All SparseML Sparsification APIs are designed to work with recipes. The files encode the instructions needed for modifying the model and/or training process as a list of modifiers. Example modifiers can be anything from setting the learning rate for the optimizer to gradual magnitude pruning. The files are written in YAML and stored in YAML or markdown files using YAML front matter. The rest of the SparseML system is coded to parse the recipe files into a native format for the desired framework and apply the modifications to the model and training pipeline.
In a recipe, modifiers must be written in a list that includes "modifiers" in its name.
The easiest ways to get or create recipes are by either using the pre-configured recipes in SparseZoo or using Sparsify's automatic creation.
However, power users may be inclined to create their own recipes by hand to enable more fine-grained control or to add in custom modifiers.
A sample recipe for pruning a model generally looks like the following:
version: 0.1.0
modifiers:
- !EpochRangeModifier
start_epoch: 0.0
end_epoch: 70.0
- !LearningRateModifier
start_epoch: 0
end_epoch: -1.0
update_frequency: -1.0
init_lr: 0.005
lr_class: MultiStepLR
lr_kwargs: {'milestones': [43, 60], 'gamma': 0.1}
- !GMPruningModifier
start_epoch: 0
end_epoch: 40
update_frequency: 1.0
init_sparsity: 0.05
final_sparsity: 0.85
mask_type: unstructured
params: ['sections.0.0.conv1.weight', 'sections.0.0.conv2.weight', 'sections.0.0.conv3.weight']
Recipes can contain multiple modifiers, each modifying a portion of the training process in a different way.
In general, each modifier will have a start and an end epoch for when the modifier should be active.
The modifiers will start at start_epoch
and run until end_epoch
.
Note that it does not run through end_epoch
.
Additionally, all epoch values support decimal values such that they can be started somewhere in the middle of an epoch.
For example, start_epoch: 2.5
will start in the middle of the second training epoch.
The most commonly used modifiers are enumerated as subsections below.
The EpochRangeModifier
controls the range of epochs for training a model.
Each supported ML framework has an implementation to enable easily retrieving this number of epochs.
Note, that this is not a hard rule and if other modifiers have a larger end_epoch
or smaller start_epoch
then those values will be used instead.
The only parameters that can be controlled for EpochRangeModifier
are the start_epoch
and end_epoch
.
Both parameters are required.
Required Parameters:
start_epoch
: The start range for the epoch (0 indexed)end_epoch
: The end range for the epoch
Example:
- !EpochRangeModifier
start_epoch: 0.0
end_epoch: 25.0
The pruning modifiers handle pruning the specified layer(s) in a given model.
The ConstantPruningModifier
enforces the sparsity structure and level for an already pruned layer(s) in a model.
The modifier is generally used for transfer learning from an already pruned model or
to enforce sparsity while quantizing.
The weights remain trainable in this setup; however, the sparsity is unchanged.
Required Parameters:
params
: The parameters in the model to prune. This can be set to a string containing__ALL__
to prune all parameters, a list to specify the targeted parameters, or regex patterns prefixed by 're:' of parameter name patterns to match. For example:['blocks.1.conv']
for PyTorch and['mnist_net/blocks/conv0/conv']
for TensorFlow. Regex can also be used to match all conv params:['re:.*conv']
for PyTorch and['re:.*/conv']
for TensorFlow.
Example:
- !ConstantPruningModifier
params: __ALL__
The GMPruningModifier
prunes the parameter(s) in a model to a
target sparsity (percentage of 0's for a layer's param/variable)
using gradual magnitude pruning.
This is done gradually from an initial to final sparsity (init_sparsity
, final_sparsity
)
over a range of epochs (start_epoch
, end_epoch
) and updated at a specific interval defined by the update_frequency
.
For example, using the following settings start_epoch: 0
, end_epoch: 5
, update_frequency: 1
,
init_sparsity: 0.05
, final_sparsity: 0.8
will do the following:
- at epoch 0 set the sparsity for the specified param(s) to 5%
- once every epoch, gradually increase the sparsity towards 80%
- by the start of epoch 5, stop pruning and set the final sparsity for the specified param(s) to 80%
Required Parameters:
params
: The parameters in the model to prune. This can be set to a string containing__ALL__
to prune all parameters, a list to specify the targeted parameters, or regex patterns prefixed by 're:' of parameter name patterns to match. For example:['blocks.1.conv']
for PyTorch and['mnist_net/blocks/conv0/conv']
for TensorFlow. Regex can also be used to match all conv params:['re:.*conv']
for PyTorch and['re:.*/conv']
for TensorFlow.init_sparsity
: The decimal value for the initial sparsity to start pruning with. Atstart_epoch
will set the sparsity for the param/variable to this value. Generally, this is kept at 0.05 (5%).final_sparsity
: The decimal value for the final sparsity to end pruning with. By the start ofend_epoch
will set the sparsity for the param/variable to this value. Generally, this is kept in a range from 0.6 to 0.95 depending on the model and layer. Anything less than 0.4 is not useful for performance.start_epoch
: The epoch to start the pruning at (0 indexed). This supports floating-point values to enable starting pruning between epochs.end_epoch
: The epoch before which to stop pruning. This supports floating-point values to enable stopping pruning between epochs.update_frequency
: The number of epochs/fractions of an epoch between each pruning step. It supports floating-point values to enable updating inside of epochs. Generally, this is set to update once per epoch (1.0
). However, if the loss for the model recovers quickly, it should be set to a lesser value. For example: set it to0.5
for once every half epoch (twice per epoch).
