[net] # Training batch=120 subdivisions=4 # Testing #batch=1 #subdivisions=1 height=224 width=224 channels=3 momentum=0.9 decay=0.0005 max_crop=256 burn_in=1000 #burn_in=100 learning_rate=0.256 policy=poly power=4 max_batches=800000 momentum=0.9 decay=0.00005 angle=7 hue=.1 saturation=.75 exposure=.75 aspect=.75 ###### STEM - downsample # conv1 [convolutional] filters=24 size=3 pad=1 stride=2 batch_normalize=1 activation=swish ###################### ###### 'r1_k3_a1_p1_s11_e1_i24_o24' ###### 3x3 - 24 # dw [convolutional] filters=24 groups=24 size=3 stride=1 pad=1 batch_normalize=1 activation=leaky # out [convolutional] filters=24 size=1 stride=1 pad=1 batch_normalize=1 activation=linear [shortcut] from=-3 activation=linear ###################### ###### 'r1_k3.5.7_a1.1_p1.1_s22_e6_i24_o32', ###### 3x3, 5x5, 7x7 - 32 - downsample ###### expand 6 = 144 # expand [convolutional] filters=144 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=leaky # dw [route] layers = -1 group_id=0 groups=3 [convolutional] batch_normalize=1 filters=48 groups=48 size=3 stride=2 pad=1 activation=leaky [route] layers = -3 group_id=1 groups=3 [convolutional] batch_normalize=1 filters=48 groups=48 size=5 stride=2 pad=1 activation=leaky [route] layers = -5 group_id=2 groups=3 [convolutional] batch_normalize=1 filters=48 groups=48 size=7 stride=2 pad=1 activation=leaky [route] layers = -1,-3,-5 # out [convolutional] filters=32 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=linear ###################### ###### 'r1_k3_a1.1_p1.1_s11_e3_i32_o32', ###### 3x3 - 32 ###### expand 3 = 96 # expand [convolutional] filters=96 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=leaky # dw [convolutional] filters=96 groups=96 size=3 stride=1 pad=1 batch_normalize=1 activation=leaky # out [convolutional] filters=32 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=linear [shortcut] from=-4 activation=linear ###################### ###### 'r1_k3.5.7.9_a1_p1_s22_e6_i32_o40_se0.5_sw', ###### 3x3, 5x5, 7x7, 9x9 - 40 - downsample ###### expand 6 = 192 # expand [convolutional] filters=192 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [route] layers = -1 group_id=0 groups=4 [convolutional] batch_normalize=1 filters=48 groups=48 size=3 stride=2 pad=1 activation=swish [route] layers = -3 group_id=1 groups=4 [convolutional] batch_normalize=1 filters=48 groups=48 size=5 stride=2 pad=1 activation=swish [route] layers = -5 group_id=2 groups=4 [convolutional] batch_normalize=1 filters=48 groups=48 size=7 stride=2 pad=1 activation=swish [route] layers = -7 group_id=3 groups=4 [convolutional] batch_normalize=1 filters=48 groups=48 size=9 stride=2 pad=1 activation=swish [route] layers = -1,-3,-5,-7 #squeeze-n-excitation [avgpool] # squeeze ratio 0.5 [convolutional] filters=96 size=1 stride=1 activation=swish # excitation [convolutional] filters=192 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=40 size=1 stride=1 pad=1 batch_normalize=1 activation=linear # R1 ###################### ###### 'r3_k3.5_a1.1_p1.1_s11_e6_i40_o40_se0.5_sw', ###### 3x3, 5x5 - 40 ###### expand 6 = 240 # expand [convolutional] filters=240 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [route] layers = -1 group_id=0 groups=2 [convolutional] batch_normalize=1 filters=120 groups=120 size=3 stride=1 pad=1 activation=swish [route] layers = -3 group_id=1 groups=2 [convolutional] batch_normalize=1 filters=120 groups=120 size=5 stride=1 pad=1 activation=swish [route] layers = -1,-3 #squeeze-n-excitation [avgpool] # squeeze ratio 0.5 [convolutional] filters=120 size=1 stride=1 activation=swish # excitation [convolutional] filters=240 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=40 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=linear [shortcut] from=-12 activation=linear # R2 ###################### ###### 'r3_k3.5_a1.1_p1.1_s11_e6_i40_o40_se0.5_sw', ###### 3x3, 5x5 - 40 ###### expand 6 = 240 # expand [convolutional] filters=240 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [route] layers = -1 group_id=0 groups=2 [convolutional] batch_normalize=1 filters=120 groups=120 size=3 stride=1 pad=1 activation=swish [route] layers = -3 group_id=1 groups=2 [convolutional] batch_normalize=1 filters=120 groups=120 size=5 stride=1 pad=1 activation=swish [route] layers = -1,-3 #squeeze-n-excitation [avgpool] # squeeze ratio 0.5 [convolutional] filters=120 size=1 stride=1 activation=swish # excitation [convolutional] filters=240 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=40 