lumin.nn.callbacks package¶
Submodules¶
lumin.nn.callbacks.callback module¶
-
class
lumin.nn.callbacks.callback.
Callback
(model=None, plot_settings=<lumin.plotting.plot_settings.PlotSettings object>)[source]¶ Bases:
lumin.nn.callbacks.abs_callback.AbsCallback
Base callback class from which other callbacks should inherit.
- Parameters
model (
Optional
[AbsModel
]) – model to refer to during trainingplot_settings (
PlotSettings
) – PlotSettings class
-
set_model
(model)[source]¶ Sets the callback’s model in order to allow the callback to access and adjust model parameters
- Parameters
model (
AbsModel
) – model to refer to during training- Return type
None
-
set_plot_settings
(plot_settings)[source]¶ Sets the plot settings for any plots produced by the callback
- Parameters
plot_settings (
PlotSettings
) – PlotSettings class- Return type
None
lumin.nn.callbacks.cyclic_callbacks module¶
-
class
lumin.nn.callbacks.cyclic_callbacks.
AbsCyclicCallback
(interp, param_range, cycle_mult=1, decrease_param=False, scale=1, model=None, nb=None, plot_settings=<lumin.plotting.plot_settings.PlotSettings object>)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Abstract class for callbacks affecting lr or mom
- Parameters
interp (
str
) – string representation of interpolation function. Either ‘linear’ or ‘cosine’.param_range (
Tuple
[float
,float
]) – minimum and maximum values for parametercycle_mult (
int
) – multiplicative factor for adjusting the cycle length after each cycle. E.g cycle_mult=1 keeps the same cycle length, cycle_mult=2 doubles the cycle length after each cycle.decrease_param (
bool
) – whether to begin by decreasing the parameter, otherwise begin by increasing itscale (
int
) – multiplicative factor for setting the initial number of epochs per cycle. E.g scale=1 means 1 epoch per cycle, scale=5 means 5 epochs per cycle.model (
Optional
[AbsModel
]) – model to refer to during trainingnb (
Optional
[int
]) – number of minibatches (iterations) to expect per epochplot_settings (
PlotSettings
) – PlotSettings class
-
on_batch_begin
(**kargs)[source]¶ Computes the new value for the optimiser parameter and returns it
- Return type
float
- Returns
new value for optimiser parameter
-
on_batch_end
(**kargs)[source]¶ Increments the callback’s progress through the cycle
- Return type
None
-
on_epoch_begin
(**kargs)[source]¶ Ensures the cycle_end flag is false when the epoch starts
- Return type
None
-
class
lumin.nn.callbacks.cyclic_callbacks.
CycleLR
(lr_range, interp='cosine', cycle_mult=1, decrease_param='auto', scale=1, model=None, nb=None, plot_settings=<lumin.plotting.plot_settings.PlotSettings object>)[source]¶ Bases:
lumin.nn.callbacks.cyclic_callbacks.AbsCyclicCallback
Callback to cycle learning rate during training according to either: cosine interpolation for SGDR https://arxiv.org/abs/1608.03983 or linear interpolation for Smith cycling https://arxiv.org/abs/1506.01186
- Parameters
lr_range (
Tuple
[float
,float
]) – tuple of initial and final LRsinterp (
str
) – ‘cosine’ or ‘linear’ interpolationcycle_mult (
int
) – Multiplicative constant for altering the cycle length after each complete cycledecrease_param (
Union
[str
,bool
]) – whether to increase or decrease the LR (effectively reverses lr_range order), ‘auto’ selects according to interpscale (
int
) – Multiplicative constant for altering the length of a cycle. 1 corresponds to one cycle = one (sub-)epochmodel (
Optional
[AbsModel
]) –Model
to alter, alternatively callset_model()
.nb (
Optional
[int
]) – Number of batches in a (sub-)epochplot_settings (
PlotSettings
) –PlotSettings
class to control figure appearance
- Examples::
>>> cosine_lr = CycleLR(lr_range=(0, 2e-3), cycle_mult=2, scale=1, ... interp='cosine', nb=100) >>> >>> cyclical_lr = CycleLR(lr_range=(2e-4, 2e-3), cycle_mult=1, scale=5, interp='linear', nb=100)
-
class
lumin.nn.callbacks.cyclic_callbacks.
