DeepLearning.AI TensorFlow Developer Professional Certificate - Course 3
- Week 1: Common Patterns in Time Series
- Trend
- Outlier
- Period
- Imputation technique
- Non-stationary time series: ts that changes behavior at certain time point
- We can only use ML for a period of time
- Stationary: behavior does not change
- Train test split (training, validation, test):
- Fixed partitioning
- Roll-forward partitioning: use daily/weekly/monthly data within the training data
- Common Metrics
- MSE
- RMSE: same scale of the original errors
- MAE/Mean Abs Deviation: does not penalize large error as much as MSE. We can use MAE if large error is not dangerous
- MAPE: P for percentage.
np.abs(errors/x_valid).mean()shows size of the errors compare to their value
- MA
- Differencing
- Week 2: Deep Neural Networks for Time Series
- Time windows: use a subset of time value as the input windows, use the next time value as the label.
- Sequence bias
- Linear Regression for Time Series
- ffnn for Time Series
- Callback for learning rate scheduler that update lr for each epoch
- Week 3: Recurrent Neural Networks for Time Series
- RNN: The input/output of RNN is 3-dim: batch size, time step, dim of each time step
- Lambda layers: help fix window dataset (2-dim input) before feed it into RNN (3 dim input)
- Week 4: Real-world time series data
- LSTM + CNN
- Sunspot dataset: https://storage.googleapis.com/laurencemoroney-blog.appspot.com/Sunspots.csv