Add gradient penalty visual debug

codes/trainer/ExtensibleTrainer.py

@@ -230,7 +230,12 @@ class ExtensibleTrainer(BaseModel):
             # Push the detached new state tensors into the state map for use with the next step.
             for k, v in new_states.items():
                 # State is immutable to reduce complexity. Overwriting existing state keys is not supported.
-                assert k not in state.keys()
+                class OverwrittenStateError(Exception):
+                    def __init__(self, k, keys):
+                        super().__init__(f'Attempted to overwrite state key: {k}.  The state should be considered '
+                                         f'immutable and keys should not be overwritten. Current keys: {keys}')
+                if k in state.keys():
+                    raise OverwrittenStateError(k, list(state.keys()))
                 state[k] = v
 
             if train_step:

codes/trainer/losses.py

@@ -72,6 +72,14 @@ class ConfigurableLoss(nn.Module):
     def forward(self, net, state):
         raise NotImplementedError
 
+    def is_stateful(self) -> bool:
+        """
+        Losses can inject into the state too. useful for when a loss computation can be used by another loss.
+        if this is true, the forward pass must return (loss, new_state). If false (the default), forward() only returns
+        the loss value.
+        """
+        return False
+
     def extra_metrics(self):
         return self.metrics
 
@@ -270,7 +278,9 @@ class DiscriminatorGanLoss(ConfigurableLoss):
         if self.gradient_penalty:
             [r.requires_grad_() for r in real]
         fake = extract_params_from_state(self.opt['fake'], state)
+        new_state = {}
         fake = [f.detach() for f in fake]
+        new_state = {}
         if self.noise:
             nreal = []
             nfake = []
@@ -313,12 +323,20 @@ class DiscriminatorGanLoss(ConfigurableLoss):
             # Apply gradient penalty. TODO: migrate this elsewhere.
             from models.stylegan.stylegan2_lucidrains import gradient_penalty
             assert len(real) == 1   # Grad penalty doesn't currently support multi-input discriminators.
-            gp = gradient_penalty(real[0], d_real)
+            gp, gp_structure = gradient_penalty(real[0], d_real, return_structured_grads=True)
             self.metrics.append(("gradient_penalty", gp.clone().detach()))
             loss = loss + gp
             self.metrics.append(("gradient_penalty", gp))
+            # The gp_structure is a useful visual debugging tool to see what areas of the generated image the disc is paying attention to.
+            gpimg = (gp_structure / (torch.std(gp_structure, dim=(-1, -2), keepdim=True) * 2)) \
+                              - torch.mean(gp_structure, dim=(-1, -2), keepdim=True) + .5
+            new_state['%s_%s_gp_structure_img' % (self.opt['fake'], self.opt['real'])] = gpimg
 
-        return loss
+        return loss, new_state
+
+    # This loss is stateful because it injects a debugging result from the GP term when enabled.
+    def is_stateful(self) -> bool:
+        return True
 
 
 # Computes a loss created by comparing the output of a generator to the output from the same generator when fed an

codes/trainer/steps.py

@@ -123,13 +123,10 @@ class ConfigurableStep(Module):
     # chunked tensors. Use grad_accum_step to dereference these steps. Should return a dict of tensors that later
     # steps might use. These tensors are automatically detached and accumulated into chunks.
     def do_forward_backward(self, state, grad_accum_step, amp_loss_id, train=True):
-        new_state = {}
+        local_state = {}  # <-- Will store the entire local state to be passed to injectors & losses.
-
+        new_state = {}  # <-- Will store state values created by this step for returning to ExtensibleTrainer.
-        # Prepare a de-chunked state dict which will be used for the injectors & losses.
-        local_state = {}
         for k, v in state.items():
             local_state[k] = v[grad_accum_step]
-        local_state.update(new_state)
         local_state['train_nets'] = str(self.get_networks_trained())
 
         # Some losses compute backward() internally. Accommodate this by stashing the amp_loss_id in env.
@@ -164,6 +161,11 @@ class ConfigurableStep(Module):
                    'before' in loss.opt.keys() and self.env['step'] > loss.opt['before'] or \
                    'every' in loss.opt.keys() and self.env['step'] % loss.opt['every'] != 0:
                     continue
+                if loss.is_stateful():
+                    l, lstate = loss(self.get_network_for_name(self.step_opt['training']), local_state)
+                    local_state.update(lstate)
+                    new_state.update(lstate)
+                else:
                     l = loss(self.get_network_for_name(self.step_opt['training']), local_state)
                 total_loss += l * self.weights[loss_name]
                 # Record metrics.