hi, I did an experiments on hair style transfer, but it seem don't work. about instagan

[leonardodora]

I think 1500 images for each class is okay.

The epoch really depends on the difficulty of the task, but I think 100 was okay to see the possibility.

thanks a lot! btw, your coding is perfect.

[sangwoomo]

Hi, is this task you want?
I thinks your your 1) real seg B are zero or 2) GAN loss is not working.

image source: https://medium.com/@crosssceneofwindff/instance-aware-image-to-image-translation-about-instagan-c508e63fe838

[leonardodora]

Hi, is this task you want?
I thinks your your 1) real seg B are zero or 2) GAN loss is not working.

image source: https://medium.com/@crosssceneofwindff/instance-aware-image-to-image-translation-about-instagan-c508e63fe838

yes, just like this. but i did experiments on real person. thanks for your really quick reply, i have checked real seg B, it is ok, just like real seg A

btw, i have modify this param from 2 to 1, is it the problem?

[leonardodora]

Hi, is this task you want?
I thinks your your 1) real seg B are zero or 2) GAN loss is not working.

image source: https://medium.com/@crosssceneofwindff/instance-aware-image-to-image-translation-about-instagan-c508e63fe838

btw, how can i check the gan loss?

epoch: 19 epoch_iter: 1620 loss: OrderedDict([('D_A', 0.28495314717292786), ('G_A', 0.44496801495552063), ('cyc_A', 1.4068127870559692), ('idt_A', 0.2712179124355316), ('ctx_A', 0.41984423995018005), ('D_B', 0.00030860648257657886), ('G_B', 1.0045173168182373), ('cyc_B', 0.3270185589790344), ('idt_B', 1.4243183135986328), ('ctx_B', 0.29253628849983215)])
epoch: 19 epoch_iter: 1640 loss: OrderedDict([('D_A', 0.21153371036052704), ('G_A', 0.224431112408638), ('cyc_A', 1.5483123064041138), ('idt_A', 0.2964574694633484), ('ctx_A', 0.19816333055496216), ('D_B', 0.001018411829136312), ('G_B', 0.9957813024520874), ('cyc_B', 0.3424926996231079), ('idt_B', 1.49729323387146), ('ctx_B', 0.20376181602478027)])
epoch: 19 epoch_iter: 1660 loss: OrderedDict([('D_A', 0.2148074209690094), ('G_A', 0.3643377125263214), ('cyc_A', 1.1582698822021484), ('idt_A', 0.4839262366294861), ('ctx_A', 0.19721737504005432), ('D_B', 0.0003931066603399813), ('G_B', 0.9965217709541321), ('cyc_B', 0.6051976680755615), ('idt_B', 1.0777069330215454), ('ctx_B', 0.5396640300750732)])

it seems like cyc_A, cyc_B have huge gap, is it responsible for the no working?

[leonardodora]

and i just noticed that the cyc_a loss would suddenly increase at epoch: 3 epoch_iter: 880
and always above 1.0 while cyc_b loss will decrease as expected
before the epoch: 3 epoch_iter: 880, the cyc_a loss seems to be fine