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=linear [shortcut] from=-12 activation=linear # R3 ###################### ###### 'r3_k3.5_a1.1_p1.1_s11_e6_i40_o40_se0.5_sw', ###### 3x3, 5x5 - 40 ###### expand 6 = 240 # expand [convolutional] filters=240 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [route] layers = -1 group_id=0 groups=2 [convolutional] batch_normalize=1 filters=120 groups=120 size=3 stride=1 pad=1 activation=swish [route] layers = -3 group_id=1 groups=2 [convolutional] batch_normalize=1 filters=120 groups=120 size=5 stride=1 pad=1 activation=swish [route] layers = -1,-3 #squeeze-n-excitation [avgpool] # squeeze ratio 0.5 [convolutional] filters=120 size=1 stride=1 activation=swish # excitation [convolutional] filters=240 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=40 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=linear [shortcut] from=-12 activation=linear ###################### ###### 'r1_k3.5.7_a1_p1_s22_e6_i40_o80_se0.25_sw', ###### 3x3, 5x5, 7x7 - 80 - downsample ###### expand 6 = 240 # expand [convolutional] filters=240 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [route] layers = -1 group_id=0 groups=3 [convolutional] batch_normalize=1 filters=80 groups=80 size=3 stride=2 pad=1 activation=swish [route] layers = -3 group_id=1 groups=3 [convolutional] batch_normalize=1 filters=80 groups=80 size=5 stride=2 pad=1 activation=swish [route] layers = -5 group_id=2 groups=3 [convolutional] batch_normalize=1 filters=80 groups=80 size=7 stride=2 pad=1 activation=swish [route] layers = -1,-3,-5 #squeeze-n-excitation [avgpool] # squeeze ratio 0.25 [convolutional] filters=60 size=1 stride=1 activation=swish # excitation [convolutional] filters=240 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=80 size=1 stride=1 pad=1 batch_normalize=1 activation=linear # R1 ###################### ###### 'r3_k3.5.7.9_a1.1_p1.1_s11_e6_i80_o80_se0.25_sw', ###### 3x3, 5x5, 7x7, 9x9 - 80 ###### expand 6 = 480 # expand [convolutional] filters=480 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [route] layers = -1 group_id=0 groups=4 [convolutional] batch_normalize=1 filters=120 groups=120 size=3 stride=1 pad=1 activation=swish [route] layers = -3 group_id=1 groups=4 [convolutional] batch_normalize=1 filters=120 groups=120 size=5 stride=1 pad=1 activation=swish [route] layers = -5 group_id=2 groups=4 [convolutional] batch_normalize=1 filters=120 groups=120 size=7 stride=1 pad=1 activation=swish [route] layers = -7 group_id=3 groups=4 [convolutional] batch_normalize=1 filters=120 groups=120 size=9 stride=1 pad=1 activation=swish [route] layers = -1,-3,-5,-7 #squeeze-n-excitation [avgpool] # squeeze ratio 0.25 [convolutional] filters=120 size=1 stride=1 activation=swish # excitation [convolutional] filters=480 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=80 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=linear [shortcut] from=-16 activation=linear # R2 ###################### ###### 'r3_k3.5.7.9_a1.1_p1.1_s11_e6_i80_o80_se0.25_sw', ###### 3x3, 5x5, 7x7, 9x9 - 80 ###### expand 6 = 480 # expand [convolutional] filters=480 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [route] layers = -1 group_id=0 groups=4 [convolutional] batch_normalize=1 filters=120 groups=120 size=3 stride=1 pad=1 activation=swish [route] layers = -3 group_id=1 groups=4 [convolutional] batch_normalize=1 filters=120 groups=120 size=5 stride=1 pad=1 activation=swish [route] layers = -5 group_id=2 groups=4 [convolutional] batch_normalize=1 filters=120 groups=120 size=7 stride=1 pad=1 activation=swish [route] layers = -7 group_id=3 groups=4 [convolutional] batch_normalize=1 filters=120 groups=120 size=9 stride=1 pad=1 activation=swish [route] layers = -1,-3,-5,-7 #squeeze-n-excitation [avgpool] # squeeze ratio 0.25 [convolutional] filters=120 size=1 stride=1 activation=swish # excitation [convolutional] filters=480 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=80 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=linear [shortcut] from=-16 activation=linear # R3 ###################### ###### 'r3_k3.5.7.9_a1.1_p1.1_s11_e6_i80_o80_se0.25_sw', ###### 3x3, 5x5, 7x7, 9x9 - 80 ###### expand 6 = 480 # expand [convolutional] filters=480 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [route] layers = -1 group_id=0 groups=4 [convolutional] batch_normalize=1 filters=120 groups=120 size=3 stride=1 pad=1 activation=swish [route] layers = -3 group_id=1 groups=4 [convolutional] batch_normalize=1 filters=120 groups=120 size=5 stride=1 