Sequence

Deep learning models for sequence related data, i.e. (language, sessions)

Installation

Only install dependencies by running or $ pip install -r requirements.txt or install as (editable) library by pip install -e .

STAMP: Short-Term Attention/Memory Priority Model forSession-based Recommendation

See paper

Run STAMP on Yoochoose 1/64:

$ python run.py stamp \
--dataset='Yoochoose 1/64' \
--embedding_dim=100 \
-e=10 \
--lr=0.001  \
--batch_size=32 \
--model=stamp \
--scale_loss_by_lengths=false \
--train_percentage=0.95

Results on test set

Dataset	Yoochoose 1/64
Measures	P@20	MRR@20
STMP	64.44	30.45
STAMP	65.36	30.84

Generating Sentences from a Continuous Space

See paper

Creating a dataset

A dataset can be created from custom data and passed as path to a pickled file to the argument parser.

from sequence.data.utils import Dataset
import pickle
import pandas as pd

df = pd.DataFrame({"sessions": ["ses_1", "ses_2", "ses_1", "ses_1", "ses_1", "ses_2"],
              "pages": ["foo", "bar", "spam", "eggs", "spam", "home"]
             })
df

sessions	pages
ses_1	foo
ses_2	bar
ses_1	spam
ses_1	eggs
ses_1	spam
ses_2	home

grouped = df.groupby("sessions").agg(list)
grouped

pages
[foo, spam, eggs, spam]
[bar, home]

dataset = Dataset(
    sentences=[path for path in grouped["pages"]],
    min_len=1,
    max_len=20
)

with open("somepath.pkl", "wb") as f:
    pickle.dump(dataset, f)

And then running a model w/: $ python run.py stamp --dataset='somepath.pkl'

Options

run.py [-h] [--logging_name LOGGING_NAME] [--batch_size BATCH_SIZE]
              [--save_every_n SAVE_EVERY_N] [--embedding_dim EMBEDDING_DIM]
              [--storage_dir STORAGE_DIR] [--tensorboard TENSORBOARD]
              [--lr LR] [-e EPOCHS] [--min_length MIN_LENGTH]
              [--max_length MAX_LENGTH] [--train_percentage TRAIN_PERCENTAGE]
              [--dataset DATASET] [--force_cpu FORCE_CPU]
              [--weight_decay WEIGHT_DECAY] [--global_step GLOBAL_STEP]
              [--continue MODEL_REGISTRY_PATH] [--optimizer OPTIMIZER]
              {vae,stamp} ...

positional arguments:
  {vae,stamp}
    vae                 Run VAE model
    stamp               Run ST(A)MP model

optional arguments:
  -h, --help            show this help message and exit
  --logging_name LOGGING_NAME
  --batch_size BATCH_SIZE
  --save_every_n SAVE_EVERY_N
                        Save every n batches
  --embedding_dim EMBEDDING_DIM
  --storage_dir STORAGE_DIR
  --tensorboard TENSORBOARD
  --lr LR
  -e EPOCHS, --epochs EPOCHS
  --min_length MIN_LENGTH
                        Minimum sequence length
  --max_length MAX_LENGTH
                        Maximum sequence length
  --train_percentage TRAIN_PERCENTAGE
  --dataset DATASET     Pickled dataset file path, or named dataset (brown,
                        treebank, Yoochoose 1/64). If none given, NLTK BROWN
                        dataset will be used
  --force_cpu FORCE_CPU
  --weight_decay WEIGHT_DECAY
  --global_step GLOBAL_STEP
                        Overwrite global step.
  --continue MODEL_REGISTRY_PATH
                        Path to existing ModelRegistry
  --optimizer OPTIMIZER
                        adam|sgd