Merge branch 'master' into eca-weights

README.md

@@ -10,7 +10,7 @@
 * Remove separate tiered (`t`) vs tiered_narrow (`tn`) ResNet model defs, all `tn` changed to `t` and `t` models removed (`seresnext26t_32x4d` only model w/ weights that was removed).
 * Support model default_cfgs with separate train vs test resolution `test_input_size`
 
-### Jan 30, 2012
+### Jan 30, 2021
 * Add initial "Normalization Free" NF-RegNet-B* and NF-ResNet model definitions based on [paper](https://arxiv.org/abs/2101.08692)
 
 ### Jan 25, 2021

timm/models/layers/activations_me.py

@@ -30,6 +30,9 @@ class SwishJitAutoFn(torch.autograd.Function):
     Inspired by conversation btw Jeremy Howard & Adam Pazske
     https://twitter.com/jeremyphoward/status/1188251041835315200
     """
+    @staticmethod
+    def symbolic(g, x):
+        return g.op("Mul", x, g.op("Sigmoid", x))
 
     @staticmethod
     def forward(ctx, x):
@@ -152,6 +155,13 @@ class HardSwishJitAutoFn(torch.autograd.Function):
         x = ctx.saved_tensors[0]
         return hard_swish_jit_bwd(x, grad_output)
 
+    @staticmethod
+    def symbolic(g, self):
+        input = g.op("Add", self, g.op('Constant', value_t=torch.tensor(3, dtype=torch.float)))
+        hardtanh_ = g.op("Clip", input, g.op('Constant', value_t=torch.tensor(0, dtype=torch.float)), g.op('Constant', value_t=torch.tensor(6, dtype=torch.float)))
+        hardtanh_ = g.op("Div", hardtanh_, g.op('Constant', value_t=torch.tensor(6, dtype=torch.float)))
+        return g.op("Mul", self, hardtanh_)
+
 
 def hard_swish_me(x, inplace=False):
     return HardSwishJitAutoFn.apply(x)