Hi everyone! The code and explanations behind this video are here - github.com/VikParuchuri/zero_to_gpt/blob/master/explanations/dense.ipynb . You can also find all the lessons in this series here - github.com/VikParuchuri/zero_to_gpt .
Hi, thank you so much for your really helpfull videos and explanations. i wanted to ask you about the code for a CNN that i saw in your github, is there a video of it? or is part of other video already published? i can´t find it. thanks.
If you went as far as to do it from scratch, then you might have just as well done it in anything other than Python. All that Python has in this field is libraries. Because it's not like Python has any advantage when it comes to linear algebra. If anything, it will be slower than most other things.
Does anyone know why for the mse function he does (actual - predicted) ** 2, but for mse_grad he writes predicted - actual? Wouldn't it matter whether you do (predicted - actual) or (actual - predicted) in the mse_grad function as this will change how you update your parameters?
Hi Vik! Please help with football video information.
Can you please tell me how to start predicting future matches? I prepared the schedule for next week and added it to the "matches.csv" file In the data, I have moving averages for the following data: "xg", "xga", "gls", "sh", "sot", "g/sh", "g/sot", "dist", "fk", "pk", "pkatt", "npxg" ", "npxg/sh". How can I run prediction now? Thanks in advance for your reply. P.S. I am writing through google translator, I hope you understand what I mean.
I think you want help on this video, right? - ua-cam.com/video/0irmDBWLrco/v-deo.html You basically take all of the training data up to the last day (today), then generate a prediction. The prediction will be for the next match. You'll need to do it without backtesting, and without dropping any rows from the end of the training data. I talk a little bit about the steps at the very end of this video - ua-cam.com/video/egTylm6C2is/v-deo.html
Vik, thanks for the reply. Yes, I watched the NBA video and noticed that at the very end there is information about what I am trying to figure out😊 But I watch the video through a translator and there is a possibility that he translates the speech from the video incorrectly, so this process is not clear to me yet. Is it possible for you to record a short video demonstrating how to do this? And what's the price? Thanks for the info 👍
Many thanks for this very comprehensive course, but I'm having a problem - when I run the program I get the following output: Epoch: 0 Train MSE: nan Valid MSE: nan Epoch: 1 Train MSE: nan Valid MSE: nan Epoch: 2 Train MSE: nan Valid MSE: nan (Truncated example). Printing 'loss' and 'epoch_loss' seperately, the following is output (after approx 4500 lines of numeric output): epoch_loss: 4789.425964748963 loss: [[-69.19245634] [-62.18411289] [ nan] [ nan] [ nan] [-63.21175768] [-63.18508556] [-62.17785529]] epoch_loss: nan loss: [[nan] [nan] Running both my own code from following this video plus your code from Github, same results. Any ideas?
There can be a lot of potential reasons for nan loss, so it's hard to know for sure. Basically, some value (weight, prediction, gradient) is too large for the numpy data type. Things I would try: - Lower the learning rate - It's possible your system defaults to a float format with a lower range - check the dtype of the numpy arrays, and switch to float64 if the dtype is something else - Are you initializing the weights the same way I am? You could try initializing them to smaller values than I did to see if anything changes. - Make sure you're using mse_grad as the loss, not mse
Hi everyone! The code and explanations behind this video are here - github.com/VikParuchuri/zero_to_gpt/blob/master/explanations/dense.ipynb . You can also find all the lessons in this series here - github.com/VikParuchuri/zero_to_gpt .
Hi, thank you so much for your really helpfull videos and explanations.
i wanted to ask you about the code for a CNN that i saw in your github, is there a video of it?
or is part of other video already published?
i can´t find it.
thanks.
Thank you for your thoroughly explanation, I have a question, how to decide which matrix to transpose during backpropagation?
this videos have so much value! Thank you!!
Glad you like them!
If you went as far as to do it from scratch, then you might have just as well done it in anything other than Python. All that Python has in this field is libraries.
