From 4a9163a864093d4a63f0cd9f40fc5b26770c447b Mon Sep 17 00:00:00 2001 From: Wenjie Du Date: Sun, 16 Jun 2024 00:05:31 +0800 Subject: [PATCH] feat: add Reformer as an imputation model; --- pypots/imputation/__init__.py | 2 + pypots/imputation/reformer/__init__.py | 25 ++ pypots/imputation/reformer/core.py | 88 +++++++ pypots/imputation/reformer/data.py | 24 ++ pypots/imputation/reformer/model.py | 331 +++++++++++++++++++++++++ tests/imputation/reformer.py | 130 ++++++++++ 6 files changed, 600 insertions(+) create mode 100644 pypots/imputation/reformer/__init__.py create mode 100644 pypots/imputation/reformer/core.py create mode 100644 pypots/imputation/reformer/data.py create mode 100644 pypots/imputation/reformer/model.py create mode 100644 tests/imputation/reformer.py diff --git a/pypots/imputation/__init__.py b/pypots/imputation/__init__.py index 3a9ba6ec..ed0fbd84 100644 --- a/pypots/imputation/__init__.py +++ b/pypots/imputation/__init__.py @@ -23,6 +23,7 @@ from .crossformer import Crossformer from .informer import Informer from .autoformer import Autoformer +from .reformer import Reformer from .dlinear import DLinear from .patchtst import PatchTST from .usgan import USGAN @@ -54,6 +55,7 @@ "DLinear", "Informer", "Autoformer", + "Reformer", "NonstationaryTransformer", "Pyraformer", "BRITS", diff --git a/pypots/imputation/reformer/__init__.py b/pypots/imputation/reformer/__init__.py new file mode 100644 index 00000000..ddd255bf --- /dev/null +++ b/pypots/imputation/reformer/__init__.py @@ -0,0 +1,25 @@ +""" +The package of the partially-observed time-series imputation model Reformer. + +Refer to the paper +`Kitaev, Nikita, Łukasz Kaiser, and Anselm Levskaya. +Reformer: The Efficient Transformer. +International Conference on Learning Representations, 2020. +`_ + +Notes +----- +This implementation is inspired by the official one https://github.com/google/trax/tree/master/trax/models/reformer and +https://github.com/lucidrains/reformer-pytorch + +""" + +# Created by Wenjie Du +# License: BSD-3-Clause + + +from .model import Reformer + +__all__ = [ + "Reformer", +] diff --git a/pypots/imputation/reformer/core.py b/pypots/imputation/reformer/core.py new file mode 100644 index 00000000..c1c70fe4 --- /dev/null +++ b/pypots/imputation/reformer/core.py @@ -0,0 +1,88 @@ +""" +The core wrapper assembles the submodules of Reformer imputation model +and takes over the forward progress of the algorithm. +""" + +# Created by Wenjie Du +# License: BSD-3-Clause + +import torch.nn as nn + +from ...nn.modules.reformer import ReformerEncoder +from ...nn.modules.saits import SaitsLoss, SaitsEmbedding + + +class _Reformer(nn.Module): + def __init__( + self, + n_steps, + n_features, + n_layers, + d_model, + n_heads, + bucket_size, + n_hashes, + causal, + d_ffn, + dropout, + ORT_weight: float = 1, + MIT_weight: float = 1, + ): + super().__init__() + + self.n_steps = n_steps + + self.saits_embedding = SaitsEmbedding( + n_features * 2, + d_model, + with_pos=False, + dropout=dropout, + ) + self.encoder = ReformerEncoder( + n_steps, + n_layers, + d_model, + n_heads, + bucket_size, + n_hashes, + causal, + d_ffn, + dropout, + ) + + # for the imputation task, the output dim is the same as input dim + self.output_projection = nn.Linear(d_model, n_features) + self.saits_loss_func = SaitsLoss(ORT_weight, MIT_weight) + + def forward(self, inputs: dict, training: bool = True) -> dict: + X, missing_mask = inputs["X"], inputs["missing_mask"] + + # WDU: the original Reformer paper isn't proposed for imputation task. Hence the model doesn't take + # the missing mask into account, which means, in the process, the model doesn't know which part of + # the input data is missing, and this may hurt the model's imputation performance. Therefore, I apply the + # SAITS embedding method to project the concatenation of features and masks into a hidden space, as well as + # the output layers to project back from the hidden space to the original space. + enc_out = self.saits_embedding(X, missing_mask) + + # Reformer encoder processing + enc_out = self.encoder(enc_out) + # project back the original data space + reconstruction = self.output_projection(enc_out) + + imputed_data = missing_mask * X + (1 - missing_mask) * reconstruction + results = { + "imputed_data": imputed_data, + } + + # if in training mode, return results with losses + if training: + X_ori, indicating_mask = inputs["X_ori"], inputs["indicating_mask"] + loss, ORT_loss, MIT_loss = self.saits_loss_func( + reconstruction, X_ori, missing_mask, indicating_mask + ) + results["ORT_loss"] = ORT_loss + results["MIT_loss"] = MIT_loss + # `loss` is always the item for backward propagating to update