Support UniFormer

[thinksoso]

Support Uniformer

• build uniformer model
• convert pretrained weight
• inference test on imagenet and update model_zoo small_plus
• update docs
• update changelog
• pytorch speed comparison

[thinksoso]

eval 结果跟论文对不上的问题已经向官方提了 issue Sense-X/UniFormer#16

[thinksoso]

测速结果

OneFlow uniformer_base time: 198.4ms (= 19835.6ms / 100, input_shape=[16, 3, 224, 224])
PyTorch uniformer_base time: 205.1ms (= 20506.7ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.03 (= 205.1ms / 198.4ms)

PyTorch uniformer_base_ls time: 211.3ms (= 21126.4ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.06 (= 211.3ms / 198.9ms)

OneFlow uniformer_small time: 83.6ms (= 8361.4ms / 100, input_shape=[16, 3, 224, 224])
PyTorch uniformer_small time: 102.1ms (= 10210.4ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.22 (= 102.1ms / 83.6ms)

OneFlow uniformer_small_plus time: 99.2ms (= 9921.4ms / 100, input_shape=[16, 3, 224, 224])
PyTorch uniformer_small_plus time: 108.9ms (= 10885.4ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.10 (= 108.9ms / 99.2ms)

OneFlow uniformer_base time: 257.9ms (= 25788.3ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch uniformer_base time: 265.4ms (= 26543.0ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.03 (= 265.4ms / 257.9ms)

✔️ Relative speed: 0.99 (= 273.9ms / 275.5ms)

OneFlow uniformer_small time: 121.7ms (= 12169.2ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch uniformer_small time: 123.5ms (= 12352.7ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.02 (= 123.5ms / 121.7ms)

OneFlow uniformer_small_plus time: 139.7ms (= 13967.8ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch uniformer_small_plus time: 137.3ms (= 13732.3ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
❌ Relative speed: 0.98 (= 137.3ms / 139.7ms)

[rentainhe]

@kaijieshi7 开杰这边double review一下，我没有问题了~

[Andy1621]

抱歉，issue已回复，精度没对上是模型代码有个参数写错了，现在的model我都测了一遍没有问题的

[Andy1621]

测速结果

OneFlow uniformer_base time: 198.4ms (= 19835.6ms / 100, input_shape=[16, 3, 224, 224]) PyTorch uniformer_base time: 205.1ms (= 20506.7ms / 100, input_shape=[16, 3, 224, 224]) ✔️ Relative speed: 1.03 (= 205.1ms / 198.4ms)

PyTorch uniformer_base_ls time: 211.3ms (= 21126.4ms / 100, input_shape=[16, 3, 224, 224]) ✔️ Relative speed: 1.06 (= 211.3ms / 198.9ms)

OneFlow uniformer_small time: 83.6ms (= 8361.4ms / 100, input_shape=[16, 3, 224, 224]) PyTorch uniformer_small time: 102.1ms (= 10210.4ms / 100, input_shape=[16, 3, 224, 224]) ✔️ Relative speed: 1.22 (= 102.1ms / 83.6ms)

OneFlow uniformer_small_plus time: 99.2ms (= 9921.4ms / 100, input_shape=[16, 3, 224, 224]) PyTorch uniformer_small_plus time: 108.9ms (= 10885.4ms / 100, input_shape=[16, 3, 224, 224]) ✔️ Relative speed: 1.10 (= 108.9ms / 99.2ms)

OneFlow uniformer_base time: 257.9ms (= 25788.3ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2) PyTorch uniformer_base time: 265.4ms (= 26543.0ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2) ✔️ Relative speed: 1.03 (= 265.4ms / 257.9ms)

✔️ Relative speed: 0.99 (= 273.9ms / 275.5ms)

OneFlow uniformer_small time: 121.7ms (= 12169.2ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2) PyTorch uniformer_small time: 123.5ms (= 12352.7ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2) ✔️ Relative speed: 1.02 (= 123.5ms / 121.7ms)

OneFlow uniformer_small_plus time: 139.7ms (= 13967.8ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2) PyTorch uniformer_small_plus time: 137.3ms (= 13732.3ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2) ❌ Relative speed: 0.98 (= 137.3ms / 139.7ms)

您好，不同模型的测速结果，oneflow官方有提供吗，对专门的测速比较挺感兴趣

[kaijieshi7]

https://github.com/Oneflow-Inc/vision/blob/main/ci/check/compare_speed_with_pytorch.py
在这里，最简单的是直接测试，这里的代码是转换引用d