mirror of
https://github.com/huggingface/transformers.git
synced 2025-07-04 21:30:07 +06:00
2ac5b9325e
* Initial commit of PatchTST model classes Co-authored-by: Phanwadee Sinthong <phsinthong@gmail.com> Co-authored-by: Nam Nguyen <namctin@gmail.com> Co-authored-by: Vijay Ekambaram <vijaykr.e@gmail.com> Co-authored-by: Ngoc Diep Do <55230119+diepi@users.noreply.github.com> Co-authored-by: Wesley Gifford <79663411+wgifford@users.noreply.github.com> * Add PatchTSTForPretraining * update to include classification Co-authored-by: Phanwadee Sinthong <phsinthong@gmail.com> Co-authored-by: Nam Nguyen <namctin@gmail.com> Co-authored-by: Vijay Ekambaram <vijaykr.e@gmail.com> Co-authored-by: Ngoc Diep Do <55230119+diepi@users.noreply.github.com> Co-authored-by: Wesley Gifford <79663411+wgifford@users.noreply.github.com> * clean up auto files * Add PatchTSTForPrediction * Fix relative import * Replace original PatchTSTEncoder with ChannelAttentionPatchTSTEncoder * temporary adding absolute path + add PatchTSTForForecasting class * Update base PatchTSTModel + Unittest * Update ForecastHead to use the config class * edit cv_random_masking, add mask to model output * Update configuration_patchtst.py * add masked_loss to the pretraining * add PatchEmbeddings * Update configuration_patchtst.py * edit loss which considers mask in the pretraining * remove patch_last option * Add commits from internal repo * Update ForecastHead * Add model weight initilization + unittest * Update PatchTST unittest to use local import * PatchTST integration tests for pretraining and prediction * Added PatchTSTForRegression + update unittest to include label generation * Revert unrelated model test file * Combine similar output classes * update PredictionHead * Update configuration_patchtst.py * Add Revin * small edit to PatchTSTModelOutputWithNoAttention * Update modeling_patchtst.py * Updating integration test for forecasting * Fix unittest after class structure changed * docstring updates * change input_size to num_input_channels * more formatting * Remove some unused params * Add a comment for pretrained models * add channel_attention option add channel_attention option and remove unused positional encoders. * Update PatchTST models to use HF's MultiHeadAttention module * Update paper + github urls * Fix hidden_state return value * Update integration test to use PatchTSTForForecasting * Adding dataclass decorator for model output classes * Run fixup script * Rename model repos for integration test * edit argument explanation * change individual option to shared_projection * style * Rename integration test + import cleanup * Fix outpu_hidden_states return value * removed unused mode * added std, mean and nops scaler * add initial distributional loss for predition * fix typo in docs * add generate function * formatting * add num_parallel_samples * Fix a typo * copy weighted_average function, edit PredictionHead * edit PredictionHead * add distribution head to forecasting * formatting * Add generate function for forecasting * Add generate function to prediction task * formatting * use argsort * add past_observed_mask ordering * fix arguments * docs * add back test_model_outputs_equivalence test * formatting * cleanup * formatting * use ACT2CLS * formatting * fix add_start_docstrings decorator * add distribution head and generate function to regression task add distribution head and generate function to regression task. Also made add PatchTSTForForecastingOutput, PatchTSTForRegressionOutput. * add distribution head and generate function to regression task add distribution head and generate function to regression task. Also made add PatchTSTForForecastingOutput, PatchTSTForRegressionOutput. * fix typos * add forecast_masking * fixed tests * use set_seed * fix doc test * formatting * Update docs/source/en/model_doc/patchtst.md Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com> * better var names * rename PatchTSTTranspose * fix argument names and docs string * remove compute_num_patches and unused class * remove assert * renamed to PatchTSTMasking * use num_labels for classification * use num_labels * use default num_labels from super class * move model_type after docstring * renamed PatchTSTForMaskPretraining * bs -> batch_size * more review fixes * use hidden_state * rename encoder layer and block class * remove commented seed_number * edit docstring * Add docstring * formatting * use past_observed_mask * doc suggestion * make fix-copies * use Args: * add docstring * add docstring * change some variable names and add PatchTST before some class names * formatting * fix argument types * fix tests * change x variable to patch_input * format * formatting * fix-copies * Update tests/models/patchtst/test_modeling_patchtst.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * move loss to forward * Update src/transformers/models/patchtst/modeling_patchtst.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update src/transformers/models/patchtst/modeling_patchtst.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update src/transformers/models/patchtst/modeling_patchtst.