Sequential | Image viewer for Mac OS X

No, you can't control the alignment or memory location of classes (reference types). You can't even get the size of a class instance in memory.

It is possible to control the size and alignment of structs (and of the fields within them). Structs are value types and work pretty much the same as in C++. If you create an array of a struct, each entry has the size of the struct, and if that is large enough, you could get what you want. But there's no guarantee that the individual entries are really distributed over cache lines. That will also depend on the size and organisation of the cache.

Note also that the address of a managed instance (whether a class or a struct) can change at runtime. The garbage collector is allowed to move instances around to compact the heap, and it will do this quite often. So there is also no guarantee that the same instance will always end up in the same cache line. It is possible to "pin" an instance while a certain block executes, but this is mostly intended when interfacing with native functions and not in a context of performance optimization.

this looks like a simplified version of what the reactor-pool does, in essence. have you considered using that with eg. a maximum size of 1?

https://github.com/reactor/reactor-pool/

https://projectreactor.io/docs/pool/0.2.7/api/reactor/pool/Pool.html

The pool is probably overkill, because it has the overhead of having to deal with multiple resources on top of multiple competing borrowers like in your case, but maybe it could provide some inspiration for you to go further.

All rows need the same data size, of course some values can be empty in csv.

This seems to be a bug in Tensorflow 2.7 when using model.save combined with the parameter save_format="tf", which is set by default. The layers RandomFlip, RandomRotation, RandomZoom, and RandomContrast are causing the problems, since they are not serializable. Interestingly, the Rescaling layer can be saved without any problems. A workaround would be to simply save your model with the older Keras H5 format model.save("test", save_format='h5'):

This is a good question... and I would have hoped there would be a practical guide somewhere. One could question if SO would be a good place to ask this question, but regardless, here's my attempt to summarize the various scale_color_*() and scale_fill_*() functions built into ggplot2. Here, we'll describe the range of functions using scale_color_*(); however, the same general rules will apply for scale_fill_*() functions.

There are 22 functions in all, but happily we can group them intelligently based on practical usage scenarios. There are three key criteria that can be used to define practically how to use each of the scale_color_*() functions:

1. Nature of the mapping data. Is the data mapped to the color aesthetic discrete or continuous? CONTINUOUS data is something that can be explained via real numbers: time, temperature, lengths - these are all continuous because even if your observations are 1 and 2, there can exist something that would have a theoretical value of 1.5. DISCRETE data is just the opposite: you cannot express this data via real numbers. Take, for example, if your observations were: "Model A" and "Model B". There is no obvious way to express something in-between those two. As such, you can only represent these as single colors or numbers.

2. The Colorspace. The color palette used to draw onto the plot. By default, ggplot2 uses (I believe) a color palette based on evenly-spaced hue values. There are other functions built into the library that use either Brewer palettes or Viridis colorspaces.

3. The level of Specification. Generally, once you have defined if the scale function is continuous and in what colorspace, you have variation on the level of control or specification the user will need or can specify. A good example of this is the functions: *_continuous(), *_gradient(), *_gradient2(), and *_gradientn().

We can start off with continuous scales. These functions are all used when applied to observations that are continuous variables (see above). The functions here can further be defined if they are either binned or not binned. "Binning" is just a way of grouping ranges of a continuous variable to all be assigned to a particular color. You'll notice the effect of "binning" is to change the legend keys from a "colorbar" to a "steps" legend.

The continuous example (colorbar legend):

The code doesn't work completely properly. Would you please explain my mistakes?

You were quite close. The main problem is that the return keyword in front of the recursive calls terminates the for loop and the entire f function prematurely. This will cause it to visit only the nodes on the first possible branch, not all of them.

The other issue is that branches might be empty at the end of the function, yet you still access [0][0]. Instead return the entire array from f, and access the first tuple on in getResult.

These two small fixes already make the function work¹:

Here's my current understanding, which could be incomplete or incorrect. A verification would be appreciated.

C++20 renamed strongly happens before to simply happens before, and introduced a new, more relaxed definition for strongly happens before, which imposes less ordering.

Simply happens before is used to reason about the presence of data races in your code. (Actually that would be the plain 'happens before', but the two are equivalent in absence of consume operations, the use of which is discouraged by the standard, since most (all?) major compilers treat them as acquires.)

The weaker strongly happens before is used to reason about the global order of seq-cst operations.

This change was introduced in proposal P0668R5: Revising the C++ memory model, which is based on the paper Repairing Sequential Consistency in C/C++11 by Lahav et al (which I didn't fully read).

The proposal explains why the change was made. Long story short, the way most compilers implement atomics on Power and ARM architectures turned out to be non-conformant in rare edge cases, and fixing the compilers had a performance cost, so they fixed the standard instead.

The change only affects you if you mix seq-cst operations with acquire-release operations on the same atomic variable (i.e. if an acquire operation reads a value from a seq-cst store, or a seq-cst operation reads a value from a release store).

If you don't mix operations in this manner, then you're not affected (i.e. can treat simply happens before and strongly happens before as equivalent).

The gist of the change is that the synchronization between a seq-cst operation and the corresponding acquire/release operation no longer affects the position of this specific seq-cst operation in the global seq-cst order, but the synchronization itself is still there.

This makes the seq-cst order for such seq-cst operations very moot, see below.

The proposal presents following example, and I'll try to explain my understanding of it:

If the hyperspectral dataset is given to you as a large image with many channels, I suppose that the classification of each pixel should depend on the pixels around it (otherwise I would not format the data as an image, i.e. without grid structure). Given this assumption, breaking up the input picture into 1x1 parts is not a good idea as you are loosing the grid structure.

I further suppose that the order of the channels is arbitrary, which implies that convolution over the channels is probably not meaningful (which you however did not plan to do anyways).

Instead of reformatting the data the way you did, you may want to create a model that takes an image as input and also outputs an "image" containing the classifications for each pixel. I.e. if you have 10 classes and take a (145, 145, 200) image as input, your model would output a (145, 145, 10) image. In that architecture you would not have any fully-connected layers. Your output layer would also be a convolutional layer.

That however means that you will not be able to keep your current architecture. That is because the tasks for MNIST/CIFAR10 and your hyperspectral dataset are not the same. For MNIST/CIFAR10 you want to classify an image in it's entirety, while for the other dataset you want to assign a class to each pixel (while most likely also using the pixels around each pixel).

Some further ideas:

• If you want to turn the pixel classification task on the hyperspectral dataset into a classification task for an entire image, maybe you can reformulate that task as "classifying a hyperspectral image as the class of it's center (or top-left, or bottom-right, or (21th, 104th), or whatever) pixel". To obtain the data from your single hyperspectral image, for each pixel, I would shift the image such that the target pixel is at the desired location (e.g. the center). All pixels that "fall off" the border could be inserted at the other side of the image.
• If you want to stick with a pixel classification task but need more data, maybe split up the single hyperspectral image you have into many smaller images (e.g. 10x10x200). You may even want to use images of many different sizes. If you model only has convolution and pooling layers and you make sure to maintain the sizes of the image, that should work out.

I managed to make this function that does exactly what you want.