temp valiadtion based on new stock + isolated seq

Author: Axle-Bucamp

hygdra_forecasting/utils/learning_rate_sheduler.py

@@ -51,7 +51,7 @@ def lrfn(self, epoch):
 
         # return ....
         lr_ramp_ep = 3 # int(0.02 * epochs)  # 30% of epochs for warm-up
-        lr_sus_ep = max(0, int(0.3 * self.epochs) - lr_ramp_ep)
+        lr_sus_ep = max(0, int(0.15 * self.epochs) - lr_ramp_ep)
         if epoch < lr_ramp_ep:  # Warm-up phase
             lr = (self.lr_max - self.lr_start) / lr_ramp_ep * epoch + self.lr_start
         elif epoch < lr_ramp_ep + lr_sus_ep:  # Sustain phase at max learning rate

train.py

@@ -7,6 +7,7 @@
     from hygdra_forecasting.dataloader.dataloader import StockDataset
     from torch import device, cuda
     from torch.utils.data import DataLoader
+    import numpy as np
 
     if cuda.is_available():
         device = device('cuda:0')
@@ -29,11 +30,30 @@
     dataset_val = StockDataset(ticker=tickers_val)
     dataloader_val = DataLoader(dataset_val, batch_size=256, shuffle=True, num_workers=1)
 
+    # temp (non distinct loss and balance) seq val on known stock
+    lenval = len(dataset_val)
+    indval = len(dataset) // 2  # Select half the dataset
+
+    # Ensure index bounds are valid 
+    if indval > 0:
+        # Select random indices without replacement
+        random_indices = np.random.choice(len(dataset), indval, replace=False)
+        
+        # Move selected data to dataset_val
+        dataset_val.data = np.concatenate((dataset_val.data, dataset.data[random_indices].copy()), axis=0)
+        dataset_val.label = np.concatenate((dataset_val.label, dataset.label[random_indices].copy()), axis=0)
+
+        # Remove selected indices from dataset
+        mask = np.ones(len(dataset), dtype=bool)
+        mask[random_indices] = False
+
+        dataset.data = dataset.data[mask]
+        dataset.label = dataset.label[mask]
+
     # Initialize your model
     input_sample, _ = dataset.__getitem__(0)
     setup_seed(20) # test liquid ? # check shaping and batch computation based on Dataset
     model = ConvCausalLTSM(input_shape=input_sample.shape)
     # LtsmAttentionforecastPred(input_shape=input_sample.shape)
-    # ConvCausalLTSM(input_shape=input_sample.shape)
     model = train_model(model, dataloader, dataloader_val, epochs=100, learning_rate=0.01, lrfn=CosineWarmup(0.01, 100).lrfn)
 

train_graph.py

@@ -7,6 +7,7 @@
     from torch.utils.data import DataLoader
     from hygdra_forecasting.utils.learning_rate_sheduler import CosineWarmup
     from torch import cuda, device, load, nn
+    import numpy as np
 
     if cuda.is_available():
         device = device('cuda:0')
@@ -30,6 +31,26 @@
     dataset_val = StockGraphDataset(ticker=tickers_val, indics=TICKERS_ETF)
     dataloader_val = DataLoader(dataset_val, batch_size=32, shuffle=True, num_workers=1)
 
+    # temp (non distinct loss and balance) seq val on known stock
+    lenval = len(dataset_val)
+    indval = len(dataset) // 2  # Select half the dataset
+
+    # Ensure index bounds are valid 
+    if indval > 0:
+        # Select random indices without replacement
+        random_indices = np.random.choice(len(dataset), indval, replace=False)
+        
+        # Move selected data to dataset_val
+        dataset_val.data = np.concatenate((dataset_val.data, dataset.data[random_indices].copy()), axis=0)
+        dataset_val.label = np.concatenate((dataset_val.label, dataset.label[random_indices].copy()), axis=0)
+
+        # Remove selected indices from dataset
+        mask = np.ones(len(dataset), dtype=bool)
+        mask[random_indices] = False
+
+        dataset.data = dataset.data[mask]
+        dataset.label = dataset.label[mask]
+
     # Initialize your model
     input_sample, _ = dataset.__getitem__(0)
     setup_seed(20)