popov/gogo2
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

cleanup, CNN fixes

Browse Source
This commit is contained in:
Dobromir Popov
2025-07-05 00:12:40 +03:00
parent ce8c00a9d1
commit 5ca7493708
18 changed files with 587 additions and 5181 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
NN/models/cnn_model.py
View File

@ -329,13 +329,13 @@ class EnhancedCNNModel(nn.Module):
x = x.unsqueeze(0)
elif len(x.shape) > 3:
# Input has extra dimensions - flatten to [batch, seq, features]
x = x.view(x.shape[0], -1, x.shape[-1])
x = x.reshape(x.shape[0], -1, x.shape[-1])
x = self._memory_barrier(x) # Apply barrier after shape changes
batch_size, seq_len, features = x.shape
# Reshape for processing: [batch, seq, features] -> [batch*seq, features]
x_reshaped = x.view(-1, features)
x_reshaped = x.reshape(-1, features)
x_reshaped = self._memory_barrier(x_reshaped)
# Input embedding
@ -343,7 +343,7 @@ class EnhancedCNNModel(nn.Module):
embedded = self._memory_barrier(embedded)
# Reshape back for conv1d: [batch*seq, channels] -> [batch, channels, seq]
embedded = embedded.view(batch_size, seq_len, -1).transpose(1, 2).contiguous()
embedded = embedded.reshape(batch_size, seq_len, -1).transpose(1, 2).contiguous()
embedded = self._memory_barrier(embedded)
# Multi-scale feature extraction - ensure each path creates independent tensors
@ -380,10 +380,10 @@ class EnhancedCNNModel(nn.Module):
# Global aggregation - create independent tensors
avg_pooled = self.global_pool(attended_features)
avg_pooled = self._memory_barrier(avg_pooled.view(avg_pooled.shape[0], -1)) # Flatten instead of squeeze
avg_pooled = self._memory_barrier(avg_pooled.reshape(avg_pooled.shape[0], -1)) # Flatten instead of squeeze
max_pooled = self.global_max_pool(attended_features)
max_pooled = self._memory_barrier(max_pooled.view(max_pooled.shape[0], -1)) # Flatten instead of squeeze
max_pooled = self._memory_barrier(max_pooled.reshape(max_pooled.shape[0], -1)) # Flatten instead of squeeze
# Combine global features - create new tensor
global_features = torch.cat([avg_pooled, max_pooled], dim=1)
@ -399,7 +399,7 @@ class EnhancedCNNModel(nn.Module):
# Combine all features for final decision (8 regime classes + 1 volatility)
# Create completely independent tensors for concatenation
vol_pred_flat = self._memory_barrier(volatility_pred.view(volatility_pred.shape[0], -1)) # Flatten instead of squeeze
vol_pred_flat = self._memory_barrier(volatility_pred.reshape(volatility_pred.shape[0], -1)) # Flatten instead of squeeze
combined_features = torch.cat([processed_features, regime_probs, vol_pred_flat], dim=1)
combined_features = self._memory_barrier(combined_features)
@ -411,15 +411,15 @@ class EnhancedCNNModel(nn.Module):
trading_probs = self._memory_barrier(F.softmax(scaled_logits, dim=1))
# Flatten confidence to ensure consistent shape
confidence_flat = self._memory_barrier(confidence.view(confidence.shape[0], -1))
volatility_flat = self._memory_barrier(volatility_pred.view(volatility_pred.shape[0], -1))
confidence_flat = self._memory_barrier(confidence.reshape(confidence.shape[0], -1))
volatility_flat = self._memory_barrier(volatility_pred.reshape(volatility_pred.shape[0], -1))
return {
'logits': self._memory_barrier(trading_logits),
'probabilities': self._memory_barrier(trading_probs),
'confidence': confidence_flat[:, 0] if confidence_flat.shape[1] > 0 else confidence_flat.view(-1)[0],
'confidence': confidence_flat[:, 0] if confidence_flat.shape[1] > 0 else confidence_flat.reshape(-1)[0],
'regime': self._memory_barrier(regime_probs),
'volatility': volatility_flat[:, 0] if volatility_flat.shape[1] > 0 else volatility_flat.view(-1)[0],
'volatility': volatility_flat[:, 0] if volatility_flat.shape[1] > 0 else volatility_flat.reshape(-1)[0],
'features': self._memory_barrier(processed_features)
}