I hope you enjoy the course :) And check out Tabnine, the FREE AI-powered code completion tool that helps you to code faster: www.tabnine.com/?.com&PythonEngineer * ---------------------------------------------------------------------------------------------------------- * This is a sponsored link. You will not have any additional costs, instead you will support me and my project. Thank you so much for the support! 🙏
at 37:00 I found after adding 2 that not all members of the tensor had exactly x+2. I tried this several times with always one of the parts of the tensor had less than x+2. Then at 37:16 you also had an anomaly. Why is this?
Incredible tutorial, thank you! Some corrections: - 1:12:02 correct gradient function in the manual gradient calculation should be `np.dot(2*x, y_predicted - y) / len(x)`, because np.dot results in a scalar and mean() has no effect of calculating the mean. (TY @Arman Seyed-Ahmadi) - 1:23:52 the optimizer is applying the gradient exactly like we do, there is no difference. The reason the PyTorch model has different predictions is because 1) you use a model with a bias, 2) the values are initialized randomly. To turn off the bias use `bias=False` in the model construction. To initialize the weight to zero use a `with torch.no_grad()` block and set `model.weight[0,0] = 0`. Then all versions result in the exact same model with the exact same predictions (as expected).
Thanks for this second comment! To add to this: nn.Linear wants to solve y = wx + b here. This 'b' is the bias, and by setting bias = False, instead it learns y = wx as we want it to. This also means that model.parameters() will yield only [w] and not [w, b] anymore, so do not forget to change that in line 52 in the video as well.
This is a fantastic tutorial, thank you for sharing this great material! There is one mistake though that needs clarification: ========================================== At 1:12:02 it is mentioned that the code with automatic differentiation does not converge as fast because "back-propagation is not as exact as the numerical gradient". This is incorrect: the reason why the convergence of the two codes are different is because there is a mistake in the gradient() function. When the dot product np.dot(2x, y_pred_y) is performed, the result is a scalar and .mean() does not do anything. Instead of doing .mean(), np.dot(2x, y_pred_y) should simply be divided by len(x) to give the correct mean gradient. After doing this, both methods give the exact same convergence history and final results.
This is one of the very few videos which is teaching Pytorch from the ground up! Beautiful work, @Python Engineer. Highly recommend it for any newbie + refresher.
Thanks for the course Patrick! It was a great refresher! BTW, at 3:42:02, in the newer versions instead of pretrained=True it is changed to weights=True.
For the feedforward part, you need to send the model to the GPU when instantiating it: model = NeuralNet(input_size, hidden_size, num_classes).to(device) if your device is 'cuda' and you forget the '.to(device)' you will get an error.
I just completed the course on ML from scratch from Python Engineer. It was a great course for someone who learned all those algorithms in the past and wants to see how they get implemented using basic python lib and numpy.
On 4:14:00, I think you should use the ground truth as the labels rather than the predicted (line 130). Because the PR curve use the ground truth and predicted score to paint
Thanks a lot for the low level explanations. At 1:01:47 when you dot product the array turns into a single scalar. So mean() returns that number(the sum), not average. When you fix it you get the exact same results as with pytorch's implementation in 1:12:00
@@phi6934 I don't remember the details right now, but just dividing the expression with the size of the tensor must do the work. In the expression put smt like .../len(x) instead of .mean()
Thanks for the awesome course! The material is extremely well curated, every minute is pure gold. I particularly liked the fact that for each subject there is a smooth transition from numpy to torch. It's perfect for someone who wants a quick and thorough deeplearning recap and get comfortable with hands-on pytorch coding.
In the Gradient Descent and Training Pipeline sections, the presenter glosses over why it takes 5x more training steps to converge. There are a couple factors: - Autograd is less aggressive than the manual gradient calculation, effectively lowering the learning rate (you can go all the way up to 0.1 after you move to torch and autograd) - nn.Linear() includes a bias by default and a non-zero initialization of the weights, making it not a direct comparison. You can get much closer by adding `bias=False` to the model initialization and by zeroing out the weigth with `model.weight.data.fill_(0.0`
at 1:01:41 he uses np.dot and when it should be np.multiply, that will make it consistent with the pytorch implementation. By doing np.dot, the items are multiplied and summed leaving just one value to which the mean function is applied, so the reason the numpy version get to 0 loss quicker is the gradient is not being averaged correctly.
Your course is great! Congratulations! I just had to do a small correction in your code in part "13. Feed Forward Net" so that I could run it on GPU. It was necessary to add the "device" (that was preciously declared) as an argument in the nn.Linear function. Without this detail it is not possible to run the code in GPU. class NeuralNet(nn.Module): def __init__(self, input_size, hidden_size, n_classes, device): super(NeuralNet,self).__init__() self.l1 = nn.Linear(input_size, hidden_size, device=device) self.relu = nn.ReLU() self.l2 = nn.Linear(hidden_size, n_classes, device=device) def forward(self, x): out = self.l1(x) out = self.relu(out) out = self.l2(out) return out
Dear with apologies kindly notice, At timestamp 1:12:05 make a correction in stating, that the backprop grad was not correct, Actually the numerical one was not correct. Because np.dot is computing a single number and then taking mean is the same number, instead use 2*x/4 in np.dot(2*x,(Y_pred-Y).mean()) to correct your numerical gradient. Using np.dot(2*x/4,(Y_pred-Y)) will produce same result as back propagated result. Mean will be usefull when W and X are matrices. Thank you
This is the best course on this topic I've seen so far. It is perfect when you want to understand what you're doing and the way things are brought is very pedagogic.