Because it's not like Python has any advantage when it comes to linear algebra. If anything, it will be slower than most other things.
You are a fantastic teacher! Subscribing. Love your pace and explanation of what and why you are doing something.
beautiful video Vik! starting my term project this week and NN are a main method i’ll be using! thank you!
Thanks, Cade! Good luck :)
Great content as usual!
Does anyone know why for the mse function he does (actual - predicted) ** 2, but for mse_grad he writes predicted - actual? Wouldn't it matter whether you do (predicted - actual) or (actual - predicted) in the mse_grad function as this will change how you update your parameters?
How would you calculate the bias?
WTF at 5:09 happened mse has only two arguments how it's taking weight and bias as input?
Thank you for your video, I believe there is an error 31:25 where you define the MSE function. Shouldn't you be taking the mean of the error squared?
I'm also having the problem with ModuleNotFoundError: No module named 'csv_data'
at the very beginning
Good job bro, JESUS IS COMING BACK VERY SOON; WATCH AND PREPARE
Thanks! One question, is this video complete?
please i want to download the dataset that you are using in this code. kindly share link of this dataset thanks you
Does this code work if there is no hidden layer? Only an input and output layer?
could you show me the link to get the dataset you used please...
i love math
@41:49 the link to the document pls
Hi Vik!
Please help with football video information.
Can you please tell me how to start predicting future matches? I prepared the schedule for next week and added it to the "matches.csv" file
In the data, I have moving averages for the following data:
"xg", "xga", "gls", "sh", "sot", "g/sh", "g/sot", "dist", "fk", "pk", "pkatt", "npxg" ", "npxg/sh".
How can I run prediction now?
Thanks in advance for your reply.
P.S. I am writing through google translator, I hope you understand what I mean.
I think you want help on this video, right? - ua-cam.com/video/0irmDBWLrco/v-deo.html
You basically take all of the training data up to the last day (today), then generate a prediction. The prediction will be for the next match. You'll need to do it without backtesting, and without dropping any rows from the end of the training data.
I talk a little bit about the steps at the very end of this video - ua-cam.com/video/egTylm6C2is/v-deo.html
Vik, thanks for the reply.
Yes, I watched the NBA video and noticed that at the very end there is information about what I am trying to figure out😊
But I watch the video through a translator and there is a possibility that he translates the speech from the video incorrectly, so this process is not clear to me yet.
Is it possible for you to record a short video demonstrating how to do this?
And what's the price?
Thanks for the info 👍
My dude, I love you!
Peace! ❤
1:06:17 We should not update the weights until we find the grads for all layers. In your code its a mistake, pls correct it.
Wow thank❤
where bias came from? 11
it's like a weight, it's adjusted in training. 11 is just an example (for video)
Python is a baby's toy
Many thanks for this very comprehensive course, but I'm having a problem - when I run the program I get the following output:
Epoch: 0 Train MSE: nan Valid MSE: nan
Epoch: 1 Train MSE: nan Valid MSE: nan
Epoch: 2 Train MSE: nan Valid MSE: nan
(Truncated example).
Printing 'loss' and 'epoch_loss' seperately, the following is output (after approx 4500 lines of numeric output):
epoch_loss: 4789.425964748963
loss: [[-69.19245634]
[-62.18411289]
[ nan]
[ nan]
[ nan]
[-63.21175768]
[-63.18508556]
[-62.17785529]]
epoch_loss: nan
loss: [[nan]
[nan]
Running both my own code from following this video plus your code from Github, same results. Any ideas?
HI! Have you solved it?
There can be a lot of potential reasons for nan loss, so it's hard to know for sure. Basically, some value (weight, prediction, gradient) is too large for the numpy data type.
Things I would try:
- Lower the learning rate
- It's possible your system defaults to a float format with a lower range - check the dtype of the numpy arrays, and switch to float64 if the dtype is something else
- Are you initializing the weights the same way I am? You could try initializing them to smaller values than I did to see if anything changes.
- Make sure you're using mse_grad as the loss, not mse