MVP (aka TSBERT)

Self-Supervised Pretraining of Time Series Models

Masked Value Predictor callback used to predict time series step values after a binary mask has been applied.

---

source

self_mask

 self_mask (o)

---

source

create_future_mask

 create_future_mask (o, r=0.15, sync=False)

---

source

create_variable_mask

 create_variable_mask (o, r=0.15)

---

source

create_subsequence_mask

 create_subsequence_mask (o, r=0.15, lm=3, stateful=True, sync=False)

t = torch.rand(16, 3, 100)
mask = create_subsequence_mask(t, sync=False)
test_eq(mask.shape, t.shape)
mask = create_subsequence_mask(t, sync=True)
test_eq(mask.shape, t.shape)
mask = create_variable_mask(t)
test_eq(mask.shape, t.shape)
mask = create_future_mask(t)
test_eq(mask.shape, t.shape)

o = torch.randn(2, 3, 4)
o[o>.5] = np.nan
test_eq(torch.isnan(self_mask(o)).sum(), 0)

t = torch.rand(16, 30, 100)
mask = create_subsequence_mask(t, r=.15) # default settings
test_eq(mask.dtype, torch.bool)
plt.figure(figsize=(10, 3))
plt.pcolormesh(mask[0], cmap='cool')
plt.title(f'sample 0 subsequence mask (sync=False) - default mean: {mask[0].float().mean().item():.3f}')
plt.show()
plt.figure(figsize=(10, 3))
plt.pcolormesh(mask[1], cmap='cool')
plt.title(f'sample 1 subsequence mask (sync=False) - default mean: {mask[1].float().mean().item():.3f}')
plt.show()

t = torch.rand(16, 30, 100)
mask = create_subsequence_mask(t, r=.5) # 50% of values masked
test_eq(mask.dtype, torch.bool)
plt.figure(figsize=(10, 3))
plt.pcolormesh(mask[0], cmap='cool')
plt.title(f'sample 0 subsequence mask (r=.5) mean: {mask[0].float().mean().item():.3f}')
plt.show()

t = torch.rand(16, 30, 100)
mask = create_subsequence_mask(t, lm=5) # average length of mask = 5 
test_eq(mask.dtype, torch.bool)
plt.figure(figsize=(10, 3))
plt.pcolormesh(mask[0], cmap='cool')
plt.title(f'sample 0 subsequence mask (lm=5) mean: {mask[0].float().mean().item():.3f}')
plt.show()

t = torch.rand(16, 30, 100)
mask = create_subsequence_mask(t, stateful=False) # individual time steps masked 
test_eq(mask.dtype, torch.bool)
plt.figure(figsize=(10, 3))
plt.pcolormesh(mask[0], cmap='cool')
plt.title(f'per sample subsequence mask (stateful=False) mean: {mask[0].float().mean().item():.3f}')
plt.show()

t = torch.rand(1, 30, 100)
mask = create_subsequence_mask(t, sync=True) # all time steps masked simultaneously
test_eq(mask.dtype, torch.bool)
plt.figure(figsize=(10, 3))
plt.pcolormesh(mask[0], cmap='cool')
plt.title(f'per sample subsequence mask (sync=True) mean: {mask[0].float().mean().item():.3f}')
plt.show()

t = torch.rand(1, 30, 100)
mask = create_variable_mask(t) # masked variables
test_eq(mask.dtype, torch.bool)
plt.figure(figsize=(10, 3))
plt.pcolormesh(mask[0], cmap='cool')
plt.title(f'per sample variable mask mean: {mask[0].float().mean().item():.3f}')
plt.show()

t = torch.rand(1, 30, 100)
mask = create_future_mask(t, r=.15, sync=True) # masked steps
test_eq(mask.dtype, torch.bool)
plt.figure(figsize=(10, 3))
plt.pcolormesh(mask[0], cmap='cool')
plt.title(f'future mask mean: {mask[0].float().mean().item():.3f}')
plt.show()

t = torch.rand(1, 30, 100)
mask = create_future_mask(t, r=.15, sync=False) # masked steps
mask = create_future_mask(t, r=.15, sync=True) # masked steps
test_eq(mask.dtype, torch.bool)
plt.figure(figsize=(10, 3))
plt.pcolormesh(mask[0], cmap='cool')
plt.title(f'future mask mean: {mask[0].float().mean().item():.3f}')
plt.show()

---

source

create_mask

 create_mask (o, r=0.15, lm=3, stateful=True, sync=False,
              subsequence_mask=True, variable_mask=False,
              future_mask=False)