CycleMom
(mom_range, interp='cosine', cycle_mult=1, decrease_param='auto', scale=1, model=None, nb=None, plot_settings=<lumin.plotting.plot_settings.PlotSettings object>)[source]¶ Bases:
lumin.nn.callbacks.cyclic_callbacks.AbsCyclicCallback
Callback to cycle momentum (beta 1) during training according to either: cosine interpolation for SGDR https://arxiv.org/abs/1608.03983 or linear interpolation for Smith cycling https://arxiv.org/abs/1506.01186 By default is set to evolve in opposite direction to learning rate, a la https://arxiv.org/abs/1803.09820
- Parameters
mom_range (
Tuple
[float
,float
]) – tuple of initial and final momentainterp (
str
) – ‘cosine’ or ‘linear’ interpolationcycle_mult (
int
) – Multiplicative constant for altering the cycle length after each complete cycledecrease_param (
Union
[str
,bool
]) – whether to increase or decrease the momentum (effectively reverses mom_range order), ‘auto’ selects according to interpscale (
int
) – Multiplicative constant for altering the length of a cycle. 1 corresponds to one cycle = one (sub-)epochmodel (
Optional
[AbsModel
]) –Model
to alter, alternatively callset_model()
nb (
Optional
[int
]) – Number of batches in a (sub-)epochplot_settings (
PlotSettings
) –PlotSettings
class to control figure appearance
- Examples::
>>> cyclical_mom = CycleMom(mom_range=(0.85 0.95), cycle_mult=1, ... scale=5, interp='linear', nb=100)
-
class
lumin.nn.callbacks.cyclic_callbacks.
OneCycle
(lengths, lr_range, mom_range=(0.85, 0.95), interp='cosine', model=None, nb=None, plot_settings=<lumin.plotting.plot_settings.PlotSettings object>)[source]¶ Bases:
lumin.nn.callbacks.cyclic_callbacks.AbsCyclicCallback
Callback implementing Smith 1-cycle evolution for lr and momentum (beta_1) https://arxiv.org/abs/1803.09820 Default interpolation uses fastai-style cosine function. Automatically triggers early stopping on cycle completion.
- Parameters
lengths (
Tuple
[int
,int
]) – tuple of number of (sub-)epochs in first and second stages of cyclelr_range (
List
[float
]) – tuple of initial and final LRsmom_range (
Tuple
[float
,float
]) – tuple of initial and final momentainterp (
str
) – ‘cosine’ or ‘linear’ interpolationmodel (
Optional
[AbsModel
]) –Model
to alter, alternatively callset_model()
nb (
Optional
[int
]) – Number of batches in a (sub-)epochplot_settings (
PlotSettings
) –PlotSettings
class to control figure appearance
- Examples::
>>> onecycle = OneCycle(lengths=(15, 30), lr_range=[1e-4, 1e-2], ... mom_range=(0.85, 0.95), interp='cosine', nb=100)
lumin.nn.callbacks.data_callbacks module¶
-
class
lumin.nn.callbacks.data_callbacks.
BinaryLabelSmooth
(coefs=0, model=None)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Callback for applying label smoothing to binary classes, based on https://arxiv.org/abs/1512.00567 Applies smoothing during both training and inference.
- Parameters
coefs (
Union
[float
,Tuple
[float
,float
]]) – Smoothing coefficients: 0->coef[0] 1->1-coef[1]. if passed float, coef[0]=coef[1]model (
Optional
[AbsModel
]) – not used, only for compatability
- Examples::
>>> lbl_smooth = BinaryLabelSmooth(0.1) >>> >>> lbl_smooth = BinaryLabelSmooth((0.1, 0.02))
-
class
lumin.nn.callbacks.data_callbacks.
SequentialReweight
(reweight_func, scale=0.1, model=None)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Caution
Experiemntal proceedure
During ensemble training, sequentially reweight training data in last validation fold based on prediction performance of last trained model. Reweighting highlights data which are easier or more difficult to predict to the next model being trained.
