glenn-jocher
commited on
Commit
•
5d4258f
1
Parent(s):
7f9bbf0
Fix MixConv2d() remove shortcut + apply depthwise (#5410)
Browse files- models/common.py +1 -1
- models/experimental.py +11 -10
- utils/torch_utils.py +1 -1
models/common.py
CHANGED
@@ -113,7 +113,7 @@ class BottleneckCSP(nn.Module):
|
|
113 |
self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False)
|
114 |
self.cv4 = Conv(2 * c_, c2, 1, 1)
|
115 |
self.bn = nn.BatchNorm2d(2 * c_) # applied to cat(cv2, cv3)
|
116 |
-
self.act = nn.
|
117 |
self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
|
118 |
|
119 |
def forward(self, x):
|
|
|
113 |
self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False)
|
114 |
self.cv4 = Conv(2 * c_, c2, 1, 1)
|
115 |
self.bn = nn.BatchNorm2d(2 * c_) # applied to cat(cv2, cv3)
|
116 |
+
self.act = nn.SiLU()
|
117 |
self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
|
118 |
|
119 |
def forward(self, x):
|
models/experimental.py
CHANGED
@@ -2,7 +2,7 @@
|
|
2 |
"""
|
3 |
Experimental modules
|
4 |
"""
|
5 |
-
|
6 |
import numpy as np
|
7 |
import torch
|
8 |
import torch.nn as nn
|
@@ -48,26 +48,27 @@ class Sum(nn.Module):
|
|
48 |
|
49 |
class MixConv2d(nn.Module):
|
50 |
# Mixed Depth-wise Conv https://arxiv.org/abs/1907.09595
|
51 |
-
def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True):
|
52 |
super().__init__()
|
53 |
-
|
54 |
if equal_ch: # equal c_ per group
|
55 |
-
i = torch.linspace(0,
|
56 |
-
c_ = [(i == g).sum() for g in range(
|
57 |
else: # equal weight.numel() per group
|
58 |
-
b = [c2] + [0] *
|
59 |
-
a = np.eye(
|
60 |
a -= np.roll(a, 1, axis=1)
|
61 |
a *= np.array(k) ** 2
|
62 |
a[0] = 1
|
63 |
c_ = np.linalg.lstsq(a, b, rcond=None)[0].round() # solve for equal weight indices, ax = b
|
64 |
|
65 |
-
self.m = nn.ModuleList(
|
|
|
66 |
self.bn = nn.BatchNorm2d(c2)
|
67 |
-
self.act = nn.
|
68 |
|
69 |
def forward(self, x):
|
70 |
-
return
|
71 |
|
72 |
|
73 |
class Ensemble(nn.ModuleList):
|
|
|
2 |
"""
|
3 |
Experimental modules
|
4 |
"""
|
5 |
+
import math
|
6 |
import numpy as np
|
7 |
import torch
|
8 |
import torch.nn as nn
|
|
|
48 |
|
49 |
class MixConv2d(nn.Module):
|
50 |
# Mixed Depth-wise Conv https://arxiv.org/abs/1907.09595
|
51 |
+
def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True): # ch_in, ch_out, kernel, stride, ch_strategy
|
52 |
super().__init__()
|
53 |
+
n = len(k) # number of convolutions
|
54 |
if equal_ch: # equal c_ per group
|
55 |
+
i = torch.linspace(0, n - 1E-6, c2).floor() # c2 indices
|
56 |
+
c_ = [(i == g).sum() for g in range(n)] # intermediate channels
|
57 |
else: # equal weight.numel() per group
|
58 |
+
b = [c2] + [0] * n
|
59 |
+
a = np.eye(n + 1, n, k=-1)
|
60 |
a -= np.roll(a, 1, axis=1)
|
61 |
a *= np.array(k) ** 2
|
62 |
a[0] = 1
|
63 |
c_ = np.linalg.lstsq(a, b, rcond=None)[0].round() # solve for equal weight indices, ax = b
|
64 |
|
65 |
+
self.m = nn.ModuleList(
|
66 |
+
[nn.Conv2d(c1, int(c_), k, s, k // 2, groups=math.gcd(c1, int(c_)), bias=False) for k, c_ in zip(k, c_)])
|
67 |
self.bn = nn.BatchNorm2d(c2)
|
68 |
+
self.act = nn.SiLU()
|
69 |
|
70 |
def forward(self, x):
|
71 |
+
return self.act(self.bn(torch.cat([m(x) for m in self.m], 1)))
|
72 |
|
73 |
|
74 |
class Ensemble(nn.ModuleList):
|
utils/torch_utils.py
CHANGED
@@ -166,7 +166,7 @@ def initialize_weights(model):
|
|
166 |
elif t is nn.BatchNorm2d:
|
167 |
m.eps = 1e-3
|
168 |
m.momentum = 0.03
|
169 |
-
elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6]:
|
170 |
m.inplace = True
|
171 |
|
172 |
|
|
|
166 |
elif t is nn.BatchNorm2d:
|
167 |
m.eps = 1e-3
|
168 |
m.momentum = 0.03
|
169 |
+
elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU]:
|
170 |
m.inplace = True
|
171 |
|
172 |
|