Yes, for example, efficientnet
class Swish(nn.Module):
def forward(self, x):
return x * torch.sigmoid(x)
class Flatten(nn.Module):
def forward(self, x):
return x.reshape(x.shape[0], -1)
class SqueezeExcitation(nn.Module):
def __init__(self, inplanes, se_planes):
super(SqueezeExcitation, self).__init__()
self.reduce_expand = nn.Sequential(
nn.Conv2d(inplanes, se_planes,
kernel_size=1, stride=1, padding=0, bias=True),
Swish(),
nn.Conv2d(se_planes, inplanes,
kernel_size=1, stride=1, padding=0, bias=True),
nn.Sigmoid()
)
def forward(self, x):
x_se = torch.mean(x, dim=(-2, -1), keepdim=True)
x_se = self.reduce_expand(x_se)
return x_se * x
class MBConv(nn.Module):
def init(self, inplanes, planes, kernel_size, stride,
expand_rate=1.0, se_rate=0.25,
drop_connect_rate=0.2):
super(MBConv, self).init()
expand_planes = int(inplanes * expand_rate)
se_planes = max(1, int(inplanes * se_rate))
self.expansion_conv = None
if expand_rate > 1.0:
self.expansion_conv = nn.Sequential(
nn.Conv2d(inplanes, expand_planes,
kernel_size=1, stride=1, padding=0, bias=False),
nn.BatchNorm2d(expand_planes, momentum=0.01, eps=1e-3),
Swish()
)
inplanes = expand_planes
self.depthwise_conv = nn.Sequential(
nn.Conv2d(inplanes, expand_planes,
kernel_size=kernel_size, stride=stride,
padding=kernel_size // 2, groups=expand_planes,
bias=False),
nn.BatchNorm2d(expand_planes, momentum=0.01, eps=1e-3),
Swish()
)
self.squeeze_excitation = SqueezeExcitation(expand_planes, se_planes)
self.project_conv = nn.Sequential(
nn.Conv2d(expand_planes, planes,
kernel_size=1, stride=1, padding=0, bias=False),
nn.BatchNorm2d(planes, momentum=0.01, eps=1e-3),
)
self.with_skip = stride == 1
self.drop_connect_rate = torch.tensor(drop_connect_rate, requires_grad=False)
def _drop_connect(self, x):
keep_prob = 1.0 - self.drop_connect_rate
drop_mask = torch.rand(x.shape[0], 1, 1, 1) + keep_prob
drop_mask = drop_mask.type_as(x)
drop_mask.floor_()
return drop_mask * x / keep_prob
def forward(self, x):
z = x
if self.expansion_conv is not None:
x = self.expansion_conv(x)
x = self.depthwise_conv(x)
x = self.squeeze_excitation(x)
x = self.project_conv(x)
# Add identity skip
if x.shape == z.shape and self.with_skip:
if self.training and self.drop_connect_rate is not None:
self._drop_connect(x)
x += z
return x
def init_weights(module):
if isinstance(module, nn.Conv2d):
nn.init.kaiming_normal_(module.weight, a=0, mode=‘fan_out’)
elif isinstance(module, nn.Linear):
init_range = 1.0 / math.sqrt(module.weight.shape[1])
nn.init.uniform_(module.weight, a=-init_range, b=init_range)
class EfficientNet(nn.Module):
def _setup_repeats(self, num_repeats):
return int(math.ceil(self.depth_coefficient * num_repeats))
def _setup_channels(self, num_channels):
num_channels *= self.width_coefficient
new_num_channels = math.floor(num_channels / self.divisor + 0.5) * self.divisor
new_num_channels = max(self.divisor, new_num_channels)
if new_num_channels < 0.9 * num_channels:
new_num_channels += self.divisor
return new_num_channels
def __init__(self, num_classes,
width_coefficient=1.0,
depth_coefficient=1.0,
se_rate=0.25,
dropout_rate=0.2,
drop_connect_rate=0.2):
super(EfficientNet, self).__init__()
self.width_coefficient = width_coefficient
self.depth_coefficient = depth_coefficient
self.divisor = 8
list_channels = [32, 16, 24, 40, 80, 112, 192, 320, 1280]
list_channels = [self._setup_channels(c) for c in list_channels]
list_num_repeats = [1, 2, 2, 3, 3, 4, 1]
list_num_repeats = [self._setup_repeats(r) for r in list_num_repeats]
expand_rates = [1, 6, 6, 6, 6, 6, 6]
strides = [1, 2, 2, 2, 1, 2, 1]
kernel_sizes = [3, 3, 5, 3, 5, 5, 3]
# Define stem:
self.stem = nn.Sequential(
nn.Conv2d(3, list_channels[0], kernel_size=3, stride=2, padding=1, bias=False),
nn.BatchNorm2d(list_channels[0], momentum=0.01, eps=1e-3),
Swish()
)
# Define MBConv blocks
blocks = []
counter = 0
num_blocks = sum(list_num_repeats)
for idx in range(7):
num_channels = list_channels[idx]
next_num_channels = list_channels[idx + 1]
num_repeats = list_num_repeats[idx]
expand_rate = expand_rates[idx]
kernel_size = kernel_sizes[idx]
stride = strides[idx]
drop_rate = drop_connect_rate * counter / num_blocks
name = "MBConv{}_{}".format(expand_rate, counter)
blocks.append((
name,
MBConv(num_channels, next_num_channels,
kernel_size=kernel_size, stride=stride, expand_rate=expand_rate,
se_rate=se_rate, drop_connect_rate=drop_rate)
))
counter += 1
for i in range(1, num_repeats):
name = "MBConv{}_{}".format(expand_rate, counter)
drop_rate = drop_connect_rate * counter / num_blocks
blocks.append((
name,
MBConv(next_num_channels, next_num_channels,
kernel_size=kernel_size, stride=1, expand_rate=expand_rate,
se_rate=se_rate, drop_connect_rate=drop_rate)
))
counter += 1
self.blocks = nn.Sequential(OrderedDict(blocks))
# Define head
self.head = nn.Sequential(
nn.Conv2d(list_channels[-2], list_channels[-1],
kernel_size=1, bias=False),
nn.BatchNorm2d(list_channels[-1], momentum=0.01, eps=1e-3),
Swish(),
nn.AdaptiveAvgPool2d(1),
Flatten(),
nn.Dropout(p=dropout_rate),
nn.Linear(list_channels[-1], num_classes)
)
self.apply(init_weights)
def forward(self, x):
f = self.stem(x)
f = self.blocks(f)
y = self.head(f)
return y
model = EfficientNet(num_classes=10,
width_coefficient=1.4, depth_coefficient=1.8,
dropout_rate=0.4)
resolution = 380
img_stats = [[0.485, 0.456, 0.406], [0.229, 0.224, 0.225]]
but I managed to save it with this library called dill
For example,
import dill
Dill routine
model_copy=dill.dumps(model)
torch.save(model_copy,‘model_ignite_original.pt’)
and I dont get any errors…
when I want to load it I do it like this
model1 = torch.load(model_name)
model=dill.loads(model1)
