DataLoader actually returns a LIST of x tensors and y tensors, not a tensor of tensors (this would be impossible unless the x and y dimensions were the same, of course).
This is why you’re amazing. You don’t just teach the framework. You teach the logic, the intuition, the best practices, and the philosophy behind deep learning. Thanks.
I am gonna use deep learning and pytorch for an NLP-related project that I am in. I have been looking for tutorials and there are thousands of them available out there! After I got bored with two of them, gave a try to yours, liked it very much! I didn't get bored, you explained details very well with fine wording. I even *finally* started using jupyter thanks to this tutorial. Now, I am going to the third episode!
I wanted to express my gratitude for sharing that incredibly insightful and valuable deep learning tutorial on neural networks. It has been a game-changer in my understanding of the topic, and I truly appreciate your informal guidance. Thank you!
The 'Base 8' as you say is because a byte has 8 bits of data. When working directly with hardware (particularly micro controllers) the most efficient way to execute or parse data is dependant on the processor itself though it will always be in multiples of 8. If that makes sense.... p.s. love the channel!!!
I was taking the fastai deep learning course in which Jeremy Howard uses fastai library which is built on pytorch. Your tutorials are really helping me to get a better understanding of the core pytorch stuff that was used in that course. Thank you😄
Great vid Harrison thanks. Regarding choosing batch size, yes typically the larger the batch size, the faster your model converges. But setting your batch size too large will result in an out_of_memory error. So it's important to calculate the largest batch size possible without exceeding your CUDA memory limit. But one thing I've never gotten a straight answer is this: "What's the largest batch size you can choose, as a function of 1) your particular GPU memory size, 2) the size (MB, KB, etc) of your training examples (maybe GBs if training very large images), and 3) the size of your model parameters all three of these must fit inside your GPU memory. So in theory there should be a formula to calculate the largest possible batch size. Something like GPU_mem_size - model_weights_params_size = remaining memory for training samples. Then take that value and divide it by the typical size of a single training sample (image file or whatever). The result of this division, call it n, is theoretically the largest batch size you can fit in your GPU. Then you would probably round down to the nearest multiple of 8. I'm probably leaving something out of this equation, but that's the general idea. Any thoughts on a straightforward approach to calculating this theoretical max batch size? Thanks Harrison (or anyone else who happens to know the answer)
You can display every image one by one (later on you will probably use a grid or something) like this: trainset = torch.utils.data.DataLoader(train, batch_size=10, shuffle=True) testset = torch.utils.data.DataLoader(test, batch_size=10, shuffle=True) batch_idx = 0 for data in trainset: print('Training batch:', batch_idx) batch_images = data[0] batch_labels = data[1] for image, label in zip(batch_images, batch_labels): plt.imshow(image.view(28, 28)) print('Image label:', label) plt.show() batch_idx += 1 Just like with OpenCV showing a plot using PyPlot leads to a pause in the execution of the program until the window with the image inside is closed. In addition to save typing for future viewers use dict.fromkeys(range(0,10), 0) to create the train_counter_dict. ;)
I understand the purpose of mnist, i am personally tired of them, but I am glad that you take this approach then move on to more complex models, you are awesome, thanks!!!
As Harrison mentioned at 25:29 If anyone is wondering how to create counter_dict using Counter, then here is the code. from collections import Counter ys = [x[1] for x in train] print(Counter(ys)) print(dict(Counter(ys)) #if you want dictionary object
Thank you for the beginner friendly great tutorials. I’m new to learning python but was able to understand and follow so far! Seeing the potential applications of basic concepts me more reason to keep learning!
Hi, there! First of all, the way you explain is amazing. Then, I tried the code of this video and I got this warning: "..\torch\csrc\utils\tensor_numpy.cpp:141: UserWarning: The given NumPy array is not writeable, and PyTorch does not support non-writeable tensors. This means you can write to the underlying (supposedly non-writeable) NumPy array using the tensor. You may want to copy the array to protect its data or make it writeable before converting it to a tensor. This type of warning will be suppressed for the rest of this program." I hope you can help this very beginner.