---

source

MVP

 MVP (r:float=0.15, subsequence_mask:bool=True, lm:float=3.0,
      stateful:bool=True, sync:bool=False, variable_mask:bool=False,
      future_mask:bool=False, custom_mask:Optional=None,
      sel_vars:Optional[list]=None, nan_to_num:int=0,
      window_size:Optional[tuple]=None, dropout:float=0.1, crit:<built-
      infunctioncallable>=None, weights_path:Optional[str]=None,
      target_dir:str='./models/MVP', fname:str='model',
      save_best:bool=True, verbose:bool=False)

Basic class handling tweaks of the training loop by changing a Learner in various events

Experiments

from tsai.data.external import get_UCR_data, check_data
from tsai.data.preprocessing import TSStandardize, TSNan2Value
from tsai.data.core import TSCategorize, get_ts_dls
from tsai.learner import ts_learner
from tsai.models.InceptionTimePlus import InceptionTimePlus

dsid = 'MoteStrain'
X, y, splits = get_UCR_data(dsid, split_data=False)
check_data(X, y, splits, False)
X[X<-1] = np.nan # This is to test the model works well even if nan values are passed through the dataloaders.

X      - shape: [1272 samples x 1 features x 84 timesteps]  type: memmap  dtype:float32  isnan: 0
y      - shape: (1272,)  type: memmap  dtype:<U1  n_classes: 2 (636 samples per class) ['1', '2']  isnan: False
splits - n_splits: 2 shape: [20, 1252]  overlap: False

# Pre-train
tfms  = [None, [TSCategorize()]]
batch_tfms = [TSStandardize(by_var=True)]
unlabeled_dls = get_ts_dls(X, splits=splits, tfms=tfms, batch_tfms=batch_tfms)
learn = ts_learner(unlabeled_dls, InceptionTimePlus, cbs=[MVP(fname=f'{dsid}', window_size=(.5, 1))]) # trained on variable window size
learn.fit_one_cycle(1, 3e-3)

epoch	train_loss	valid_loss	time
0	1.270972	1.194974	00:06

learn = ts_learner(unlabeled_dls, InceptionTimePlus, cbs=[MVP(weights_path=f'models/MVP/{dsid}.pth')])
learn.fit_one_cycle(1, 3e-3)

epoch	train_loss	valid_loss	time
0	0.837741	1.200484	00:07

learn.MVP.show_preds(sharey=True) # these preds are highly inaccurate as the model's been trained for just 1 epoch for testing purposes

# Fine-tune
tfms  = [None, [TSCategorize()]]
batch_tfms = [TSStandardize(by_var=True), TSNan2Value()]
labeled_dls = get_ts_dls(X, y, splits=splits, tfms=tfms, batch_tfms=batch_tfms, bs=64)
learn = ts_learner(labeled_dls, InceptionTimePlus, pretrained=True, weights_path=f'models/MVP/{dsid}.pth', metrics=accuracy)
learn.fit_one_cycle(1)

epoch	train_loss	valid_loss	accuracy	time
0	0.773015	0.744267	0.460863	00:09

tfms  = [None, [TSCategorize()]]
batch_tfms = [TSStandardize(by_var=True), TSNan2Value()]
unlabeled_dls = get_ts_dls(X, splits=splits, tfms=tfms, batch_tfms=batch_tfms, bs=64)
fname = f'{dsid}_test'
mvp = MVP(subsequence_mask=True, sync='random', variable_mask=True, future_mask=True, fname=fname)
learn = ts_learner(unlabeled_dls, InceptionTimePlus, metrics=accuracy, cbs=mvp) # Metrics will not be used!

/Users/nacho/opt/anaconda3/envs/py37torch113/lib/python3.7/site-packages/ipykernel_launcher.py:42: UserWarning: Only future_mask will be used

tfms  = [None, [TSCategorize()]]
batch_tfms = [TSStandardize(by_var=True)]
unlabeled_dls = get_ts_dls(X, splits=splits, tfms=tfms, batch_tfms=batch_tfms, bs=64)
fname = f'{dsid}_test'
mvp = MVP(subsequence_mask=True, sync='random', variable_mask=True, future_mask=True, custom_mask=partial(create_future_mask, r=.15),
                fname=fname)
learn = ts_learner(unlabeled_dls, InceptionTimePlus, metrics=accuracy, cbs=mvp) # Metrics will not be used!

/Users/nacho/opt/anaconda3/envs/py37torch113/lib/python3.7/site-packages/ipykernel_launcher.py:40: UserWarning: Only custom_mask will be used

try: os.remove("models/MVP/MoteStrain.pth")
except OSError: pass
try: os.remove("models/MVP/model.pth")
except OSError: pass