pytorch

1
import torch
2

3

4
class TorchTensorEngine:
5
    def rand(self, shape, device=None, dtype=None, requires_grad=False):
6
        return torch.rand(
7
            shape, device=device, dtype=dtype, requires_grad=requires_grad
8
        )
9

10
    def randn(self, shape, device=None, dtype=None, requires_grad=False):
11
        return torch.randn(
12
            shape, device=device, dtype=dtype, requires_grad=requires_grad
13
        )
14

15
    def nchw_rand(self, shape, device=None, requires_grad=False):
16
        return self.rand(shape, device=device, requires_grad=requires_grad)
17

18
    def reset(self, _):
19
        pass
20

21
    def rand_like(self, v):
22
        return torch.rand_like(v)
23

24
    def numpy(self, t):
25
        return t.cpu().numpy()
26

27
    def mul(self, t1, t2):
28
        return t1 * t2
29

30
    def add(self, t1, t2):
31
        return t1 + t2
32

33
    def batch_norm(self, data, mean, var, training):
34
        return torch.nn.functional.batch_norm(data, mean, var, training=training)
35

36
    def instance_norm(self, data):
37
        return torch.nn.functional.instance_norm(data)
38

39
    def layer_norm(self, data, shape):
40
        return torch.nn.functional.layer_norm(data, shape)
41

42
    def sync_cuda(self):
43
        torch.cuda.synchronize()
44

45
    def backward(self, tensors, grad_tensors, _):
46
        torch.autograd.backward(tensors, grad_tensors=grad_tensors)
47

48
    def sum(self, data, dims):
49
        return torch.sum(data, dims)
50

51
    def softmax(self, data, dim=None, dtype=None):
52
        return torch.nn.functional.softmax(data, dim, dtype)
53

54
    def cat(self, inputs, dim=0):
55
        return torch.cat(inputs, dim=dim)
56

57
    def clamp(self, data, min, max):
58
        return torch.clamp(data, min=min, max=max)
59

60
    def relu(self, data):
61
        return torch.nn.functional.relu(data)
62

63
    def tanh(self, data):
64
        return torch.tanh(data)
65

66
    def max_pool2d(self, data, kernel_size, stride=1):
67
        return torch.nn.functional.max_pool2d(data, kernel_size, stride=stride)
68

69
    def avg_pool2d(self, data, kernel_size, stride=1):
70
        return torch.nn.functional.avg_pool2d(data, kernel_size, stride=stride)
71

72
    def conv2d_layer(self, ic, oc, kernel_size, groups=1):
73
        return torch.nn.Conv2d(ic, oc, kernel_size, groups=groups)
74

75
    def matmul(self, t1, t2):
76
        return torch.matmul(t1, t2)
77

78
    def to_device(self, module, device):
79
        return module.to(device)
80