Could you check the writeable flag via print(arr.flags['WRITEABLE'])? If it’s set to False, you should create a copy of it as suggested by the warning message.
The batch size (more specifically stochastic gradient descent) has nothing to do with overfitting. Lower batch size (stochastic gradient descent in general) can lead to less local minima problems and at the same time you can train much faster because it is better to get a less accurate estimate of the gradient but apply it more often with a bit smaller learning rate (to make up for the "noisy" gradient estimates).
This is controversial. Yann LeCun, sourcing a paper from 2018, says you shouldn't use a batch size bigger than 32. But other papers say you can use bigger and bigger batch sizes without any problem with speed and generalization, as long you use some techniques such as warming up the learning rate (start it small and increase it gradually). Source: Géron, Hands-on Machine Learning, 2nd Edition.
@@rodrigomaximo1034 we agree that the local minima problems (sharpness) and the lack of exploration cause lack of generalization. Is lack of generalization the same as overfitting?
Other way to check if the data is balanced: counter_dict = {} for data in train: if data[1] in counter_dict: counter_dict[data[1]] += 1 else: counter_dict[data[1]] = 1 total = sum(counter_dict.values()) for k,v in counter_dict.items(): counter_dict[k] = v/total print(counter_dict)
Small info for beginners and advanced users who just missed. If your python version is less than 3.6, then you will need to use this piece of code for the last cell. for i in range(10): print('{}: {}'.format(i, counter_dict[i])) To get your python version: import sys print(sys.version)
The whole Base 8 thing... I think it's actually just powers of 2. This helps internally when doing the matrix multiplication operations. I think it's easier to distribute that computational load.
You could also (in theory; don't know whether pytorch provides support for this) make sure that examples of each class is presented to the network with equal probability in each epoch if your dataset is not so well balanced. Its still better to utilize extra data of a class if it has significant more examples rather than discarding the extra data.
17:26 Correction: data is a list - the first element is a tensor of tensors (the pixel information of 10 samples) while the second element is a tensor of integers representing the labels (the corresponding digits) of those 10 sample images.
@@romariocarvalhoneto7831 batch_size might be a misleading name since, in this video, 10 is not related to the batch size, but rather to the number of labels/outcomes. 🙃
This was super helpful, thank you so much. Is it nessecary to suffle the train and test datasets in DataLoader when we've already shuffled them in defining train and test?
Great video. I have learned so much with just these two videos. I do have question. Regarding in second: 21:00. Why is it that when you print (y), you get the same number as when you imshow the value that you passed to x (data[0][0])? Does this tells me that x (data[0][0]) and y (data[1][0]) have the same value? I don't quite understand that part.
I like the cut of your jib sir! Great videos, it's nice to see someone confident, knows what they're doing and is a natural teacher. It feels like you're primarily did it to share your knowledge and made videos because that supported the main goal. Sounds silly but there's a lot of guff on RUclips that is first and foremost a video to watch Some feedback though, code review style: You should have explained what that 1 was all about. Everything else made perfect sense, but hand-waving over that sorta left me with a gap where I was getting distracted thinking about wtf it meant and had to rewind. In fact it's still bugging me. Gonna have to look it up. The powers of two thing is 'cause everything is transistors and 0 or 1. So they're even numbers, they align nicely with memory, are faster with less waste etc. Grumbling about magic numbers in other people's code, then putting them in yours.. tsk tsk. Would have been better to make width and height variables to encourage this new generation of ML barbarians to give things decent names. Could have explained that "rectified linear" is just "floor(0)", had to look it up and frowned at the whole of mathematics for their hyperpathalogical superfluoustic jargonization. The stuff about scaling the outputs kinda lost both of us too. Finally, religious issue: embrace functional! You get the view vs reshape thing, but overwrite variables and prefer classes with mutable state. Functions are much cleaner other than passing parameters, better to test, easier to reuse. Might cut down on the copypasta in the AI world if people are reusing functions. Great vids though like I said, pace is great, feels natural and the plot thing and your efficiency almost made me *want* to use a jupyter notebook as my interpreter. I probably would if I wasn't a snob about unit tests 😂
I have 2 questions. 1:what is the ideal state that the data set has to be in before applying the transform.compose() method. here the mnist data was surely clean. 2: while balancing what should be the general deviation percentage amongst all labels?
