use millions of colors

dataloader.py

@@ -4,8 +4,8 @@ from torch.utils.data import DataLoader, TensorDataset
 from utils import extract_colors, preprocess_data
 
 
-def create_dataloader(N: int = 50, **kwargs):
-    rgb_tensor, _ = extract_colors()
+def create_dataloader(N: int = 1e8, **kwargs):
+    rgb_tensor = torch.rand((int(N), 3), dtype=torch.float32)
     rgb_tensor = preprocess_data(rgb_tensor)
     # Creating a dataset and data loader
     dataset = TensorDataset(rgb_tensor, torch.zeros(len(rgb_tensor)))
@@ -36,7 +36,8 @@ def create_named_dataloader(N: int = 0, **kwargs):
 
 if __name__ == "__main__":
     batch_size = 4
-    train_dataloader = create_dataloader(batch_size=batch_size, shuffle=True)
+    train_dataloader = create_dataloader(N=1e6, batch_size=batch_size, shuffle=True)
+    print(len(train_dataloader.dataset))
     train_dataloader_with_names = create_named_dataloader(
         batch_size=batch_size, shuffle=True
     )

experiments.csv

@@ -1,25 +0,0 @@
-,batch_size,alpha,learning_rate
-0,32.0,0.3,0.0001
-1,32.0,0.3,0.01
-2,32.0,0.9,1e-06
-3,32.0,0.7,0.001
-4,64.0,0.5,0.001
-5,64.0,0.1,1e-06
-6,32.0,0.1,0.001
-7,128.0,0.5,1e-06
-8,128.0,0.7,0.001
-9,128.0,0.9,1e-05
-10,128.0,0.1,1e-06
-11,128.0,0.3,1e-06
-12,64.0,0.3,0.01
-13,64.0,0.1,1e-06
-14,128.0,0.5,0.001
-15,32.0,0.3,1e-05
-16,32.0,0.7,1e-06
-17,32.0,0.3,1e-06
-18,64.0,0.3,0.0001
-19,64.0,0.3,1e-06
-20,128.0,0.5,1e-05
-21,32.0,0.1,0.01
-22,64.0,0.1,1e-05
-23,64.0,0.3,0.001

main.py

@@ -7,7 +7,7 @@ import torch
 from pytorch_lightning.callbacks import EarlyStopping
 
 from callbacks import SaveImageCallback
-from dataloader import create_named_dataloader
+from dataloader import create_dataloader
 from model import ColorTransformerModel
 
 
@@ -65,20 +65,20 @@ if __name__ == "__main__":
     early_stop_callback = EarlyStopping(
         monitor="hp_metric",  # Metric to monitor
         min_delta=1e-5,  # Minimum change in the monitored quantity to qualify as an improvement
-        patience=24,  # Number of epochs with no improvement after which training will be stopped
+        patience=5,  # Number of epochs with no improvement after which training will be stopped
         mode="min",  # Mode can be either 'min' for minimizing the monitored quantity or 'max' for maximizing it.
         verbose=True,
     )
 
     save_img_callback = SaveImageCallback(
-        save_interval=0,
+        save_interval=1,
         final_dir="out",
     )
 
     # Initialize data loader with parsed arguments
     # named_data_loader also has grayscale extras. TODO: remove unnamed
-    train_dataloader = create_named_dataloader(
-        N=0,
+    train_dataloader = create_dataloader(
+        N=1e8,
         batch_size=args.bs,
         shuffle=True,
         num_workers=args.num_workers,

model.py

@@ -122,7 +122,7 @@ class ColorTransformerModel(pl.LightningModule):
             lr=self.hparams.learning_rate,
         )
         lr_scheduler = ReduceLROnPlateau(
-            optimizer, mode="min", factor=0.05, patience=10, cooldown=20, verbose=True
+            optimizer, mode="min", factor=0.05, patience=5, cooldown=10, verbose=True
         )
         return {
             "optimizer": optimizer,

search.py

@@ -24,7 +24,7 @@ alpha_values = [0, 1, 2]
 widths = [64, 128, 256, 512]
 # learning_rate_values = [5e-4]
 batch_size_values = [32, 64, 128]
-max_epochs_values = [500]
+max_epochs_values = [50]
 seeds = list(range(21, 1992))
 
 # Generate all possible combinations of hyperparameters