prelu

Author: nihui

tools/pnnx/tests/onnx/test_onnx_activation_ops.py

@@ -12,7 +12,10 @@
 from onnxscript import opset19 as op
 
 @script()
-def Model(x: FLOAT["C","H","W"]):
+def Model(x: FLOAT["N","C","W"], y: FLOAT["N","C","H","W"], z: FLOAT["N","C","D","H","W"]):
+
+    prelu_slope = op.RandomNormal(seed=0.0, shape=[12])
+
     return (
         op.Celu(x),
         op.Clip(x, min=0, max=2.3),
@@ -23,29 +26,65 @@ def Model(x: FLOAT["C","H","W"]):
         op.LeakyRelu(x),
         op.LogSoftmax(x),
         op.Mish(x),
-        # op.PRelu(x),
+        op.PRelu(x, op.Unsqueeze(prelu_slope, axes=[1])),
         op.Relu(x),
         op.Selu(x),
         op.Sigmoid(x),
         op.Softmax(x),
         op.Softplus(x),
         # op.Swish(x),
+
+        op.Celu(y),
+        op.Clip(y, min=0, max=2.3),
+        op.Elu(y),
+        # op.Gelu(y),
+        op.HardSigmoid(y),
+        op.HardSwish(y),
+        op.LeakyRelu(y),
+        op.LogSoftmax(y),
+        op.Mish(y),
+        op.PRelu(y, op.Unsqueeze(prelu_slope, axes=[1,2])),
+        op.Relu(y),
+        op.Selu(y),
+        op.Sigmoid(y),
+        op.Softmax(y),
+        op.Softplus(y),
+        # op.Swish(y),
+
+        # op.Celu(z),
+        op.Clip(z, min=0, max=2.3),
+        # op.Elu(z),
+        # op.Gelu(z),
+        op.HardSigmoid(z),
+        op.HardSwish(z),
+        op.LeakyRelu(z),
+        # op.LogSoftmax(z),
+        op.Mish(z),
+        op.PRelu(z, op.Unsqueeze(prelu_slope, axes=[1,2,3])),
+        op.Relu(z),
+        # op.Selu(z),
+        op.Sigmoid(z),
+        # op.Softmax(z),
+        # op.Softplus(z),
+        # op.Swish(z),
         )
 
 def test():
     # save onnx
     onnx.save(Model.to_model_proto(), "test_onnx_activation_ops.onnx")
 
     torch.manual_seed(0)
-    x = torch.rand(3, 4, 5)
+    x = torch.rand(1, 12, 64)
+    y = torch.rand(1, 12, 48, 64)
+    z = torch.rand(1, 12, 21, 28, 44)
 
     # ort inference
     sess = ort.InferenceSession("test_onnx_activation_ops.onnx")
-    a = tuple(torch.from_numpy(out) for out in sess.run(None, {"x": x.numpy()}))
+    a = tuple(torch.from_numpy(out) for out in sess.run(None, {"x": x.numpy(), "y": y.numpy(), "z": z.numpy()}))
 
     # onnx to pnnx and ncnn
     import os
-    os.system("../../src/pnnx test_onnx_activation_ops.onnx inputshape=[3,4,5] inputshape2=[13,14,15]")
+    os.system("../../src/pnnx test_onnx_activation_ops.onnx inputshape=[1,12,64],[1,12,48,64],[1,12,21,28,44] inputshape2=[7,12,22],[8,12,33,11],[9,12,9,12,13] fp16=0")
 
     # pnnx inference
     import test_onnx_activation_ops_pnnx