Good video. Could you please clarify for loop during balance computation? I would expect that data in trainset would be returned in shuffled way but you use it like normal array
Hi sentdex, thanks for pointout out the balance issue. Do you know how to balance the dataset for unsupervised learning where you don't know if it's balanced or not (likely not)?
Are there NN models that are trained to construct (or suggest parameters for) other NN models (e.g., # and types of layers, # of nodes, connectedness, loss functions, etc.) based on the type and size of the data that one has? Is scale the main issue, is there a recurrence problem in even beginning to approach this, would the problem come with generalizing the model enough to take all types of data as an input, or something else?
I have been trying to learn about image classification for a while. I have looked at several methods of classification but what I really need to know to get started is how do I prepare my images for classification. I am photographing tiny sea shells and have a few unique categories but don't know how because all the videos I have seen use either the Iris set, the digital images set, or other established data sets. But I haven't seen any videos on how to use original data. Maybe I haven't looked far enough for videos or other information.
I'm not quite understanding how numbers of layers/neurons can be trial and error past a certain point, for example GPT, that's a lot of parameters to change randomly
Hi I am getting the following error : "stack expects each tensor to be equal size, but got [60000, 28, 28] at entry 0 and [60000] at entry 1" when I try to print the data by : for data in trainset: print(data) break Could u guide me where I am doing wrong?
Xs,ys=data and if ys is printed it returns a kind of list of tensors. But when iterating again in ys using y, the output should be just the label value right ? how does it return all the number of 1s,3s and all. For that case Xs should have been used , but here y has been used. its just confusing me a lot. Why do we iterate through ys instead of Xs?
I have this strange doubt, while I was running the code on colab. first, I ran the total and count dict part, and then in the next cell, I ran the for loop to count the values. Now when I did this I got the answers in 10k and % in 100+. but when I ran the count_dict and for loop in single-cell it worked correctly, I wanna know the reason behind it. Please help me out!
If you encounter an issue during this step (21:11) where the kernel keeps dying, insert this snippet of code before importing matplotlib: """Python import os os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE" """
![ARGENTINA vs. BOLIVIA [6-0] | RESUMEN | ELIMINATORIAS SUDAMERICANAS | FECHA 10](http://i.ytimg.com/vi/C3wPtKNbAOA/mqdefault.jpg)
![Build Your First Pytorch Model In Minutes! [Tutorial + Code]](/img/1.gif)
DataLoader actually returns a LIST of x tensors and y tensors, not a tensor of tensors (this would be impossible unless the x and y dimensions were the same, of course).
Thanks for the correction!
Actually it is not a list it is a DataLoader object. It's more like a generator.
though, if you meant the training examples then you are right those are lists.
This is why you’re amazing. You don’t just teach the framework. You teach the logic, the intuition, the best practices, and the philosophy behind deep learning. Thanks.
Your 30 min videos feels like 5 min while my 1 hr lecture feels like 3 hrs of pure boredom.
"You came back!"
Yeah, no shit, you beautiful human
I am gonna use deep learning and pytorch for an NLP-related project that I am in. I have been looking for tutorials and there are thousands of them available out there! After I got bored with two of them, gave a try to yours, liked it very much! I didn't get bored, you explained details very well with fine wording. I even *finally* started using jupyter thanks to this tutorial. Now, I am going to the third episode!
Loving this series. And please do a neural network from scratch series after this
It’s like you can tell the future
Instead of manually typing out the dictionary for counter_dict you could also do:
counter_dict = {x:0 for x in range(10)}
or even more sexy, use a defaultdict
Thanks a lot for posting the video series in an easy to understand manner with lots of explanation! Many thanks.
