or are there other benefits?

I'm working on a binary classification dataset and applying xgBoost model to the problem. Once the model is ready, I plot the feature importance and one of the trees resulting from the underlying random forests. Please find these plots below.

Questions

• If I take a test set of say 10 datapoints, would the importance of features vary from datapoint to datapoint for computation of that datapoints predict_proba score?
• Taking analogy from CNNs class activation map which varies from datapoint to datapoint, does the ordering and relative importance of each feature remain the same when model runs on multiple datapoints or does it vary?