[Bug Fix] Set random seed for torch and fix invalid torch cv samples. (#302)

* Fix invalid torch cv samples.

* Set random seed.

Author: Xreki

graph_net/torch/test_compiler.py

@@ -12,6 +12,7 @@
 from contextlib import contextmanager
 import time
 import json
+import random
 import numpy as np
 import platform
 from graph_net.torch.backend.graph_compiler_backend import GraphCompilerBackend
@@ -33,6 +34,15 @@
 }
 
 
+def set_seed(random_seed):
+    random.seed(random_seed)
+    np.random.seed(random_seed)
+    torch.manual_seed(random_seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(random_seed)
+        torch.cuda.manual_seed_all(random_seed)
+
+
 def load_class_from_file(
     args: argparse.Namespace, class_name: str, device: str
 ) -> Type[torch.nn.Module]:
@@ -226,6 +236,7 @@ def test_single_model(args):
         flush=True,
     )
 
+    runtime_seed = 1024
     eager_failure = False
     expected_out = None
     eager_types = []
@@ -239,6 +250,8 @@ def test_single_model(args):
             file=sys.stderr,
             flush=True,
         )
+
+        torch.manual_seed(runtime_seed)
         expected_out = eager_model_call()
         if not isinstance(expected_out, tuple):
             expected_out = (expected_out,)
@@ -270,6 +283,7 @@ def test_single_model(args):
         else:
             compiled_model = compiler(model)
 
+        torch.manual_seed(runtime_seed)
         compiled_model_call = lambda: compiled_model(**input_dict)
         compiled_stats = measure_performance(compiled_model_call, args, compiler)
         print(
@@ -480,6 +494,9 @@ def is_single_model_dir(model_dir):
 
 def main(args):
     assert os.path.isdir(args.model_path)
+
+    initalize_seed = 123
+    set_seed(random_seed=initalize_seed)
     if is_single_model_dir(args.model_path):
         test_single_model(args)
     else:

samples/mmseg/SegNeXt-b/weight_meta.py

@@ -2455,7 +2455,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block2_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 0.477
-    std = 1.551
+    std = 0.551
     data = None
 
 
@@ -4341,7 +4341,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 0.335
-    std = 1.516
+    std = 0.516
     data = None
 
 
@@ -4451,7 +4451,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 0.365
-    std = 1.702
+    std = 0.702
     data = None
 
 
@@ -4701,7 +4701,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 0.504
-    std = 3.092
+    std = 1.092
     data = None
 
 
@@ -4811,7 +4811,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 0.585
-    std = 3.823
+    std = 1.823
     data = None
 
 
@@ -5061,7 +5061,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 0.845
-    std = 6.773
+    std = 0.773
     data = None
 
 
@@ -5171,7 +5171,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 0.955
-    std = 7.576
+    std = 0.576
     data = None
 
 
@@ -5421,7 +5421,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 1.469
-    std = 14.217
+    std = 1.217
     data = None
 
 
@@ -5531,7 +5531,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 1.598
-    std = 14.929
+    std = 0.929
     data = None
 
 
@@ -5781,7 +5781,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 2.097
-    std = 20.441
+    std = 1.441
     data = None
 
 
@@ -5891,7 +5891,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 2.233
-    std = 20.387
+    std = 1.387
     data = None
 
 
@@ -6141,7 +6141,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 3.568
-    std = 38.412
+    std = 1.412
     data = None
 
 
@@ -6251,7 +6251,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 3.865
-    std = 39.828
+    std = 1.828
     data = None
 
 
@@ -6500,8 +6500,8 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     shape = [320]
     dtype = "torch.float32"
     device = "cuda:0"
-    mean = 6.457
-    std = 76.325
+    mean = 1.457
+    std = 1.325
     data = None
 
 
@@ -6510,7 +6510,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     shape = [320]
     dtype = "torch.float32"
     device = "cuda:0"
-    mean = 1.964
+    mean = 0.564
     std = 0.359
     data = None
 
@@ -6611,7 +6611,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 6.959
-    std = 79.050
+    std = 1.050
     data = None
 
 
@@ -6861,7 +6861,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block3_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 14.760
-    std = 197.511
+    std = 1.511
     data = None
 
 
@@ -7667,7 +7667,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block4_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 2.187
-    std = 41.543
+    std = 1.543
     data = None
 
 
@@ -7777,7 +7777,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block4_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 2.757
-    std = 49.021
+    std = 1.021
     data = None
 
 
@@ -8027,7 +8027,7 @@ class Program_weight_tensor_meta_L_self_modules_backbone_modules_block4_modules_
     dtype = "torch.float32"
     device = "cuda:0"
     mean = 7.591
-    std = 111.366
+    std = 1.366
     data = None