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update src/transformers/models/patchtst/modeling_patchtst.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update src/transformers/models/patchtst/modeling_patchtst.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * formatting * fix a bug when pre_norm is set to True * output_hidden_states is set to False as default * set pre_norm=True as default * format docstring * format * output_hidden_states is None by default * add missing docs * better var names * docstring: remove default to False in output_hidden_states * change labels name to target_values in regression task * format * fix tests * change to forecast_mask_ratios and random_mask_ratio * change mask names * change future_values to target_values param in the prediction class * remove nn.Sequential and make PatchTSTBatchNorm class * black * fix argument name for prediction * add output_attentions option * add output_attentions to PatchTSTEncoder * formatting * Add attention output option to all classes * Remove PatchTSTEncoderBlock * create PatchTSTEmbedding class * use config in PatchTSTPatchify * Use config in PatchTSTMasking class * add channel_attn_weights * Add PatchTSTScaler class * add output_attentions arg to test function * format * Update doc with image patchtst.md * fix-copies * rename Forecast <-> Prediction * change name of a few parameters to match with PatchTSMixer. * Remove *ForForecasting class to match with other time series models. * make style * Remove PatchTSTForForecasting in the test * remove PatchTSTForForecastingOutput class * change test_forecast_head to test_prediction_head * style * fix docs * fix tests * change num_labels to num_targets * Remove PatchTSTTranspose * remove arguments in PatchTSTMeanScaler * remove arguments in PatchTSTStdScaler * add config as an argument to all the scaler classes * reformat * Add norm_eps for batchnorm and layernorm * reformat. * reformat * edit docstring * update docstring * change variable name pooling to pooling_type * fix output_hidden_states as tuple * fix bug when calling PatchTSTBatchNorm * change stride to patch_stride * create PatchTSTPositionalEncoding class and restructure the PatchTSTEncoder * formatting * initialize scalers with configs * edit output_hidden_states * style * fix forecast_mask_patches doc string --------- Co-authored-by: Gift Sinthong <gift.sinthong@ibm.com> Co-authored-by: Nam Nguyen <namctin@gmail.com> Co-authored-by: Vijay Ekambaram <vijaykr.e@gmail.com> Co-authored-by: Ngoc Diep Do <55230119+diepi@users.noreply.github.com> Co-authored-by: Wesley Gifford <79663411+wgifford@users.noreply.github.com> Co-authored-by: Wesley M. Gifford <wmgifford@us.ibm.com> Co-authored-by: nnguyen <nnguyen@us.ibm.com> Co-authored-by: Ngoc Diep Do <diiepy@gmail.com> Co-authored-by: Kashif Rasul <kashif.rasul@gmail.com> Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com> Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
3.4 KiB
3.4 KiB
PatchTST
Overview
The PatchTST model was proposed in A Time Series is Worth 64 Words: Long-term Forecasting with Transformers by Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, Jayant Kalagnanam.
The abstract from the paper is the following:
We propose an efficient design of Transformer-based models for multivariate time series forecasting and self-supervised representation learning. It is based on two key components: (i) segmentation of time series into subseries-level patches which are served as input tokens to Transformer; (ii) channel-independence where each channel contains a single univariate time series that shares the same embedding and Transformer weights across all the series. Patching design naturally has three-fold benefit: local semantic information is retained in the embedding; computation and memory usage of the attention maps are quadratically reduced given the same look-back window; and the model can attend longer history. Our channel-independent patch time series Transformer (PatchTST) can improve the long-term forecasting accuracy significantly when compared with that of SOTA Transformer-based models. We also apply our model to self-supervised pre-training tasks and attain excellent fine-tuning performance, which outperforms supervised training on large datasets. Transferring of masked pre-trained representation on one dataset to others also produces SOTA forecasting accuracy.
Tips:
The model can also be used for time series classification and time series regression. See the respective [PatchTSTForClassification] and [PatchTSTForRegression] classes.
At a high level the model vectorizes time series into patches of a given size and encodes them via a Transformer which then outputs the prediction length forecasts:
This model was contributed by namctin, gsinthong, diepi, vijaye12, wmgifford, and kashif.
The original code can be found here.
PatchTSTConfig
autodoc PatchTSTConfig
PatchTSTModel
autodoc PatchTSTModel - forward
PatchTSTForPrediction
autodoc PatchTSTForPrediction - forward
PatchTSTForClassification
autodoc PatchTSTForClassification - forward
PatchTSTForPretraining
autodoc PatchTSTForPretraining - forward
PatchTSTForRegression
autodoc PatchTSTForRegression - forward