pytorch 常用函数
目录
通过索引赋值
a = torch.zeros([5,5])
index = (torch.LongTensor([0,1]),torch.LongTensor([1,2]))
a.index_put_((index), torch.Tensor([1,1]))
a[index] = torch.Tensor([4,4])
print(a)
tensor([[0., 4., 0., 0., 0.],
[0., 0., 4., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.]])
pytorch之tensor按索引赋值，三种方法[https://blog.csdn.net/qq_41368074/article/details/106986753]
注意，上面的这个：index是一个tensor负责x，一个tensor负责y。下面的是（0,3） （1,3） （2,2）
a = torch.zeros([5,5])
index = (torch.LongTensor([0,1,2]),torch.LongTensor([3,3,2]))#index = (torch.LongTensor([0,1]),torch.LongTensor([1,2]),torch.LongTensor([3,3]))
a.index_put_((index), torch.Tensor([1,-100,-3]))
# a[index] = torch.Tensor([4,4])
tensor([[ 0., 0., 0., 1., 0.],
[ 0., 0., 0., -100., 0.],
[ 0., 0., -3., 0., 0.],
[ 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0.]])
打印网络参数--大小
需要先安装torchsummary
from torchsummary import summary
summary(model, input_size=(3, 324, 324), device='cpu')
if __name__ == "__main__":
model = ResNet(n_classes=1000, n_blocks=[3, 4, 23, 3])
model.eval()
image = torch.randn(1, 3, 224, 224)
print(model)
print("input:", image.shape)
print("output:", model(image).shape)
from torchsummary import summary
summary(model, input_size=(3, 324, 324), device='cpu')
打印如下：
---------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 162, 162] 9,408
BatchNorm2d-2 [-1, 64, 162, 162] 128
ReLU-3 [-1, 64, 162, 162] 0
MaxPool2d-4 [-1, 64, 82, 82] 0
Conv2d-5 [-1, 64, 82, 82] 4,096
BatchNorm2d-6 [-1, 64, 82, 82] 128
ReLU-7 [-1, 64, 82, 82] 0
Conv2d-8 [-1, 64, 82, 82] 36,864
BatchNorm2d-9 [-1, 64, 82, 82] 128
ReLU-10 [-1, 64, 82, 82] 0
Conv2d-11 [-1, 256, 82, 82] 16,384
BatchNorm2d-12 [-1, 256, 82, 82] 512
Conv2d-13 [-1, 256, 82, 82] 16,384
BatchNorm2d-14 [-1, 256, 82, 82] 512
_Bottleneck-15 [-1, 256, 82, 82] 0
Conv2d-16 [-1, 64, 82, 82] 16,384
BatchNorm2d-17 [-1, 64, 82, 82] 128
ReLU-18 [-1, 64, 82, 82] 0
Conv2d-19 [-1, 64, 82, 82] 36,864
。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。
。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。
ReLU-313 [-1, 512, 11, 11] 0
Conv2d-314 [-1, 2048, 11, 11] 1,048,576
BatchNorm2d-315 [-1, 2048, 11, 11] 4,096
Conv2d-316 [-1, 2048, 11, 11] 2,097,152
BatchNorm2d-317 [-1, 2048, 11, 11] 4,096
_Bottleneck-318 [-1, 2048, 11, 11] 0
Conv2d-319 [-1, 512, 11, 11] 1,048,576
BatchNorm2d-320 [-1, 512, 11, 11] 1,024
ReLU-321 [-1, 512, 11, 11] 0
Conv2d-322 [-1, 512, 11, 11] 2,359,296
BatchNorm2d-323 [-1, 512, 11, 11] 1,024
ReLU-324 [-1, 512, 11, 11] 0
Conv2d-325 [-1, 2048, 11, 11] 1,048,576
BatchNorm2d-326 [-1, 2048, 11, 11] 4,096
Identity-327 [-1, 2048, 11, 11] 0
_Bottleneck-328 [-1, 2048, 11, 11] 0
Conv2d-329 [-1, 512, 11, 11] 1,048,576
BatchNorm2d-330 [-1, 512, 11, 11] 1,024
ReLU-331 [-1, 512, 11, 11] 0
Conv2d-332 [-1, 512, 11, 11] 2,359,296
BatchNorm2d-333 [-1, 512, 11, 11] 1,024
ReLU-334 [-1, 512, 11, 11] 0
Conv2d-335 [-1, 2048, 11, 11] 1,048,576
BatchNorm2d-336 [-1, 2048, 11, 11] 4,096
Identity-337 [-1, 2048, 11, 11] 0
_Bottleneck-338 [-1, 2048, 11, 11] 0
AdaptiveAvgPool2d-339 [-1, 2048, 1, 1] 0
Flatten-340 [-1, 2048] 0
Linear-341 [-1, 1000] 2,049,000
================================================================
Total params: 44,549,160
Trainable params: 44,549,160
