1. import torch
2. import torchvision
3.  
4. model = torchvision.models.resnet101(pretrained=True)
5. model = model.cuda()
6.  
7. model.eval()
8.  
9. from ptflops import get_model_complexity_info
10.  
11. import torch2trt
12.  
13. #x = torch.randn(6, 3, 928, 1600, device="cuda:0", dtype=torch.float32)
14.            
15. import time
16. timer = []
17. input = torch.randn(6, 3, 928, 1600, device="cuda:0", dtype=torch.float32)
18. trtmodel = torch2trt.torch2trt(model,
19.                                inputs=[input],
20.                                fp16_mode=False,
21.                                max_batch_size=6,
22.                                max_workspace_size=int(4e9))
23.  
24. for i in range(1000):
25.     s_time = time.time()
26.     trtmodel(input)
27.     torch.cuda.synchronize()
28.     e_time = time.time()
29.     timer.append(1000 * (e_time - s_time))
30.     print("{}".format(1000 * (e_time - s_time)))
31.     #print("{}".format(timer[-1]))
32. avg_time = sum(timer[500:])/len(timer[500:])
33. print("resnet_trt with 5 3 928 1600 {}".format(avg_time))
34.  
35. timer = []
36. input = torch.randn(1, 3, 224, 224, device="cuda:0", dtype=torch.float32)
37. for i in range(1000):
38.     model(input)
39.  
40. for i in range(1000):
41.     s_time = time.time()
42.     model(input)
43.     torch.cuda.synchronize()
44.     e_time = time.time()
45.     timer.append(1000 * (e_time - s_time))
46.     print("{}".format(1000 * (e_time - s_time)))
47. avg_time = sum(timer)/len(timer)
48. print("resnet_torch with 1 3 224 224 {}".format(avg_time))
49.  
50. timer = []
51. input = torch.randn(6, 3, 224, 224, device="cuda:0", dtype=torch.float32)
52. for i in range(1000):
53.     s_time = time.time()
54.     model(input)
55.     torch.cuda.synchronize()
56.     e_time = time.time()
57.     timer.append(1000 * (e_time - s_time))
58.     print("{}".format(1000 * (e_time - s_time)))
59. avg_time = sum(timer)/len(timer)
60. print("resnet_torch with 6 3 224 224 {}".format(avg_time))
61.  
62. timer = []
63. input = torch.randn(1, 3, 928, 1600, device="cuda:0", dtype=torch.float32)
64. for i in range(1000):
65.     s_time = time.time()
66.     model(input)
67.     torch.cuda.synchronize()
68.     e_time = time.time()
69.     timer.append(1000 * (e_time - s_time))
70.     print("{}".format(1000 * (e_time - s_time)))
71. avg_time = sum(timer)/len(timer)
72. print("resnet_torch with 1 3 928 1600 {}".format(avg_time))
73.  
74. timer = []
75. input = torch.randn(6, 3, 928, 1600, device="cuda:0", dtype=torch.float32)
76. for i in range(1000):
77.     s_time = time.time()
78.     model(input)
79.     torch.cuda.synchronize()
80.     e_time = time.time()
81.     timer.append(1000 * (e_time - s_time))
82.     print("{}".format(1000 * (e_time - s_time)))
83. avg_time = sum(timer)/len(timer)
84. print("resnet_torch with 6 3 928 1600 {}".format(avg_time))
85.  
86.  
87. timer = []
88. input = torch.randn(1, 3, 224, 224, device="cuda:0", dtype=torch.float32)
89. trtmodel = torch2trt.torch2trt(model,
90.                                inputs=[input],
91.                                fp16_mode=False,
92.                                max_batch_size=6,
93.                                max_workspace_size=int(4e9))
94.  
95. for i in range(1000):
96.     s_time = time.time()
97.     trtmodel(input)
98.     torch.cuda.synchronize()
99.     e_time = time.time()
100.     timer.append(1000 * (e_time - s_time))
101.     print("{}".format(1000 * (e_time - s_time)))
102. avg_time = sum(timer)/len(timer)
103. print("resnet_trt with 1 3 224 224 {}".format(avg_time))
104.  
105. timer = []
106. input = torch.randn(6, 3, 224, 224, device="cuda:0", dtype=torch.float32)
107. trtmodel = torch2trt.torch2trt(model,
108.                                inputs=[input],
109.                                fp16_mode=False,
110.                                max_batch_size=6,
111.                                max_workspace_size=int(4e9))
112.  
113. for i in range(1000):
114.     s_time = time.time()
115.     trtmodel(input)
116.     torch.cuda.synchronize()
117.     e_time = time.time()
118.     timer.append(1000 * (e_time - s_time))
119.     print("{}".format(1000 * (e_time - s_time)))
120. avg_time = sum(timer)/len(timer)
121. print("resnet_trt with 6 3 224 224 {}".format(avg_time))
122.  
123. timer = []
124. input = torch.randn(1, 3, 928, 1600, device="cuda:0", dtype=torch.float32)
125. trtmodel = torch2trt.torch2trt(model,
126.                                inputs=[input],
127.                                fp16_mode=False,
128.                                max_batch_size=6,
129.                                max_workspace_size=int(4e9))
130.  
131. for i in range(1000):
132.     s_time = time.time()
133.     trtmodel(input)
134.     torch.cuda.synchronize()
135.     e_time = time.time()
136.     timer.append(1000 * (e_time - s_time))
137.     print("{}".format(1000 * (e_time - s_time)))
138. avg_time = sum(timer)/len(timer)
139. print("resnet_trt with 1 3 928 1600 {}".format(avg_time))
140.  
141.