the model + results["loss"] = loss + + return results diff --git a/pypots/imputation/reformer/data.py b/pypots/imputation/reformer/data.py new file mode 100644 index 00000000..63f29969 --- /dev/null +++ b/pypots/imputation/reformer/data.py @@ -0,0 +1,24 @@ +""" +Dataset class for Reformer. +""" + +# Created by Wenjie Du +# License: BSD-3-Clause + +from typing import Union + +from ..saits.data import DatasetForSAITS + + +class DatasetForReformer(DatasetForSAITS): + """Actually Reformer uses the same data strategy as SAITS, needs MIT for training.""" + + def __init__( + self, + data: Union[dict, str], + return_X_ori: bool, + return_y: bool, + file_type: str = "hdf5", + rate: float = 0.2, + ): + super().__init__(data, return_X_ori, return_y, file_type, rate) diff --git a/pypots/imputation/reformer/model.py b/pypots/imputation/reformer/model.py new file mode 100644 index 00000000..47c21664 --- /dev/null +++ b/pypots/imputation/reformer/model.py @@ -0,0 +1,331 @@ +""" +The implementation of Reformer for the partially-observed time-series imputation task. + +""" + +# Created by Wenjie Du +# License: BSD-3-Clause + +from typing import Union, Optional + +import numpy as np +import torch +from torch.utils.data import DataLoader + +from .core import _Reformer +from .data import DatasetForReformer +from ..base import BaseNNImputer +from ...data.checking import key_in_data_set +from ...data.dataset import BaseDataset +from ...optim.adam import Adam +from ...optim.base import Optimizer + + +class Reformer(BaseNNImputer): + """The PyTorch implementation of the Reformer model. + Reformer is originally proposed by Kitaev et al. in :cite:`kitaev2020reformer`. + + Parameters + ---------- + n_steps : + The number of time steps in the time-series data sample. + + n_features : + The number of features in the time-series data sample. + + n_layers : + The number of layers in the Reformer model. + + d_model : + The dimension of the model. + + n_heads : + The number of heads in each layer of Reformer. + + bucket_size : + Average size of qk per bucket, 64 was recommended in paper. + + n_hashes : + 4 is permissible per author, 8 is the best but slower. + + causal : + Auto-regressive or not. + + d_ffn : + The dimension of the feed-forward network. + The window size of moving average. + + dropout : + The dropout rate for the model. + + ORT_weight : + The weight for the ORT loss, the same as SAITS. + + MIT_weight : + The weight for the MIT loss, the same as SAITS. + + batch_size : + The batch size for training and evaluating the model. + + epochs : + The number of epochs for training the model. + + patience : + The patience for the early-stopping mechanism. Given a positive integer, the training process will be + stopped when the model does not perform better after that number of epochs. + Leaving it default as None will disable the early-stopping. + + optimizer : + The optimizer for model training. + If not given, will use a default Adam optimizer. + + num_workers : + The number of subprocesses to use for data loading. + `0` means data loading will be in the main process, i.e. there won't be subprocesses. + + device : + The device for the model to run on. It can be a string, a :class:`torch.device` object, or a list of them. + If not given, will try to use CUDA devices first (will use the default CUDA device if there are multiple), + then CPUs, considering CUDA and CPU are so far the main devices for people to train ML models. + If given a list of devices, e.g. ['cuda:0', 'cuda:1'], or [torch.device('cuda:0'), torch.device('cuda:1')] , the + model will be parallely trained on the multiple devices (so far only support parallel training on CUDA devices). + Other devices like Google TPU and Apple Silicon accelerator MPS may be added in the future. + + saving_path : + The path for automatically saving model checkpoints and tensorboard files (i.e. loss values recorded during + training into a tensorboard file). Will not save if not given. + + model_saving_strategy : + The strategy to save model checkpoints. It has to be one of [None, "best", "better", "all"]. + No model will be saved when it is set as None. + The "best" strategy will only automatically save the best model after the training finished. + The "better" strategy will automatically save the model during training whenever the model performs + better than in previous epochs. + The "all" strategy will save every model after each epoch training. + + verbose : + Whether to print out the training logs during the training process. + """ + + def __init__( + self, + n_steps: int, + n_features: int, + n_layers: int, + d_model: int, + n_heads: int, + bucket_size: int, + n_hashes: int, + causal: bool, + d_ffn: int, + dropout: float = 0, + ORT_weight: float = 1, + MIT_weight: float = 1, + batch_size: int = 32, + epochs: int = 100, + patience: int = None, + optimizer: Optional[Optimizer] = Adam(), + num_workers: int = 0, + device: Optional[Union[str, torch.device, list]] = None, + saving_path: str = None, + model_saving_strategy: Optional[str] = "best", + verbose: bool = True, + ): + super().