文件列表:

.dockerignore (95, 2018-05-30)
Dockerfile (595, 2018-05-30)
LICENSE (1050, 2018-05-30)
Makefile (824, 2018-05-30)
config.h (33, 2018-05-30)
data (0, 2018-05-30)
data\corpus (0, 2018-05-30)
data\corpus\dummy.data (74, 2018-05-30)
data\saves (0, 2018-05-30)
docker-compose.yml (1254, 2018-05-30)
main (0, 2018-05-30)
main\modes.cpp (2273, 2018-05-30)
main\svdistic.cpp (4768, 2018-05-30)
models (0, 2018-05-30)
models\base (0, 2018-05-30)
models\base\base.cpp (3239, 2018-05-30)
models\base\base.h (1393, 2018-05-30)
models\svd (0, 2018-05-30)
models\svd\helpers.cpp (1523, 2018-05-30)
models\svd\model.cpp (1008, 2018-05-30)
models\svd\model.h (739, 2018-05-30)
models\svdpp (0, 2018-05-30)
models\svdpp\helpers.cpp (2958, 2018-05-30)
models\svdpp\model.cpp (1363, 2018-05-30)
models\svdpp\model.h (871, 2018-05-30)
models\timesvdpp (0, 2018-05-30)
models\timesvdpp\helpers.cpp (1002, 2018-05-30)
models\timesvdpp\model.cpp (852, 2018-05-30)
models\timesvdpp\model.h (585, 2018-05-30)
remote_setup.sh (510, 2018-05-30)
types.h (573, 2018-05-30)
utils (0, 2018-05-30)
utils\pipes.cpp (1195, 2018-05-30)
utils\pipes.h (941, 2018-05-30)

# SVDistic Optimized-for-speed Eigen implementations of SVD, SVD++ and TimeSVD++ algorithms.  [](http://maintained.tech/) ## Requirements This application is fully Dockerized for easier usage. We suggest having at least Docker 18.03.1-ce installed. If you wish to do not wish to deploy through Docker, you must have Eigen installed. Find installation instructions here: . ## Speed. SVDistic was benchmarked on task of training an SVD and SVD++ model across 95 million examples (from the Netflix dataset) with 500,000 users and 18,000 products using 100 latent factors. Each model was allocated a single 2.3 GHz Intel Xeon E5-2686 v4 (Broadwell) core and each occupied a maximum of 1.7 GB of memory. The SVD++ model completed a full epoch in 32.3 seconds. The SVD model completed a full epoch in 11.3 seconds. ## Usage Usage instructions for the actual Svdistic program. ``` Usage: ./svdistic Options: required settings are flaired with [r] -model_id STRING: name of the model -fname STRING: file name of data in data/corpus/ -n_product INT: number of products -n_user INT: number of users -n_example INT: number of examples to process -report_freq INT: frequency of epoch reports -n_epochs INT: number of epochs to run training for -reg_weight FLOAT: weight regularization strength -reg_bias FLOAT: bias regularization strength -lr FLOAT: learning rate -lr_decay FLOAT: learning rate decay ``` Add your data files to /data/corpus and note the filename as command line arguments to the program. For training and validation, your data files must match the data format specified in the following section with three valid columns denoting user id, product id and ranking. For inference, your data file must still meet the data format, but fill in whatever you want for ranking. ## Docker Use Docker for painless hyperparameter optimization. Simply update the Dockerfile to download your dataset into data/corpus and add your desired hyperparameter cases to docker-compose. ## Data format. ### Input data. The standard data format is a csv file, where each entry corresponding to a ranking. Do not add a header line. Each line consists of 3 pairs: `user_index,product_index,score`. To calculate the user index, transform user ids to indices ranging from 0 to the number of users. The same process applies to product indexing. Scores are expected to be integer values, although we weakly support float rankings. Please note that **entries must be ordered by user index**. Entries with the same user index should be consecutive in the data file. ### Weight dumps. Weights dumps are a column-major iteration through matrix values. Every entry is separated by a newline.

近期下载者：

相关文件：

评论：[我要评论] [举报此文件]

收藏者：