- Parameters
reweight_func (
Callable
[[Tensor
,Tensor
],Tensor
]) – callable function returning a tensor of same shape as targets, ideally quantifying model-prediction performancescale (
float
) – multiplicative factor for rescaling returned tensor of reweight_funcmodel (
Optional
[AbsModel
]) –Model
to provide predictions, alternatively callset_model()
- Examples::
>>> seq_reweight = SequentialReweight( ... reweight_func=nn.BCELoss(reduction='none'), scale=0.1)
-
on_train_end
(fy, val_id, **kargs)[source]¶ Reweighs the validation fold once training is finished
- Parameters
fy (
FoldYielder
) – FoldYielder providing the training and validation datafold_id – Fold index which was used for validation
- Return type
None
-
class
lumin.nn.callbacks.data_callbacks.
SequentialReweightClasses
(reweight_func, scale=0.1, model=None)[source]¶ Bases:
lumin.nn.callbacks.data_callbacks.SequentialReweight
Caution
Experiemntal proceedure
Version of
SequentialReweight
designed for classification, which renormalises class weights to original weight-sum after reweighting During ensemble training, sequentially reweight training data in last validation fold based on prediction performance of last trained model. Reweighting highlights data which are easier or more difficult to predict to the next model being trained.- Parameters
reweight_func (
Callable
[[Tensor
,Tensor
],Tensor
]) – callable function returning a tensor of same shape as targets, ideally quantifying model-prediction performancescale (
float
) – multiplicative factor for rescaling returned tensor of reweight_funcmodel (
Optional
[AbsModel
]) –Model
to provide predictions, alternatively callset_model()
- Examples::
>>> seq_reweight = SequentialReweight( ... reweight_func=nn.BCELoss(reduction='none'), scale=0.1)
-
class
lumin.nn.callbacks.data_callbacks.
BootstrapResample
(n_folds, bag_each_time=False, reweight=True, model=None)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Callback for bootstrap sampling new training datasets from original training data during (ensemble) training.
- Parameters
n_folds (
int
) – the number of folds present in trainingFoldYielder
bag_each_time (
bool
) – whether to sample a new set for each sub-epoch or to use the same sample each timereweight (
bool
) – whether to reweight the sampleed data to mathch the weight sum (per class) of the original datamodel (
Optional
[AbsModel
]) – not used, only for compatability
- Examples::
>>> bs_resample BootstrapResample(n_folds=len(train_fy))
-
on_epoch_begin
(by, **kargs)[source]¶ Resamples training data for new epoch
- Parameters
by (
BatchYielder
) – BatchYielder providing data for the upcoming epoch- Return type
None
-
class
lumin.nn.callbacks.data_callbacks.
FeatureSubsample
(cont_feats, model=None)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Callback for training a model on a random sub-sample of the range of possible input features. Only sub-samples continuous features. Number of continuous inputs infered from model. Associated
Model
will automatically mask its inputs during inference; simply provide inputs with the same number of columns as trainig data.- Parameters
cont_feats (
List
[str
]) – list of all continuous features in input data. Order must match.model (
Optional
[AbsModel
]) –Model
being trained, alternatively callset_model()
- Examples::
>>> feat_subsample = FeatureSubsample(cont_feats=['pT', 'eta', 'phi'])
-
on_epoch_begin
(by, **kargs)[source]¶ Masks input data to remove non-selected features
- Parameters
by (
BatchYielder
) – BatchYielder providing data for the upcoming epoch- Return type
None
lumin.nn.callbacks.loss_callbacks module¶
-
class
lumin.nn.callbacks.loss_callbacks.
GradClip
(clip, clip_norm=True, model=None)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Callback for clipping gradients by norm or value.
- Parameters
clip (
float
) – value to clip atclip_norm (
bool
) – whether to clip according to norm (torch.nn.utils.clip_grad_norm_) or value (torch.nn.utils.clip_grad_value_)model (
Optional
[AbsModel
]) –Model
with parameters to clip gradients, alternatively callset_model()
- Examples::
>>> grad_clip = GradClip(1e-5)
lumin.nn.callbacks.model_callbacks module¶
-
class
lumin.nn.callbacks.model_callbacks.
