DELTAS | Inference Code for DELTAS : Depth Estimation | Machine Learning library

Just upgraded the version to Delta.io 1.0.0 and it stopped happening.

The problem here is that the path contains both colons and question marks and you're likely on a Windows system. Windows has made a deliberate decision not to allow these characters in path names, so there's no way to check this file out there.

You have some options:

• Use a different operating system, like macOS or Linux.
• Check the file out using the Windows Subsystem for Linux, which doesn't suffer from these limitations.
• Ask the party responsible for the repository to change the path name or do it yourself in another environment, then check out the fixed repo.
• Ask Microsoft to fix this problem.
• Don't use this repo.

It looks like you made some commits that have the big file, then you added some more commits in which you've replaced the big file with the LFS-indirection file. What this means is that you need to remove the commits that have the big files.

Remember that when you use Git-LFS, you're using two external web sites to store your data. First, you send the big files to some LFS storage site. These big files are not (yet) committed. Then, with the real files safely stored elsewhere, you create and push commits that store only the information needed to retrieve the big files from the LFS storage site. So the commits, in the form of a Git repository, exist somewhere (e.g., on GitHub) but don't have any big files in any of them. Instead, they have small files that have instructions that say don't use this file, use a big file you get with the LFS big-file-swapper-replacer-trick, here's how to get the big file. The LFS wrappers around Git intercept Git's attempt to put the small replacement file into your working tree, and put the big file there instead.

(Note that this means that anyone who clones your repository with regular Git gets, and sees, these weird small indirection files. They must set up Git-LFS themselves so that they can have the same wrapper program intercept the attempt to put the small files out for viewing or editing.)

But if you do this in the wrong order, you first commit the big files. Then you send the big files to a second site, then you remove the big files and put in the small "here's how to get the real file" files, and make another commit that contains these small files. You still have the commit with the big files in it!

Git is designed to add new commits, and not to take old commits away. This makes it very hard to remove any commit that has a big file in it. To do that automatically requires specialized tools. Fortunately, GitHub already have instructions and tools for you, to tell you how to do this. See them here. The short version is to use git lfs migrate, if you have a modern Git-LFS. See also this other GitHub page.

There is misprint in instructions on https://sites.google.com/junsun.net/how-to-run-cuttlefish/home. AOSP building commands should be used: $ source build/envsetup.sh

As explained in "Cloning succeded but checkout failed due to invalid path. What is the path problem?", Windows cannot handle any file named AUX.

I suspect this is the issue here, considering your output:

If anyone stumbles upon this with any of the same queries, here are the answers:

1 - Sigmoid should work but it restricts the output between 0 and 1 and since this was a regression problem, you could design your problem to get a normalized output and scale it to the right number. Or just use ReLU.

2 - Yes. The type of activation on the output node changes the type of output you get.

3 - Of course it will. The derivative of the activation gradient is multiplied with the backpropagating global gradient to give the local gradient. Changing the activation function will change the derivative too.

4 - Maybe try to understand the Math behind backpropagation. It's basic chain rule and once you get it, you won't get confused like this.

A good real-world analogy is tricky here, but suppose you have a cement delivery service. You delivered, to some customer, 1000 cubic yards of cement, which have now set. Your customer calls you up and says they did not want that much cement: they only wanted ten cubic yards. So you have your trucks deliver them another 10 yards of cement. Did you improve their situation, or just make it worse?

Now, the reason for the above analogy is simple enough: Git commits only ever add new stuff to the repository. If you put a giant file into some commit, that giant file is in that commit forever. If someone didn't want that giant file, and you make a new commit that lacks the giant file, well, now you have two commits: one with the giant file, and one without. The second commit depends on, and therefore requires, the first commit.

No matter what else you do, if you keep adding on new commits, you're just making everything worse. What you need is not more cement (er, commits). You need some kind of jackhammer and cement-removal service (something to remove commits from your repository).

That's not all, though! While we already see that listing a file in .gitignore has no effect on any existing commit—no existing commit can ever be changed, at all; you just have to get them hauled away and stop using them entirely to get rid of them—it also has no good effect on new commits, at least not yet. If a file is tracked, it has no effect on that file in future commits either.

So:

• Your first job is to remove some commit(s). We generally do this with git reset. The git reset command is destructive! Be very careful here: you probably should work on a clone of your repository, rather than the original.

• Then, having removed the bad commit(s), you may need to explicitly remove the large files, so that they are no longer tracked. A file is tracked, in Git, if the file is in Git's index. How do you know if a file is in Git's index? Well, you know because it's tracked. This is kind of a circular problem, as you can see. :-) But if you definitely want it not-tracked, you can run:

I have tested this using Django 3.2.2 and a Postgres docker container, and I was able to simply save the date string (e.g. "2007-02-17") without any extra type conversions. When retrieving the saved model, Django returns the date field as a datetime.date object.

Your command line shows that you can create a branch for the commit with