Training in mixed precision implementation

Mixed precision training

This module allows the forward and backward passes of your neural net to be done in fp16 (also known as half precision). This is particularly important if you have an NVIDIA GPU with tensor cores, since it can speed up your training by 200% or more.


To train your model in mixed precision you just have to call Learner.to_fp16, which converts the model and modifies the existing Learner to add MixedPrecision.


to_fp16(learn:Learner, loss_scale:float=None, max_noskip:int=1000, dynamic:bool=False, clip:float=None, flat_master:bool=False) → Learner

Put learn in FP16 precision mode.

For example:

path = untar_data(URLs.MNIST_SAMPLE)
data = ImageDataBunch.from_folder(path)
model = simple_cnn((3,16,16,2))
learn = Learner(data, model, metrics=[accuracy]).to_fp16()
Total time: 00:03

epoch train_loss valid_loss accuracy
1 0.139469 0.115246 0.963199

Details about mixed precision training are available in NVIDIA's documentation. We will just summarize the basics here.

The only parameter you may want to tweak is loss_scale. This is used to scale the loss up, so that it doesn't underflow fp16, leading to loss of accuracy (this is reversed for the final gradient calculation after converting back to fp32). Generally the default 512 works well, however. You can also enable or disable the flattening of the master parameter tensor with flat_master=True, however in our testing the different is negligible.

Internally, the callback ensures that all model parameters (except batchnorm layers, which require fp32) are converted to fp16, and an fp32 copy is also saved. The fp32 copy (the master parameters) is what is used for actually updating with the optimizer; the fp16 parameters are used for calculating gradients. This helps avoid underflow with small learning rates.

All of this is implemented by the following Callback.

class MixedPrecision[source]

MixedPrecision(learn:Learner, loss_scale:float=None, max_noskip:int=1000, dynamic:bool=False, clip:float=None, flat_master:bool=False) :: LearnerCallback

Base class for creating callbacks for a Learner.

Callback methods

You don't have to call the following functions yourself - they're called by fastai's Callback system automatically to enable the class's functionality.


on_backward_begin(last_loss:Rank0Tensor, **kwargs:Any) → Rank0Tensor

Scale gradients up by self.loss_scale to prevent underflow.



Convert the gradients back to FP32 and divide them by the scale.


on_loss_begin(last_output:Tensor, **kwargs:Any) → Tensor

Convert half precision output to FP32 to avoid reduction overflow.



Update the params from master to model and zero grad.



Prepare the master model.