pad=1 activation=swish [route] layers = -5 group_id=2 groups=4 [convolutional] batch_normalize=1 filters=120 groups=120 size=7 stride=1 pad=1 activation=swish [route] layers = -7 group_id=3 groups=4 [convolutional] batch_normalize=1 filters=120 groups=120 size=9 stride=1 pad=1 activation=swish [route] layers = -1,-3,-5,-7 #squeeze-n-excitation [avgpool] # squeeze ratio 0.25 [convolutional] filters=120 size=1 stride=1 activation=swish # excitation [convolutional] filters=480 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=80 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=linear [shortcut] from=-16 activation=linear ###################### ###### 'r1_k3_a1_p1_s11_e6_i80_o120_se0.5_sw', ###### 3x3 - 120 ###### expand 6 = 480 # expand [convolutional] filters=480 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [convolutional] filters=480 groups=480 size=3 stride=1 pad=1 batch_normalize=1 activation=swish #squeeze-n-excitation [avgpool] # squeeze ratio 0.5 [convolutional] filters=240 size=1 stride=1 activation=swish # excitation [convolutional] filters=480 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=120 size=1 stride=1 pad=1 batch_normalize=1 activation=linear [shortcut] from=-8 activation=linear # R1 ###################### ###### 'r3_k3.5.7.9_a1.1_p1.1_s11_e3_i120_o120_se0.5_sw', ###### 3x3, 5x5, 7x7, 9x9 - 120 ###### expand 6 = 720 # expand [convolutional] filters=720 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [route] layers = -1 group_id=0 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=3 stride=1 pad=1 activation=swish [route] layers = -3 group_id=1 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=5 stride=1 pad=1 activation=swish [route] layers = -5 group_id=2 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=7 stride=1 pad=1 activation=swish [route] layers = -7 group_id=3 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=9 stride=1 pad=1 activation=swish [route] layers = -1,-3,-5,-7 #squeeze-n-excitation [avgpool] # squeeze ratio 0.5 [convolutional] filters=360 size=1 stride=1 activation=swish # excitation [convolutional] filters=720 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=120 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=linear [shortcut] from=-16 activation=linear # R2 ###################### ###### 'r3_k3.5.7.9_a1.1_p1.1_s11_e3_i120_o120_se0.5_sw', ###### 3x3, 5x5, 7x7, 9x9 - 120 ###### expand 6 = 720 # expand [convolutional] filters=720 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [route] layers = -1 group_id=0 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=3 stride=1 pad=1 activation=swish [route] layers = -3 group_id=1 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=5 stride=1 pad=1 activation=swish [route] layers = -5 group_id=2 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=7 stride=1 pad=1 activation=swish [route] layers = -7 group_id=3 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=9 stride=1 pad=1 activation=swish [route] layers = -1,-3,-5,-7 #squeeze-n-excitation [avgpool] # squeeze ratio 0.5 [convolutional] filters=360 size=1 stride=1 activation=swish # excitation [convolutional] filters=720 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=120 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=linear [shortcut] from=-16 activation=linear # R3 ###################### ###### 'r3_k3.5.7.9_a1.1_p1.1_s11_e3_i120_o120_se0.5_sw', ###### 3x3, 5x5, 7x7, 9x9 - 120 ###### expand 6 = 720 # expand [convolutional] filters=720 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [route] layers = -1 group_id=0 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=3 stride=1 pad=1 activation=swish [route] layers = -3 group_id=1 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=5 stride=1 pad=1 activation=swish [route] layers = -5 group_id=2 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=7 stride=1 pad=1 activation=swish [route] layers = -7 group_id=3 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=9 stride=1 pad=1 activation=swish [route] layers = -1,-3,-5,-7 #squeeze-n-excitation [avgpool] # squeeze ratio 0.5 [convolutional] filters=360 size=1 stride=1 activation=swish # excitation [convolutional] filters=720 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=120 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=linear [shortcut] from=-16 activation=linear ###################### ###### 'r1_k3.5.7.9_a1_p1_s22_e6_i120_o200_se0.5_sw', ###### 3x3, 5x5, 7x7, 9x9 - 200 - downsample ###### expand 6 = 720 # expand [convolutional] filters=720 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [route] layers = -1 group_id=0 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=3 stride=2 pad=1 activation=swish [route] layers = -3 group_id=1 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=5 stride=2 pad=1 activation=swish [route] layers = -5 group_id=2 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=7 stride=2 pad=1 activation=swish [route] layers = -7 group_id=3 groups=4 [convolutional] batch_normalize=1 filters=180 groups=180 size=9 stride=2 pad=1 activation=swish [route] layers = -1,-3,-5,-7 #squeeze-n-excitation [avgpool] # squeeze ratio 0.5 [convolutional] filters=360 size=1 stride=1 activation=swish # excitation [convolutional] filters=720 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=200 size=1 stride=1 pad=1 batch_normalize=1 activation=linear # R1 ###################### ###### 'r3_k3.5.7.9_a1_p1.1_s11_e6_i200_o200_se0.5_sw', ###### 3x3, 5x5, 7x7, 9x9 - 200 ###### expand 6 = 1200 # expand [convolutional] filters=1200 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [route] layers = -1 group_id=0 groups=4 [convolutional] batch_normalize=1 filters=300 groups=300 size=3 stride=1 pad=1 activation=swish [route] layers = -3 group_id=1 groups=4 [convolutional] batch_normalize=1 filters=300 groups=300 size=5 stride=1 pad=1 activation=swish [route] layers = -5 group_id=2 groups=4 [convolutional] batch_normalize=1 filters=300 groups=300 size=7 stride=1 pad=1 activation=swish [route] layers = -7 group_id=3 groups=4 [convolutional] batch_normalize=1 filters=300 groups=300 size=9 stride=1 pad=1 activation=swish [route] layers = -1,-3,-5,-7 #squeeze-n-excitation [avgpool] # squeeze ratio 0.5 [convolutional] filters=600 size=1 stride=1 activation=swish # excitation [convolutional] filters=1200 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=200 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=linear [shortcut] from=-16 activation=linear # R2 ###################### ###### 'r3_k3.5.7.9_a1_p1.1_s11_e6_i200_o200_se0.5_sw', ###### 3x3, 5x5, 7x7, 9x9 - 200 ###### expand 6 = 1200 # expand [convolutional] filters=1200 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [route] layers = -1 group_id=0 groups=4 [convolutional] batch_normalize=1 filters=300 groups=300 size=3 stride=1 pad=1 activation=swish [route] layers = -3 group_id=1 groups=4 [convolutional] batch_normalize=1 filters=300 groups=300 size=5 stride=1 pad=1 activation=swish [route] layers = -5 group_id=2 groups=4 [convolutional] batch_normalize=1 filters=300 groups=300 size=7 stride=1 pad=1 activation=swish [route] layers = -7 group_id=3 groups=4 [convolutional] batch_normalize=1 filters=300 groups=300 size=9 stride=1 pad=1 activation=swish [route] layers = -1,-3,-5,-7 #squeeze-n-excitation [avgpool] # squeeze ratio 0.5 [convolutional] filters=600 size=1 stride=1 activation=swish # excitation [convolutional] filters=1200 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=200 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=linear [shortcut] from=-16 activation=linear # R3 ###################### ###### 'r3_k3.5.7.9_a1_p1.1_s11_e6_i200_o200_se0.5_sw', ###### 3x3, 5x5, 7x7, 9x9 - 200 ###### expand 6 = 1200 # expand [convolutional] filters=1200 size=1 stride=1 pad=1 batch_normalize=1 activation=swish # dw [route] layers = -1 group_id=0 groups=4 [convolutional] batch_normalize=1 filters=300 groups=300 size=3 stride=1 pad=1 activation=swish [route] layers = -3 group_id=1 groups=4 [convolutional] batch_normalize=1 filters=300 groups=300 size=5 stride=1 pad=1 activation=swish [route] layers = -5 group_id=2 groups=4 [convolutional] batch_normalize=1 filters=300 groups=300 size=7 stride=1 pad=1 activation=swish [route] layers = -7 group_id=3 groups=4 [convolutional] batch_normalize=1 filters=300 groups=300 size=9 stride=1 pad=1 activation=swish [route] layers = -1,-3,-5,-7 #squeeze-n-excitation [avgpool] # squeeze ratio 0.5 [convolutional] filters=600 size=1 stride=1 activation=swish # excitation [convolutional] filters=1200 size=1 stride=1 activation=logistic # multiply channels [scale_channels] from=-4 # out [convolutional] filters=200 groups=2 size=1 stride=1 pad=1 batch_normalize=1 activation=linear [shortcut] from=-16 activation=linear # _conv_head [convolutional] filters=1536 size=1 stride=1 pad=1 batch_normalize=1 activation=swish [avgpool] [dropout] probability=.25 [convolutional] filters=1000 size=1 stride=1 pad=0 activation=linear [softmax] groups=1 #[cost] #type=sse