mqbench.fake_quantize package
Submodules
mqbench.fake_quantize.dorefa
- class mqbench.fake_quantize.dorefa.DoReFaFakeQuantize(observer, **observer_kwargs)[source]
Bases:
QuantizeBase
- fake_quant_enabled: Tensor
- forward(X)[source]
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Module
instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- observer_enabled: Tensor
mqbench.fake_quantize.dsq
- class mqbench.fake_quantize.dsq.DSQFakeQuantize(observer, alpha=0.4, **observer_kwargs)[source]
Bases:
QuantizeBase
- fake_quant_enabled: Tensor
- forward(X)[source]
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Module
instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- observer_enabled: Tensor
- class mqbench.fake_quantize.dsq.FakeQuantizeDSQPerchannel[source]
Bases:
Function
- static forward(ctx, x, scale, zero_point, quant_min, quant_max, ch_axis, alpha)[source]
Performs the operation.
This function is to be overridden by all subclasses.
It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).
The context can be used to store tensors that can be then retrieved during the backward pass.
- class mqbench.fake_quantize.dsq.FakeQuantizeDSQPertensor[source]
Bases:
Function
- static forward(ctx, x, scale, zero_point, quant_min, quant_max, alpha)[source]
Performs the operation.
This function is to be overridden by all subclasses.
It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).
The context can be used to store tensors that can be then retrieved during the backward pass.
mqbench.fake_quantize.fixed
- class mqbench.fake_quantize.fixed.FixedFakeQuantize(observer, **observer_kwargs)[source]
Bases:
QuantizeBase
This is actually torch.quantization.FakeQuantize.
- extra_repr()[source]
Set the extra representation of the module
To print customized extra information, you should re-implement this method in your own modules. Both single-line and multi-line strings are acceptable.
- fake_quant_enabled: Tensor
- forward(X)[source]
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Module
instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- observer_enabled: Tensor
mqbench.fake_quantize.lsq
- class mqbench.fake_quantize.lsq.FakeQuantizeLearnablePerchannelAffine[source]
Bases:
Function
- static forward(ctx, x, scale, zero_point, ch_axis, quant_min, quant_max, grad_factor)[source]
Performs the operation.
This function is to be overridden by all subclasses.
It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).
The context can be used to store tensors that can be then retrieved during the backward pass.
- class mqbench.fake_quantize.lsq.LearnableFakeQuantize(observer, scale=1.0, zero_point=0.0, use_grad_scaling=True, **observer_kwargs)[source]
Bases:
QuantizeBase
This is an extension of the FakeQuantize module in fake_quantize.py, which supports more generalized lower-bit quantization and support learning of the scale and zero point parameters through backpropagation. For literature references, please see the class _LearnableFakeQuantizePerTensorOp. In addition to the attributes in the original FakeQuantize module, the _LearnableFakeQuantize module also includes the following attributes to support quantization parameter learning.
- extra_repr()[source]
Set the extra representation of the module
To print customized extra information, you should re-implement this method in your own modules. Both single-line and multi-line strings are acceptable.
- fake_quant_enabled: Tensor
- forward(X)[source]
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Module
instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- observer_enabled: Tensor
mqbench.fake_quantize.nnie
- class mqbench.fake_quantize.nnie.NNIEFakeQuantize(observer, **observer_kwargs)[source]
Bases:
QuantizeBase
- fake_quant_enabled: Tensor
- forward(X)[source]
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Module
instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- observer_enabled: Tensor
- class mqbench.fake_quantize.nnie.NNIEQuantizeFunc[source]
Bases:
Function
- static backward(ctx, grad_output)[source]
Defines a formula for differentiating the operation.
This function is to be overridden by all subclasses.
It must accept a context
ctx
as the first argument, followed by as many outputs didforward()
return, and it should return as many tensors, as there were inputs toforward()
. Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input.The context can be used to retrieve tensors saved during the forward pass. It also has an attribute
ctx.needs_input_grad
as a tuple of booleans representing whether each input needs gradient. E.g.,backward()
will havectx.needs_input_grad[0] = True
if the first input toforward()
needs gradient computated w.r.t. the output.
- static forward(ctx, x, data_max)[source]
Performs the operation.
This function is to be overridden by all subclasses.
It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).
The context can be used to store tensors that can be then retrieved during the backward pass.
mqbench.fake_quantize.pact
- class mqbench.fake_quantize.pact.PACTFakeQuantize(observer, alpha=6.0, **observer_kwargs)[source]
Bases:
QuantizeBase
- extra_repr()[source]
Set the extra representation of the module
To print customized extra information, you should re-implement this method in your own modules. Both single-line and multi-line strings are acceptable.
- fake_quant_enabled: Tensor
- forward(X)[source]
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Module
instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- observer_enabled: Tensor
mqbench.fake_quantize.quantize_base
- class mqbench.fake_quantize.quantize_base.QuantizeBase(observer=<class 'torch.quantization.observer.MovingAverageMinMaxObserver'>, **observer_kwargs)[source]
Bases:
FakeQuantizeBase
This is an extension of the FakeQuantize module in fake_quantize.py, which supports more generalized lower-bit quantization and support learning of the scale and zero point parameters through backpropagation. For literature references, please see the class _LearnableFakeQuantizePerTensorOp. In addition to the attributes in the original FakeQuantize module, the _LearnableFakeQuantize module also includes the following attributes to support quantization parameter learning.
- extra_repr()[source]
Set the extra representation of the module
To print customized extra information, you should re-implement this method in your own modules. Both single-line and multi-line strings are acceptable.
- fake_quant_enabled: Tensor
- observer_enabled: Tensor