bale-go
commented
1 year ago Unfortunately, TSiTPlus does not work for multi-step forecasting either. Despite being mentioned as a working model in the description.
Closed CHonChou closed 1 year ago
bale-go
commented
1 year ago Unfortunately, TSiTPlus does not work for multi-step forecasting either. Despite being mentioned as a working model in the description.
MOREDataset
commented
1 year ago TSTPlus works fine for me. The only problems I get are with mWDNPlus and XceptionTimePlus..
bale-go
commented
1 year ago TSTPlus works for me too. It is the TSiTPlus that does not run.
bale-go
commented
1 year ago I solved the issue with TSiTPlus. The function "forward" in class "_TSiTBackbone" needed an extra line (x = x.contiguous()) after transposing and before returning the tensor x. I can make a PR with the fix if you'd like.
oguiza
commented
1 year ago Hi @CHonChou, @MOREDataset, @bale-go, could any of you please post the full stack trace of this type of error?
bale-go
commented
1 year ago Minimal reproducible example for TSiTPlus using tsai-0.3.2:
from tsai.all import *
ts = get_forecasting_time_series("Sunspots").values
X, y = SlidingWindow(60, horizon=3)(ts)
splits = TimeSplitter(235)(y)
batch_tfms = TSStandardize()
fcst = TSForecaster(X, y, splits=splits, path='models', batch_tfms=batch_tfms, bs=512, arch=TSiTPlus, metrics=mae, cbs=ShowGraph())
fcst.fit_one_cycle(50, 1e-3)
fcst.export("fcst.pkl")Output:
Dataset: Sunspots
downloading data...
...data downloaded. Path = data/forecasting/Sunspots.csv
use_token set to False as c_out == 1
epoch train_loss valid_loss mae time
Traceback (most recent call last):--------------------------------------------------------------------------------| 0.00% [0/5 00:00<?]
File "mre.py", line 7, in
bale-go
commented
1 year ago With XceptionTimePlus:
from tsai.all import *
ts = get_forecasting_time_series("Sunspots").values
X, y = SlidingWindow(60, horizon=3)(ts)
splits = TimeSplitter(235)(y)
batch_tfms = TSStandardize()
fcst = TSForecaster(X, y, splits=splits, path='models', batch_tfms=batch_tfms, bs=512, arch=XceptionTimePlus, metrics=mae, cbs=ShowGraph())
fcst.fit_one_cycle(50, 1e-3)
fcst.export("fcst.pkl")Dataset: Sunspots
downloading data...
...data downloaded. Path = data/forecasting/Sunspots.csv
Traceback (most recent call last):
File "mre.py", line 6, in
MOREDataset
commented
1 year ago Hi @oguiza, I was wondering if there are any updates on these issues..
oguiza
commented
1 year ago No, not yet. But I'll prioritize it and will try to fix it within the next few days.
oguiza
commented
1 year ago Hi @CHonChou, @bale-go, @MOREDataset, I have an update on this issue. The models that currently support multi-horizon forecasting are:
TSTPlusTSiTPlusPatchTSTInceptionTimePlusXCMPlusLSTMPlusGRUPlusmWDNPlusMiniRocketPlusXceptionTimePlus
(Edit: you should only try this functionality with models ending in Plus).
To avoid issues, you'll need to temporarily pip install from GitHub until tsai release 0.3.6 is available. pip install git+https://github.com/timeseriesAI/tsai.gitAll of them work with this code for training:
from tsai.basics import *
fcst_history = 60
fcst_horizon = 3
df = get_forecasting_time_series("Sunspots")
X, y = prepare_forecasting_data(df, fcst_history=fcst_history, fcst_horizon=fcst_horizon, dtype=np.float32)
splits = get_forecasting_splits(df, fcst_history=fcst_history, fcst_horizon=fcst_horizon, test_size=235)
batch_tfms = TSStandardize()
fcst = TSForecaster(X, y, splits=splits, path='models', batch_tfms=batch_tfms, bs=512, arch="TSiTPlus", metrics=mae, cbs=ShowGraph())
fcst.fit_one_cycle(50, 1e-3)
fcst.export("fcst.pkl")(the advantage of passing the arch as a string is that you don't need to import it beforehand).
And this one for inference:
from tsai.inference import load_learner
fcst = load_learner("models/fcst.pkl")
preds, *_ = fcst.get_X_preds(X[splits[1]])If you try them let me know if you find any issues.
MOREDataset
commented
1 year ago Hi @oguiza, THANK YOU VERY MUCH for your continuous hard work on this amazing library!
I can confirm that the models TSTPlus + LSTMPlus + MiniRocketPlus + mWDNPlus + XCMPlus + GRUPlus + XceptionTimePlus + TSiTPlus work now for multivariate multihorizon forecasting!
