Cut the model to save time

[DeepKnowledge1]

the Resnet has 4 layers, so if you want to use only 3 layers (1-3), you can cut the model, this will save the time

I tried it in my case, I used only 2 layers, so the new model contains only layer1 and layer2.

you do not need to compute the whole forward pass with all layers, so the inference time was decreased

[xfby2016]

Could you share your code？Thank you！

[DeepKnowledge1]

@xfby2016, I hope you can get the idea,

You can also see this blog

`def CutModel():

pretrained_model = resnet18(pretrained=True, progress=True)
NetFeatureSize = OrderedDict([('layer1', [64]), ('layer2', [128]), ('layer3', [256]), ('layer4', [512])])
Feature_layers = 'layer1,layer2'   
layer_ids = Feature_layers.split(',') ## 
output_layer =  max(layer_ids)

layers = list(pretrained_model ._modules.keys())
layer_count = 0
for l in layers:
            if l != output_layer:
                layer_count += 1
            else:
                break
            
            
for i in range(1,len(layers)-layer_count):
    pretrained_model ._modules.pop(layers[-i])
newmodel= nn.Sequential(pretrained_model ._modules)


return newmodel,NetFeatureSize`

`
model,NetFeatureSize= CutModel()
layers = [] train_outputs = OrderedDict() test_outputs = OrderedDict() layer_ids = Feature_layers.split(',')

# set model's intermediate outputs
outputs = []
def hook(module, input, output):
    outputs.append(output)

for layer in layer_ids:
    layers.append(layer)
    train_outputs[layer]= []
    test_outputs[layer]= []
    if  layer == 'layer1':
        model.layer1[-1].register_forward_hook(hook)
    elif  layer == 'layer2':
        model.layer2[-1].register_forward_hook(hook)
    elif  layer == 'layer3':
        model.layer3[-1].register_forward_hook(hook)
    elif  layer == 'layer4':
        model.layer4[-1].register_forward_hook(hook)

`

I will upload the whole stuff when it finished, there are many changes and improvements. I tried to make simple, hope this can help you

[QQR1]

Thanks for your share！Do you have the idea for accelerating Mahalanobis distance calculation?

[DeepKnowledge1]

@QQR1 please have a look

https://github.com/xiahaifeng1995/PaDiM-Anomaly-Detection-Localization-master/issues/8#issuecomment-806567383

[Yangly0]

A nice idea, thanks for your code!