I wanted to express my gratitude for sharing that incredibly insightful and valuable deep learning tutorial on neural networks. It has been a game-changer in my understanding of the topic, and I truly appreciate your informal guidance. Thank you!
The 'Base 8' as you say is because a byte has 8 bits of data. When working directly with hardware (particularly micro controllers) the most efficient way to execute or parse data is dependant on the processor itself though it will always be in multiples of 8. If that makes sense.... p.s. love the channel!!!
I was taking the fastai deep learning course in which Jeremy Howard uses fastai library which is built on pytorch. Your tutorials are really helping me to get a better understanding of the core pytorch stuff that was used in that course. Thank you😄
Doing the same here! :-)
One video a day? This is madness!
I love it
having a million subs is cool.
having a million subs who are following you to learn technical & niche content is very cool.
Of course i came back! 🤩
same :D
Just finished all the 8 videos in this playlist. Loved it. Hope you make more of these pytorch videos.
Great vid Harrison thanks. Regarding choosing batch size, yes typically the larger the batch size, the faster your model converges. But setting your batch size too large will result in an out_of_memory error. So it's important to calculate the largest batch size possible without exceeding your CUDA memory limit.
But one thing I've never gotten a straight answer is this:
"What's the largest batch size you can choose, as a function of
1) your particular GPU memory size,
2) the size (MB, KB, etc) of your training examples (maybe GBs if training very large images), and
3) the size of your model parameters
all three of these must fit inside your GPU memory. So in theory there should be a formula to calculate the largest possible batch size. Something like GPU_mem_size - model_weights_params_size = remaining memory for training samples. Then take that value and divide it by the typical size of a single training sample (image file or whatever). The result of this division, call it n, is theoretically the largest batch size you can fit in your GPU. Then you would probably round down to the nearest multiple of 8.
I'm probably leaving something out of this equation, but that's the general idea.
Any thoughts on a straightforward approach to calculating this theoretical max batch size? Thanks Harrison (or anyone else who happens to know the answer)
I am totally new to Deep Learning and Pytorch. Your explanations are awesome, I understand better now! :D Thanks so much!!
You can display every image one by one (later on you will probably use a grid or something) like this:
trainset = torch.utils.data.DataLoader(train, batch_size=10, shuffle=True)
testset = torch.utils.data.DataLoader(test, batch_size=10, shuffle=True)
batch_idx = 0
for data in trainset:
print('Training batch:', batch_idx)
batch_images = data[0]
batch_labels = data[1]
for image, label in zip(batch_images, batch_labels):
plt.imshow(image.view(28, 28))
print('Image label:', label)
plt.show()
batch_idx += 1
Just like with OpenCV showing a plot using PyPlot leads to a pause in the execution of the program until the window with the image inside is closed.
In addition to save typing for future viewers use dict.fromkeys(range(0,10), 0) to create the train_counter_dict. ;)
Man thanks for all of this. Usually I can't follow tutorials but with you its so easy. Keep up the good work :)
I understand the purpose of mnist, i am personally tired of them, but I am glad that you take this approach then move on to more complex models, you are awesome, thanks!!!
As Harrison mentioned at 25:29 If anyone is wondering how to create counter_dict using Counter, then here is the code.
from collections import Counter
ys = [x[1] for x in train]
print(Counter(ys))
print(dict(Counter(ys)) #if you want dictionary object
Similarly you can do:
from collections import Counter
Counter(trainset.dataset.targets.tolist())
You're so correct when you say that data acquisition, preprocessing and all that jazz take up about 90% of the time. and it is sooooo frustrating
You uploaded it fast! I’m sooo happyyy! Next one sooon please!!
Awesome work. Thank you! Keep them coming. It's good for the soul!🙏
Thank you for the beginner friendly great tutorials. I’m new to learning python but was able to understand and follow so far! Seeing the potential applications of basic concepts me more reason to keep learning!
