add LeViT

[kaijieshi7]

Add LeViT

• build model
• update init.py in models
• convert pretrained weight
• inference test on imagenet and update model_zoo
• update docs
• update readme
• update changelog
• pytorch speed comparison

[kaijieshi7]

OneFlow LeViT_128S time: 34.5ms (= 3447.5ms / 100, input_shape=[16, 3, 224, 224])
PyTorch LeViT_128S time: 40.7ms (= 4071.0ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.18 (= 40.7ms / 34.5ms)

OneFlow LeViT_128 time: 45.5ms (= 4551.8ms / 100, input_shape=[16, 3, 224, 224])
PyTorch LeViT_128 time: 45.8ms (= 4576.1ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.01 (= 45.8ms / 45.5ms)

OneFlow LeViT_192 time: 56.6ms (= 5663.3ms / 100, input_shape=[16, 3, 224, 224])
PyTorch LeViT_192 time: 49.0ms (= 4904.7ms / 100, input_shape=[16, 3, 224, 224])
❌ Relative speed: 0.87 (= 49.0ms / 56.6ms)

OneFlow LeViT_256 time: 60.8ms (= 6081.3ms / 100, input_shape=[16, 3, 224, 224])
PyTorch LeViT_256 time: 49.4ms (= 4942.2ms / 100, input_shape=[16, 3, 224, 224])
❌ Relative speed: 0.81 (= 49.4ms / 60.8ms)

OneFlow LeViT_384 time: 89.8ms (= 8976.2ms / 100, input_shape=[16, 3, 224, 224])
PyTorch LeViT_384 time: 55.5ms (= 5551.0ms / 100, input_shape=[16, 3, 224, 224])
❌ Relative speed: 0.62 (= 55.5ms / 89.8ms)

OneFlow LeViT_128S time: 15.2ms (= 1517.8ms / 100, input_shape=[16, 3, 224, 224], backward is disabled, ddp, world size=2)
PyTorch LeViT_128S time: 9.7ms (= 967.1ms / 100, input_shape=[16, 3, 224, 224], backward is disabled, ddp, world size=2)
❌ Relative speed: 0.64 (= 9.7ms / 15.2ms)

---

Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.

---

OneFlow LeViT_128 time: 15.4ms (= 1544.3ms / 100, input_shape=[16, 3, 224, 224], backward is disabled, ddp, world size=2)
PyTorch LeViT_128 time: 10.1ms (= 1011.0ms / 100, input_shape=[16, 3, 224, 224], backward is disabled, ddp, world size=2)
❌ Relative speed: 0.65 (= 10.1ms / 15.4ms)

---

Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.

---

OneFlow LeViT_192 time: 19.2ms (= 1918.2ms / 100, input_shape=[16, 3, 224, 224], backward is disabled, ddp, world size=2)
PyTorch LeViT_192 time: 9.8ms (= 975.5ms / 100, input_shape=[16, 3, 224, 224], backward is disabled, ddp, world size=2)
❌ Relative speed: 0.51 (= 9.8ms / 19.2ms)

---

Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.

---

OneFlow LeViT_256 time: 22.5ms (= 2250.0ms / 100, input_shape=[16, 3, 224, 224], backward is disabled, ddp, world size=2)
PyTorch LeViT_256 time: 12.2ms (= 1216.7ms / 100, input_shape=[16, 3, 224, 224], backward is disabled, ddp, world size=2)
❌ Relative speed: 0.54 (= 12.2ms / 22.5ms)

---

Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.

---

OneFlow LeViT_384 time: 31.9ms (= 3191.4ms / 100, input_shape=[16, 3, 224, 224], backward is disabled, ddp, world size=2)
PyTorch LeViT_384 time: 11.8ms (= 1177.6ms / 100, input_shape=[16, 3, 224, 224], backward is disabled, ddp, world size=2)
❌ Relative speed: 0.37 (= 11.8ms / 31.9ms)

---

Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.

---

[kaijieshi7]

这个速度误差好大啊，大家有没有什么想法

[rentainhe]

这个速度误差好大啊，大家有没有什么想法

我看看，你搬的速度有点快233

[rentainhe]

.div的速度可能有问题

[rentainhe]

.view还有一些slice的操作速度上可能也有差距

[rentainhe]

修复一下冲突就可以合并了 @kaijieshi7

[kaijieshi7]

修复一下冲突就可以合并了 @kaijieshi7

OK