1602_Squeeze Net
圖：
網絡描述：
Squeeze Net 發表于ICLR-2017，作者分别來自Berkeley和Stanford，Squeeze Net不是模型壓縮技術，而是 “design strategies for CNN architectures with few parameters” 。 Squeeze Net是Han等提出的一種輕量且高效的CNN模型，它參數比AlexNet少50x，但模型性能（accuracy）與AlexNet接近。SqueezeNet的核心在于Fire module，Fire module 由兩層構成，分别是squeeze層+expand層，如上圖所示，squeeze層是一個1x1卷積核的卷積層，expand層是1x1和3x3卷積核的卷積層，expand層中，把1x1 和3x3 得到的feature map 進行concat。
特點，優點：
(1) 提出了新的網絡架構Fire Module，通過減少參數來進行模型壓縮
(2) 使用其他方法對提出的Squeeze Net模型進行進一步壓縮，
(3) 對參數空間進行了探索，主要研究了壓縮比和3∗3，3∗3卷積比例的影響
(4) 更高效的分布式訓練，小模型參數小，網絡通信量減少；
(5) 便于模型更新，模型小，用戶端程式容易更新；
(6) 利于部署在特定硬體如FPGA，因為其記憶體受限。是以研究小模型是很有現實意義的。
其他方法：使用以下三個政策來減少Squeeze Net設計參數
1.使用1∗1卷積代替3∗3 卷積：參數減少為原來的1/9
2.減少輸入通道數量：這一部分使用squeeze layers來實作
3.将池化層操作延後，可以給卷積層提供更大的激活圖：更大的激活圖保留了更多的資訊，可以提供更高的分類準确率
代碼：
keras實作：
def fire_module(x, fire_id, squeeze=16, expand=64):
s_id = 'fire' + str(fire_id) + '/'
if K.image_data_format() == 'channels_first':
channel_axis = 1
else:
channel_axis = 3
x = Convolution2D(squeeze, (1, 1), padding='valid', name=s_id + sq1x1)(x)
x = Activation('relu', name=s_id + relu + sq1x1)(x)
left = Convolution2D(expand, (1, 1), padding='valid', name=s_id + exp1x1)(x)
left = Activation('relu', name=s_id + relu + exp1x1)(left)
right = Convolution2D(expand, (3, 3), padding='same', name=s_id + exp3x3)(x)
right = Activation('relu', name=s_id + relu + exp3x3)(right)
x = concatenate([left, right], axis=channel_axis, name=s_id + 'concat')
return x
# Original SqueezeNet from paper.
def SqueezeNet(include_top=True, weights='imagenet',
input_tensor=None, input_shape=None,
pooling=None,
classes=1000):
"""Instantiates the SqueezeNet architecture.
"""
if weights not in {'imagenet', None}:
raise ValueError('The `weights` argument should be either '
'`None` (random initialization) or `imagenet` '
'(pre-training on ImageNet).')
if weights == 'imagenet' and classes != 1000:
raise ValueError('If using `weights` as imagenet with `include_top`'
' as true, `classes` should be 1000')
input_shape = _obtain_input_shape(input_shape,
default_size=227,
min_size=48,
data_format=K.image_data_format(),
require_flatten=include_top)
if input_tensor is None:
img_input = Input(shape=input_shape)
else:
if not K.is_keras_tensor(input_tensor):
img_input = Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
x = Convolution2D(64, (3, 3), strides=(2, 2), padding='valid', name='conv1')(img_input)
x = Activation('relu', name='relu_conv1')(x)
x = MaxPooling2D(pool_size=(3, 3), strides=(2, 2), name='pool1')(x)
x = fire_module(x, fire_id=2, squeeze=16, expand=64)
x = fire_module(x, fire_id=3, squeeze=16, expand=64)
x = MaxPooling2D(pool_size=(3, 3), strides=(2, 2), name='pool3')(x)
x = fire_module(x, fire_id=4, squeeze=32, expand=128)
x = fire_module(x, fire_id=5, squeeze=32, expand=128)
x = MaxPooling2D(pool_size=(3, 3), strides=(2, 2), name='pool5')(x)
x = fire_module(x, fire_id=6, squeeze=48, expand=192)
x = fire_module(x, fire_id=7, squeeze=48, expand=192)
x = fire_module(x, fire_id=8, squeeze=64, expand=256)
x = fire_module(x, fire_id=9, squeeze=64, expand=256)
if include_top:
