augmentations | Rails plugin that provides a simple API | Application Framework library

I am attempting to train Deeplab Resnet V3 to perform semantic segmentation on a custom dataset. I had been working on my local machine however my GPU is just a small Quadro T1000 so I decided to move my model onto Google Colab to take advantage of their GPU instances and get better results.

Whilst I get the speed increase I was hoping for, I am getting wildly different training losses on colab compared to my local machine. I have copied and pasted the exact same code so the only difference I can find would be in the dataset. I am using the exact same dataset except for the one on colab is a copy of the local dataset on Google Drive. I have noticed that Drive orders file differently on Windows but I can't see how this is a problem since I randomly shuffle the dataset. I understand that these random splitting can cause small differences in the outputs however a difference of about 10x in the training losses doesn't make sense.

I have also tried running the version on colab with different random seeds, different batch sizes, different train_test_split parameters, and changing the optimizer from SGD to Adam, however, this still causes the model to converge very early at a loss of around 0.5.

Here is my code:

The answer on Confused about static dictionary in a type, in F# finished with one advice: and just in general: try to use fewer classes and more modules and functions; they're more idiomatic in F# and lead to fewer problems in general

Which is a great point, but my 30 years of OO just don't want to give up classes just yet (although I was fighting against C++ like crazy when we moved away from C...)

so let's take a practical real world object:

I am trying to patch the augmentor function from the clodsa package to read custom named .json files instead of adhering to their format, so instead of /annotations.json make new_name.json:

I am new to typescript and this might be a noob question.

I want to extend global variable provided by nodejs.

As per this blog I wrote this code and it is working

I am quite new to the field of semantic segmentation and have recently tried to run the code provided on this paper: Transfer Learning for Brain Tumor Segmentation that was made available on GitHub. It is a semantic segmentation task that uses the BraTS2020 dataset, comprising of 4 modalities, T1, T1ce, T2 and FLAIR. The author utilised a transfer learning approach using Resnet34 weights.

Due to hardware constraints, I had to half the batch size from 24 to 12. However, after training the model, I noticed a significant drop in performance, with the Dice Score (higher is better) of the 3 classes being only around 5-19-11 as opposed to the reported result of 78-87-82 in the paper. The training and validation accuracies however, seem to be performing normally, just that the model does not perform well on test data, I selected the model that was produced before overfitting (validation loss starts increasing but training loss still decreasing) but yielded equally bad results.

So far I have tried:

1. Decreasing the learning rate from 1e-3 to 1e-4, yielded similar results
2. Increased the number of batches fed to the model per training epoch to 200 batches per epoch, to match the number of iterations ran in the paper since I effectively halved the batch size - (100 batches per epoch, batch size of 24)

I noticed that image augmentations were applied to the training and validation dataset to increase the robustness of the model training. Do these augmentations need to be performed on the test set in order to make predictions? There are no resizing transforms, transforms that are present are Gaussian Blur and Noise, change in brightness intensity, rotations, elastic deformation, and mirroring, all implemented using the example here.

I'd greatly appreciate help on these questions:

1. By doubling the number of batches per epoch, it effectively matches the number of iterations performed as in the original paper since the batch size is halved. Is this the correct approach?

2. Does the test set data need to be augmented similarly to the training data in order to perform predictions? (Note: no resizing transformations were performed)

For object detection, I'm using detectron2. I want to fix the input image size so I made my customized dataloader:

I was reading the Deep Learning in Python book and wanted to understand more on the what happens when you define the steps_per_epoch and batch size. The example they use consists of 4000 images of dogs and cats, with 2000 for training, 1000 for validation, and 1000 for testing. They provide two examples of their model. One with image augmentation one without. I am confused on why they changed the batch size between the two cases.

I have understood the process is that 1 epoch is 1 pass over the entire training dataset. The batch size determines how many of the images are shown per one step. When we change the batch size, we change the number of images to be learned from. For their first example with 2000 images in training, a batch size of 20, 100 steps per epoch is logical and what they use. It will take 100 steps to see 2000 images, completing an epoch. On their next example, they implement more augmentations than re-scaling the image (6 total of rotation changes, zooms, shears, etc), the batch size increases to 32, but they left steps per epoch at 100. I assumed with the increase in batch size, steps_per_epoch declines and in this case to be 63 (round up from 62.5). Why do they leave steps_per_epoch the same in this case? In the end, does the model not see some training data or see too much data given the rounding issue?

A bit about my specific use case: I have a plugin that's designed to be integrated with Unreal Engine projects, and in order to demonstrate how to do this to users of the plugin, I've integrated it with one of Unreal's freely available sample games as an example. The integration is very specific to the game, as it does things like modifying the menu to allow the user to interact with my plugin easily.

However, in an ideal world I'd like to be able to:

1. Provide integrations with the sample game across multiple different Unreal Engine versions. At a minimum this would include 3 currently existing versions of Unreal (4.24, 4.25 and 4.26), but would extend to potentially N different future versions. This essentially makes the integration code "boilerplate", as it's required for functionality in each sample game version, but doesn't vary at all across versions.
2. Be able to maintain the bulk of this integration code from one place. I don't want to have to make identical modifications in each of the sample game integrations every time I change something, as juggling multiple parallel codebases like this is a lot of work and increases the probability of bugs.

This is almost a problem that could be solved with code patches: the integration code fits into the same functions/classes in the same files regardless of which version of the sample game I'm using. However, the contents of the sample game files themselves aren't exactly the same across engine versions, so a patch that says "insert this hunk into this file at this line" won't always get it right. There is also the theoretical possibility that a more substantial change is introduced into the sample game in future which could require me to change my integration in that case (though this hasn't happened yet - changes appear to be minimal across minor engine versions).

What is the best way to attack this problem? One particularly horrible way I can think of (but one which demonstrates the concept) would be to separate each chunk of the integration into a separate file, and then #include "Chunk1.inc", #include "Chunk2.inc", ... directly into the relevant classes and functions in each version of the sample game.

I'm new to machine learning and program. Now I'm trying to develop YOLACT AI using my own data. However, when I run train.py, I get the following error and cannot learn. What can I do to overcome this error?`