Nice video Patrick and big congratulations on 10k subscribers, there's been a lot hard work for you to get to that point! I'm sure the best is yet to come too :) I know we make quite similar videos, but I am very happy that is the case because it drives me to make better videos and I learn a lot from you as well 👊 Also the more people doing videos about PyTorch, TensorFlow and machine learning in general the better it will be for people wanting to learn about these things which is ultimately the goal.
Thank you! I just discovered your channel a few weeks back. Yes some topics are very similar, but I think this is good. I really enjoy your teaching style, and I hope you keep going! I'm sure you will also gain more followers soon! Would you be interested in some kind of collaboration in the future? Best, Patrick
Great Tutorial. Thanks! Keep doing your jab. I would appreciate more if you could add a small part to the video to explain how you can implement "many to many" case too.
Thank you for you great videos. 15:02 little question : aren't we suppose to initialize the cell_state and the hidden_state (t-1) at each epoch instead of each lstm_cell (inner loop) ? Otherwise, the cell_state which is supposed to play the memory role will be useless... Thanks !
from the out, _. The _ has the hidden state of last time step, no need to use the out in the forward propagation{out[:,-1,:]}. For lstm out, (hn, cn) = self.lstm(x,(h0,c0)) and for rnn/gru out, hn = self.rnn(x,h0) / out , hn= self.gru(x,h0). for forward progagation just use this last hidden state out=self.fc(hn). I am referring to 11:40.
@10:33, the shape of output tensor from the RNN model that is mentioned in Pytorch documentation is (seq_length, batch, hidden_size) whereas @11:03 the shape of output tensor from the RNN model is indicated as (batch_size, seq_length, hidden_size) . Which one is correct ? The output shape that is mentioned in the Pytorch documentation or the one that is in this tutorial ?
Both can be correct, by default it’s the one mentioned in the docs. But we use the argument batch_first=True here which turns it around. I usually prefer to have the batches as first dimension
If the sequence I am feeding the lstm is not the whole item. E.g I want to feed an lstm a sequence of 11 frames, one current frame and 10 previous frames. The total number of frames in the item is ~155. Each batch of 11 frames needs to output 1 frame. I want to use the hidden state of previous input to aid the learning of the next frame
why did you take the first output of the RNN and then do [:-1:] on it, why didnt you just take the second output of the RNN, which is the hidden state of the last sequence element?
This is what a teacher sounds like when the first grader still doesn't get it after explaining 5 times, but if he starts yelling the parents will complain
i tried hard but my loss doesn't reduce...my code is same as yours. when i run your code its good but when i run my code, loss remains same. do you know why so?
your videos are really helpful. For some weird reason, the torchsummary summary states that there are no learnable parameters in my rnn layers.. that must be a bug..or am i doing it wrong? XD
Great tutorial sir, thank you... i feel this one is also intermediate friendly... since i am moving from keras. BTW outside of the tutorial, what is the IDE theme, that looks great i like it !
i have a question, loss = criterion(outputs, labels) expected scalar type Long but found Float. Does anybody can help me to solve this problem? thank you.
At 10:29 the documentation says that the output has shape [seq_len, batch, hidden_size] but at 11:00 you write that the output has shape [batch_size, seq_length, hidden_size] and proceed to code accordingly. Can you please confirm whether this is a bug in the video and code?
Ah, I figured it out. At 6:25 we see that setting batch_first = True causes the output shape to become [batch_size, seq_length, hidden_size], so the video and code are correct.
Can someone please help me with an LSTM neural network model? I have to use the phm8 NASA dataset. I have preprocessed the data but I am not sure how to proceed. Please let me know!
Thanks for the great videos :) I have a quick question. At 4:52, why do we define self.hidden_size and self.num_layers, but don't do the same with the other inputs of the __init__ function?
It's very nice tutorial i have ever come across sir Could you make a video on Signature Verification by using ResNet neural network and triplet loss functions? please must reply me. Tha nk you
Hi Sir: just notice in your github of pytorchtutorial 13_feedforward.py, @ line 98 _, predicted = torch.max(outputs.data, 1). Since this line is under torch.no_grad(), I think _, predicted = torch(outputs, 1) is okay.
The more the code you explain, the more I love this channel. Just amazing. Keep it up.
