fix: sagemaker import issue
Browse files
block.py
CHANGED
@@ -7,6 +7,7 @@ from functools import partial
|
|
7 |
from typing import Optional
|
8 |
|
9 |
import torch
|
|
|
10 |
import torch.nn as nn
|
11 |
import torch.nn.functional as F
|
12 |
from torch import Tensor
|
@@ -21,6 +22,64 @@ except ImportError:
|
|
21 |
layer_norm_fn, RMSNorm = None, None
|
22 |
|
23 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
class Block(nn.Module):
|
25 |
def __init__(
|
26 |
self,
|
|
|
7 |
from typing import Optional
|
8 |
|
9 |
import torch
|
10 |
+
import torch.fx
|
11 |
import torch.nn as nn
|
12 |
import torch.nn.functional as F
|
13 |
from torch import Tensor
|
|
|
22 |
layer_norm_fn, RMSNorm = None, None
|
23 |
|
24 |
|
25 |
+
def stochastic_depth(
|
26 |
+
input: Tensor, p: float, mode: str, training: bool = True
|
27 |
+
) -> Tensor:
|
28 |
+
"""
|
29 |
+
Implements the Stochastic Depth from `"Deep Networks with Stochastic Depth"
|
30 |
+
<https://arxiv.org/abs/1603.09382>`_ used for randomly dropping residual
|
31 |
+
branches of residual architectures.
|
32 |
+
Args:
|
33 |
+
input (Tensor[N, ...]): The input tensor or arbitrary dimensions with the first one
|
34 |
+
being its batch i.e. a batch with ``N`` rows.
|
35 |
+
p (float): probability of the input to be zeroed.
|
36 |
+
mode (str): ``"batch"`` or ``"row"``.
|
37 |
+
``"batch"`` randomly zeroes the entire input, ``"row"`` zeroes
|
38 |
+
randomly selected rows from the batch.
|
39 |
+
training: apply stochastic depth if is ``True``. Default: ``True``
|
40 |
+
Returns:
|
41 |
+
Tensor[N, ...]: The randomly zeroed tensor.
|
42 |
+
"""
|
43 |
+
if p < 0.0 or p > 1.0:
|
44 |
+
raise ValueError(f"drop probability has to be between 0 and 1, but got {p}")
|
45 |
+
if mode not in ["batch", "row"]:
|
46 |
+
raise ValueError(f"mode has to be either 'batch' or 'row', but got {mode}")
|
47 |
+
if not training or p == 0.0:
|
48 |
+
return input
|
49 |
+
|
50 |
+
survival_rate = 1.0 - p
|
51 |
+
if mode == "row":
|
52 |
+
size = [input.shape[0]] + [1] * (input.ndim - 1)
|
53 |
+
else:
|
54 |
+
size = [1] * input.ndim
|
55 |
+
noise = torch.empty(size, dtype=input.dtype, device=input.device)
|
56 |
+
noise = noise.bernoulli_(survival_rate)
|
57 |
+
if survival_rate > 0.0:
|
58 |
+
noise.div_(survival_rate)
|
59 |
+
return input * noise
|
60 |
+
|
61 |
+
|
62 |
+
torch.fx.wrap("stochastic_depth")
|
63 |
+
|
64 |
+
|
65 |
+
class StochasticDepth(nn.Module):
|
66 |
+
"""
|
67 |
+
See :func:`stochastic_depth`.
|
68 |
+
"""
|
69 |
+
|
70 |
+
def __init__(self, p: float, mode: str) -> None:
|
71 |
+
super().__init__()
|
72 |
+
self.p = p
|
73 |
+
self.mode = mode
|
74 |
+
|
75 |
+
def forward(self, input: Tensor) -> Tensor:
|
76 |
+
return stochastic_depth(input, self.p, self.mode, self.training)
|
77 |
+
|
78 |
+
def __repr__(self) -> str:
|
79 |
+
s = f"{self.__class__.__name__}(p={self.p}, mode={self.mode})"
|
80 |
+
return s
|
81 |
+
|
82 |
+
|
83 |
class Block(nn.Module):
|
84 |
def __init__(
|
85 |
self,
|