h_1, H) + (criterion_identity(h_2, H) * 0.5)
identity_loss = (Gl2m_id_loss + Gm2h_id_loss * 0.67) * 0.5
#total_loss
loss_G = identity_loss * 0.5 + G_loss + cycle_loss * 10
loss_G.backward()
optimizer_G.step()
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
optimizer_Dm.zero_grad()
#real
Dm_loss = criterion_GAN(Dm(M), target_real)
#fake
m_1 = ReplayBuffer().push_and_pop(m_1)
m_2 = ReplayBuffer().push_and_pop(m_2)
m_3 = ReplayBuffer().push_and_pop(m_3)
Dm_f_loss = criterion_GAN(Dm(m_1.detach()), target_fake)
Dm_f_loss += criterion_GAN(Dm(m_2.detach()), target_fake) * 0.5
Dm_f_loss += criterion_GAN(Dm(m_3.detach()), target_fake) * 0.25
# Total loss
loss_Dm = (Dm_f_loss + Dm_loss) * 0.5
loss_Dm.backward()
optimizer_Dm.step()
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
optimizer_Dh.zero_grad()
#real
Dh_loss = criterion_GAN(Dh(H), target_real)
#fake
h_1 = ReplayBuffer().push_and_pop(h_1)
h_2 = ReplayBuffer().push_and_pop(h_2)
G_loss = (Gc2d_loss + Gd2c_loss) * 0.5
#cycle_loss
cdc_cycle_loss = criterion_cycle(cr, c)
dcd_cycle_loss = criterion_cycle(dr, d)
cycle_loss = (dcd_cycle_loss + cdc_cycle_loss) * 0.5
#total_loss
loss_G = cycle_loss * 10 + G_loss + identity_loss * 0.5
loss_G.backward()
optimizer_G.step()
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
optimizer_Dc.zero_grad()
#real
Dc_c_loss = criterion_GAN(Dc(c), target_real)
#fake
cf = ReplayBuffer().push_and_pop(cf)
Dc_cf = Dc(cf.detach())
Dc_cf_loss = criterion_GAN(Dc_cf, target_fake)
# Total loss
loss_Dc = (Dc_c_loss + Dc_cf_loss) * 0.5
loss_Dc.backward()
optimizer_Dc.step()
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
optimizer_Dd.zero_grad()
#real
Dd_d_loss = criterion_GAN(Dd(d), target_real)
#fake
df = ReplayBuffer().push_and_pop(df)
Dd_df = Dd(df.detach())
Dd_df_loss = criterion_GAN(Dd_df, target_fake)
# Total loss