Thank you Harrison for your wonderful tutorial, and getting me started in the right direction!!
one of the best explanations that I have ever seen man thanks
to change editor font size, go to settings > notebook > user preferences >
{
"codeCellConfig": {"fontSize": 24}
}
or whatever.
Hi, there! First of all, the way you explain is amazing. Then, I tried the code of this video and I got this warning: "..\torch\csrc\utils\tensor_numpy.cpp:141: UserWarning: The given NumPy array is not writeable, and PyTorch does not support non-writeable tensors. This means you can write to the underlying (supposedly non-writeable) NumPy array using the tensor. You may want to copy the array to protect its data or make it writeable before converting it to a tensor. This type of warning will be suppressed for the rest of this program." I hope you can help this very beginner.
I am having the same issue.
Could somebody please take a look at this problem? It's quite discouraging for beginners like us.
@@chuckchen2851 Hey, did you fix it somehow?
Could you check the writeable flag via print(arr.flags['WRITEABLE'])?
If it’s set to False, you should create a copy of it as suggested by the warning message.
The batch size (more specifically stochastic gradient descent) has nothing to do with overfitting. Lower batch size (stochastic gradient descent in general) can lead to less local minima problems and at the same time you can train much faster because it is better to get a less accurate estimate of the gradient but apply it more often with a bit smaller learning rate (to make up for the "noisy" gradient estimates).
This is controversial. Yann LeCun, sourcing a paper from 2018, says you shouldn't use a batch size bigger than 32. But other papers say you can use bigger and bigger batch sizes without any problem with speed and generalization, as long you use some techniques such as warming up the learning rate (start it small and increase it gradually). Source: Géron, Hands-on Machine Learning, 2nd Edition.
@@rodrigomaximo1034 we agree that the local minima problems (sharpness) and the lack of exploration cause lack of generalization. Is lack of generalization the same as overfitting?
Jeez man, you're a freakin hero! looking forward for the next video of this series :)
Your Coffee Cup collection is really nice and intriguing.
To change the code font size you should be able to go to settings -> JupyterLab Theme -> Increase Code Font Size
Mate as always awesome, I hope you realise the content you create is better than most of the paid stuff.
Other way to check if the data is balanced:
counter_dict = {}
for data in train:
if data[1] in counter_dict:
counter_dict[data[1]] += 1
else:
counter_dict[data[1]] = 1
total = sum(counter_dict.values())
for k,v in counter_dict.items():
counter_dict[k] = v/total
print(counter_dict)
I obviously came back you are so good!
running into exactly the problems you are talking about. thank you!!!
Can't wait to see Neural Networking with numpy alone!
very good videos to pick up pytorch without too much background talking. thanks.
Small info for beginners and advanced users who just missed.
If your python version is less than 3.6, then you will need to use this piece of code for the last cell.
for i in range(10):
print('{}: {}'.format(i, counter_dict[i]))
To get your python version:
import sys
print(sys.version)
The whole Base 8 thing... I think it's actually just powers of 2. This helps internally when doing the matrix multiplication operations. I think it's easier to distribute that computational load.
good stuff, I m running through your Pytorch tutorials, thank you
It's Rocking Dude!
Just Love what u are doin!
You could also (in theory; don't know whether pytorch provides support for this) make sure that examples of each class is presented to the network with equal probability in each epoch if your dataset is not so well balanced. Its still better to utilize extra data of a class if it has significant more examples rather than discarding the extra data.
for imbalanced datasets you can try oversampling techniques to achieve that, you can try SMOTE or ADASYN for that matter.
Remember troop: usually, it's just usually! 13:39
count={}
for xs,ys in trainset:
for y in ys:
try:
count[int(y)]+=1
except:
count[int(y)]=1
print(count)
Reduced some labor in creating Count dictionary :)
You explain complex stuff so easily
UserWarning: The given NumPy array is not writeable, and PyTorch does not support non-writeable tensors. ....help
same problem over here ... - somebody having a solution?
