Intermediate Deep Learning with PyTorch
Build image classifiers, LSTM and GRU forecasters and multi-output models in PyTorch, and keep training stable with initialization and batch norm.
Training deep learning models reliably in PyTorch means dealing with unstable gradients, choosing an optimizer and matching the network design to the kind of data at hand. You define datasets and models as Python classes and steady training with weight initialization, ELU activations and batch normalization. You then build convolutional networks for image classification, LSTM and GRU networks for forecasting, and models that combine two inputs or predict two targets with weighted losses.
What You Will Achieve
- Write PyTorch Dataset, DataLoader and model classes and train them with different optimizers
- Keep gradients stable through careful weight initialization, ELU activations and batch normalization
- Train a multi-class image classifier with data augmentation and check its metrics for each class
- Forecast electricity consumption with a recurrent network and design models with two inputs or two outputs
Program Overview
Neural Network Training, Optimizers and Gradients in PyTorch
Revisits training and evaluation with object-oriented PyTorch code, compares optimizers and treats the unstable gradients that disrupt neural network training.
- Object-oriented datasets, data loaders and models
- Training loops, optimizers and model evaluation
- Vanishing and exploding gradients
- ReLU vs. ELU and batch normalization
Convolutional and Recurrent Networks for Images and Sequences
Handles image and sequential data in PyTorch, building convolutional classifiers with augmentation and recurrent forecasting models compared across three cell types.
- Image datasets and choosing augmentations
- Convolutional layers and multi-class evaluation
- Generating sequences and sequential datasets
- RNN, LSTM and GRU forecasting networks
Multi-Input and Multi-Output Architectures
Designs models that take more than one input or produce more than one output, and shows how loss weighting balances the tasks.
- Two-input datasets and models
- Two-output data loaders and architectures
- Evaluating multi-output models
- Loss weighting
Certificate Awarded (Paid)
Job Outlook
These PyTorch skills are used in machine learning and data science work that involves image classification, time series forecasting and models drawing on more than one data source.
Good for: Learners who want guided, exercise-based practice building image classifiers, recurrent forecasters and multi-task models in PyTorch code.
Keep in mind: Learning rate scheduling, dropout and regularization get no lessons of their own; the chapters center on gradients, CNNs, RNNs and multi-output models.
Bottom line: Four chapters of exercises that take PyTorch work beyond a basic training loop into image, sequence and multi-task models.
Frequently Asked Questions
Which PyTorch libraries do the exercises rely on?
The exercises use torch.nn to define layers and models, torch.optim for optimizers, torchvision for image data, torchmetrics for evaluation metrics and PIL for loading images.
What task is used to practice recurrent networks?
The sequence chapter has you prepare time series data, generate training sequences and then train and evaluate a recurrent model that predicts electricity consumption, comparing plain RNN, LSTM and GRU cells along the way.
How are models with several outputs trained and evaluated?
You build a two-output dataset, data loader and model, train it on both tasks at once and evaluate each output. The chapter then shows how weighting the losses balances the importance of the different tasks.


There are no reviews yet.