blob: 17dbb488de0b02dc5e8bba3a2e38c7ee2548d8d3 (
plain)
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
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
|
op {
graph_op_name: "SparseApplyMomentum"
in_arg {
name: "var"
description: <<END
Should be from a Variable().
END
}
in_arg {
name: "accum"
description: <<END
Should be from a Variable().
END
}
in_arg {
name: "lr"
description: <<END
Learning rate. Must be a scalar.
END
}
in_arg {
name: "grad"
description: <<END
The gradient.
END
}
in_arg {
name: "indices"
description: <<END
A vector of indices into the first dimension of var and accum.
END
}
in_arg {
name: "momentum"
description: <<END
Momentum. Must be a scalar.
END
}
out_arg {
name: "out"
description: <<END
Same as "var".
END
}
attr {
name: "use_locking"
description: <<END
If `True`, updating of the var and accum tensors will be protected
by a lock; otherwise the behavior is undefined, but may exhibit less
contention.
END
}
attr {
name: "use_nesterov"
description: <<END
If `True`, the tensor passed to compute grad will be
var - lr * momentum * accum, so in the end, the var you get is actually
var - lr * momentum * accum.
END
}
summary: "Update relevant entries in \'*var\' and \'*accum\' according to the momentum scheme."
description: <<END
Set use_nesterov = True if you want to use Nesterov momentum.
That is for rows we have grad for, we update var and accum as follows:
$$accum = accum * momentum + grad$$
$$var -= lr * accum$$
END
}
|