17:26 Correction: data is a list - the first element is a tensor of tensors (the pixel information of 10 samples) while the second element is a tensor of integers representing the labels (the corresponding digits) of those 10 sample images.
counter_dict = {i:0 for i in range(10)}
batch_size = 10
counter_dict = {x:0 for x in (range(batch_size))}
@@romariocarvalhoneto7831 batch_size might be a misleading name since, in this video, 10 is not related to the batch size, but rather to the number of labels/outcomes. 🙃
musa kurt kardeş, yapıyosun bu sporu
gave a thumbs up just for "ya came back!", but this tutorial deserves the thumbs up anyway :)
this is a great series. thank you very much :)
This was super helpful, thank you so much.
Is it nessecary to suffle the train and test datasets in DataLoader when we've already shuffled them in defining train and test?
I just love your videos. Thanks so much!
you CAME *BACK*
Great tutorial; it's going to save me a lot of time; thank you...... ...... ......
sentdex is the bestman,
Thanks
If you haven't solve the font size problam go to Tools->Settings->Editor and than just pick a font size.
Great video. I have learned so much with just these two videos.
I do have question. Regarding in second: 21:00. Why is it that when you print (y), you get the same number as when you imshow the value that you passed to x (data[0][0])? Does this tells me that x (data[0][0]) and y (data[1][0]) have the same value? I don't quite understand that part.
Yeahhhhhh i came back!!! Its so real. Great tutorials
this course is excellent
So if in the case that 1 would be 45 percent of your outputs, how would you balance it? Just remove the ones out?
Hit CTRL+ or CTRL- to make your fonts bigger or smaller inside of a browser
I like the cut of your jib sir! Great videos, it's nice to see someone confident, knows what they're doing and is a natural teacher. It feels like you're primarily did it to share your knowledge and made videos because that supported the main goal. Sounds silly but there's a lot of guff on RUclips that is first and foremost a video to watch
Some feedback though, code review style:
You should have explained what that 1 was all about. Everything else made perfect sense, but hand-waving over that sorta left me with a gap where I was getting distracted thinking about wtf it meant and had to rewind. In fact it's still bugging me. Gonna have to look it up.
The powers of two thing is 'cause everything is transistors and 0 or 1. So they're even numbers, they align nicely with memory, are faster with less waste etc.
Grumbling about magic numbers in other people's code, then putting them in yours.. tsk tsk. Would have been better to make width and height variables to encourage this new generation of ML barbarians to give things decent names.
Could have explained that "rectified linear" is just "floor(0)", had to look it up and frowned at the whole of mathematics for their hyperpathalogical superfluoustic jargonization. The stuff about scaling the outputs kinda lost both of us too.
Finally, religious issue: embrace functional! You get the view vs reshape thing, but overwrite variables and prefer classes with mutable state. Functions are much cleaner other than passing parameters, better to test, easier to reuse. Might cut down on the copypasta in the AI world if people are reusing functions.
Great vids though like I said, pace is great, feels natural and the plot thing and your efficiency almost made me *want* to use a jupyter notebook as my interpreter. I probably would if I wasn't a snob about unit tests 😂
I have 2 questions. 1:what is the ideal state that the data set has to be in before applying the transform.compose() method. here the mnist data was surely clean. 2: while balancing what should be the general deviation percentage amongst all labels?
Good video. Could you please clarify for loop during balance computation? I would expect that data in trainset would be returned in shuffled way but you use it like normal array
Can the '1' in x.shape be considered as the number of channels???
1 for grayscale, and 3 for RGB??
What if the data is not balanced? what do we need to do? could you explain that too?
Hi sentdex, thanks for pointout out the balance issue. Do you know how to balance the dataset for unsupervised learning where you don't know if it's balanced or not (likely not)?
Are there NN models that are trained to construct (or suggest parameters for) other NN models (e.g., # and types of layers, # of nodes, connectedness, loss functions, etc.) based on the type and size of the data that one has? Is scale the main issue, is there a recurrence problem in even beginning to approach this, would the problem come with generalizing the model enough to take all types of data as an input, or something else?
awesome work brother.