Thanks 😊 really glad you enjoy it
I love you. Very clear explanation. I have been looking for this content for a while
thank you!
This video gives me a very clear picture of implementing RNN with Pytorch. I really appreciate it!
Man.. You are really good in explaining.. Finally understood RNN, LSTM and GRU implementation from your video and the official documentation.
Thanks so much! I watched through all of your pytorch tutorial, and it is the best pytorch tutorial on youtube!!
thanks! Glad you like it :)
1시간 동안 인터넷 돌아다녀도 LSTM 차원 안 맞아서 고생했는데 이거 보고 좀 풀렸네. 굿
Very nice tutorial for Pytorch. Thanks for the initial code.
i am sure you will get 10K very soon and 100K and then 1M :) keep the vibe dude
Thank you :)
Nice video Patrick and big congratulations on 10k subscribers, there's been a lot hard work for you to get to that point! I'm sure the best is yet to come too :) I know we make quite similar videos, but I am very happy that is the case because it drives me to make better videos and I learn a lot from you as well 👊 Also the more people doing videos about PyTorch, TensorFlow and machine learning in general the better it will be for people wanting to learn about these things which is ultimately the goal.
Thank you! I just discovered your channel a few weeks back. Yes some topics are very similar, but I think this is good. I really enjoy your teaching style, and I hope you keep going! I'm sure you will also gain more followers soon! Would you be interested in some kind of collaboration in the future?
Best,
Patrick
@@patloeber For sure, if we find something we can collaborate on! :)
@AladdinPersson Bros so did you collaborate? :)
From freeecodecamp!! Thanks for your content man!
Great! Thank you
your example is clear !! thanks
Your tutorials are incredible! thank you so much!
glad you like it!
I was waiting for you to upload some videos on deep learning ...thanks so much !!!
Glad you like it :)
Great Tutorial. Thanks! Keep doing your jab.
I would appreciate more if you could add a small part to the video to explain how you can implement "many to many" case too.
Danke Schön Patrick
!
Glad you like it!
perfect ourses!!!!!!!
Thank you for you great videos.
15:02 little question : aren't we suppose to initialize the cell_state and the hidden_state (t-1) at each epoch instead of each lstm_cell (inner loop) ? Otherwise, the cell_state which is supposed to play the memory role will be useless...
Thanks !
Bumping this question ;D
Why would you need to retain memory across batches in an image classification task?
great work. Thank you so much
Do you have tutorial of hyperparameters for RNN.? That would be great!!!!
from the out, _. The _ has the hidden state of last time step, no need to use the out in the forward propagation{out[:,-1,:]}. For lstm out, (hn, cn) = self.lstm(x,(h0,c0)) and for rnn/gru out, hn = self.rnn(x,h0) / out , hn= self.gru(x,h0).
for forward progagation just use this last hidden state out=self.fc(hn).
I am referring to 11:40.
You're the best!
@10:33, the shape of output tensor from the RNN model that is mentioned in Pytorch documentation is (seq_length, batch, hidden_size) whereas @11:03 the shape of output tensor from the RNN model is indicated as (batch_size, seq_length, hidden_size) . Which one is correct ? The output shape that is mentioned in the Pytorch documentation or the one that is in this tutorial ?
Both can be correct, by default it’s the one mentioned in the docs. But we use the argument batch_first=True here which turns it around. I usually prefer to have the batches as first dimension
Thanks a lot
you save my life, best wishes for you ^_^
Thanks 😊
This is useful ,thanks a lot
Glad to hear that
REALLY GOOD !
thank you!
super understandable!
great tutorial. thank you!
Thanks !
thank you for the video! how would I go about using the previous hidden state from the previous batch in the new input?
Hmm why would you want to do that?
If the sequence I am feeding the lstm is not the whole item. E.g I want to feed an lstm a sequence of 11 frames, one current frame and 10 previous frames. The total number of frames in the item is ~155. Each batch of 11 frames needs to output 1 frame. I want to use the hidden state of previous input to aid the learning of the next frame
amazinggg stuff!
thanks!
Nice!
The h0 and c0 init should be in __init__, not forward(). right?
