recreate iris looking result. weird though.

losses.py

@@ -40,10 +40,12 @@ def preservation_loss(inputs, outputs, target_inputs=None, target_outputs=None):
     # print(rgb_norm)
 
     # Calculate 1D Space Norm (modulo 1 to account for circularity)
-    transformed_norm = circle_norm(outputs, target_outputs) * 2
+    transformed_norm = circle_norm(outputs, target_outputs)  # * 2
 
     diff = torch.pow(rgb_norm - transformed_norm, 2)
-    N = torch.count_nonzero(rgb_norm)
+    N = len(outputs)
+    N = (N * (N - 1)) / 2
+    # N = torch.count_nonzero(rgb_norm)
     return torch.sum(diff) / N
 
 

model.py

@@ -31,7 +31,8 @@ class ColorTransformerModel(L.LightningModule):
         d = self.hparams.depth
         bias = self.hparams.bias
         if self.hparams.loop:
-            midlayers = [nn.Linear(w, w, bias=bias), t()] * d
+            midlayers = []
+            midlayers += [nn.Linear(w, w, bias=bias), t()] * d
         else:
             midlayers = sum(
                 [
@@ -64,8 +65,8 @@ class ColorTransformerModel(L.LightningModule):
         p_loss = preservation_loss(
             inputs,
             outputs,
-            target_inputs=rgb_tensor,
+            # target_inputs=rgb_tensor,
-            target_outputs=self.forward(rgb_tensor),
+            # target_outputs=self.forward(rgb_tensor),
         )
         alpha = self.hparams.alpha
 

newsearch.py

@@ -27,19 +27,19 @@ learning_rate_values = [1e-3]
 # learning_rate_values = [5e-4]
 
 # alpha_values = [0, .25, 0.5, 0.75, 1]  # alpha = 0 is unsupervised. alpha = 1 is supervised.
-alpha_values = [0.9]
+alpha_values = [0]
 # widths = [2**k for k in range(4, 13)]
 # depths = [1, 2, 4, 8, 16]
-widths, depths = [512], [8]
+widths, depths = [512], [4]
 
 batch_size_values = [256]
 max_epochs_values = [100]
 seeds = list(range(21, 1992))
 optimizers = [
     # "Adagrad",
-    # "Adam",
+    "Adam",
     # "SGD",
-    "AdamW",
+    # "AdamW",
     # "LBFGS",
     # "RAdam",
     # "RMSprop",
@@ -80,7 +80,7 @@ python newmain.py fit \
 --model.depth {d} \
 --model.bias true \
 --model.loop true \
---model.transform relu \
+--model.transform tanh \
 --trainer.min_epochs 10 \
 --trainer.max_epochs {me} \
 --trainer.log_every_n_steps 3 \