# It's not obvious where to cut the network...
# Could do the 8th or 9th layer... some work recommends cutting earlier layers.
x = Dropout(0.5, name='drop9')(x)
x = Convolution2D(classes, (1, 1), padding='valid', name='conv10')(x)
x = Activation('relu', name='relu_conv10')(x)
x = GlobalAveragePooling2D()(x)
x = Activation('softmax', name='loss')(x)
else:
if pooling == 'avg':
x = GlobalAveragePooling2D()(x)
elif pooling=='max':
x = GlobalMaxPooling2D()(x)
elif pooling==None:
pass
else:
raise ValueError("Unknown argument for 'pooling'=" + pooling)
# Ensure that the model takes into account
# any potential predecessors of `input_tensor`.
if input_tensor is not None:
inputs = get_source_inputs(input_tensor)
else:
inputs = img_input
model = Model(inputs, x, name='squeezenet')
# load weights
if weights == 'imagenet':
if include_top:
weights_path = get_file('squeezenet_weights_tf_dim_ordering_tf_kernels.h5',
WEIGHTS_PATH,
cache_subdir='models')
else:
weights_path = get_file('squeezenet_weights_tf_dim_ordering_tf_kernels_notop.h5',
WEIGHTS_PATH_NO_TOP,
cache_subdir='models')
model.load_weights(weights_path)
if K.backend() == 'theano':
layer_utils.convert_all_kernels_in_model(model)
if K.image_data_format() == 'channels_first':
if K.backend() == 'tensorflow':
warnings.warn('You are using the TensorFlow backend, yet you '
'are using the Theano '
'image data format convention '
'(`image_data_format="channels_first"`). '
'For best performance, set '
'`image_data_format="channels_last"` in '
'your Keras config '
'at ~/.keras/keras.json.')
return model
pytorch實作：
class Fire(nn.Module):
def __init__(self, inplanes, squeeze_planes,
expand1x1_planes, expand3x3_planes):
super(Fire, self).__init__()
self.inplanes = inplanes
self.squeeze = nn.Conv2d(inplanes, squeeze_planes, kernel_size=1)
self.squeeze_activation = nn.ReLU(inplace=True)
self.expand1x1 = nn.Conv2d(squeeze_planes, expand1x1_planes,
kernel_size=1)
self.expand1x1_activation = nn.ReLU(inplace=True)
self.expand3x3 = nn.Conv2d(squeeze_planes, expand3x3_planes,
kernel_size=3, padding=1)
self.expand3x3_activation = nn.ReLU(inplace=True)
def forward(self, x):
x = self.squeeze_activation(self.squeeze(x))
return torch.cat([
self.expand1x1_activation(self.expand1x1(x)),
self.expand3x3_activation(self.expand3x3(x))
], 1)
class SqueezeNet(nn.Module):
def __init__(self, version='1_0', num_classes=1000):
super(SqueezeNet, self).__init__()
self.num_classes = num_classes
if version == '1_0':
self.features = nn.Sequential(
nn.Conv2d(3, 96, kernel_size=7, stride=2),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True),
Fire(96, 16, 64, 64),
Fire(128, 16, 64, 64),
Fire(128, 32, 128, 128),
nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True),
Fire(256, 32, 128, 128),
Fire(256, 48, 192, 192),
Fire(384, 48, 192, 192),
Fire(384, 64, 256, 256),
nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True),
Fire(512, 64, 256, 256),
)
elif version == '1_1':
self.features = nn.Sequential(
nn.Conv2d(3, 64, kernel_size=3, stride=2),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True),
Fire(64, 16, 64, 64),
Fire(128, 16, 64, 64),
nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True),
Fire(128, 32, 128, 128),
Fire(256, 32, 128, 128),
nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True),
Fire(256, 48, 192, 192),
Fire(384, 48, 192, 192),
Fire(384, 64, 256, 256),
Fire(512, 64, 256, 256),
)
else:
# FIXME: Is this needed? SqueezeNet should only be called from the
# FIXME: squeezenet1_x() functions
# FIXME: This checking is not done for the other models
raise ValueError("Unsupported SqueezeNet version {version}:"
"1_0 or 1_1 expected".format(version=version))
# Final convolution is initialized differently from the rest
final_conv = nn.Conv2d(512, self.num_classes, kernel_size=1)
self.classifier = nn.Sequential(
nn.Dropout(p=0.5),
final_conv,
nn.ReLU(inplace=True),
nn.AdaptiveAvgPool2d((1, 1))
)
for m in self.modules():
if isinstance(m, nn.Conv2d):
if m is final_conv:
init.normal_(m.weight, mean=0.0, std=0.01)
else:
init.kaiming_uniform_(m.weight)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, x):
x = self.features(x)
x = self.classifier(x)
return torch.flatten(x, 1)