If z is a scalar then z.backward() is defined (and I understand the computation), while if z is not a scalar then z.backward() is not defined unless you provide appropriate inputs. However, it was not entirely clear to me what computation is occurring when we do z.backward(x) for example (where x is appropriate). This subject matter is around 33:00.
What is happening is that PyTorch is assuming that you have provided the intermediate gradients i.e. (dLoss/dz), then using these intermediate gradients PyTorch is able to compute the gradients further downstream and backward step is successful.
I've taken a graduate course in deep learning and neural, and have watched other tutorials here and there, but this is by far the most helpful one. Granted, all the previous materials have probably contributed, but the way you teach is unparalleled!
Nice tutorial ! @1:11:40 at line # 37. Instead of using "w -= learning_rate * w.grad" , I used expanded form "w = w - learning_rate * w.grad" and thought it would be same. But in this case 'w.grad' return 'None'. w.require_grad is False and hence error. Though "w -= learning_rate * w.grad" is same as "w.data = w.data - learning_rate * w.grad". It seems torch Tensor ( with require_grad True) have some overridden "__iadd__" implementation.
1:12:09 It's because the gradient in your formula is not correct, not because pytorch's backpropogation calculation. You should put the ".mean()" into the brackets of "np.dot()".
Someone has probably mentioned this already, but on line 23 at 1:04:08 .mean() is not doing anything since taking the dot product already returned a scalar. This is just dividing by one. Instead, you should be dividing by len(x) or len(y), or there may be another more efficient way to get the same result.
a probable mistake: Leaky ReLU isn't used for solving the problem of vanishing gradient problem but Dead Neurons problem. Which can happen when you use ReLU activation functions.
Excellent series. Using this to review what I've learned and to also learn PyTorch, thank you for this. The only thing I'd change is that you add an upward inflection to the end of most of your sentences which is a bit jarring (makes it sound like every sentence is a question).
Thanks for this incredible resource. FYI I believe the gradient function computed at 1:01:38 is incorrect. I'm pretty sure it should be: def gradient(x, y, y_predicted): return ((y_predicted-y)*2*x).mean()
Update: Note a subtle detail, if in with torch.no_grad() you use w = instead of w -= a new w variable will be created with requires_grad = False, which is fixed by w.requires_grad = True Original: Using pytorch 1.11, and go figure @1:11 w.grad.zero_() errors, instead I had to put w.requires_grad = True
Absolutely great. But what was missing for me was how then to use a trained model. Conspicuous in its absence was how at the end to feed data into a trained model and get the answer it was trained to give. Is there another video that explains this?
How is it that for a feed forward neural network we zero the gradients first before computing gradients and updating weights @3:08:35, whereas in the case of linear/logistic regression, we zero the gradients after computing them and updating the weights @1:36:19 @1:52:41. Intuitively, this should not make any difference, but i wanted to confirm if that truely is the case. Is this just a nomenclature thingy?
This tutorial is supppppppppppper great! The best deep learning tutorial I've ever watched. Thank you so much. I enjoined the tutorial that I didn't want it to stop! I look forward to seeing more great videos like this from this channel
At 1:01:55 you are taking the mean of a scalar, which doesn't do anything. Since you have 4 data points only this effectively means that your learning_rate was multiplied by 4. This is the reason why it seems to work better than PyTorch: this particular case is so well behaved that to speed up is sufficient to take larger steps.
How do you make the pytorch program displayed in the "OUTPUT" of the console under vscode? What plug-ins need to be installed? What are the shortcut keys? I can only run the code in "TERMINAL", which annoys me.
Great tutorial! one small point regarding CNN - CIFAR10 While calculation accuracy , its better to use for i in range(len(labels)): than for i in range(len(lbatch_size)): since if last set of batch_size less than original batch_size given then it will throw index bound error
This is an error I have found Time: 1:01:55 According to the equation,we actually need to find 1/N ,where N represents the number of term(here 4).According to the code,we are computing mean after converting the rest of the code to a dot product,which contains just a value.So instead of dividing with the desired value(4),we are dividing with 1.
Hello, why do we need to put the batch_size argument in test_loader = torch.utils.data.DataLoader(dataset = test_dataset, batch_size = batch_size, shuffle = False). There is no real point using batches for the test set or am I missing something very important?
Super good tutorial, this really made my day - many thanks !!! In the 05_gradients_torch, the difference in results from 05_gradients_numpy is because the derivative function should return 1/N * np.dot(2*x, y_pred-y) where N = 4. Then the results are exactly equal.
Great tutorial on the tensorboard part. But you made a mistake when plotting the precision and recall curve when you append predicted into the labels where it should have been labels1 - the actual labels of the eval batch
Wow. This course is awesome. An end to end of everything. I was wondering why I need to learn about Tensorboard and JSON files (other series) for using Torch. This was very useful to me.
Note at 2:08: `dataiter.next()` is no throwing an AttributeError: '_MultiProcessingDataLoaderIter' object has no attribute 'next'... I changed that line to `data = next(dataiter)`
1:42:10 Is there a reason you used `X, y` instead of `X, Y`? I believe it should `Y` as we're dealing with a tensor of dependent variables right? It would be `y` if we were dealing with a scalar though
I don’t understand why we don’t have to reshape mnist dataset label? Isn’t that 1d tensor with only one row? Why we don’t have to reshape it to 2d tensor 100,1 tensor