epoch: 3 epoch_iter: 760 loss: OrderedDict([('D_A', 0.28271806240081787), ('G_A', 0.5144460201263428), ('cyc_A', 0.9399356842041016), ('idt_A', 0.5795601606369019), ('ctx_A', 0.8255990743637085), ('D_B', 0.2468995451927185), ('G_B', 0.38095200061798096), ('cyc_B', 0.6753614544868469), ('idt_B', 0.6988549828529358), ('ctx_B', 0.29284632205963135)])
epoch: 3 epoch_iter: 780 loss: OrderedDict([('D_A', 0.2870528995990753), ('G_A', 0.2630198895931244), ('cyc_A', 0.8024196624755859), ('idt_A', 0.6109607219696045), ('ctx_A', 0.6542154550552368), ('D_B', 0.3109867572784424), ('G_B', 0.2686704993247986), ('cyc_B', 0.9172786474227905), ('idt_B', 0.6556259989738464), ('ctx_B', 0.5361669063568115)])
epoch: 3 epoch_iter: 800 loss: OrderedDict([('D_A', 0.24323898553848267), ('G_A', 0.34255099296569824), ('cyc_A', 0.8689618110656738), ('idt_A', 0.5951452851295471), ('ctx_A', 0.5810612440109253), ('D_B', 0.263603538274765), ('G_B', 0.1840197741985321), ('cyc_B', 0.8152450323104858), ('idt_B', 0.6198552846908569), ('ctx_B', 0.5905605554580688)])
epoch: 3 epoch_iter: 820 loss: OrderedDict([('D_A', 0.4089164137840271), ('G_A', 0.3948238790035248), ('cyc_A', 0.5638543367385864), ('idt_A', 0.3886997699737549), ('ctx_A', 0.3770771324634552), ('D_B', 0.6841108202934265), ('G_B', 1.0538008213043213), ('cyc_B', 0.49967172741889954), ('idt_B', 0.43776753544807434), ('ctx_B', 0.39757239818573)])
epoch: 3 epoch_iter: 840 loss: OrderedDict([('D_A', 0.24178935587406158), ('G_A', 0.3354800343513489), ('cyc_A', 0.5423353910446167), ('idt_A', 0.5768237709999084), ('ctx_A', 0.5036472082138062), ('D_B', 1.1370093822479248), ('G_B', 1.6951065063476562), ('cyc_B', 0.6251733899116516), ('idt_B', 0.48071300983428955), ('ctx_B', 0.4486328661441803)])
epoch: 3 epoch_iter: 860 loss: OrderedDict([('D_A', 0.2573254108428955), ('G_A', 0.17050953209400177), ('cyc_A', 2.20874285697937), ('idt_A', 0.664345920085907), ('ctx_A', 0.4304313659667969), ('D_B', 15.340181350708008), ('G_B', 13.359832763671875), ('cyc_B', 2.5855371952056885), ('idt_B', 2.1018173694610596), ('ctx_B', 0.7736623883247375)])
epoch: 3 epoch_iter: 880 loss: OrderedDict([('D_A', 0.2343287616968155), ('G_A', 0.22161865234375), ('cyc_A', 2.010279893875122), ('idt_A', 0.531023383140564), ('ctx_A', 0.3765827417373657), ('D_B', 1.0707379579544067), ('G_B', 0.46550095081329346), ('cyc_B', 2.1977295875549316), ('idt_B', 1.9820410013198853), ('ctx_B', 0.5809430480003357)])
epoch: 3 epoch_iter: 900 loss: OrderedDict([('D_A', 0.2524489760398865), ('G_A', 0.38755548000335693), ('cyc_A', 1.8325910568237305), ('idt_A', 0.7078045010566711), ('ctx_A', 0.4501146376132965), ('D_B', 0.2241455465555191), ('G_B', 0.5823928713798523), ('cyc_B', 2.1543965339660645), ('idt_B', 1.6840202808380127), ('ctx_B', 0.7103649377822876)])
epoch: 3 epoch_iter: 920 loss: OrderedDict([('D_A', 0.26689115166664124), ('G_A', 0.21848547458648682), ('cyc_A', 2.143144130706787), ('idt_A', 0.7535605430603027), ('ctx_A', 0.3836369216442108), ('D_B', 0.22803309559822083), ('G_B', 0.449783056974411), ('cyc_B', 2.3941097259521484), ('idt_B', 2.0672645568847656), ('ctx_B', 0.9657135009765625)])

[sangwoomo]

How much data are you using? I think the segmentation is collapsed in some reason, hence the cycle-consistency loss goes large. Does it work seem to work before epoch 3?

[leonardodora]

How much data are you using? I think the segmentation is collapsed in some reason, hence the cycle-consistency loss goes large. Does it work seem to work before epoch 3?

i use about 6000 images for each class. And I checked the seg is ok, but the class label is kind of mess,
so i change the project to long2short hair. the model is training, i use 1500 images for each class. could you please tell me about how many epoch it results in good(looks working) result

[sangwoomo]

I think 1500 images for each class is okay.

The epoch really depends on the difficulty of the task, but I think 100 was okay to see the possibility.

[leonardodora]

hi, I have opt my datasets and epoches reach to 100, but instagan seems don't work. (fake_A_seg is all 255. like this.

so I wonder how fine datasets I should feed to train this model
btw, I have trained on jeans2skirt_ccp and it works.

[sangwoomo]

I'm not sure why it happens. I think your task is harder than that I did.
I did not use any regularizers in my experiments, but maybe you need some additional regularizers to avoid the segmentation collapse.

[Jessicall]

@leonardodora Hi! Did your task work?I also want to use InstaGAN to experiment on hair transfer,but I
don't if this method works?