wdhorton

wdhorton

Building wheel torch-0.5.0a0+1b7172a
running install
running build_deps
+ USE_CUDA=0
+ USE_ROCM=0
+ USE_NNPACK=0
+ USE_MKLDNN=0
+ USE_GLOO_IBVERBS=0
+ FULL_CAFFE2=0
+ [[ 9 -gt 0 ]]

wdhorton

params = []

# callback for storing the params of the model after each epoch
class SaveModelParams(Callback):
    def __init__(self, model):
        self.model = model

    def on_epoch_end(self, metrics):
        params.append([p.data.cpu().numpy() for p in self.model.parameters()])

wdhorton

def collect_bn_modules(module, bn_modules):
    if isinstance(module, torch.nn.modules.batchnorm._BatchNorm):
        bn_modules.append(module)

def fix_batchnorm(swa_model, train_dl):
    """
    During training, batch norm layers keep track of a running mean and
    variance of the previous layer's activations. Because the parameters
    of the SWA model are computed as the average of other models' parameters,
    the SWA model never sees the training data itself, and therefore has no

wdhorton

class SWA(Callback):
    def __init__(self, model, swa_model, swa_start):
        super().__init__()
        self.model,self.swa_model,self.swa_start=model,swa_model,swa_start
        
    def on_train_begin(self):
        self.epoch = 0
        self.swa_n = 0

    def on_epoch_end(self, metrics):

wdhorton