search

PaddlePaddle / PaddleTS

Awesome Easy-to-Use Deep Time Series Modeling based on PaddlePaddle, including comprehensive functionality modules like TSDataset, Analysis, Transform, Models, AutoTS, and Ensemble, etc., supporting versatile tasks like time series forecasting, representation learning, and anomaly detection, etc., featured with quick tracking of SOTA deep models.
Apache License 2.0
482 stars 118 forks source link

使用TFT模型进行时序预测的output_quantiles维度问题 #480

Closed IgnoranceSmile closed 5 months ago

IgnoranceSmile commented 6 months ago

是否output_quantiles只支持长度为1和3，测试在其他长度下均会报错。

示例代码： import pandas as pd import numpy as np from paddlets import TSDataset

x = np.linspace(-np.pi, np.pi, 200) sinx = np.sin(x) * 4 + np.random.randn(200)

df = pd.DataFrame( { 'time_col': pd.date_range('2022-01-01', periods=200, freq='1h'), 'value': sinx } ) target_dataset = TSDataset.load_from_dataframe( df, #Also can be path to the CSV file time_col='time_col', target_cols='value', freq='1h' ) new_line = pd.Series( np.array(range(210)), index=pd.date_range('2022-01-01', periods=210, freq='1h') ) target_dataset.set_column( column='new_b', value=new_line, type='known_cov' ) train_ts, val_ts = target_dataset.split(0.7)

from paddlets.models.forecasting import TransformerModel, TFTModel, RNNBlockRegressor

model = TFTModel( in_chunk_len = 10, out_chunk_len = 10, output_quantiles = [0.1, 0.2] ) model.fit(train_ts)

报错：

ValueError Traceback (most recent call last) /tmp/ipykernel_20005/1436158038.py in 6 output_quantiles = [0.1, 0.2] 7 ) ----> 8 model.fit(train_ts)

/usr/local/lib/python3.7/dist-packages/paddlets/models/forecasting/dl/paddle_base_impl.py in fit(self, train_tsdataset, valid_tsdataset) 344 self._check_multi_tsdataset(valid_tsdataset) 345 train_dataloader, valid_dataloaders = self._init_fit_dataloaders(train_tsdataset, valid_tsdataset) --> 346 self._fit(train_dataloader, valid_dataloaders) 347 348 def _fit(

/usr/local/lib/python3.7/dist-packages/paddlets/models/forecasting/dl/paddle_base_impl.py in _fit(self, train_dataloader, valid_dataloaders) 370 # Call the on_epoch_begin method of each callback before the epoch starts. 371 self._callback_container.on_epoch_begin(epoch_idx) --> 372 self._train_epoch(train_dataloader) 373 374 # Predict for each eval set.

/usr/local/lib/python3.7/dist-packages/paddlets/models/forecasting/dl/paddle_base_impl.py in _train_epoch(self, train_loader) 438 self._callback_container.on_batch_begin(batch_idx) 439 X, y = self._prepare_X_y(data) --> 440 batch_logs = self._train_batch(X, y) ... 301 comment = OpProtoHolder.instance().get_op_proto(op_type).comment

ValueError: (InvalidArgument) Broadcast dimension mismatch. Operands could not be broadcast together with the shape of X = [3] and the shape of Y = [121, 10, 1, 2]. Received [3] in X is not equal to [2] in Y at i:3. [Hint: Expected x_dims_array[i] == y_dims_array[i] || x_dims_array[i] <= 1 || y_dims_array[i] <= 1 == true, but received x_dims_array[i] == y_dims_array[i] || x_dims_array[i] <= 1 || y_dims_array[i] <= 1:0 != true:1.] (at /paddle/paddle/phi/kernels/funcs/common_shape.h:84) [operator < elementwise_mul > error]

Sunting78 commented 6 months ago

您好，我们跟进复现下您的问题。

Sunting78 commented 6 months ago

您好，这个是因为损失函数使用了默认的3维度的quantiles。如果希望使用2维度，可以修改这行代码的quantiles。这个问题我们也会排期修复。https://github.com/PaddlePaddle/PaddleTS/blob/a3f275c51a1c6d6803ec2af269c049a29039ae59/paddlets/models/forecasting/dl/distributions/likelihood.py#L167