Fasterrcnn sample code test 1: make anchor_ generator = None

Focus on RPN One column ：

import torch
import torchvision
from torchvision.models.detection import FasterRCNN
from torchvision.models.detection.rpn import AnchorGenerator

#  Definition backbone
backbone = torchvision.models.mobilenet_v2().features

backbone.out_channels = 1280

##！！！！ RPN, Pay attention here 
#anchor_generator = AnchorGenerator(sizes=((32, 64, 128, 256, 512),), aspect_ratios=((0.5, 1, 2),))
anchor_generator = None

#  Define which feature Will be used for RoIAlign
roi_pooler = torchvision.ops.MultiScaleRoIAlign(featmap_names=['0'],
                                                output_size=7,
                                                sampling_ratio=2)

#  Integrate into the model 

model = FasterRCNN(backbone=backbone, num_classes=10, rpn_anchor_generator=anchor_generator,
                   box_roi_pool=roi_pooler)

model.eval()

#  Compile a data 
x = [torch.rand(3, 500, 689), torch.rand(3, 400, 654)]


a = model(x)

print(a)

——————————————————————————————————————————

##！！！！ RPN, Pay attention here 
#anchor_generator = AnchorGenerator(sizes=((32, 64, 128, 256, 512),), aspect_ratios=((0.5, 1, 2),))
anchor_generator = None

Generally, you can set it yourself anchor_generator, You can also use the system default , Reference resources fast-rcnn Source code

def _default_anchorgen():
    anchor_sizes = ((32,), (64,), (128,), (256,), (512,))
    aspect_ratios = ((0.5, 1.0, 2.0),) * len(anchor_sizes)
    return AnchorGenerator(anchor_sizes, aspect_ratios)

        if rpn_anchor_generator is None:
            rpn_anchor_generator = _default_anchorgen()

——————————————————————————————————————————————
But I use the default , An error occurred

  assert condition, message
AssertionError: Anchors should be Tuple[Tuple[int]] because each feature map could potentially have different sizes and aspect ratios. 
There needs to be a match between the number of feature maps passed and the number of sizes / aspect ratios specified.

---

speculation ： default def _default_anchorgen(), This is FPN Of , Multiscale generation .

    anchor_sizes = ((32,), (64,), (128,), (256,), (512,))
    aspect_ratios = ((0.5, 1.0, 2.0),) * len(anchor_sizes)

(32,), (64,), (128,), (256,), (512,) Each represents five feature_map Upper value size ;
aspect_ratios =((0.5, 1.0, 2.0), (0.5, 1.0, 2.0), (0.5, 1.0, 2.0), (0.5, 1.0, 2.0), (0.5, 1.0, 2.0))
This is what I guess ........ Haven't seen the anchor box generation , provisional