On the other hand, for Stochastic Gradient Descent (SGD), you compute the weight update serially for each individual image; so SGD cannot be parallelized and is therefore slower than batch gradient descent. Also SGD weight updates are noisy compared to batch weight updates, since each is computed from a single example. Batch updates, being averaged over (say) bs=64 examples, are sqrt(64) = 8 times less noisy.
1 Like
I think Jeremy’s ideas are good. In fact as this decease spreads by touch it should be compulsory to wear when shopping for food, for the shelf life of most products is less that 72 hrs.
Maybe I am bit dramatic
How can one determine the available options for the various functions and methods offered in the fastai2 library?
There are many settings that are done by string and others that are given by some imported keyword.
How can I know, for a given context, which are my options? (Shift-Tab usually won’t help here).
For example,
- in the Datablock API, how can I know which block types are available?
- Which transforms are available?
- Which architectures?
- In the callbacks, which monitor values are acceptable?
Etc, Etc.
The ability to quickly know what’s available will improve significantly the learning rate for fastai, I think…
Thanks!
when i add fastai2 in the requiremetns.txt file it fails to build that environment on heroku as it’s more than 500MB build (maybe this is limitation on free account). Is there a way to only install the required bits from fastai2 so it’s less than 500MB?
3 Likes
The documentation is what you’re looking for here.
For instance a simple ‘Datablock’ search gets you everything about it, the blocks, methods etc.
DataBlock Docs
@miwojc what is the size of it? Do you need all the functionality?
Make a new thread and tag me in it, along with the people who hearted your post. Although, please note that I haven’t done too much with the Heroku platform as I run my own servers.
[Possible suboptimal option:] Your experience would probably be best with the entire install, but you probably don’t actually need any of the notebook or fancy parts - and might be fine if you are purely running a deployment you don’t intend on working on anymore - but I think it would defeat the learning experience in the class.
You can also look at https://forums.fast.ai/t/setup-help/65529 to see how other platforms have setup their images. If you get it working, post your image script there too so others can use it who would prefer Heroku. If you just want a working solution though, the paperspace platform may be a comparable substitute: https://forums.fast.ai/t/platform-paperspace-free-paid-options/65515 .
2 Likes
@christophercao1 if you’re interested in going deeper on the subject, I gave a talk last year called “CUDA in Your Python” that goes into different ways of integrating with the GPU from Python code, it may be useful to you https://www.youtube.com/watch?v=c9Ezk6d3IuY
2 Likes
yeah, on heroku i am only interested in deploying the app, similar to what Jeremy showed with binder. so i have trained model, exported to .pkl file. now i want to deploy the app to heroku using stramlit. the issue i have is that heroku limits build size to 500MB. When i put fastai2 in requirements.txt, heroku will build with pip install fastai2. Which installs everyting fastai2 has (nlp, tabular, collabe, vision) and all dependencies which is more than 500MB. I found the way to install torch for cpu only which was just below the limit, but when i uploaded export.pkl file it went up above limit again 
What i want to find out - if it’s possible to do something like pip install --no-deps on heroku, but so far i haven’t managed to find that…
2 Likes
I created a tree classifier comparing two similar types of trees. They are a little too similar and model didn’t work well. I change one of the trees, updated the code throughout, and restarted the kernal. But when I run:
fns = get_image_files(path)
fns
the output remains the old jpgs of trees. Does anybody know why the model’s not downloading the new tree_type?
1 Like
While running 04_mnist_basics clean on Colab I have had a wired behavior. I run the whole notebook without an issue. However, when I try to run, just before Jargon recap title, the simple_net, learn = Learner…, and learn.fit, I get the following error:
Is it correct that will apply augmentation each time when asking about data by data loader ? Can we define the probability for each transformation?
I have a problem in Google colab, when I write: path= Path(‘images’) for example it gives me an error saying path is not defined. What did I miss?
Is it saying that Path is not defined?
Maybe you are missing
from fastai2.vision.all import *
The Darkest Hour is Just Before Dawn
In a rough visual model of epidemic evolution (see the figure below) the cumulative distribution function (CDF) of deaths (blue curve) follows a logistic function. On a given day the CDF is the total number of deaths since the beginning of the epidemic, normalized by the total number of deaths at the end of the epidemic. So toward the end of the epidemic, the CDF approaches 1.
The United States is currently moving through the steeply rising part of the CDF (blue curve).
The derivative of the logistic function is the distribution of deaths, which has a peaked, symmetric shape. Its value at a given day is the deaths on that day, normalized by the total number of deaths in the epidemic
Bending the curverefers to the falling away of the slope of the CDF (blue curve) after the peak in daily deaths (red curve) at T = 0.0.
Note that the
peak in daily deaths(red curve) happens exactly when the CDF is rising most rapidly, i.e. when the blue curve reaches its steepest slope!
That’s why the epidemiologists tell us that “social distancing is working” even though we can’t see it in the numbers yet.
5 Likes
I think we are saying effectively the same thing. I am saying the function for the daily fraction of total deaths is a logistic distribution function, and was clarifying that it isn’t an exponential like the OP originally said.
For my web deployment application I used a bird Kaggle dataset found here
You can find my web app here: https://birds.smbtraining.com/bird_name
Here is a URL to try it out: https://bdn-data.s3.amazonaws.com/blogs.dir/112/files/2017/07/Loon-by-Nick-Leadley.jpg
Just paste it into the URL box.
10 Likes
Nice job, it looks great!
1 Like
Not yet… Will do.
1 Like
although it would need some caution or awareness that you might create a feedback loop?