{% raw %}

{% endraw %} {% raw %}

`create_subsequence_mask`[source]

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

{% endraw %} {% raw %}

`create_variable_mask`[source]

create_variable_mask(o, r=0.15)

{% endraw %} {% raw %}

`create_future_mask`[source]

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

{% endraw %} {% raw %}

`natural_mask`[source]

natural_mask(o)

Applies natural missingness in a batch to non-nan values in the next sample

{% endraw %} {% raw %}

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)

{% endraw %} {% raw %}

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

{% endraw %} {% raw %}

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()

{% endraw %} {% raw %}

t = torch.rand(16, 30, 100)
mask = create_subsequence_mask(t, r=.3) # 30% 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=.3) mean: {mask[0].float().mean().item():.3f}')
plt.show()

{% endraw %} {% raw %}

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()

{% endraw %} {% raw %}

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()

{% endraw %} {% raw %}

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()

{% endraw %} {% raw %}

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()

{% endraw %} {% raw %}

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()

{% endraw %} {% raw %}

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()

{% endraw %} {% raw %}

`create_mask`[source]

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

{% endraw %} {% raw %}

`class` `MVP`[source]

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, nan_to_num:int=0, dropout:float=0.1, crit:callable=None, weights_path:Optional[str]=None, target_dir:str='./data/MVP', fname:str='model', save_best:bool=True, verbose:bool=False) :: Callback

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

{% endraw %} {% raw %}

{% endraw %}

Experiments

{% raw %}

from fastai.data.transforms import *
from tsai.data.all import *
from tsai.models.utils import *
from tsai.models.layers import *
from tsai.learner import *
from tsai.models.TSTPlus import *
from tsai.models.InceptionTimePlus import *

{% endraw %} {% raw %}

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]

{% endraw %} {% raw %}

tfms  = [None, [Categorize()]]
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}')])
learn.fit_one_cycle(1, 3e-3)

epoch	train_loss	valid_loss	time
0	1.013921	1.186778	00:17

best epoch:   0  val_loss: 1.186778 - pretrained weights_path='data/MVP/MoteStrain.pth'

{% endraw %} {% raw %}

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

weights from data/MVP/MoteStrain.pth successfully transferred!

epoch	train_loss	valid_loss	time
0	0.689358	1.160601	00:19

best epoch:   0  val_loss: 1.160601 - pretrained weights_path='data/MVP/model.pth'

{% endraw %} {% raw %}

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

{% endraw %} {% raw %}

tfms  = [None, [Categorize()]]
batch_tfms = [TSStandardize(by_var=True), Nan2Value()]
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'data/MVP/{dsid}.pth', metrics=accuracy)
learn.fit_one_cycle(1)

weights from data/MVP/MoteStrain.pth successfully transferred!

epoch	train_loss	valid_loss	accuracy	time
0	0.697812	0.699655	0.442492	00:18

{% endraw %} {% raw %}

tfms  = [None, [Categorize()]]
batch_tfms = [TSStandardize(by_var=True), Nan2Value()]
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/anaconda3/lib/python3.7/site-packages/ipykernel_launcher.py:37: UserWarning: Only future_mask will be used

{% endraw %} {% raw %}

tfms  = [None, [Categorize()]]
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/anaconda3/lib/python3.7/site-packages/ipykernel_launcher.py:35: UserWarning: Only custom_mask will be used

{% endraw %}

create_subsequence_mask[source]

create_variable_mask[source]

create_future_mask[source]

natural_mask[source]

create_mask[source]

class MVP[source]

Experiments

`create_subsequence_mask`[source]

`create_variable_mask`[source]

`create_future_mask`[source]

`natural_mask`[source]

`create_mask`[source]

`class` `MVP`[source]