SWA
(start_epoch, renewal_period=-1, model=None, val_fold=None, cyclic_callback=None, update_on_cycle_end=None, verbose=False, plot_settings=<lumin.plotting.plot_settings.PlotSettings object>)[source]¶ Bases:
lumin.nn.callbacks.model_callbacks.AbsModelCallback
Callback providing Stochastic Weight Averaging based on (https://arxiv.org/abs/1803.05407) This adapted version allows the tracking of a pair of average models in order to avoid having to hardcode a specific start point for averaging:
Model average x0 will begin to be tracked start_epoch (sub-)epochs/cycles after training begins.
cycle_since_replacement is set to 1
Renewal_period (sub-)epochs/cycles later, a second average x1 will be tracked.
At the next renewal period, the performance of x0 and x1 will be compared on data contained in val_fold.
- If x0 is better than x1:
x1 is replaced by a copy of the current model
cycle_since_replacement is increased by 1
renewal_period is multiplied by cycle_since_replacement
- Else:
x0 is replaced by x1
x1 is replaced by a copy of the current model
cycle_since_replacement is set to 1
renewal_period is set back to its original value
Additonally, will optionally (default True) lock-in to any cyclical callbacks to only update at the end of a cycle.
- Parameters
start_epoch (
int
) – (sub-)epoch/cycle to begin averagingrenewal_period (
int
) – How often to check performance of averages, and renew tracking of least performantmodel (
Optional
[AbsModel
]) –Model
to provide parameters, alternatively callset_model()
val_fold (
Optional
[Dict
[str
,ndarray
]]) – Dictionary containing inputs, targets, and weights (or None) as Numpy arrayscyclic_callback (
Optional
[AbsCyclicCallback
]) – Optional for any cyclical callback which is runningupdate_on_cycle_end (
Optional
[bool
]) – Whether to lock in to the cyclic callback and only update at the end of a cycle. Default yes, if cyclic callback present.verbose (
bool
) – Whether to print out update information for testing and operation confirmationplot_settings (
PlotSettings
) –PlotSettings
class to control figure appearance
- Examples::
>>> swa = SWA(start_epoch=5, renewal_period=5)
-
get_loss
()[source]¶ Evaluates SWA model and returns loss
- Return type
float
- Returns
Loss on validation fold for oldest SWA average
-
class
lumin.nn.callbacks.model_callbacks.
AbsModelCallback
(model=None, val_fold=None, cyclic_callback=None, update_on_cycle_end=None, plot_settings=<lumin.plotting.plot_settings.PlotSettings object>)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Abstract class for callbacks which provide alternative models during training
- Parameters
model (
Optional
[AbsModel
]) –Model
to provide parameters, alternatively callset_model()
val_fold (
Optional
[Dict
[str
,ndarray
]]) – Dictionary containing inputs, targets, and weights (or None) as Numpy arrayscyclic_callback (
Optional
[AbsCyclicCallback
]) – Optional for any cyclical callback which is runningupdate_on_cycle_end (
Optional
[bool
]) – Whether to lock in to the cyclic callback and only update at the end of a cycle. Default yes, if cyclic callback present.plot_settings (
PlotSettings
) –PlotSettings
class to control figure appearance
lumin.nn.callbacks.opt_callbacks module¶
-
class
lumin.nn.callbacks.opt_callbacks.
LRFinder
(nb, lr_bounds=[1e-07, 10], model=None, plot_settings=<lumin.plotting.plot_settings.PlotSettings object>)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Callback class for Smith learning-rate range test (https://arxiv.org/abs/1803.09820)
- Parameters
nb (
int
) – number of batches in a (sub-)epochlr_bounds (
Tuple
[float
,float
]) – tuple of initial and final LRmodel (
Optional
[AbsModel
]) –Model
to alter, alternatively callset_model()
plot_settings (
PlotSettings
) –PlotSettings
class to control figure appearance
-
on_batch_end
(loss, **kargs)[source]¶ Records loss and increments LR
- Parameters
loss (
float
) – training loss for most recent batch