Implementation of GloVe using TensorFlow estimator API.
The trainer module in this repository also allows for cloud model training and evaluation on Google Cloud Platform. Please refer to cloud.
ENV_NAME=glove-tensorflow
# clone repo
git clone git@github.com:yxtay/glove-tensorflow.git && cd recommender-tensorflow
# create and activate conda environment
conda env create -n ${ENV_NAME} -y python=3.7
conda activate ${ENV_NAME}
# install requirements
# make install-requirments
pip install -r requirements/main.txt -r requirements/dev.txt
You may also use accompanying docker commands to avoid environment setup.
The text8 dataset is used for demonstration purposes.
The following script downloads the data, processes it to prepare the vocabulary and cooccurrence matrix. The data is serialised to csv
.
# make data
python -m src.data.text8
With Docker
# make docker-data
docker run --rm -w=/home \
--mount type=bind,source=$(pwd),target=/home \
continuumio/anaconda3:2019.10 \
python -m src.data.text8
Sample data
row_token_id | col_token_id | count | value | row_token | col_token | neg_weight | glove_weight | glove_value |
---|---|---|---|---|---|---|---|---|
6125 | 38 | 24 | 16.9500 | altogether | not | 0.6421 | 0.3428 | 2.83027 |
18 | 1571 | 176 | 74.1000 | was | prominent | 7.5889 | 1.0000 | 4.30542 |
91 | 372 | 19 | 5.4500 | th | society | 3.1999 | 0.2877 | 1.69562 |
432 | 541 | 12 | 5.9000 | numbers | note | 0.6461 | 0.2038 | 1.77495 |
1304 | 285 | 25 | 11.1667 | na | europe | 0.4112 | 0.3535 | 2.41293 |
32 | 18 | 2312 | 723.2000 | be | was | 406.5180 | 1.0000 | 6.58369 |
2247 | 1154 | 136 | 46.5833 | html | www | 0.0740 | 1.0000 | 3.84124 |
710 | 229 | 18 | 9.0500 | cannot | point | 0.8569 | 0.2763 | 2.20276 |
467 | 3756 | 12 | 5.2000 | style | width | 0.0911 | 0.2038 | 1.64866 |
80 | 543 | 35 | 20.6333 | over | lost | 2.6989 | 0.4550 | 3.02691 |
Usage
usage: text8.py [-h] [--url URL] [--dest DEST] [--vocab-size VOCAB_SIZE]
[--coverage COVERAGE] [--context-size CONTEXT_SIZE] [--reset]
[--log-path LOG_PATH]
Download, extract and prepare text8 data.
optional arguments:
-h, --help show this help message and exit
--url URL url of text8 data (default:
http://mattmahoney.net/dc/text8.zip)
--dest DEST destination directory for downloaded and extracted
files (default: data)
--vocab-size VOCAB_SIZE
maximum size of vocab (default: None)
--coverage COVERAGE token coverage to set token count cutoff (default:
0.9)
--context-size CONTEXT_SIZE
size of context window (default: 5)
--reset whether to recompute interactions
--log-path LOG_PATH path of log file (default: main.log)
# make train
python -m trainer.estimator
With Docker
# make docker-train
docker run --rm -w=/home \
--mount type=bind,source=$(pwd),target=/home \
tensorflow/tensorflow:2.1.0-py3 \
python -m trainer.estimator
Usage
usage: estimator.py [-h] [--train-csv TRAIN_CSV] [--vocab-txt VOCAB_TXT]
[--row-name ROW_NAME] [--col-name COL_NAME]
[--target-name TARGET_NAME] [--weight-name WEIGHT_NAME]
[--pos-name POS_NAME] [--neg-name NEG_NAME]
[--job-dir JOB_DIR] [--disable-datetime-path]
[--embedding-size EMBEDDING_SIZE] [--l2-reg L2_REG]
[--neg-factor NEG_FACTOR] [--optimizer OPTIMIZER]
[--learning-rate LEARNING_RATE] [--batch-size BATCH_SIZE]
[--train-steps TRAIN_STEPS]
[--steps-per-epoch STEPS_PER_EPOCH] [--top-k TOP_K]
optional arguments:
-h, --help show this help message and exit
--train-csv TRAIN_CSV
path to the training csv data (default:
data/interaction.csv)
--vocab-txt VOCAB_TXT
path to the vocab txt (default: data/vocab.txt)
--row-name ROW_NAME row id name (default: row_token)
--col-name COL_NAME column id name (default: col_token)
--target-name TARGET_NAME
target name (default: glove_value)
--weight-name WEIGHT_NAME
weight name (default: glove_weight)
--pos-name POS_NAME positive name (default: value)
--neg-name NEG_NAME negative name (default: neg_weight)
--job-dir JOB_DIR job directory (default: checkpoints/glove)
--disable-datetime-path
flag whether to disable appending datetime in job_dir
path (default: False)
--embedding-size EMBEDDING_SIZE
embedding size (default: 64)
--l2-reg L2_REG scale of l2 regularisation (default: 0.01)
--neg-factor NEG_FACTOR
negative loss factor (default: 1.0)
--optimizer OPTIMIZER
name of optimzer (default: Adam)
--learning-rate LEARNING_RATE
learning rate (default: 0.001)
--batch-size BATCH_SIZE
batch size (default: 1024)
--train-steps TRAIN_STEPS
number of training steps (default: 16384)
--steps-per-epoch STEPS_PER_EPOCH
number of steps per checkpoint (default: 16384)
--top-k TOP_K number of similar items (default: 20)
# make train MODEL_NAME=keras
python -m trainer.keras
# make train MODEL_NAME=logistic_matrix_factorisation
python -m trainer.logistic_matrix_factorisation
You may inspect model training metrics with Tensorboard.