I think for the batch_size they use base 8 because 1 byte is 8 bits.
thanks for the vid as well
One way to increase the font is to use Pycharm instead
Lol xD
for resizing jupyter lab, I pressed ctrl and used the scroll wheel
Working with tensors, can't count to 10. Now, I know you are for real.
Anyone can work with tensors. Counting to 10? whewph. I wouldn't even know where to begin =/
I have been trying to learn about image classification for a while. I have looked at several methods of classification but what I really need to know to get started is how do I prepare my images for classification. I am photographing tiny sea shells and have a few unique categories but don't know how because all the videos I have seen use either the Iris set, the digital images set, or other established data sets. But I haven't seen any videos on how to use original data. Maybe I haven't looked far enough for videos or other information.
You're breath taking
24:37 counter_dict = {k:0 for k in range(10)}
add this to your setting overrides of advance setting editor to increase your font size:
{"codeCellConfig": {
"autoClosingBrackets": true,
"fontFamily": null,
"fontSize": 20,
"lineHeight": null,
"lineNumbers": false,
"lineWrap": "off",
"matchBrackets": true,
"readOnly": false,
"insertSpaces": true,
"tabSize": 4,
"wordWrapColumn": 80
},
"markdownCellConfig": {
"autoClosingBrackets": false,
"fontFamily": null,
"fontSize": 20,
"lineHeight": null,
"lineNumbers": false,
"lineWrap": "on",
"matchBrackets": false,
"readOnly": false,
"insertSpaces": true,
"tabSize": 4,
"wordWrapColumn": 80
},
"rawCellConfig": {
"autoClosingBrackets": false,
"fontFamily": null,
"fontSize": 20,
"lineHeight": null,
"lineNumbers": false,
"lineWrap": "on",
"matchBrackets": false,
"readOnly": false,
"insertSpaces": true,
"tabSize": 4,
"wordWrapColumn": 80
}}
I'm not quite understanding how numbers of layers/neurons can be trial and error past a certain point, for example GPT, that's a lot of parameters to change randomly
something is wrong with the data files from MNIST. They are in a UBYTE format. So your examples don't work. Please help!
So I want to master one, pytorch or tensorflow, which would you advise me to go for sir?
Hi I am getting the following error : "stack expects each tensor to be equal size, but got [60000, 28, 28] at entry 0 and [60000] at entry 1" when I try to print the data by : for data in trainset:
print(data)
break
Could u guide me where I am doing wrong?
A dictionary comprehension would have worked great to define your counter_dict!
counter_dict = {i:0 for i in range(10)}
Xs,ys=data
and if ys is printed it returns a kind of list of tensors. But when iterating again in ys using y, the output should be just the label value right ? how does it return all the number of 1s,3s and all. For that case Xs should have been used , but here y has been used. its just confusing me a lot. Why do we iterate through ys instead of Xs?
you used MNIST dataset but what if I have JSON data locally and how to load such local dataset
I love how not even one percent disliked this video... That is quite uncommon to see on YT...
I have this strange doubt, while I was running the code on colab. first, I ran the total and count dict part, and then in the next cell, I ran the for loop to count the values. Now when I did this I got the answers in 10k and % in 100+. but when I ran the count_dict and for loop in single-cell it worked correctly, I wanna know the reason behind it. Please help me out!
Thanks Harrison .
Why use Jupyter Lab?
Why the picture of 3 is in color if the dataset is black and white, I wonder?
Like in the beginning so i don't forget! :-)
Your a hero
My gpu ram is so small I can only fit 1 batch size is that even a good idea?
sorry i'm late to the problem but would a low tech solution work? can you just zoom your browser in?
shouldn't we use stratified sampling for splitting?
If you encounter an issue during this step (21:11) where the kernel keeps dying, insert this snippet of code before importing matplotlib:
"""Python
import os
os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE"
"""
hey Sentdex: What if we use reinforcement deep q learning model and needed to input the data from the sensors.