L6_Augmentation and Advanced Computer vision

Flashcards based on Deep Learning Lecture 6, covering data augmentation, modern convnets, and other computer vision tasks.

Why is data typically divided into batches in deep learning?

Data is divided into batches to accommodate memory constraints. Large datasets often exceed the capacity of GPU or RAM. Processing data in batches allows iterative updates to model parameters, optimizing resource utilization and enabling training on large datasets. Analogy: Processing a large book chapter by chapter instead of all at once due to cognitive limitations.

What is the opportunity of using both CPU and GPU in deep learning pipelines?

Leveraging both CPU and GPU facilitates parallel processing. GPU accelerates numerical computations in deep learning models, while CPU handles data preprocessing and orchestration. This parallelization maximizes throughput and reduces training time. Analogy: A chef (GPU) cooking food while a sous chef (CPU) prepares ingredients simultaneously.

What is the benefit of prefetching in data loading?

Prefetching minimizes GPU idle time by proactively loading data. While the GPU processes the current batch, the CPU prepares the subsequent batch, ensuring continuous data flow. This optimization reduces pipeline latency and enhances overall training efficiency. Analogy: Ordering the next drink while finishing the current one to avoid waiting.

What are some Keras convenience functions for 'standard' data?

Keras provides functions such as keras.utils.image_dataset_from_directory, keras.utils.timeseries_dataset_from_array, keras.utils.text_dataset_from_directory, and keras.utils.audio_dataset_from_directory to simplify loading of standard data types from directories or arrays. These utilities streamline data ingestion for common deep learning tasks.

What does the .map() method do on a TensorFlow Dataset?

The .map() method applies a user-defined transformation function to each element in a TensorFlow Dataset. This enables preprocessing, augmentation, or feature engineering operations to be performed efficiently on the dataset. The mapped function transforms each element independently, creating a new transformed dataset. Analogy: Using a cookie cutter to transform each piece of dough into a specific shape.

Does the Dataset .map() method modify data in-place?

No, the .map() method does not modify the original Dataset; it returns a new Dataset with transformed elements. The original dataset remains unchanged, preserving the integrity of the source data. Analogy: Photocopying a document; the original remains unchanged.

What does the .prefetch() method do on a TensorFlow Dataset?

The .prefetch() method optimizes data loading by preparing elements from the input dataset in advance. This asynchronous prefetching allows the current element to be processed while subsequent elements are being prepared, minimizing GPU starvation and improving training throughput. Analogy: A restaurant prepping ingredients ahead of time to ensure the chef always has what they need.

What is data augmentation used for?

Data augmentation enhances model generalization by generating artificially modified duplicates of the training data. This technique mitigates overfitting and improves model robustness by exposing it to a wider range of variations. Transformations include rotations, scaling, color shifts, and more. Analogy: Showing a child different variations of the same object to improve understanding.

What is keras.layers.RandAugment?

The keras.layers.RandAugment layer implements random data augmentation by applying a set of randomly selected transformations to the input data. This automated augmentation strategy optimizes the training process and obviates the need for manual selection of augmentations. To keep with our dog analogy, RandAugment will randomly pick different ways it modifies the dog pictures.

What is the Keras Functional API?

The Keras Functional API offers a flexible paradigm for constructing deep learning models with arbitrary layer connections and topologies. It supports multi-input, multi-output models, shared layers, and complex network architectures beyond sequential models. Analogy: Building with LEGOs instead of following a fixed instruction manual.

What are some of the advantages of non-sequential networks?

Non-sequential networks facilitate advanced model architectures with multiple inputs, multiple outputs, arbitrary layer connections, and networks-inside-the-network. These capabilities enable the creation of sophisticated models tailored to specific tasks and data structures. For example, a model might take in an image and some numerical data, and depending on the data apply different image processing layers.

What is a key feature of Inception networks?

Inception networks utilize parallel convolutional layers with varying kernel sizes to capture image features at multiple scales. This multi-scale feature extraction enables the model to discern both fine-grained details and high-level patterns. Analogy: Viewing an image from up close to see individual pixels, or from further away to get the general gist of the image.

What is a key feature of Residual Networks?

Residual Networks incorporate skip connections that bypass one or more layers, enabling the model to learn residual mappings and mitigate the vanishing gradient problem. These connections facilitate the flow of information and improve the training of deep networks. Analogy: Having a 'fast track' option in addition to your normal processing track.

What is the main idea behind skip connections in residual networks?

Skip connections in residual networks combat vanishing gradients and prevent layer stagnation. These connections facilitate gradient flow and enable the effective training of deep networks by ensuring that each layer contributes meaningfully to the overall learning process. Analogy: Wearing a seatbelt so even if the car stops suddenly, you'll still be fine.

What is a DenseNet?

A DenseNet is a network architecture that employs dense connections between all layers, fostering feature reuse and enhancing information flow throughout the network. This dense connectivity promotes robust feature extraction and improves model performance. Analogy: The network is like a super-efficient team where everyone communicates directly with everyone else.

What is the main component of Xception networks?

Xception networks primarily use depthwise separable convolution layers. As opposed to applying convolution to all channels in an image, deptwise separable convolution first applies convolution to one channel at a time, then mixes the channels. In effect, the network decouples feature location from feature interrelation.

What are the Keras Applications?

Keras Applications are a collection of pre-trained models for computer vision tasks, offering ready-to-use architectures for feature extraction, transfer learning, and fine-tuning. These models provide a convenient starting point for various computer vision applications. Analogy: The Keras Applications are like blueprints of famous buildings, which can readily be used and adapted.

What should you do if you want to use Keras Applications?

Before using a Keras Application, preprocess the input images using its corresponding preprocessing function (e.g., keras.applications.Xception.preprocess_input()). This vital step ensures compatibility between the input data and the pre-trained weights of the model. Different Keras Applications may be trained on different types of data, so using this convenience function is essential.

What is the purpose of segmentation in computer vision?

In computer vision, segmentation involves classifying each pixel in an image, enabling fine-grained visual tasks such as object localization, scene understanding, and autonomous navigation. For instance, self-driving cars need to be able to single out every street sign and pedestrian in its cameras. Analogy: Segmentation is like coloring-in a black-and-white drawing to highlight the different parts.

What is an encoder-decoder structure?

An encoder-decoder structure downsamples the input to find large-scale patterns, then upsamples again to the correct dimensions, as is typically used in segmentation problems. Analogy: Think of it like compressing a file (encoding) to make it smaller, then decompressing it (decoding) back to its original size.