Food Vision: Transfer Learning and Fine Tuning

Objectives:

• To classify a dish from a given food image across 101 different food classes
• To build a model by applying transfer learning and fine-tuning with TensorFlow
• To train the model with 75000+ images of the Food101 dataset

Steps:

1. Data pre-processing
• Import and explore datasets
• Create preprocessing functions
• Batch and prepare datasets
• Create modelling callbacks
• Setup mixed precision training
2. Building a model
• Build a feature extraction model using transfer learning
• Fine-tune the feature extraction model
3. Training a model
• Train the model using trained dataset
4. Model evaluation
• Make predictions on food images using validation dataset
• Evaluate for precision, recall, and f1 scores
• Create a confusion matrix, and find the most wrong predictions
5. Deployment
• Make predictions on custom food images with prediction probabilities

Outcomes:

SkimLit: Natural Language Procesisng

Objectives:

• To classify medical abstract senstences into different sentence classes using NLP models
• To build and train a model with 200,000+ labelled randomized abstracts of PubMed 200k RCT dataset
• To evaluate the model with prediction probabilities

Steps:

1. Data pre-processing
• Download and visualize text dataset
• Get lists of sentences
• Make numeric labels
2. Model 0: baseline model
• Building a model
• Model evaluation
3. Model 1: Conv1D with token embeddings
• Convert text into numbers using text vectorization
• Creating custom text embedding
• Creating datasets
• Building a model
• Model evaluation
4. Model 2: Feature extraction with pretrained token embeddings
• Building a model
• Model evaluation
5. Model 3: Conv1D with character embeddings
• Creating character-level token vectorizer instance
• Creating character-level embedding
• Building a model
• Model evaluation
6. Model 4: hybrid embeddings (token embeddings + character embeddings)
• Building a model
• Combining token and character data into dataset
• Model evaluation
7. Model 5: Transfer learning (pretrained token + character + positional embeddings)
• Create positional embeddings
• Building a model
• Create tribid embedding dataset
• Model evaluation
8. Comparision performance
• Make preditions and evaluation
• Evaluate model on test dataset
• Find most wrong
9. Make example prediction

Outcomes:

BitPredict: Time-Series Forecasting with TensorFlow

Objectives:

• To predict the price of Bitcoin based on historical data with time series forecasting
• To build an ensemble model from several trained time series forecasting models
• To evaluate the model with performance metric of MAE

Steps:

1. Data pre-processing
• Get the series data
• Format and visualize datasets
2. Model 0: Naive forecast (baseline)
3. Model 1: Dense model (window=7, horizon=1)
4. Model 2: Dense model (window=30, horizon=1)
5. Model 3: Dense model (window=30, horizon=7)
6. Model 4: Conv1D
7. Model 5: RNN (LSTM)
8. Model 6: Dense (multivariate time series)
9. Model 7: N-BEATS algorithm
10. Model 8: Ensemble model
11. Model 9: Dense on full data
12. Compare models

Outcomes:

Pre-Trained ImageNet Models in TensorFlow and Keras

Objectives:

• To classify images using a pre-trained ImageNet model

Steps:

1. Data pre-processing
• Download assests
• Read and display sample images
2. Setup pre-trained model
• Load a pre-trained model
• Preprocess the input images and add batch dimension
• Make model prediction
• Decode the predictions to find the class name and confidence scores
3. Display the results

Outcomes:

Image Classification: Feedforward Networks in Keras

Objectives:

• To classify hand-written digits using feed-forward multilayer network architecture for multinomial classification

Steps:

1. Load and split the MNIST dataset
2. Pre-process dataset
• Input feature transformation and normalization
• One-Hot label encoding
3. Build and train a model using DNN architecture
4. Evaluate a model with loss and accuracy matrics
5. Predict the test dataset and display the results

Outcomes:

Traffic Sign Recognition: Transfer Learning and Fine-tuning

Objectives:

• To classify traffic sign images using fine-tuning pre-trained VGG-16 architecture model in TensorFlow and Keras

Steps:

1. Data pre-processing
• Download and extract dataset
• Create train and validation datasets
• Create and pre-process test dataset
2. Model inference using VGG-16 for Fine-Tuning
• Load VGG-16 convolutional base and and instantiate the weights
• Freeze the initial layers in convolutional base
• Add classifier to complete the model
• Compile and train the model
3. Model evaluation
• Make predictions on validation dataset
4. Deployment
• Make predictions on test dataset
• Display the results

Outcomes: