Tiny-SSD:VGG-16をSqueeze-netで置き換える
19301 ワード
前言:この作者の考え方は何もなくて、ネットを変えただけで、しかし意外にも彼の中のパラメータはただ勝手に書いただけで、完全に成立するわけではありませんて、もしこれらのパラメータをSSDの中でVGG 16のように設計すれば、正しいコンパイル(とてもつらい)ができて、ネットをここに残して、直接VGG 16の部分を置き換えることができます.最近kittiデータセットを使用してテストとyolo V 3の比較を行う予定です
テスト関数:
# coding=utf-8
"""Keras implementation of SSD."""
import keras.backend as K
from keras.layers import Activation
from keras.layers import AtrousConvolution2D
from keras.layers import Conv2D
from keras.layers import Dense
from keras.layers import Flatten
from keras.layers import GlobalAveragePooling2D
from keras.layers import Input
from keras.layers import MaxPooling2D
from keras.layers import merge
from keras.layers import Reshape
from keras.layers import ZeroPadding2D
from keras.models import Model, Sequential
from ssd_layers import Normalize
from ssd_layers import PriorBox
from keras.models import Sequential
from keras.layers import Dense, Flatten, Dropout, Concatenate
from keras.layers.convolutional import Conv2D, MaxPooling2D
import numpy as np
def SqueezeNet(inputs, nb_classes=21):
""" Keras Implementation of SqueezeNet(arXiv 1602.07360)
@param nb_classes: total number of final categories
Arguments:
inputs -- shape of the input images (channel, cols, rows)
"""
img_size = (inputs[0], inputs[1])
input_img = (Input(shape=inputs))
conv1 = Conv2D(
64, (3, 3), activation='relu', kernel_initializer='glorot_uniform',
strides=(1, 1), name='conv1', padding='same',
data_format="channels_last")(input_img)
# maxpool1
maxpool1 = MaxPooling2D(
pool_size=(2, 2), strides=(2, 2), name='maxpool1',
data_format="channels_last")(conv1)
# fire1
fire1_squeeze = Conv2D(
15, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
name='fire1_squeeze',
data_format="channels_last")(maxpool1)
fire1_expand1 = Conv2D(
49, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire1_expand1',
data_format="channels_last")(fire1_squeeze)
fire1_expand2 = Conv2D(
53, (3, 3), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire1_expand2',
data_format="channels_last")(fire1_squeeze)
merge1 = Concatenate(axis=3)([fire1_expand1, fire1_expand2])
maxpool2 = MaxPooling2D(
pool_size=(2, 2), strides=(2, 2), name='maxpool2',
data_format="channels_last")(merge1)
# fire2
fire2_squeeze = Conv2D(
15, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
name='fire2_squeeze',
data_format="channels_last")(maxpool2)
fire2_expand1 = Conv2D(
54, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire2_expand1',
data_format="channels_last")(fire2_squeeze)
fire2_expand2 = Conv2D(
52, (3, 3), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire2_expand2',
data_format="channels_last")(fire2_squeeze)
merge2 = Concatenate(axis=3)([fire2_expand1, fire2_expand2])
# maxpool3
maxpool3 = MaxPooling2D(
pool_size=(3, 3), strides=(2, 2), padding='same', name='maxpool3',
data_format="channels_last")(merge2)
# fire3
fire3_squeeze = Conv2D(
29, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
name='fire3_squeeze',
data_format="channels_last")(maxpool3)
fire3_expand1 = Conv2D(
92, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire3_expand1',
data_format="channels_last")(fire3_squeeze)
fire3_expand2 = Conv2D(
94, (3, 3), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire3_expand2',
data_format="channels_last")(fire3_squeeze)
merge3 = Concatenate(axis=3)([fire3_expand1, fire3_expand2])
# fire4
fire4_squeeze = Conv2D(
29, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
name='fire4_squeeze',
data_format="channels_last")(merge3)
fire4_expand1 = Conv2D(
90, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
name='fire4_expand1',
data_format="channels_last")(fire4_squeeze)
fire4_expand2 = Conv2D(
83, (3, 3), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire4_expand2',
data_format="channels_last")(fire4_squeeze)
merge4 = Concatenate(axis=3)([fire4_expand1, fire4_expand2])
# maxpool4
maxpool4 = MaxPooling2D(
pool_size=(2, 2), strides=(2, 2), name='maxpool4', padding='same',
data_format="channels_last")(merge4)
# fire5
fire5_squeeze = Conv2D(
44, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
name='fire5_squeeze',
data_format="channels_last")(maxpool4)
fire5_expand1 = Conv2D(
166, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire5_expand1',
data_format="channels_last")(fire5_squeeze)
fire5_expand2 = Conv2D(
161, (3, 3), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire5_expand2',
data_format="channels_last")(fire5_squeeze)
merge5 = Concatenate(axis=3)([fire5_expand1, fire5_expand2])
# fire6
fire6_squeeze = Conv2D(
45, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
name='fire6_squeeze',
data_format="channels_last")(merge5)
fire6_expand1 = Conv2D(
155, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire6_expand1',
data_format="channels_last")(fire6_squeeze)
fire6_expand2 = Conv2D(
146, (3, 3), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire6_expand2',
data_format="channels_last")(fire6_squeeze)
merge6 = Concatenate(axis=3)([fire6_expand1, fire6_expand2])
# fire7
fire7_squeeze = Conv2D(
49, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
name='fire7_squeeze',
data_format="channels_last")(merge6)
fire7_expand1 = Conv2D(
163, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire7_expand1',
data_format="channels_last")(fire7_squeeze)
fire7_expand2 = Conv2D(
171, (3, 3), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire7_expand2',
data_format="channels_last")(fire7_squeeze)
merge7 = Concatenate(axis=3)([fire7_expand1, fire7_expand2])
# fire8
fire8_squeeze = Conv2D(
25, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
name='fire8_squeeze',
data_format="channels_last")(merge7)
fire8_expand1 = Conv2D(
29, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire8_expand1',
data_format="channels_last")(fire8_squeeze)
fire8_expand2 = Conv2D(
54, (3, 3), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire8_expand2',
data_format="channels_last")(fire8_squeeze)
merge8 = Concatenate(axis=3)([fire8_expand1, fire8_expand2])
# maxpool9
maxpool9 = MaxPooling2D(
pool_size=(3, 3), strides=(2, 2), padding='same', name='maxpool9',
data_format="channels_last")(merge8)
# fire9
fire9_squeeze = Conv2D(
37, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
name='fire9_squeeze',
data_format="channels_last")(maxpool9)
fire9_expand1 = Conv2D(
45, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire9_expand1',
data_format="channels_last")(fire9_squeeze)
fire9_expand2 = Conv2D(
56, (3, 3), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire9_expand2',
data_format="channels_last")(fire9_squeeze)
merge9 = Concatenate(axis=3)([fire9_expand1, fire9_expand2])
# maxpool10
maxpool10 = MaxPooling2D(
pool_size=(2, 2), strides=(2, 2), name='maxpool10',
data_format="channels_last")(merge9)
# fire10
fire10_squeeze = Conv2D(
38, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire10_squeeze',
data_format="channels_last")(maxpool10)
fire10_expand1 = Conv2D(
41, (1, 1), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire10_expand1',
data_format="channels_last")(fire10_squeeze)
fire10_expand2 = Conv2D(
44, (3, 3), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='fire10_expand2',
data_format="channels_last")(fire10_squeeze)
merge10 = Concatenate(axis=3)([fire10_expand1, fire10_expand2])
# cov12-1
conv12_1 = Conv2D(
51, (3, 3), activation='relu', kernel_initializer='glorot_uniform',
padding='same', strides=(2, 2), name='conv12_1',
data_format='channels_last')(merge10)
# padding_1
padding_1 = ZeroPadding2D((1, 1), data_format='channels_last')(conv12_1)
# conv12_2
conv12_2 = Conv2D(
46, (3, 3), activation='relu', kernel_initializer='glorot_uniform',
name='conv12_2',
data_format='channels_last')(padding_1)
# conv13_1
conv13_1 = Conv2D(
55, (3, 3), activation='relu', kernel_initializer='glorot_uniform',
padding='same', name='conv13_1',
data_format='channels_last')(conv12_2)
# padding_2
# padding_2 = ZeroPadding2D((1,1),data_format='channels_last')(conv13_1)
# conv13_2
conv13_2 = Conv2D(
85, (3, 3), activation='relu', kernel_initializer='glorot_uniform',
name='conv13_2',
data_format='channels_last')(conv13_1)
# Prediction from Fire4
num_priors = 3
fire4_norm = Normalize(20, name='fire4_norm')(merge4)
fire4_norm_mbox_loc = Conv2D(
num_priors*4, (3, 3), name='fire4_norm_mbox_loc', padding='same',
data_format='channels_last')(fire4_norm)
fire4_mbox_norm_loc_flat = Flatten()(fire4_norm_mbox_loc)
name = 'fire4_norm_mbox_conf'
if nb_classes != 21:
name += '_{}'.format(nb_classes)
fire4_norm_mbox_conf = Conv2D(
num_priors*nb_classes, (3, 3), name=name, padding='same',
data_format='channels_last')(fire4_norm)
fire4_norm_mbox_conf_flat = Flatten()(fire4_norm_mbox_conf)
fire4_norm_mbox_priorbox = PriorBox(img_size, 30.0, aspect_ratios=[2],
variances=[0.1, 0.1, 0.2, 0.2],
name='fire4_norm_mbox_priorbox')(fire4_norm)
fire4_priorbox_flatten = Flatten()(fire4_norm_mbox_priorbox)
# Prediction from Fire8
num_priors = 6
fire8_mbox_loc = Conv2D(
num_priors*4, (3, 3), name='fire8_mbox_loc', padding='same',
data_format='channels_last')(merge8)
fire8_mbox_loc_flat = Flatten()(fire8_mbox_loc)
name = 'fire8_mbox_conf'
if nb_classes != 21:
name += '_{}'.format(nb_classes)
fire8_mbox_conf = Conv2D(
num_priors*nb_classes, (3, 3), name=name, padding='same',
data_format='channels_last')(merge8)
fire8_mbox_conf_flat = Flatten()(fire8_mbox_conf)
fire8_mbox_priorbox = PriorBox(img_size, 60.0, max_size=114.0, aspect_ratios=[2, 3],
variances=[0.1, 0.1, 0.2, 0.2],
name='fire8_mbox_priorbox')(merge8)
# Prediction from Fire9
num_priors = 6
fire9_mbox_loc = Conv2D(
num_priors*4, (3, 3), name='fire9_mbox_loc', padding='same',
data_format='channels_last')(merge9)
fire9_mbox_loc_flat = Flatten()(fire9_mbox_loc)
name = 'fire9_mbox_conf'
if nb_classes != 21:
name += '_{}'.format(nb_classes)
fire9_mbox_conf = Conv2D(
num_priors*nb_classes, (3, 3), name=name, padding='same',
data_format='channels_last')(merge9)
fire9_mbox_conf_flat = Flatten()(fire9_mbox_conf)
fire9_mbox_priorbox = PriorBox(img_size, 114.0, max_size=168.0, aspect_ratios=[2, 3],
variances=[0.1, 0.1, 0.2, 0.2],
name='fire9_mbox_priorbox')(merge9)
# Prediction from Fire10
num_priors = 6
fire10_mbox_loc = Conv2D(
num_priors*4, (3, 3), name='fire10_mbox_loc', padding='same',
data_format='channels_last')(merge10)
fire10_mbox_loc_flat = Flatten()(fire10_mbox_loc)
