#
# Copyright 2016 The BigDL Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import torch
from bigdl.chronos.forecaster.base_forecaster import BasePytorchForecaster
from bigdl.chronos.model.VanillaLSTM_pytorch import model_creator, optimizer_creator, loss_creator
[docs]class LSTMForecaster(BasePytorchForecaster):
"""
Example:
>>> #The dataset is split into x_train, x_val, x_test, y_train, y_val, y_test
>>> # 1. Initialize Forecaster directly
>>> forecaster = LSTMForecaster(past_seq_len=24,
input_feature_num=2,
output_feature_num=2,
...)
>>>
>>> # 2. Initialize Forecaster from from_tsdataset
>>> forecaster = LSTMForecaster.from_tsdataset(tsdata, **kwargs)
>>> forecaster.fit(tsdata, epochs=2, ...)
>>> forecaster.to_local() # if you set distributed=True
>>> test_pred = forecaster.predict(x_test)
>>> test_eval = forecaster.evaluate((x_test, y_test))
>>> forecaster.save({ckpt_name})
>>> forecaster.load({ckpt_name})
"""
def __init__(self,
past_seq_len,
input_feature_num,
output_feature_num,
hidden_dim=32,
layer_num=1,
normalization=True,
decomposition_kernel_size=0,
dropout=0.1,
optimizer="Adam",
loss="mse",
lr=0.001,
metrics=["mse"],
seed=None,
distributed=False,
workers_per_node=1,
distributed_backend="ray"):
"""
Build a LSTM Forecast Model.
:param past_seq_len: Specify the history time steps (i.e. lookback).
:param input_feature_num: Specify the feature dimension.
:param output_feature_num: Specify the output dimension.
:param hidden_dim: int or list, Specify the hidden dim of each lstm layer.
The value defaults to 32.
:param layer_num: Specify the number of lstm layer to be used. The value
defaults to 1.
:param normalization: bool, Specify if to use normalization trick to
alleviate distribution shift. It first subtractes the last value
of the sequence and add back after the model forwarding.
:param decomposition_kernel_size: int, Specify the kernel size in moving
average. The decomposition method will be applied if and only if
decomposition_kernel_size is greater than 1, which first decomposes
the raw sequence into a trend component by a moving average kernel
and a remainder(seasonal) component. Then, two models are applied
to each component and sum up the two outputs to get the final
prediction. This value defaults to 0.
:param dropout: int or list, Specify the dropout close possibility
(i.e. the close possibility to a neuron). This value defaults to 0.1.
:param optimizer: Specify the optimizer used for training. This value
defaults to "Adam".
:param loss: str or pytorch loss instance, Specify the loss function
used for training. This value defaults to "mse". You can choose
from "mse", "mae", "huber_loss" or any customized loss instance
you want to use.
:param lr: Specify the learning rate. This value defaults to 0.001.
:param metrics: A list contains metrics for evaluating the quality of
forecasting. You may only choose from "mse" and "mae" for a
distributed forecaster. You may choose from "mse", "mae",
"rmse", "r2", "mape", "smape" or a callable function for a
non-distributed forecaster. If callable function, it signature
should be func(y_true, y_pred), where y_true and y_pred are numpy
ndarray.
:param seed: int, random seed for training. This value defaults to None.
:param distributed: bool, if init the forecaster in a distributed
fashion. If True, the internal model will use an Orca Estimator.
If False, the internal model will use a pytorch model. The value
defaults to False.
:param workers_per_node: int, the number of worker you want to use.
The value defaults to 1. The param is only effective when
distributed is set to True.
:param distributed_backend: str, select from "ray" or
"horovod". The value defaults to "ray".
