update to new temperature approach

AnimationTransformer.py

@@ -1,5 +1,6 @@
 import math
 import time
+import random
 
 import torch
 import torch.nn as nn
@@ -170,33 +171,39 @@ def fit(model, optimizer, loss_function, train_dataloader, val_dataloader, epoch
     return train_loss_list, validation_loss_list
 
 
-def predict(model, source_sequence, sos_token: torch.Tensor, device, max_length=32, eos_scaling=1, backpropagate=False, showResult= True):
+def predict(model, source_sequence, sos_token: torch.Tensor, device, max_length=32, eos_scaling=1, backpropagate=False, showResult= True, temperature=1):
     if backpropagate:
         model.train()
     else:
-        model.eval()
+        model.eval()   
 
     source_sequence = source_sequence.float().to(device)
     y_input = torch.unsqueeze(sos_token, dim=0).float().to(device)
-
+    #print(source_sequence, source_sequence.unsqueeze(0))
     i = 0
     while i < max_length:
         # Get source mask
+        #print(y_input, y_input.unsqueeze(0))
         prediction = model(source_sequence.unsqueeze(0), y_input.unsqueeze(0),  # un-squeeze for batch
                            # tgt_mask=get_tgt_mask(y_input.size(0)).to(device),
                            src_key_padding_mask=create_pad_mask(source_sequence.unsqueeze(0)).to(device))
-
         next_embedding = prediction[0, -1, :]  # prediction on last token
+        
         pred_deep_svg, pred_type, pred_parameters = dataset_helper.unpack_embedding(next_embedding, dim=0)
         #print(pred_deep_svg, pred_type, pred_parameters)
         pred_deep_svg, pred_type, pred_parameters = pred_deep_svg.to(device), pred_type.to(device), pred_parameters.to(
             device)
 
+        pred_type = pred_type / temperature
+
         # === TYPE ===
         # Apply Softmax
         type_softmax = torch.softmax(pred_type, dim=0)
         type_softmax[0] = type_softmax[0] * eos_scaling  # Reduce EOS
-        animation_type = torch.argmax(type_softmax, dim=0)
+
+        indices = torch.argsort(type_softmax, descending=True)
+        animation_type = random.choice(indices[:3])
+        #animation_type = torch.argmax(type_softmax, dim=0)
 
         # Break if EOS is most likely
         if animation_type == 0:
@@ -222,6 +229,7 @@ def predict(model, source_sequence, sos_token: torch.Tensor, device, max_length=
 
         # === SEQUENCE ===
         y_new = torch.concat([closest_token[:-26], pred_type.to(device), pred_parameters], dim=0)
+        #y_new = torch.concat([pred_deep_svg, pred_type.to(device), pred_parameters], dim=0)
         y_input = torch.cat((y_input, y_new.unsqueeze(0)), dim=0)
 
         # === INFO PRINT ===

animationPipeline.py

@@ -15,10 +15,10 @@ def animateLogo(path : str, targetPath : str):
     except Exception as e:
         print(f"An error occurred: {e}")
     #transformer
-    NUM_HEADS = 6 # Dividers of 282: {1, 2, 3, 6, 47, 94, 141, 282}
-    NUM_ENCODER_LAYERS = 2
-    NUM_DECODER_LAYERS = 8
-    DROPOUT=0.1
+    NUM_HEADS = 47 # Dividers of 282: {1, 2, 3, 6, 47, 94, 141, 282}
+    NUM_ENCODER_LAYERS = 6
+    NUM_DECODER_LAYERS = 4
+    DROPOUT=0.21
     # CONSTANTS
     FEATURE_DIM = 282
 
@@ -34,7 +34,7 @@ def animateLogo(path : str, targetPath : str):
         use_positional_encoder=True
     ).to(device)
 
-    model.load_state_dict(torch.load("models/animation_transformer.pth", map_location=torch.device('cpu')), strict=False)
+    model.load_state_dict(torch.load("models/animation_transformer2.pth", map_location=torch.device('cpu')), strict=False)
 
     df = compute_embedding(path, "models/deepSVG_hierarchical_ordered.pth.tar")
     df = df.drop("animation_id", axis=1)
@@ -46,7 +46,7 @@ def animateLogo(path : str, targetPath : str):
 
     sos_token = torch.zeros(282)
     sos_token[256] = 1
-    result = predict(model, inp, sos_token=sos_token, device=device, max_length=inp.shape[0], eos_scaling=1)
+    result = predict(model, inp, sos_token=sos_token, device=device, max_length=inp.shape[0], eos_scaling=0.5, temperature=100)
     result = pd.DataFrame(result[1:, -26:].cpu().detach().numpy())
     result = pd.DataFrame({"model_output" : [row.tolist() for index, row in result.iterrows()]})
     result["animation_id"] = range(len(result))

models/{animation_transformer.pth → animation_transformer2.pth}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:12ae92d0b1a5ada8a8681122f76ea7c4e6b3fdf0169dd4b3a5d908899e563f86
-size 60658902
+oid sha256:e63638f545c6f925a1a6d31578d507834de8ed30b71db2a0762c86859c597c44
+size 69927679