Merge remote-tracking branch 'origin/master'

# Conflicts:
#	models/transformer_model.py
#	multi_run.py

_parameters.ini

@@ -31,7 +31,7 @@ activation = gelu
 use_bias = True
 use_norm = True
 use_residual = True
-dropout = 0.2
+dropout = 0.21
 
 lat_dim = 32
 patch_size = 8
@@ -48,7 +48,7 @@ use_norm = True
 dropout = 0.2
 lat_dim = 32
 features = 64
-filters = [16, 32, 64, 128]
+filters = [16, 32, 64]
 
 [VisualTransformer]
 weight_init = xavier_normal_

models/cnn_baseline.py

@@ -14,10 +14,12 @@ class CNNBaseline(CombinedModelMixins,
                   LightningBaseModule
                   ):
 
-    def __init__(self, in_shape, n_classes, weight_init, activation, use_bias, use_norm, dropout, lat_dim, features,
+    def __init__(self, in_shape, n_classes, weight_init, activation,
+                 use_bias, use_norm, dropout, lat_dim, features,
                  filters,
                  lr, weight_decay, sto_weight_avg, lr_warm_restart_epochs, opt_reset_interval,
-                 loss, scheduler):
+                 loss, scheduler, lr_scheduler_parameter
+                 ):
 
         # TODO: Move this to parent class, or make it much easieer to access....
         a = dict(locals())

models/transformer_model.py

@@ -21,7 +21,7 @@ class VisualTransformer(CombinedModelMixins,
                         ):
 
     def __init__(self, in_shape, n_classes, weight_init, activation,
-                 embedding_size, heads, attn_depth, patch_size, use_residual,
+                 embedding_size, heads, attn_depth, patch_size, use_residual, variable_length,
                  use_bias, use_norm, dropout, lat_dim, loss, scheduler, mlp_dim, head_dim,
                  lr, weight_decay, sto_weight_avg, lr_scheduler_parameter, opt_reset_interval):
 
@@ -75,7 +75,7 @@ class VisualTransformer(CombinedModelMixins,
             nn.Linear(self.embed_dim, self.params.lat_dim),
             nn.GELU(),
             nn.Dropout(self.params.dropout),
-            nn.Linear(self.params.lat_dim, self.params.n_classes),
+            nn.Linear(self.params.lat_dim, n_classes),
             nn.Softmax()
         )
 
@@ -91,12 +91,15 @@ class VisualTransformer(CombinedModelMixins,
         tensor = rearrange(tensor, 'b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1=p, p2=p)
         b, n, _ = tensor.shape
 
-        # mask
-        lengths = torch.count_nonzero(tensor, dim=-1)
-        mask = (lengths == torch.zeros_like(lengths))
-        # CLS-token awareness
-        # mask = torch.cat((torch.zeros(b, 1), mask), dim=-1)
-        # mask = repeat(mask, 'b n -> b h n', h=self.params.heads)
+        if self.params.variable_length and mask is None:
+            # mask
+            lengths = torch.count_nonzero(tensor, dim=-1)
+            mask = (lengths == torch.zeros_like(lengths))
+            # CLS-token awareness
+            # mask = torch.cat((torch.zeros(b, 1), mask), dim=-1)
+            # mask = repeat(mask, 'b n -> b h n', h=self.params.heads)
+        else:
+            mask = mask
 
         tensor = self.patch_to_embedding(tensor)
 

util/module_mixins.py

@@ -152,7 +152,7 @@ class TestMixin:
             class_names = {val: key for val, key in ['negative', 'positive']}
 
         df = pd.DataFrame(data=dict(filename=[Path(x).name for x in sorted_y.keys()],
-                                    prediction=y_max.cpu().numpy()))
+                                    prediction=[class_names[x.item()] for x in y_max.cpu()]))
         result_file = Path(self.logger.log_dir / 'predictions.csv')
         if result_file.exists():
             try:

util/optimizer_mixin.py

@@ -26,7 +26,7 @@ class OptimizerMixin:
         optimizer_dict.update(optimizer=optimizer)
 
         if self.params.scheduler == CosineAnnealingWarmRestarts.__name__:
-            scheduler = CosineAnnealingWarmRestarts(optimizer, self.params.lr_scheduler_parameter)
+            scheduler = CosineAnnealingWarmRestarts(optimizer, T_0=self.params.lr_scheduler_parameter)
         elif self.params.scheduler == LambdaLR.__name__:
             lr_reduce_ratio = self.params.lr_scheduler_parameter
             scheduler = LambdaLR(optimizer, lr_lambda=lambda epoch: lr_reduce_ratio ** epoch)