I was looking at the activation maps of vgg19 in pytorch. I found that all the values of the maps are positive even before I applied the ReLU.
This seems very strange to me... If this would be correct (could be that I not used the register_forward_hook method correctly?) why would one then apply ReLu at all?
This is my code to produce this:
import torch
import torchvision
import torchvision.models as models
import torchvision.transforms as transforms
from torchsummary import summary
import os, glob
import matplotlib.pyplot as plt
import numpy as np
# settings:
batch_size = 4
# load the model
model = models.vgg19(pretrained=True)
summary(model.cuda(), (3, 32, 32))
model.cpu()
# how to preprocess??? See here:
# https://discuss.pytorch.org/t/how-to-preprocess-input-for-pre-trained-networks/683/2
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
transform = transforms.Compose(
[transforms.ToTensor(),
normalize])
# build data loader
trainset = torchvision.datasets.CIFAR10(root='./data', train=True,
download=True, transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size,
shuffle=True, num_workers=2)
# show one image
dataiter = iter(trainloader)
images, labels = dataiter.next()
# set a hook
activation = {}
def get_activation(name):
def hook(model, input, output):
activation[name] = output.detach()
return hook
# hook at the first conv layer
hook = model.features[0].register_forward_hook(get_activation("firstConv"))
model(images)
hook.remove()
# show results:
flatted_feat_maps = activation["firstConv"].detach().numpy().flatten()
print("All positiv??? --> ",np.all(flatted_feat_maps >= 0))
plt.hist(flatted_feat_maps)
plt.show()
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 32, 32] 1,792
ReLU-2 [-1, 64, 32, 32] 0
Conv2d-3 [-1, 64, 32, 32] 36,928
ReLU-4 [-1, 64, 32, 32] 0
MaxPool2d-5 [-1, 64, 16, 16] 0
Conv2d-6 [-1, 128, 16, 16] 73,856
ReLU-7 [-1, 128, 16, 16] 0
Conv2d-8 [-1, 128, 16, 16] 147,584
ReLU-9 [-1, 128, 16, 16] 0
MaxPool2d-10 [-1, 128, 8, 8] 0
Conv2d-11 [-1, 256, 8, 8] 295,168
ReLU-12 [-1, 256, 8, 8] 0
Conv2d-13 [-1, 256, 8, 8] 590,080
ReLU-14 [-1, 256, 8, 8] 0
Conv2d-15 [-1, 256, 8, 8] 590,080
ReLU-16 [-1, 256, 8, 8] 0
Conv2d-17 [-1, 256, 8, 8] 590,080
ReLU-18 [-1, 256, 8, 8] 0
MaxPool2d-19 [-1, 256, 4, 4] 0
Conv2d-20 [-1, 512, 4, 4] 1,180,160
ReLU-21 [-1, 512, 4, 4] 0
Conv2d-22 [-1, 512, 4, 4] 2,359,808
ReLU-23 [-1, 512, 4, 4] 0
Conv2d-24 [-1, 512, 4, 4] 2,359,808
ReLU-25 [-1, 512, 4, 4] 0
Conv2d-26 [-1, 512, 4, 4] 2,359,808
ReLU-27 [-1, 512, 4, 4] 0
MaxPool2d-28 [-1, 512, 2, 2] 0
Conv2d-29 [-1, 512, 2, 2] 2,359,808
ReLU-30 [-1, 512, 2, 2] 0
Conv2d-31 [-1, 512, 2, 2] 2,359,808
ReLU-32 [-1, 512, 2, 2] 0
Conv2d-33 [-1, 512, 2, 2] 2,359,808
ReLU-34 [-1, 512, 2, 2] 0
Conv2d-35 [-1, 512, 2, 2] 2,359,808
ReLU-36 [-1, 512, 2, 2] 0
MaxPool2d-37 [-1, 512, 1, 1] 0
AdaptiveAvgPool2d-38 [-1, 512, 7, 7] 0
Linear-39 [-1, 4096] 102,764,544
ReLU-40 [-1, 4096] 0
Dropout-41 [-1, 4096] 0
Linear-42 [-1, 4096] 16,781,312
ReLU-43 [-1, 4096] 0
Dropout-44 [-1, 4096] 0
Linear-45 [-1, 1000] 4,097,000
================================================================
Total params: 143,667,240
Trainable params: 143,667,240
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.01
Forward/backward pass size (MB): 5.25
Params size (MB): 548.05
Estimated Total Size (MB): 553.31
----------------------------------------------------------------
Could it be that I somehow did not use the register_forward_hook correctly?
CodePudding user response：
You should clone the output in
def get_activation(name):
def hook(model, input, output):
activation[name] = output.detach().clone() #
return hook
Note that Tensor.detach only detaches the tensor from the graph, but both tensors will still share the same underlying storage.
Returned Tensor shares the same storage with the original one. In-place modifications on either of them will be seen, and may trigger errors in correctness checks. IMPORTANT NOTE: Previously, in-place size / stride / storage changes (such as resize_ / resize_as_ / set_ / transpose_) to the returned tensor also update the original tensor. Now, these in-place changes will not update the original tensor anymore, and will instead trigger an error. For sparse tensors: In-place indices / values changes (such as zero_ / copy_ / add_) to the returned tensor will not update the original tensor anymore, and will instead trigger an error.