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 1.20
Forward/backward pass size (MB): 894.98
Params size (MB): 169.94
Estimated Total Size (MB): 1066.12
----------------------------------------------------------------
print(model)效果如下：
ResNet(
(layer1): _Stem(
(conv1): _ConvBnReLU(
(conv): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
(bn): BatchNorm2d(64, eps=1e-05, momentum=0.0010000000000000009, affine=True, track_running_stats=True)
(relu): ReLU()
)
(pool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=True)
)
(layer2): _ResLayer(
(block1): _Bottleneck(
(reduce): _ConvBnReLU(
(conv): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=1e-05, momentum=0.0010000000000000009, affine=True, track_running_stats=True)
(relu): ReLU()
)
(conv3x3): _ConvBnReLU(
(conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=1e-05, momentum=0.0010000000000000009, affine=True, track_running_stats=True)
(relu): ReLU()
)
(increase): _ConvBnReLU(
(conv): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.0010000000000000009, affine=True, track_running_stats=True)
)
(shortcut): _ConvBnReLU(
(conv): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.0010000000000000009, affine=True, track_running_stats=True)
)
)
(block2): _Bottleneck(
(reduce): _ConvBnReLU(
(conv): Conv2d(256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=1e-05, momentum=0.0010000000000000009, affine=True, track_running_stats=True)
(relu): ReLU()
)
(conv3x3): _ConvBnReLU(
(conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=1e-05, momentum=0.0010000000000000009, affine=True, track_running_stats=True)
(relu): ReLU()
)
(increase): _ConvBnReLU(
(conv): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.0010000000000000009, affine=True, track_running_stats=True)
)
(shortcut): Identity()
)
由于类别更改导致没法用预训练模型
# Model setup
model = DeepLabV2_ResNet101_MSC(n_classes=CONFIG.DATASET.N_CLASSES)
state_dict = torch.load(CONFIG.MODEL.INIT_MODEL) ## 这是预训练模型
/# 找到与类别数相关的层名，比如最后的全连接层肯定与类别数量相关，把预训练的相应的层名改了
/#设置load_state_dict(new_state_dict, strict=False)的strict=False即可！
/# model.base.load_state_dict(new_state_dict, strict=False)
import collections
new_state_dict = collections.OrderedDict()
for k, v in state_dict.items():
name = k.replace('base.','')
if 'aspp' in name:
name = name + '_2'
new_state_dict[name] = v
print(" Init:", CONFIG.MODEL.INIT_MODEL)
for m in model.base.state_dict().keys():
if m not in new_state_dict.keys():
print(" Skip init:", m)
model.base.load_state_dict(new_state_dict, strict=False) # model.base.load_state_dict(state_dict, strict=False) # to skip ASPP
model = nn.DataParallel(model)
model.to(device)
# Loss definition
criterion = nn.CrossEntropyLoss(ignore_index=CONFIG.DATASET.IGNORE_LABEL)
criterion.to(device)
主要参考的 [https://github.com/kazuto1011/deeplab-pytorch]
conf = labels[best_truth_idx]，label的size=3,best_truth_idx的size=4,居然可以跑的通!!测试了一下,原来是下标索引赋值啊!!
labels = torch.tensor([1,2,3])
best_truth_idx = torch.tensor([1, 0, 0, 2])
conf = labels[best_truth_idx]
print(conf)
#打印: tensor([2, 1, 1, 3])
若best_truth_idx = torch.tensor([1, 0, 0, 3])就会报错,超过下标越界
RuntimeError: index 3 is out of bounds for dim with size 3
所以,索引的不能超过labels的size
好记性不如烂键盘---点滴、积累、进步！
