What’s going on, guys? Hope you’re doing amazing in this video. I want to show you how to implement a bi-directional lsdm in Pi Torch. So what we’re gonna do is? I have some code set up here, which it’s just some very, very simple code, Some training loop check accuracy and the loading demonstate. I said this is from a previous video. If you want to check this out and what we want to do Is we want to create a bi-directional Elysium? And what we’re gonna do is we’re gonna do class. Let’s call it bi-directional. Brm and we’re gonna inherit an end module and we’re gonna do define in it. And, yeah, let me just mention here that so what we have Is We have the amnesty, the set and Emily’s data set we can view as having 28 sequences, which each has 28 features, so the image is a 28 by 28 grid, so we can kind of unroll it row by row and we can feed that to the RNN sequentially so that each we have to Nate Time steps at each time step just takes a single row of the of the image, and then we just have some number of layers that we’re gonna use. We have the hidden size of the RNN, just some hyper parameters and number of classes, learning rate, etc, but we’re gonna send in to the bi-directional Ren. It’s just input size, the hidden size, the number of layers and the number of classes were going to do super BR9 self and then dot in it and then we’re gonna do self that hidden size is just hidden size self that num layers. It’s just naam layers. Now we’re gonna define our LST M and we’re going to N N Dot LST n input size and then hidden size number of layers. That’s just sort of the order where we send it in and then we’re gonna do Batch first equals true for for the M meanest data set, which has batches as they as the first access and then just to make it bi-directional. All we have to do is bi-directional equals true so easy, you know? I guess, and then we’re gonna do self That fully connected. Is Anand that linear, and we’re gonna do hidden size to number of classes and then so we know forward just one thing that we need to keep in mind. Is that well we need to define our hidden state and also the cell state that we’re gonna send into lsdm and, yeah, so let’s do that. H0 will be our hidden state. Torch that zeros and we’re gonna do self dot numb layers and then we’re going to times that by two and why we do that is that so we have seven number of layers, but we also have, let’s see, so we have, so we have several number layers, and then we need to times it by two since we have one of them going forward and one of them going backward, but they’re all going to get concatenated into the same for the same specific hidden state, so we need to just expand this this tensor by two in this axis and then we’re gonna do exit size of zero, just the number of examples we send in into the batch size and then self dot hidden size and then just dot two device. Okay, and actually, we need to do that here. As well, so the nonlinear will take the hidden size by times two. Since for the same argument, we have one going forward, one going backward and they’re gonna get concatenated and that’s a sort of the hidden hidden state for that particular time sequence, just some background from bio direction or LCM. I’m sort of assuming in this video that you know the theory behind it and this is just a sort of the implementation of it and, yeah, so we can just copy this for the cell state, and then we all need. We just need to sort of run the LST M and H 0 comma cell state in a tuple and what would be the output here is the is, this is the hidden state and then the cell state, but we’re not really gonna use those so we can just remove, or we. We don’t have to look at that output and then we’re just gonna do out is self that FC of out and we’re just gonna take the last hidden state to send into the Tillinger layer and then we’re just gonna return out. And, yeah, all we need to do here. Is we need to initialize the network with the BR9 and this should be ready to run so? I’m gonna run this for I. Guess to epochs and let’s see what kind of results we get so training for two epochs. We get nine about 97% accuracy. Which is, I guess not too great, but yeah, we just train for two epoch, so you could. I guess expand this as well and you would probably get a lot better accuracy, but yeah. That was, that’s pretty much it for the bi-directional. STM if you have any questions, leave them in the comment otherwise. Thank you so much for watching your video, and I hope to see you in the next one [Music].