วีดีโอ

Nice work! Could you post your code to GitHub and send me a link? I will then take a look over the weekend and get back to you with a few suggestions.

Hi, there could be a couple of reasons you are getting NaN (Not a Number). Try to check your code vs the solution which you can find here: colab.research.google.com/github/LeakyAI/FirstNeuralNet/blob/main/FirstNeuralNetworkSolution.ipynb

so it trains the model so that it can mimic the mystery function while minimizing loss right

That's exactly right! If the loss is going down, it means the neural network is learning how to mimic the mystery function. As a next step, I would encourage you to learn more about training models, especially why there is a need to split our dataset into a separate training, validation and testing set. This is to ensure our model generalizes as it trains and does not overfit or underfit. In this example, we have just trained the model directly on the data. In the realworld, you will want to always ensure your model is trained to generalize the data being used.

Thank you so much for your kind words, we are working on new content and hope to release more soon. Happy learning!

Hey Maestro2050, I would encourage you to learn more about deep learning in general, and then try different projects. Hugging Face is a great place to go as it hosts a massive open source repository of fantastic neural network models that can do amazing things (detect items in an image, convert text from one language to another, create a ChatGPT style chatbot etc.). I would explore that as a next step and build some projects in areas that are you passionate about.

Hi, it will depend a lot on your data. If your data is naturally unlabeled, then you will want to choose an unsupervised approach and kmeans is a great start. It will cluster results it believes to be similar. Lots of parameters can be specified for kmeans. In this tutorial, the neural network is taking in labeled data (input data + number of lemonades sold as the label) so we are using a supervised approach. In this case, we train a simple neural network where we submit several known example combinations (inputs + number lemonades as the label) for the neural network to train on. If you have data similar to this (ie, known inputs and associated outputs), then you can opt to try to model with a neural network in a similar fashion. If your label is not regression (as with number of lemonades in this case), you can also convert the network instead to predict classes (classification) using a similar approach. Hope that helps!

I am using torch version 1.13.1

Hi! Looks like the inputs/outputs might be mis-aligned for some reason, could you check that you have the following code when creating the input variable x and output/label variable y? # Create our PyTorch tensors and move to CPU or GPU if available # Extract the inputs and create a PyTorch tensor x (inputs) inputs = ['Weekend','Sunny','Warm','BigSign','Price'] x = torch.tensor(df[inputs].values,dtype=torch.float, device=device) # Extract the outputs and create a PyTorch tensor y (outputs) outputs = ['NumberSold'] y = torch.tensor(df[outputs].values,dtype=torch.float, device=device)

You can also check your code vs the solution posted here: github.com/LeakyAI/FirstNeuralNet/blob/main/FirstNeuralNetworkSolution.ipynb

Hi, there are a lot of components to make an effective ANN using tabular data from sources like excel. First, you must have enough data so the model is able to properly train. There must be a correlation between the inputs and desired forecasted variable for the model to train properly. Normally, if you are building a model that predicts a value (regression), you do not want to have an activation function (sigmoid, tanh etc.) in the last layer so the model is free to predict a value and not be constrained. However, if your model is used for classification instead (for example, sentiment analysis or predicting TRUE/FALSE etc.) then a sigmoid activation function would be appropriate in the last layer.

Dear Hiwhy, this portion of your code here looks ok to me, your error could be up earlier in your project and and likely happening around the definition of x and y. If you could send me the entire project code I would be happy to take a further look.

Dear Katalin, in this beginner tutorial you are using a very simple synthetic dataset which allows you to focus instead on building and training of a simple neural network. The code creates the dataset using a simple formula to make it easy for the neural network to learn from the data and start making predictions. In the course (find it on leaky.ai), you will tackle real world datasets which require a lot more upfront work before you can use them with neural networks (cleaning, standardizing values, dealing with empty fields, non-linearities etc.). You will learn those techniques in the course and also how to build more advanced neural networks that can tackle real world, complex datasets.

Glad you liked it!

Great question, yes indeed you should always break up your data into at least a training and a test dataset (usually, we also use a validation set). We cover a lot of this in the hands-on course itself: www.leaky.ai/introduction-to-ai-programming-with-pytorch-course-beginner How you do this will depend a bit on how your your data is structured. A simple approach uses Python slicing: # Assume you have a dataset split into x (input) and y (output/label) import numpy as np # Build a list of indices and shuffle them randomly indices = list(range(len(x))) np.random.shuffle(indices) # Allocate 80% of the data for the training set trainingPercent = .8 split = int(total*trainingPercent) # Now simply slice the dataset into train and validation train_set = Dataset(x[indices[:split]], y[indices[:split]]) test_set = Dataset(x[indices[split:]], y[indices[split:]]) You can also use sklearn, they have a great function that does this for you: from sklearn.model_selection import train_test_split You can find a good example of how to use it here: towardsdatascience.com/how-to-split-data-into-three-sets-train-validation-and-test-and-why-e50d22d3e54c

Thank you so much for your kind words and feedback! Yea, PyTorch can be tough to learn so I try to take a more simplified step-by-step approach to teaching. There are also more tutorials on the website: www.leaky.ai/ PyTorch Dataloaders, for example, are very powerful and heavily used when building PyTorch projects. You can checkout the Dataloader tutorial to get started as a next step (www.leaky.ai/dataloader). Have a great day!

Thank you so much for your kind words, we are absolutely thrilled that you were able to benefit from the tutorial! We hope to have more videos available soon and you can also find additional tutorials on the website (www.leaky.ai). Have a wonderful day and good luck with your A.I. journey!

Hey Mazin, that is great that you are trying to apply this to a new dataset! Keep in mind, the model is very simple here and therefore will usually work with limited and simple datasets. In general, NaN values can come from different sources including your new dataset (inputs) but also the gradient calculations that are calculated during training potentially. Try to simplify the dataset itself and see if it eliminates the issue. If the issue is not the dataset, you can try to reduce the training loop to one iteration and then experiment with the model size/shape itself (adding additional layers and possibly broadening the width of the layers). Once you have eliminated the NaN, you can then gradually bring back the other components of the dataset and also increase your training iterations. Hope that helps!

@@leakyai i am using your dataset, and still getting Total Loss: nan How can i fix the issue?

@@sohanzz Yea I have the same problem. I am using a different dataset and get Nan as total loss as well

NaN is the python code for 'Not a Number' so somewhere in the calculation it becomes something different. check the type by using "Type(variable)", then you can backtrack where the error comes from.

Thank you so much for your feedback!

Thank you so much for letting me know! So glad you found it helpful! If there are other topics you would like, please do send them over as it will help us prioritize the content!

Thank you so much for sharing, that is awesome! If you want, you can now attempt the “Predict Sales” tutorial where you will use a neural network to predict lemonade sales by training it on a synthetic dataset. You can find that video here th-cam.com/video/SugVN2t3M8I/w-d-xo.html or you can follow along the written version here www.leaky.ai/projectlemonade. If you have any questions, please don’t hesitate to post them below the videos, we are always here to help! Thanks for watching!

Thank you so much Alejandro for your feedback! Great suggestion, we will put it on our list of future videos to get done! If possible, please help us get the word out for the channel by sharing it with anyone you know that might benefit from the free tutorials! Thanks for watching!