自定义网络剪枝问题请教

[tonyskypc]

作者现有的剪枝，经过测试非常有效，现有对网络修改后，再运行剪枝失败，请指教：# parameters
nc: 1 # number of classes
depth_multiple: 0.33 # model depth multiple
width_multiple: 0.50 # layer channel multiple

anchors

anchors:

• [4,4, 5,5, 5,5] # P2/4
• [5,5, 5,5, 6,6] # P3/8
• [6,6, 7,7, 8,8] # P4/16
  #- [116,90, 156,198, 373,326] # P5/16
  #- 4,4, 5,5, 5,5, 5,5, 5,5, 6,6, 6,6, 7,7, 8,8

YOLOv5 backbone

backbone:

[from, number, module, args]

[from, number, module, args]

[ [ -1, 1, Conv, [ 64, 6, 2, 2, 1, True, 1.0 ] ], # 0-P1/2
[ -1, 1, Conv, [ 128, 3, 2, None, 1, True, 1.0 ] ], # 1-P2/4
[ -1, 3, C3, [ 128, True, 1, [ 0.5, 0.5 ], [ 1.0, 1.0, 1.0 ], 1.0 ] ],
[ -1, 1, Conv, [ 256, 3, 2, None, 1, True, 1.0 ] ], # 3-P3/8
[ -1, 6, C3, [ 256, True, 1, [ 0.5, 0.5 ], [ 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 ], 1.0 ] ],
[ -1, 1, Conv, [ 512, 3, 2, None, 1, True, 1.0 ] ], # 5-P4/16
[ -1, 9, C3, [ 512, True, 1, [ 0.5, 0.5 ], [ 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 ], 1.0 ] ],

[-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 去掉３２倍特征图输出

[ -1, 3, C3, [ 512, True, 1, [ 0.5, 0.5 ], [ 1.0, 1.0, 1.0 ], 1.0 ] ],
[ -1, 1, SPPF, [ 512, 5, 0.5 ] ],  # 9

]

YOLOv5 v6.0 head

head:
[[-1, 1, Conv, [512, 1, 1, None, 1, True, 1.0]],
[-1, 1, nn.Upsample, [None, 2, 'nearest']],
[[-1, 4], 1, Concat, [1]], # cat backbone P3
[-1, 3, C3, [512, False, 1, [0.5, 0.5], [1.0, 1.0, 1.0], 1.0]], # 17 (P3/8-small)

[ -1, 1, Conv, [256, 1, 1, None, 1, True, 1.0]],
[ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ],
[ [ -1, 2 ], 1, Concat, [ 1 ] ], # cat backbone P2 160*160
[ -1, 3, C3, [256, False, 1, [0.5, 0.5], [1.0, 1.0, 1.0], 1.0]], # 16 (P2/4-small)

[-1, 1, Conv, [256, 3, 2, None, 1, True, 1.0]],
[[-1, 13], 1, Concat, [1]], # cat head P2
[-1, 3, C3, [256, False, 1, [0.5, 0.5], [1.0, 1.0, 1.0], 1.0]], # 20 (P3/16-medium)

[-1, 1, Conv, [256, 3, 2, None, 1, True, 1.0]],
[[-1, 9], 1, Concat, [1]], # cat head P4
[-1, 3, C3, [512, False, 1, [0.5, 0.5], [1.0, 1.0, 1.0], 1.0]], # 22 (P4/16-large)

[[16, 19, 22], 1, Detect, [nc, anchors]], # Detect(P2, P3, P4)
]

[Cydia2018]

请问对网络结构修改了哪些部分？报错信息是什么？

[tonyskypc]