Example:
- !GMPruningModifier
params: ['blocks.1.conv']
init_sparsity: 0.05
final_sparsity: 0.8
start_epoch: 5.0
end_epoch: 20.0
update_frequency: 1.0
The QuantizationModifier
sets the model to run with
quantization aware training (QAT).
QAT emulates the precision loss of int8 quantization during training so weights can be
learned to limit any accuracy loss from quantization.
Once the QuantizationModifier
is enabled, it cannot be disabled (no end_epoch
).
Quantization zero points are set to be asymmetric for activations and symmetric for weights.
Currently only available in PyTorch.
Notes:
- ONNX exports of PyTorch QAT models will be QAT models themselves (emulated quantization).
To convert your QAT ONNX model to a fully quantizerd model use
the script
scripts/pytorch/model_quantize_qat_export.py
or the functionneuralmagicML.pytorch.quantization.quantize_qat_export
. - If performing QAT on a sparse model, you must preserve sparsity during QAT by
applying a
ConstantPruningModifier
or have already used aGMPruningModifier
withleave_enabled
set to True.
Required Parameters:
start_epoch
: The epoch to start QAT. This supports floating-point values to enable starting pruning between epochs.
Example:
- !QuantizationModifier
start_epoch: 0.0
The learning rate modifiers set the learning rate (LR) for an optimizer during training. If you are using an Adam optimizer, then generally, these are not useful. If you are using a standard stochastic gradient descent optimizer, then these give a convenient way to control the LR.
The SetLearningRateModifier
sets the learning rate (LR) for the optimizer to a specific value at a specific point
in the training process.
Required Parameters:
start_epoch
: The epoch in the training process to set thelearning_rate
value for the optimizer. This supports floating-point values to enable setting the LR between epochs.learning_rate
: The floating-point value to set as the learning rate for the optimizer atstart_epoch
.
Example:
- !SetLearningRateModifier
start_epoch: 5.0
learning_rate: 0.1
The LearningRateModifier
sets schedules for controlling the learning rate for an optimizer during training.
If you are using an Adam optimizer, then generally, these are not useful.
If you are using a standard stochastic gradient descent optimizer, then these give a convenient way to control the LR.
Provided schedules to choose from are the following:
ExponentialLR
: Multiplies the learning rate by agamma
value every epoch. To use this one,lr_kwargs
should be set to a dictionary containinggamma
. For example:{'gamma': 0.9}
StepLR
: Multiplies the learning rate by agamma
value after a certain epoch period defined bystep
. To use this one,lr_kwargs
must be set to a dictionary containinggamma
andstep_size
. For example:{'gamma': 0.9, 'step_size': 2.0}
MultiStepLR
: Multiplies the learning rate by agamma
value at specific epoch points defined bymilestones
. To use this one,lr_kwargs
must be set to a dictionary containinggamma
andmilestones
. For example:{'gamma': 0.9, 'milestones': [2.0, 5.5, 10.0]}
Required Parameters:
start_epoch
: The epoch to start modifying the LR at (0 indexed). This supports floating-point values to enable starting pruning between epochs.end_epoch
: The epoch to stop modifying the LR before. This supports floating-point values to enable stopping pruning between epochs.lr_class
: The LR class to use, one of [ExponentialLR
,StepLR
,MultiStepLR
].lr_kwargs
: The named arguments for thelr_class
.init_lr
: [Optional] The initial LR to set atstart_epoch
and to use for creating the schedules. If not given, the optimizer's current LR will be used at startup.
Example:
- !LearningRateModifier
start_epoch: 0.0
end_epoch: 25.0
lr_class: MultiStepLR
lr_kwargs:
gamma: 0.9
milestones: [2.0, 5.5, 10.0]
init_lr: 0.1
The TrainableParamsModifier
controls the params that are marked as trainable for the current optimizer.
This is generally useful when transfer learning to easily mark which parameters should or should not be frozen/trained.
Required Parameters:
params
: The names of parameters to mark as trainable or not. This can be set to a string containing__ALL__
to mark all parameters, a list to specify the targeted parameters, or regex patterns prefixed by 're:' of parameter name patterns to match. For example:['blocks.1.conv']
for PyTorch and['mnist_net/blocks/conv0/conv']
for TensorFlow. Regex can also be used to match all conv params:['re:.*conv']
for PyTorch and['re:.*/conv']
for TensorFlow.
Example:
- !TrainableParamsModifier
params: __ALL__
The SetWeightDecayModifier
sets the weight decay (L2 penalty) for the optimizer to a
specific value at a specific point in the training process.
Required Parameters:
start_epoch
: The epoch in the training process to set theweight_decay
value for the optimizer. This supports floating-point values to enable setting the weight decay between epochs.weight_decay
: The floating-point value to set as the weight decay for the optimizer atstart_epoch
.
Example:
- !SetWeightDecayModifier
start_epoch: 5.0
weight_decay: 0.0