Course notebooks and milestone projects for Daniel Bourke's Udemy Course.
Milestone Projects are as follows:
- FoodVision - Classification of Kaggle's Food101 Dataset.
- SkimLit - Reconstructing & reformatting PudMed abstracts to make them more readable.
- BitPredict - Time-series forecasting the price of bitcoin.
Dataset: Food 101
Models:
- Baseline feature extractor based on EfficientNet B0, trained using mixed precision training: Accuracy 70%.
- Fine-tuned EfficientNet B0 last 3 layers: Accuracy 74%.
- Fine-tuned EfficientNet B0 last 3 layers and using an augmented dataset: Accuracy 79%.
- Fine-tuned EfficientNet B0 all layers while using an augmented dataset, and an adpative learning rate: Accuracy 79%.
Dataset: PudMed RCT
ETL:
- Preprocess each abstract and append into a dataframe with a line per sentennce in abstract & relevant metadata (line number, total number of lines & classification (e.g. objective, method, conclusion etc))).
- One hot encode label values, line numbers and total lines.
- Vectorise and create embeddings (USE, GloVe, BERT, custom embedding layers).
- Batch and prefect datasets to optimise training speeds.
Models:
- Basline TF-IDF Multinomial Naive-Bayes Classifier: Accuracy 72%
- Conv1D (Token embedding layer): Accuracy 79%.
- Feature Extractor (USE embeddings): Accuracy 71%.
- Conv1D (Character-level embeddings): Accuracy 61%.
- Conv1D (Hybrid embedding layer: Token & Character level): Accuracy 72%.
- Transfer learning (Tribrid embedding layers: Token, Character, and Positional Embeddings): Accuracy 84%.
Dataset: BTC data
Models:
- Niave-Bayes Classifier.
- Dense (Window size: 7 days, Horizon: 1 day).
- Dense (Window size: 7 days, Horizon: 1 day).
- Dense (Window size: 30 days, Horizon: 1 day).
- Dense (Window size: 30 days, Horizon: 7 days).
- 1D CNN.
- Bi-directional LSTM.
- Multivariate Dense.
- NBEATS Algorithm.
- Ensemble Model (Series of Dense models using NBEATS data pipeline).