New TensorFlow TFDWConv() module

[glenn-jocher]

Analog to DWConv() module

yolov5/models/common.py

Lines 53 to 57 in 8aa2085

	class DWConv(Conv):
	# Depth-wise convolution class
	def __init__(self, c1, c2, k=1, s=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups
	super().__init__(c1, c2, k, s, g=math.gcd(c1, c2), act=act)

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced TensorFlow compatibility for YOLOv5 layers.

📊 Key Changes

• Refined the Conv and DepthwiseConv2D layer parameter definitions for clarity and consistency.
• Added an activations function to map PyTorch activation functions to their TensorFlow equivalents.
• Reshaped the pixel shuffle operation to be more concise.
• Included explicit handling of the activation functions for depthwise convolutions and regular convolutions.

🎯 Purpose & Impact

• The changes aim to make the codebase more maintainable by improving readability and structure.
• Users can expect more reliable cross-framework model behavior, with YOLOv5's TensorFlow version functioning closer to its PyTorch counterpart.
• The update ensures activation functions are properly translated between PyTorch and TensorFlow, potentially leading to performance improvements and fewer discrepancies. 🚀

[glenn-jocher]

@zldrobit this is my TF DWConv() implementation. Seems to work correctly. Tested with https://github.com/ultralytics/yolov5/releases/download/v6.1/yolov5s-dw.pt that has one layer converted to DWConv.

[glenn-jocher]

@zldrobit I noticed that the TF Conv2d docs show a groups argument. Do you know if TF 2.8 now supports this argument? It looks like maybe still not supported in https://stackoverflow.com/questions/66857534/group-conv-in-tensorflow-2

[glenn-jocher]

@zldrobit tests show a problem with TFDWConv(). To reproduce replace a few Conv() layers with DWConv() layers in yolov5s.yaml, i.e.:

# Parameters
nc: 80  # number of classes
depth_multiple: 0.33  # model depth multiple
width_multiple: 0.50  # layer channel multiple
anchors:
  - [10,13, 16,30, 33,23]  # P3/8
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32

# YOLOv5 v6.0 backbone
backbone:
  # [from, number, module, args]
  [[-1, 1, Conv, [64, 6, 2, 2]],  # 0-P1/2
   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
   [-1, 3, C3, [128]],
   [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
   [-1, 6, C3, [256]],
   [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
   [-1, 9, C3, [512]],
   [-1, 1, Conv, [1024, 3, 2]],  # 7-P5/32
   [-1, 3, C3, [1024]],
   [-1, 1, SPPF, [1024, 5]],  # 9
  ]

# YOLOv5 v6.0 head
head:
  [[-1, 1, Conv, [512, 1, 1]],
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
   [-1, 3, C3, [512, False]],  # 13

   [-1, 1, Conv, [256, 1, 1]],
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
   [-1, 3, C3, [256, False]],  # 17 (P3/8-small)

   [-1, 1, DWConv, [256, 5, 2]],
   [[-1, 14], 1, Concat, [1]],  # cat head P4
   [-1, 3, C3, [512, False]],  # 20 (P4/16-medium)

   [-1, 1, DWConv, [512, 5, 2]],
   [[-1, 10], 1, Concat, [1]],  # cat head P5
   [-1, 3, C3, [1024, False]],  # 23 (P5/32-large)

   [[17, 20, 23], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  ]

and then train a model and observe lower val mAP on TF exports. No errors. Not sure where the problem lies. EDIT: fixed in #7845

!python train.py --weights yolov5s.pt --cfg yolov5s.yaml --epochs 300

!python export.py --weights runs/train/exp/weights/best.pt --include saved_model
!python val.py --weights runs/train/exp/weights/best.pt
!python val.py --weights runs/train/exp/weights/best_saved_model

[glenn-jocher]

@zldrobit issue resolved in #7845

All new TF modules are validated! :)