Skip to content

Latest commit

 

History

History
167 lines (104 loc) · 6.8 KB

nn.rst

File metadata and controls

167 lines (104 loc) · 6.8 KB
Raw

Training, Prediction, and Hyper-parameter Search for Neural Networks

For users who are just getting started, see:

If you have been familiar with the basic operations, see:

Using CLI via an Example

Step 1. Data Preparation

Create a data sub-directory within LibMultiLabel and go to this sub-directory.

mkdir -p data/rcv1
cd data/rcv1

Download the RCV1 :ref:`libmultilabel-format` dataset from LIBSVM Data by the following commands.

wget -O train.txt.bz2 https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/multilabel/rcv1_topics_train.txt.bz2
wget -O test.txt.bz2 https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/multilabel/rcv1_topics_test.txt.bz2

Uncompress data files and change the directory back to LibMultiLabel.

bzip2 -d *.bz2
cd ../..

See Dataset Formats here if you want to use your own dataset.

Step 2. Training and Prediction via an Example

Train a CNN model and predict the test set by an example config. Use --cpu to run the program on the cpu.

python3 main.py --config example_config/rcv1/kim_cnn.yml

Training and (Optional) Prediction

In the training procedure, you can build a model from scratch or start from some pre-obtained information.

python3 main.py --config CONFIG_PATH \
                [--checkpoint_path CHECKPOINT_PATH] \
                [--embed_file EMBED_NAME_OR_EMBED_PATH] \
                [--vocab_file VOCAB_CSV_PATH]
  • config: configure parameters in a yaml file. See python3 main.py --help.

If a model was trained before by this package, the training procedure can start with it.

  • checkpoint_path: specify the path to a pre-trained model.

To use your own word embeddings or vocabulary set, specify the following parameters:

  • embed_file: choose one of the pretrained embeddings defined in torchtext or specify the path to your word embeddings with each line containing a word followed by its vectors. Example:
the 0.04656 0.21318 -0.0074364 ...
a -0.29712 0.094049 -0.096662 ...
an -0.3206 0.43316 -0.086867 ...
  • vocab_file: set the file path to a predefined vocabulary set that contains lines of words.
the
a
an

For validation, you can evaluate the model with a set of evaluation metrics. Set monitor_metrics to define what you want to print on the screen. The argument val_metric is the metric for selecting the best model. Namely, the model occurred at the epoch with the best validation metric is returned after training. If you do not specify a validation set in the configuration file via val_file or a training-validation split ratio via val_size, we will split the training data into training and validation set with an 80-20 split. Example lines in a configuration file:

monitor_metrics: [P@1, P@3, P@5]
val_metric: P@1

If test_file is specified, the model with the highest val_metric will be used to predict the test set.

Prediction

To deploy/evaluate a model (i.e., a pre-obtained checkpoint), you can predict a test set by the following command.

python3 main.py --eval \
                --config CONFIG_PATH \
                --checkpoint_path CHECKPOINT_PATH \
                --test_file TEST_DATA_PATH \
                --save_k_predictions K \
                --predict_out_path PREDICT_OUT_PATH
  • Use --save_k_predictions to save the top K predictions for each instance in the test set. K=100 if not specified.
  • Use --predict_out_path to specify the file for storing the predicted top-K labels/scores.

Hyper-parameter Search

Parameter selection is known to be extremely important in machine learning practice; see a powerful reminder in "this paper". Here we leverage Ray Tune, which is a python library for hyper-parameter tuning, to select parameters. Due to the dependency of Ray Tune, first make sure your python version is not greater than 3.8. Then, install the related packages with:

pip3 install -Ur requirements_parameter_search.txt

We provide a program search_params.py to demonstrate how to run LibMultiLabel with Ray Tune. An example is as follows:

python3 search_params.py --config example_config/rcv1/cnn_tune.yml \
                         --search_alg basic_variant
  • config: configure all parameters in a yaml file. You can define a continuous, a discrete, or other types of search space (see a list here). An example of configuring the parameters is presented as follows:
dropout: ['grid_search', [0.2, 0.4, 0.6, 0.8]] # grid search
num_filter_per_size: ['choice', [350, 450, 550]] # discrete
learning_rate: ['uniform', 0.2, 0.8] # continuous
activation: tanh # not for hyper-parameter search
  • search_alg: specify a search algorithm considered in Ray Tune. We support basic_variant (e.g., grid/random), bayesopt, and optuna. You can also define search_alg in the config file. For example, if you want to run grid search over learning_rate, the config is like this:
search_alg: basic_variant
learning_rate: ['grid_search', [0.2, 0.4, 0.6, 0.8]]

After the search process, the program applies the best hyper-parameters to obtain the final model. The re-training process by default adds the validation set for training. Our empirical analysis shows that this setting improves test results. If you do not want to incorporate the validation data for training, you can specify the option no_merge_train_val. In either case, the optimization starts from scratch and runs for the number of epochs that leads to the best validation results in the hyper-parameter search.

For more information on this section, please refer to :ref:`Parameter Selection for Neural Networks`