1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
|
import torch
import torchvision.models.resnet as resnet
import torch.nn as nn
import numpy as np
import sys
if __name__ == "__main__":
assert len(sys.argv) == 3
halide_output_file = sys.argv[1]
seed = int(sys.argv[2])
np.random.seed(seed)
input_data = np.random.rand(1, 224, 224, 3).astype(np.float32)
input_data = input_data.transpose(0, 3, 1, 2)
image = torch.from_numpy(input_data)
net = resnet.resnet50(pretrained=True)
net.eval()
result = net(image)
# PyTorch resnet50 model doesn't compute softmax of the output.
softmax = nn.Softmax(dim=1)
ref_output = softmax(result)
halide_output = np.fromfile(halide_output_file, dtype=np.float32)
ref_output = ref_output.cpu().detach().numpy().reshape(halide_output.shape)
np.testing.assert_almost_equal(np.argmax(halide_output), np.argmax(ref_output))
np.testing.assert_almost_equal(halide_output, ref_output, decimal=2)
print("Success!")
|