However, there is a problem with PatchTST yielding the following error:
PatchTST
oguiza
commented
1 year ago Hi @MOREDataset, I would need a code snippet to reproduce the issue. This code works for me:
from tsai.basics import *
fcst_history = 60
fcst_horizon = 3
df = get_forecasting_time_series("Sunspots")
X, y = prepare_forecasting_data(df, fcst_history=fcst_history, fcst_horizon=fcst_horizon, dtype=np.float32)
splits = get_forecasting_splits(df, fcst_history=fcst_history, fcst_horizon=fcst_horizon, test_size=235)
batch_tfms = TSStandardize()
fcst = TSForecaster(X, y, splits=splits, path='models', batch_tfms=batch_tfms, bs=512, arch="PatchTST", metrics=mae, cbs=ShowGraph())
fcst.fit_one_cycle(50, 1e-3)
fcst.export("fcst.pkl")HI @oguiza,
Here is the snippet of the code I used:
models = {'PatchTST'}#'TSTPlus','XCMPlus','LSTMPlus','GRUPlus','MiniRocketPlus'}
for key in models:
exec('model_name = {}'.format(key))
splits = get_splits(y_train, valid_size=.2, stratify=False, random_state=42, shuffle=True)
tfms = [None, [ToFloat(), ToNumpyTensor()]]
dsets = TSDatasets(X_train, y_train, tfms=tfms, splits=splits, inplace=True)
dls = TSDataLoaders.from_dsets(dsets.train, dsets.valid, bs=[64, 128])
par = {}
#Model learning
learn = TSForecaster(X_train, y_train, splits=splits, tfms=tfms, bs=1042, arch=model_name, arch_config=par, metrics=[mae, rmse], cbs=[ShowGraph(),SaveModel()])
lr = learn.lr_find()
learning_rate = lr.valley
learn.fit_one_cycle(freeze_epoch, learning_rate)I don't see any difference between both codes though!
oguiza
commented
1 year ago Hi @MOREDataset, I've tested your code with some dummy data and it works AS LONG AS the input and output have the same number of features/ channels:
from tsai.models.PatchTST import PatchTST
X_train = np.random.rand(1000, 3, 100)
y_train = np.random.rand(1000, 3, 20)
freeze_epoch = 1
models = {'PatchTST'}#'TSTPlus','XCMPlus','LSTMPlus','GRUPlus','MiniRocketPlus'}
for key in models:
exec('model_name = {}'.format(key))
splits = get_splits(y_train, valid_size=.2, stratify=False, random_state=42, shuffle=True)
tfms = [None, [ToFloat(), ToNumpyTensor()]]
dsets = TSDatasets(X_train, y_train, tfms=tfms, splits=splits, inplace=True)
dls = TSDataLoaders.from_dsets(dsets.train, dsets.valid, bs=[64, 128])
par = {}
#Model learning
learn = TSForecaster(X_train, y_train, splits=splits, tfms=tfms, bs=1042, arch=model_name, arch_config=par, metrics=[mae, rmse], cbs=[ShowGraph(),SaveModel()])
lr = learn.lr_find()
learning_rate = lr.valley
learn.fit_one_cycle(freeze_epoch, learning_rate)However, it will fail if you use something like this:
X_train = np.random.rand(1000, 3, 100)
y_train = np.random.rand(1000, 1, 20)The reason for that is that this is the way the original model works. I have in mind a PatchTSTPlus that will allow the option to predict time series with different # features. But it still is not available. So I don't consider this a bug. This works as designed. There's a separate issue (#713) to address this enhancement. Based on that, I'd propose that this issue is closed.
Note: @MOREDataset , your code could be simplified if you use something like this:
X_train = np.random.rand(1000, 3, 100)
y_train = np.random.rand(1000, 3, 20)
splits = get_splits(y_train, valid_size=.2, stratify=False, random_state=42, shuffle=True)
tfms = [None, [ToFloat(), ToNumpyTensor()]]
learn = TSForecaster(X_train, y_train, splits=splits, tfms=tfms, bs=1042, arch="PatchTST", metrics=[mae, rmse], cbs=[ShowGraph(),SaveModel()])
lr = learn.lr_find()
learning_rate = lr.valley
learn.fit_one_cycle(1, learning_rate)You can pass a string as an arch to TSClassifier, TSRegressor, and TSForecaster.
MOREDataset
commented
1 year ago Hi @oguiza,
Thank you for such a detailed response. Yes, I doubled checked the size of the input, and the X_train and y_train have different features.
l will make sure to follow your highlighted issue.
Thank you again.
oguiza
commented
1 year ago Closed as the initial issue is resolved.
Hello, I get an error when I use many-Plus models for multi-horizon forecasting: TypeError: init() got an unexpected keyword argument 'custom_head' my code as follows：
those models include： (For multi-horizon forecast, have no custom_head) FCN FCNPlus InceptionTime MLP RNN LSTM GRU RNN_FCN MRNN_FCN LSTM_FCN MLSTM_FCN GRU_FCN MGRU_FCN ResCNN ResNet ResNetPlus TCN XceptionTime XceptionTimePlus ResNetPlus - need seq_len RNN_FCNPlus - cannot multiply MRNN_FCNPlus - cannot multiply LSTM_FCNPlus - cannot multiply MLSTM_FCNPlus - cannot multiply GRU_FCNPlus - cannot multiply MGRU_FCNPlus - cannot multiply