# make tensorboard
CHECKPOINTS_DIR=checkpoints
tensorboard --logdir ${CHECKPOINTS_DIR}
With Docker
# make docker-tensorboard
CHECKPOINTS_DIR=checkpoints
docker run --rm -w=/home -p 6006:6006 \
--mount type=bind,source=$(pwd),target=/home \
tensorflow/tensorflow:2.1.0-py3 \
tensorboard --logdir ${CHECKPOINTS_DIR}
Access Tensorboard on your browser
The trained and serialised model may be served with TensorFlow Serving.
# make serving MODEL_NAME=glove
JOB_DIR=checkpoints/glove_estimator
MODEL_NAME=glove
docker run --rm -p 8500:8500 -p 8501:8501 \
--mount type=bind,source=$(pwd)/${JOB_DIR}/export/exporter,target=/models/${MODEL_NAME} \
-e MODEL_NAME=${MODEL_NAME} -t tensorflow/serving:2.1.0
Model signature
# make saved-model-cli JOB_DIR=checkpoints/glove_estimator/export/exporter/1582880583
JOB_DIR=checkpoints/glove_estimator/export/exporter/1582880583
saved_model_cli show --all --dir ${JOB_DIR}
MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:
signature_def['serving_default']:
The given SavedModel SignatureDef contains the following input(s):
inputs['col_token'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: col_token:0
inputs['row_token'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: row_token:0
The given SavedModel SignatureDef contains the following output(s):
outputs['input_embedding'] tensor_info:
dtype: DT_FLOAT
shape: (-1, 64)
name: predictions/row_embedding/embedding_lookup/Identity_1:0
outputs['input_string'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: predictions/input_string_lookup/LookupTableFindV2:0
outputs['top_k_similarity'] tensor_info:
dtype: DT_FLOAT
shape: (-1, 20)
name: predictions/top_k_sim:0
outputs['top_k_string'] tensor_info:
dtype: DT_STRING
shape: (-1, 20)
name: predictions/top_k_string_lookup/LookupTableFindV2:0
Method name is: tensorflow/serving/predict
Once served, you may query the model with the following command.
Sample request
# make query MODEL_NAME=glove
MODEL_NAME=glove
curl -X POST \
http://localhost:8501/v1/models/${MODEL_NAME}:predict \
-d '{"instances": [{"row_token": "man", "col_token": "man"}]}'
Sample response
{
"predictions": [
{
"input_embedding": [
-0.39519611,
-0.000384220504,
0.360801637,
0.71601522,
-0.425830722,
-0.259146929,
-0.13219744,
0.307031065,
0.695665479,
-0.504015446,
...
],
"input_string": "man",
"top_k_similarity": [
0.99999994,
0.725774705,
0.707765281,
0.693533063,
0.679038405,
0.646895647,
0.642417192,
0.640380502,
0.63178885,
0.631023884,
...
],
"top_k_string": [
"man",
"person",
"god",
"woman",
"young",
"movie",
"great",
"good",
"himself",
"son",
...
]
}
]
}
For cloud model training and evaluation, please refer to cloud.
- Mahoney, M. (2006). Large Text Compression Benchmark. Retrieved from http://mattmahoney.net/dc/text.html.
- Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. GloVe: Global Vectors for Word Representation. [pdf][bib]