"""
# config setting
self.data_config = {
"past_seq_len": past_seq_len,
"future_seq_len": 1, # lstm model only supports 1 step prediction
"input_feature_num": input_feature_num,
"output_feature_num": output_feature_num
}
self.model_config = {
"hidden_dim": hidden_dim,
"layer_num": layer_num,
"dropout": dropout,
"seed": seed,
"normalization": normalization,
"decomposition_kernel_size": decomposition_kernel_size
}
self.loss_config = {
"loss": loss
}
self.optim_config = {
"lr": lr,
"optim": optimizer
}
# model creator settings
self.model_creator = model_creator
self.optimizer_creator = optimizer_creator
if isinstance(loss, str):
self.loss_creator = loss_creator
else:
def customized_loss_creator(config):
return config["loss"]
self.loss_creator = customized_loss_creator
# distributed settings
self.distributed = distributed
self.remote_distributed_backend = distributed_backend
self.local_distributed_backend = "subprocess"
self.workers_per_node = workers_per_node
# other settings
self.lr = lr
self.metrics = metrics
self.seed = seed
# nano setting
current_num_threads = torch.get_num_threads()
self.thread_num = current_num_threads
self.optimized_model_thread_num = current_num_threads
if current_num_threads >= 24:
self.num_processes = max(1, current_num_threads//8) # 8 is a magic num
else:
self.num_processes = 1
self.use_ipex = False
self.onnx_available = True
self.quantize_available = True
self.checkpoint_callback = True
self.use_hpo = True
self.optimized_model_output_tensor = True
super().__init__()
[docs] @classmethod
def from_tsdataset(cls, tsdataset, past_seq_len=None, **kwargs):
'''
Build a LSTM Forecaster Model.
:param tsdataset: Train tsdataset, a bigdl.chronos.data.tsdataset.TSDataset instance.
:param past_seq_len: Specify the history time steps (i.e. lookback).
Do not specify the 'past_seq_len' if your tsdataset has called
the 'TSDataset.roll' method or 'TSDataset.to_torch_data_loader'.
:param kwargs: Specify parameters of Forecaster,
e.g. loss and optimizer, etc. More info, please refer to
LSTMForecaster.__init__ methods.
:return: A LSTM Forecaster Model.
'''
from bigdl.chronos.data.tsdataset import TSDataset
from bigdl.nano.utils.common import invalidInputError
invalidInputError(isinstance(tsdataset, TSDataset),
f"We only supports input a TSDataset, but get{type(tsdataset)}.")
def check_time_steps(tsdataset, past_seq_len):
if tsdataset.lookback is not None and past_seq_len is not None:
return tsdataset.lookback == past_seq_len
return True
invalidInputError(not tsdataset._has_generate_agg_feature,
"We will add su`pport for 'gen_rolling_feature' method later.")
if tsdataset.lookback is not None: # calling roll or to_torch_data_loader
past_seq_len = tsdataset.lookback
output_feature_num = len(tsdataset.roll_target)
input_feature_num = len(tsdataset.roll_feature) + output_feature_num
elif past_seq_len is not None: # initialize only
past_seq_len = past_seq_len if isinstance(past_seq_len, int)\
else tsdataset.get_cycle_length()
output_feature_num = len(tsdataset.target_col)
input_feature_num = len(tsdataset.feature_col) + output_feature_num
else:
invalidInputError(False,
"Forecaster needs 'past_seq_len' to specify "
"the history time step of training.")
invalidInputError(check_time_steps(tsdataset, past_seq_len),
"tsdataset already has history time steps and "
"differs from the given past_seq_len "
f"Expected past_seq_len to be {tsdataset.lookback}, "
f"but found {past_seq_len}.",
fixMsg="Do not specify past_seq_len "
"or call tsdataset.roll method again and specify time step.")
if tsdataset.id_sensitive:
_id_list_len = len(tsdataset._id_list)
input_feature_num *= _id_list_len
output_feature_num *= _id_list_len
return cls(past_seq_len=past_seq_len,
input_feature_num=input_feature_num,
output_feature_num=output_feature_num,
**kwargs)