name = 'fire10_mbox_conf'
if nb_classes != 21:
name += '_{}'.format(nb_classes)
fire10_mbox_conf = Conv2D(
nb_classes*num_priors, (3, 3), name=name, padding='same',
data_format='channels_last')(merge10)
fire10_mbox_conf_flat = Flatten()(fire10_mbox_conf)
fire10_mbox_priorbox = PriorBox(img_size, 168.0, max_size=222.0, aspect_ratios=[2, 3],
variances=[0.1, 0.1, 0.2, 0.2],
name='fire10_mbox_priorbox')(merge10)
# Prediction from Conv12_2
num_priors = 6
conv12_maxpool = MaxPooling2D(pool_size=(1, 1), data_format="channels_last")(conv12_2)
conv12_mbox_loc = Conv2D(
num_priors*4, (3, 3), name='conv12_mbox_loc', padding='same',
data_format='channels_last')(conv12_maxpool)
conv12_mbox_loc_flat = Flatten()(conv12_mbox_loc)
name = 'conv12_mbox_conf'
if nb_classes != 21:
name += '_{}'.format(nb_classes)
conv12_mbox_conf = Conv2D(
num_priors*nb_classes, (3, 3), name=name, padding='same',
data_format='channels_last')(conv12_maxpool)
conv12_mbox_conf_flat = Flatten()(conv12_mbox_conf)
conv12_mbox_priorbox = PriorBox(img_size, 222.0, max_size=276.0, aspect_ratios=[2, 3],
variances=[0.1, 0.1, 0.2, 0.2],
name='conv12_mbox_priorbox')(conv12_maxpool)
# pool6
# pool6 = GlobalAveragePooling2D(name='pool6')(conv8_2)
# Prediction from Conv13_2
conv13_maxpool = MaxPooling2D(pool_size=(1, 1), data_format="channels_last")(conv13_2)
num_priors = 6
conv13_mbox_loc = Conv2D(
num_priors*4, (3, 3), name='conv13_mbox_loc', padding='same',
data_format='channels_last')(conv13_maxpool)
conv13_mbox_loc_flat = Flatten()(conv13_mbox_loc)
name = 'conv13_mbox_conf'
if nb_classes != 21:
name += '_{}'.format(nb_classes)
conv13_mbox_conf = Conv2D(
num_priors*nb_classes, (3, 3), name=name, padding='same',
data_format='channels_last')(conv13_maxpool)
conv13_mbox_conf_flat = Flatten()(conv13_mbox_conf)
conv13_mbox_priorbox = PriorBox(img_size, 276.0, max_size=330.0, aspect_ratios=[2, 3],
variances=[0.1, 0.1, 0.2, 0.2],
name='conv13_mbox_priorbox')(conv13_maxpool)
# Gather all predictions
mbox_loc = Concatenate(axis=1)([fire4_mbox_norm_loc_flat,
fire8_mbox_loc_flat,
fire9_mbox_loc_flat,
fire10_mbox_loc_flat,
conv12_mbox_loc_flat,
conv13_mbox_loc_flat])
mbox_conf = Concatenate(axis=1)([fire4_norm_mbox_conf_flat,
fire8_mbox_conf_flat,
fire9_mbox_conf_flat,
fire10_mbox_conf_flat,
conv12_mbox_conf_flat,
conv13_mbox_conf_flat])
# fire4_mbox_priorbox_reshape = Reshape((-1,8),name = ' fire4_mbox_priorbox_reshape')(fire4_norm_mbox_priorbox)
mbox_priorbox = Concatenate(axis=1)([fire4_norm_mbox_priorbox,
fire8_mbox_priorbox,
fire9_mbox_priorbox,
fire10_mbox_priorbox,
conv12_mbox_priorbox,
conv13_mbox_priorbox])
# dense = Dense(4096,activation='relu')(flatten_bbox)
num_boxes = mbox_loc._keras_shape[-1] // 4
if hasattr(mbox_loc, '_keras_shape'):
num_boxes = mbox_loc._keras_shape[-1] // 4
elif hasattr(mbox_loc, 'int_shape'):
num_boxes = K.int_shape(mbox_loc)[-1] // 4
mbox_loc_final = Reshape((num_boxes, 4), name='mbox_loc_final')(mbox_loc)
mbox_conf_logits = Reshape((num_boxes, nb_classes), name='mbox_conf_logits')(mbox_conf)
mbox_conf_final = Activation('softmax', name='mbox_conf_final')(mbox_conf_logits)
predictions = Concatenate(axis=2, name='preditions')([mbox_loc_final, mbox_conf_final, mbox_priorbox])
return Model(inputs=input_img, outputs=predictions)
テスト関数:
model = SqueezeNet((300,300,3), 4)
model.compile(loss='categorical_crossentropy',optimizer='sgd',metrics=['accuracy'])
model.summary()