3 channel image restoration about pixelrl HOT 1 OPEN

If you want to apply different actions on the 3 channels, I think the easiest way is to make different output layers for each channel.
For example, in init function in denoise/MyFCN.py:
conv7_pi_r=chainerrl.policies.SoftmaxPolicy(L.Convolution2D( 64, n_actions, 3, stride=1, pad=1, nobias=False, initialW=w)),
conv7_pi_g=chainerrl.policies.SoftmaxPolicy(L.Convolution2D( 64, n_actions, 3, stride=1, pad=1, nobias=False, initialW=w)),
conv7_pi_b=chainerrl.policies.SoftmaxPolicy(L.Convolution2D( 64, n_actions, 3, stride=1, pad=1, nobias=False, initialW=w)),
conv7_V_r=L.Convolution2D( 64, 1, 3, stride=1, pad=1, nobias=False, initialW=net.layer18.W.data, initial_bias=net.layer18.b.data),
conv7_V_g=L.Convolution2D( 64, 1, 3, stride=1, pad=1, nobias=False, initialW=net.layer18.W.data, initial_bias=net.layer18.b.data),
conv7_V_b=L.Convolution2D( 64, 1, 3, stride=1, pad=1, nobias=False, initialW=net.layer18.W.data, initial_bias=net.layer18.b.data),
and in pi_and_v function:
pout_r = self.conv7_pi_r(h_pi)
pout_g = self.conv7_pi_g(h_pi)
pout_b = self.conv7_pi_b(h_pi)
vout_r = self.conv7_V_r(h_V)
vout_g = self.conv7_V_g(h_V)
vout_b = self.conv7_V_b(h_V)
return pout_r, pout_g, pout_b, vout_r, vout_g, vout_b

Then, modify the pixelwise_a3c.py. You can independently treat pout_r, pout_g, and pout_b in the same manner as pout. The pi_loss is replaced with the sum of pi_loss_r, pi_loss_g, and pi_loss_b. Similarly, you can treat vout_r, vout_g, and vout_b. In this way, you don't have to modify the SoftmaxPolicy layer.

from pixelrl.