[-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 去掉３２倍特征图输出 去掉了，在head里面，对p2检测头做了上采样，为了提取更小目标的特征。p5检测头输出也去了，只有p2，p3,p4(4,8,16)输出
Reversing anchor order
Model Summary: 266 layers, 1397744 parameters, 1397744 gradients, 12.2 GFLOPs
Reversing anchor order
Model Summary: 266 layers, 1748606 parameters, 1748606 gradients, 13.3 GFLOPs
Reversing anchor order
Model Summary: 266 layers, 1414902 parameters, 1414902 gradients, 10.3 GFLOPs
Reversing anchor order
Model Summary: 266 layers, 1591100 parameters, 1591100 gradients, 12.7 GFLOPs
Reversing anchor order
Model Summary: 266 layers, 1608152 parameters, 1608152 gradients, 13.5 GFLOPs
Reversing anchor order
Model Summary: 266 layers, 1355248 parameters, 1355248 gradients, 12.0 GFLOPs
0%| | 0/134 [00:00<?, ?it/s]
Traceback (most recent call last):
File "prune_eagleeye.py", line 126, in
rand_prune_and_eval(model,ignore_idx,opt)
File "prune_eagleeye.py", line 67, in rand_prune_and_eval
mAP = ABE(compact_model)
File "/root/share175/sport_trains/shot_ball/detect_shot/yolov5-multibackbone-compression/utils/adaptive_bn.py", line 126, in call
pred = compact_model(imgs) # forward
File "/root/miniconda3/envs/Pytorch180/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/root/share175/sport_trains/shot_ball/detect_shot/yolov5-multibackbone-compression/models/yolo.py", line 128, in forward
return self._forward_once(x, profile, visualize) # single-scale inference, train
File "/root/share175/sport_trains/shot_ball/detect_shot/yolov5-multibackbone-compression/models/yolo.py", line 152, in _forward_once
x = m(x) # run
File "/root/miniconda3/envs/Pytorch180/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/root/share175/sport_trains/shot_ball/detect_shot/yolov5-multibackbone-compression/models/yolo.py", line 55, in forward
x[i] = self.mi # conv
File "/root/miniconda3/envs/Pytorch180/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/root/miniconda3/envs/Pytorch180/lib/python3.8/site-packages/torch/nn/modules/conv.py", line 399, in forward
return self._conv_forward(input, self.weight, self.bias)
File "/root/miniconda3/envs/Pytorch180/lib/python3.8/site-packages/torch/nn/modules/conv.py", line 395, in _conv_forward
return F.conv2d(input, weight, bias, self.stride,
RuntimeError: Given groups=1, weight of size [18, 128, 1, 1], expected input[16, 121, 480, 480] to have 128 channels, but got 121 channels instead
Exception in thread Thread-3:
Traceback (most recent call last):
File "/root/miniconda3/envs/Pytorch180/lib/python3.8/threading.py", line 932, in _bootstrap_inner
self.run()
File "/root/miniconda3/envs/Pytorch180/lib/python3.8/threading.py", line 870, in run
self._target(*self._args, **self._kwargs)
File "/root/miniconda3/envs/Pytorch180/lib/python3.8/site-packages/torch/utils/data/_utils/pin_memory.py", line 25, in _pin_memory_loop
r = in_queue.get(timeout=MP_STATUS_CHECK_INTERVAL)
File "/root/miniconda3/envs/Pytorch180/lib/python3.8/multiprocessing/queues.py", line 116, in get
return _ForkingPickler.loads(res)
File "/root/miniconda3/envs/Pytorch180/lib/python3.8/site-packages/torch/multiprocessing/reductions.py", line 282, in rebuild_storage_fd
fd = df.detach()
File "/root/miniconda3/envs/Pytorch180/lib/python3.8/multiprocessing/resource_sharer.py", line 57, in detach
with _resource_sharer.get_connection(self._id) as conn:
File "/root/miniconda3/envs/Pytorch180/lib/python3.8/multiprocessing/resource_sharer.py", line 87, in get_connection

[Cydia2018]

你好，已解决。

[tonyskypc]

非常即时，非常棒，通过测试，之前的错误已经解决了，希望提供一个捐赠渠道，感谢作者的辛苦。