Update false positive case.

hxzd5568 · Nov 1, 2022 · ea8965b · ea8965b
1 parent 8047004
commit ea8965b
Showing 1 changed file with 12 additions and 10 deletions.
diff --git a/bug/false_positive/conv_bn_tan.py b/bug/false_positive/conv_bn_tan.py
@@ -4,13 +4,12 @@
 
 code = """
 #[version = "0.0.5"]
-def @main(%x0: Tensor[(14, 1, 11, 7), float16], %x2: Tensor[(5, 1, 3, 1), float16], %x3: Tensor[(5), float16], %x4: Tensor[(5), float16], %x5: Tensor[(5), float16], %x6: Tensor[(5), float16]) -> (Tensor[(14, 5, 5, 3), float16],) {
-  %0 = nn.avg_pool2d(%x0, pool_size=[3, 3], strides=[2, 1], dilation=[1, 2], padding=[0, 0, 0, 0], count_include_pad=True) /* from_string */ /* ty=Tensor[(14, 1, 5, 3), float16] */;
-  %1 = nn.conv2d(%0, %x2, strides=[1, 2], padding=[1, 1, 1, 2], dilation=[1, 2], channels=5, kernel_size=[3, 1]) /* from_string */ /* ty=Tensor[(14, 5, 5, 3), float16] */;
-  %2 = nn.batch_norm(%1, %x3, %x4, %x5, %x6, center=False, scale=False) /* from_string */ /* ty=(Tensor[(14, 5, 5, 3), float16], Tensor[(5), float16], Tensor[(5), float16]) */;
-  %3 = %2.0;
-  %4 = tan(%3) /* from_string */ /* ty=Tensor[(14, 5, 5, 3), float16] */;
-  (%4,)
+def @main(%x0: Tensor[(14, 1, 11, 7), float16], %x2: Tensor[(5, 1, 3, 1), float16], 
+    %x3: Tensor[(5), float16], %x4: Tensor[(5), float16], %x5: Tensor[(5), float16], 
+    %x6: Tensor[(5), float16]) -> Tensor[(14, 5, 5, 4), float16] {
+  %0 = nn.conv2d(%x0, %x2, strides=[2, 2], padding=[0, 0, 0, 0], dilation=[1, 1]);
+  %1 = nn.batch_norm(%0, %x3, %x4, %x5, %x6, axis=1, center=False, scale=False);
+  tan(%1.0)
 }
 """
 
@@ -31,7 +30,10 @@ def @main(%x0: Tensor[(14, 1, 11, 7), float16], %x2: Tensor[(5, 1, 3, 1), float1
 gmod = GraphModule(lib['default'](cpu()))
 gmod.run(**inputs)
 opt_out = gmod.get_output(0).numpy()
+
 diff = np.abs(opt_out - ref_out)
-print(f'{np.unravel_index(np.nanargmax(diff), ref_out.shape)}: {np.nanmax(diff)}')
-rel_diff = diff / ref_out
-print(f'{np.unravel_index(np.nanargmax(rel_diff), ref_out.shape)}: {np.nanmax(rel_diff)}')
+max_idx = np.unravel_index(np.nanargmax(diff), ref_out.shape)
+print('pos:', max_idx)
+ref_elem, opt_elem = ref_out[max_idx], opt_out[max_idx]
+print('ref:', ref_elem, 'opt:', opt_elem)
+print('abs_err:', np.abs(ref_elem - opt_elem), 'rel_err:', np.abs(ref_elem - opt_elem) / ref_elem)