I agree, we should want to use memory since we're using RNNs, but here it's for image classification, so in the end memory is useless.
do you have hyperparameter turning？
Optuna. that would be great
Hi, do you think you will make videos on reinforcement learning ?
Yes! It’s a complex topic but I definitely want to make tutorials about this in the future!
@@patloeber Awesome !
great work. Any best practice ideas to figure out the relevant tensor shapes at various steps? That is challenging to me. THanks
You’re tooo goooooodddddd
Thank you!
When u changed the code for LSTM in __init__, didnt you forget to change in forward to LSTM? Seems like you run GRU again.
why did you take the first output of the RNN and then do [:-1:] on it, why didnt you just take the second output of the RNN, which is the hidden state of the last sequence element?
thanks!
glad you like it!
hi, how can we use pygad lib with pytorch? specially for optimization of RNNs
how could it be done with an rgb image? [Batch_size,3,W,H]
More I watch more I want to watch.
That's nice to hear!
This is what a teacher sounds like when the first grader still doesn't get it after explaining 5 times, but if he starts yelling the parents will complain
Great
Good video, but wouldn't the classification be better if you'd connect all the outputs of the recurrent layer to the classes via the linear layer?
i tried hard but my loss doesn't reduce...my code is same as yours. when i run your code its good but when i run my code, loss remains same. do you know why so?
your videos are really helpful. For some weird reason, the torchsummary summary states that there are no learnable parameters in my rnn layers.. that must be a bug..or am i doing it wrong? XD
Great tutorial sir, thank you... i feel this one is also intermediate friendly... since i am moving from keras. BTW outside of the tutorial, what is the IDE theme, that looks great i like it !
i realize Im kind of off topic but does anybody know of a good website to watch new tv shows online ?
I Think I used Monokai or Night Owl
Can you talk about Transformer architecture/seq2seq model???
Hi, can I see the codes of this video from the first ?
Thanks in advance
i have a question,
loss = criterion(outputs, labels)
expected scalar type Long but found Float.
Does anybody can help me to solve this problem? thank you.
At 10:29 the documentation says that the output has shape [seq_len, batch, hidden_size] but at 11:00 you write that the output has shape [batch_size, seq_length, hidden_size] and proceed to code accordingly. Can you please confirm whether this is a bug in the video and code?
Ah, I figured it out. At 6:25 we see that setting batch_first = True causes the output shape to become [batch_size, seq_length, hidden_size], so the video and code are correct.
Yep! Be very careful with this argument and the correct order :D
Can someone please help me with an LSTM neural network model? I have to use the phm8 NASA dataset. I have preprocessed the data but I am not sure how to proceed. Please let me know!
Thanks for the great videos :) I have a quick question. At 4:52, why do we define self.hidden_size and self.num_layers, but don't do the same with the other inputs of the __init__ function?
because they are needed in the forward function as well, and the others aren't
@@patloeber Ok, thanks!
Can you do a tutorial on the Attention and Transformer model.
Should definitely be added in the future
It's very nice tutorial i have ever come across
sir Could you make a video on Signature Verification by using ResNet neural network and triplet loss functions?
please must reply me. Tha nk you
Thanks! Will have a look at that...
Bro what is your opinion about pycaret (ml library)
Havent' used it yet. But maybe it's worth checking it out
Bro .....can you add nn.sequential ....in upcoming videos😁
I'll look into that
hey what screen recorder do you use please tell me
QuickTime (but on Windows or Linux I recomment OBS)
@@patloeber Thanks for replying
@@patloeber Could you tell me which video editor are you using?
DaVinci Resolve
@@patloeber Oh thanks for replying
Can we have some small video on Meshed-,Memory-Transformers pleeeaasseee 🥺🥺🥺
thanks, sir for this tutorial. can you give us a tutorial about object detection (Yolo) in the future thanks, sir.
Yes this is already on my list
Hi Sir: just notice in your github of pytorchtutorial 13_feedforward.py, @ line 98 _, predicted = torch.max(outputs.data, 1).
Since this line is under torch.no_grad(), I think _, predicted = torch(outputs, 1) is okay.
yep I think you are right
why is no one using a real data set, iris, mnist and other dataset serve no purpose in real world.
wasted tutorial. copied from pytorch/example github. that too u removed dropout. what a waste
Thanks a lot ! Great work !
glad you like it!