__init__( + batch_size, + epochs, + patience, + num_workers, + device, + saving_path, + model_saving_strategy, + verbose, + ) + + self.n_steps = n_steps + self.n_features = n_features + # model hype-parameters + self.n_heads = n_heads + self.n_layers = n_layers + self.d_model = d_model + self.bucket_size = bucket_size + self.n_hashes = n_hashes + self.causal = causal + self.d_ffn = d_ffn + self.dropout = dropout + self.ORT_weight = ORT_weight + self.MIT_weight = MIT_weight + + # set up the model + self.model = _Reformer( + self.n_steps, + self.n_features, + self.n_layers, + self.d_model, + self.n_heads, + self.bucket_size, + self.n_hashes, + self.causal, + self.d_ffn, + self.dropout, + self.ORT_weight, + self.MIT_weight, + ) + self._send_model_to_given_device() + self._print_model_size() + + # set up the optimizer + self.optimizer = optimizer + self.optimizer.init_optimizer(self.model.parameters()) + + def _assemble_input_for_training(self, data: list) -> dict: + ( + indices, + X, + missing_mask, + X_ori, + indicating_mask, + ) = self._send_data_to_given_device(data) + + inputs = { + "X": X, + "missing_mask": missing_mask, + "X_ori": X_ori, + "indicating_mask": indicating_mask, + } + + return inputs + + def _assemble_input_for_validating(self, data: list) -> dict: + return self._assemble_input_for_training(data) + + def _assemble_input_for_testing(self, data: list) -> dict: + indices, X, missing_mask = self._send_data_to_given_device(data) + + inputs = { + "X": X, + "missing_mask": missing_mask, + } + + return inputs + + def fit( + self, + train_set: Union[dict, str], + val_set: Optional[Union[dict, str]] = None, + file_type: str = "hdf5", + ) -> None: + # Step 1: wrap the input data with classes Dataset and DataLoader + training_set = DatasetForReformer( + train_set, return_X_ori=False, return_y=False, file_type=file_type + ) + training_loader = DataLoader( + training_set, + batch_size=self.batch_size, + shuffle=True, + num_workers=self.num_workers, + ) + val_loader = None + if val_set is not None: + if not key_in_data_set("X_ori", val_set): + raise ValueError("val_set must contain 'X_ori' for model validation.") + val_set = DatasetForReformer( + val_set, return_X_ori=True, return_y=False, file_type=file_type + ) + val_loader = DataLoader( + val_set, + batch_size=self.batch_size, + shuffle=False, + num_workers=self.num_workers, + ) + + # Step 2: train the model and freeze it + self._train_model(training_loader, val_loader) + self.model.load_state_dict(self.best_model_dict) + self.model.eval() # set the model as eval status to freeze it. + + # Step 3: save the model if necessary + self._auto_save_model_if_necessary(confirm_saving=True) + + def predict( + self, + test_set: Union[dict, str], + file_type: str = "hdf5", + ) -> dict: + """Make predictions for the input data with the trained model. + + Parameters + ---------- + test_set : dict or str + The dataset for model validating, should be a dictionary including keys as 'X', + or a path string locating a data file supported by PyPOTS (e.g. h5 file). + If it is a dict, X should be array-like of shape [n_samples, sequence length (n_steps), n_features], + which is time-series data for validating, can contain missing values, and y should be array-like of shape + [n_samples], which is classification labels of X. + If it is a path string, the path should point to a data file, e.g. a h5 file, which contains + key-value pairs like a dict, and it has to include keys as 'X' and 'y'. + + file_type : + The type of the given file if test_set is a path string. + + Returns + ------- + file_type : + The dictionary containing the clustering results and latent variables if necessary. + + """ + # Step 1: wrap the input data with classes Dataset and DataLoader + self.model.eval() # set the model as eval status to freeze it. + test_set = BaseDataset( + test_set, + return_X_ori=False, + return_X_pred=False, + return_y=False, + file_type=file_type, + ) + test_loader = DataLoader( + test_set, + batch_size=self.batch_size, + shuffle=False, + num_workers=self.num_workers, + ) + imputation_collector = [] + + # Step 2: process the data with the model + with torch.no_grad(): + for idx, data in enumerate(test_loader): + inputs = self._assemble_input_for_testing(data) + results = self.model.forward(inputs, training=False) + imputation_collector.append(results["imputed_data"]) + + # Step 3: output collection and return + imputation = torch.cat(imputation_collector).cpu().detach().numpy() + result_dict = { + "imputation": imputation, + } + return result_dict + + def impute( + self, + test_set: Union[dict, str], + file_type: str = "hdf5", + ) -> np.ndarray: + """Impute missing values in the given data with the trained model. + + Parameters + ---------- + test_set : + The data samples for testing, should be array-like of shape [n_samples, sequence length (n_steps), + n_features], or a path string locating a data file, e.g. h5 file. + + file_type : + The type of the given file if X is a path string. + + Returns + ------- + array-like, shape [n_samples, sequence length (n_steps), n_features], + Imputed data. + """ + + result_dict = self.predict(test_set, file_type=file_type) + return result_dict["imputation"] diff --git a/tests/imputation/reformer.py b/tests/imputation/reformer.py new file mode 100644 index 00000000..15b3d749 --- /dev/null +++ b/tests/imputation/reformer.py @@ -0,0 +1,130 @@ +""" +Test cases for Reformer imputation model. +""" + +# Created by Wenjie Du +# License: BSD-3-Clause + + +import os.path +import unittest + +import numpy as np +import pytest + +from pypots.imputation import Reformer +from pypots.optim import Adam +from pypots.utils.logging import logger +from pypots.utils.metrics import calc_mse +from tests.global_test_config import ( + DATA, + EPOCHS, + DEVICE, + TRAIN_SET, + VAL_SET, + TEST_SET, + GENERAL_H5_TRAIN_SET_PATH, + GENERAL_H5_VAL_SET_PATH, + GENERAL_H5_TEST_SET_PATH, + RESULT_SAVING_DIR_FOR_IMPUTATION, + check_tb_and_model_checkpoints_existence, +) + + +class TestReformer(unittest.TestCase): + logger.info("Running tests for an imputation model Reformer...") + + # set the log and model saving path + saving_path = os.path.join(RESULT_SAVING_DIR_FOR_IMPUTATION, "Reformer") + model_save_name = "saved_reformer_model.pypots" + + # initialize an Adam optimizer + optimizer = Adam(lr=0.001, weight_decay=1e-5) + + # initialize a Reformer model + reformer = Reformer( + DATA["n_steps"], + DATA["n_features"], + n_layers=2, + d_model=32, + n_heads=2, + bucket_size=4, + n_hashes=4, + causal=True, + d_ffn=32, + dropout=0, + epochs=EPOCHS, + saving_path=saving_path, + optimizer=optimizer, + device=DEVICE, + ) + + @pytest.mark.xdist_group(name="imputation-reformer") + def test_0_fit(self): + self.reformer.fit(TRAIN_SET, VAL_SET) + + @pytest.mark.xdist_group(name="imputation-reformer") + def test_1_impute(self): + imputation_results = self.reformer.predict(TEST_SET) + assert not np.isnan( + imputation_results["imputation"] + ).any(), "Output still has missing values after running impute()." + + test_MSE = calc_mse( + imputation_results["imputation"], + DATA["test_X_ori"], + DATA["test_X_indicating_mask"], + ) + logger.info(f"Reformer test_MSE: {test_MSE}") + + @pytest.mark.xdist_group(name="imputation-reformer") + def test_2_parameters(self): + assert hasattr(self.reformer, "model") and self.reformer.model is not None + + assert ( + hasattr(self.reformer, "optimizer") and self.reformer.optimizer is not None + ) + + assert hasattr(self.reformer, "best_loss") + self.assertNotEqual(self.reformer.best_loss, float("inf")) + + assert ( + hasattr(self.reformer, "best_model_dict") + and self.reformer.best_model_dict is not None + ) + + @pytest.mark.xdist_group(name="imputation-reformer") + def test_3_saving_path(self): + # whether the root saving dir exists, which should be created by save_log_into_tb_file + assert os.path.exists( + self.saving_path + ), f"file {self.saving_path} does not exist" + + # check if the tensorboard file and model checkpoints exist + check_tb_and_model_checkpoints_existence(self.reformer) + + # save the trained model into file, and check if the path exists + saved_model_path = os.path.join(self.saving_path, self.model_save_name) + self.reformer.save(saved_model_path) + + # test loading the saved model, not necessary, but need to test + self.reformer.load(saved_model_path) + + @pytest.mark.xdist_group(name="imputation-reformer") + def test_4_lazy_loading(self): + self.reformer.fit(GENERAL_H5_TRAIN_SET_PATH, GENERAL_H5_VAL_SET_PATH) + imputation_results = self.reformer.predict(GENERAL_H5_TEST_SET_PATH) + assert not np.isnan( + imputation_results["imputation"] + ).any(), "Output still has missing values after running impute()." + + test_MSE = calc_mse( + imputation_results["imputation"], + DATA["test_X_ori"], + DATA["test_X_indicating_mask"], + ) + logger.info(f"Lazy-loading Reformer test_MSE: {test_MSE}") + + +if __name__ == "__main__": + unittest.main()