辅导ECE 802-602、讲解Python程序、辅导LSTM RNNs、讲解Python编程

ECE 802-602 Fall 2018

Mini-project#3

Due: Friday, November 9, 2019, by 12 mid-nite (can be extended at a grade penalty (loss) of 5%/day for

up to 5 days.

Exploring code(s)/model(s) for LSTM RNNs. The description here will focus on the KERAS 2.0

Library. Any equivalent examples from other libraries will be accepted.

Keras 2.0 Examples used:

(1) Your assigned/chosen example from miniproject_2, e.g., imdb_bidirectional_lstm.py, or another

example with (added) LSTM layer(s).

(2) pre-trained-word-embeddings.py

10+20% (30% total)

1) Setting the baseline: Run each example at default (or best) values. Use the (performance) results

as a baseline for reference. (Typically, run until convergence or greater than 10 but less than 100

epochs). You may plot the metric, vs. epochs, to visualize settling of the validation accuracy

curves. You will use the same epochs & plots for the next part as well).

35% each (70% total)

2) Construct your best networks: For each example, augment the network model to construct layers

of convnets followed by one or more layers of LSTM RNN. Your goal is to achieve better metric

performance of your augmented revised network for each over the baseline. Define your own

best final network model (for each example) and show its performance vis-à-vis the baseline

example. Use reasonable common epochs to make the comparison (typically till convergence or

greater than 10 but no more than 100 epochs). (In each case, you will receive full credit if your

accuracy is no less than the default example, and would receive additional 1% for every 1%

accuracy above the baseline example, up to 3%).

Early Planning Note:

Your final project is to use your final defined (best) architecture for one of the examples, using the

20news or newly publically available benchmark dataset) but using a modified LSTM variants

(supplied). For the final project, the assignment of which example LSTM variant model will be

randomly assigned. Preserve/save your results of this Assignment #3 as you may use it as part of the full

comparison with the final project model in your final_project paper report.

A brief summary of your results should be in the form of a mini-report highlighting the key findings or

within the jupyter-file. Sample topics may include

-Summary description of each case

-Lessons learned: what worked and failed. What new things have you learned/discovered if any that you

might share with me and/or the class.

-submit your report files via (MSU) Gitlab