work with order execution - we are forced to do limit orders over the API
This commit is contained in:
Dobromir Popov
@ -807,6 +807,17 @@ class DQNAgent:
|
||||
if isinstance(expected_dim, tuple):
|
||||
expected_dim = np.prod(expected_dim)
|
||||
|
||||
# Debug: Check what dimensions we're actually seeing
|
||||
if sanitized_states:
|
||||
actual_dims = [len(state) for state in sanitized_states[:5]] # Check first 5
|
||||
logger.debug(f"DQN State dimensions - Expected: {expected_dim}, Actual samples: {actual_dims}")
|
||||
|
||||
# If all states have a consistent dimension different from expected, use that
|
||||
unique_dims = list(set(len(state) for state in sanitized_states))
|
||||
if len(unique_dims) == 1 and unique_dims[0] != expected_dim:
|
||||
logger.warning(f"All states have dimension {unique_dims[0]} but expected {expected_dim}. Using actual dimension.")
|
||||
expected_dim = unique_dims[0]
|
||||
|
||||
# Filter out states with wrong dimensions and fix them
|
||||
valid_states = []
|
||||
valid_next_states = []
|
||||
@ -1076,162 +1087,165 @@ class DQNAgent:
|
||||
# Zero gradients
|
||||
self.optimizer.zero_grad()
|
||||
|
||||
# Forward pass with amp autocasting
|
||||
with torch.cuda.amp.autocast():
|
||||
# Get current Q values and extrema predictions
|
||||
current_q_values, current_extrema_pred, current_price_pred, hidden_features, current_advanced_pred = self.policy_net(states)
|
||||
current_q_values = current_q_values.gather(1, actions.unsqueeze(1)).squeeze(1)
|
||||
|
||||
# Get next Q values from target network
|
||||
with torch.no_grad():
|
||||
next_q_values, next_extrema_pred, next_price_pred, next_hidden_features, next_advanced_pred = self.target_net(next_states)
|
||||
next_q_values = next_q_values.max(1)[0]
|
||||
# Forward pass with amp autocasting
|
||||
import warnings
|
||||
with warnings.catch_warnings():
|
||||
warnings.simplefilter("ignore", FutureWarning)
|
||||
with torch.cuda.amp.autocast():
|
||||
# Get current Q values and extrema predictions
|
||||
current_q_values, current_extrema_pred, current_price_pred, hidden_features, current_advanced_pred = self.policy_net(states)
|
||||
current_q_values = current_q_values.gather(1, actions.unsqueeze(1)).squeeze(1)
|
||||
|
||||
# Check for dimension mismatch and fix it
|
||||
if rewards.shape[0] != next_q_values.shape[0]:
|
||||
# Log the shape mismatch for debugging
|
||||
logger.warning(f"Shape mismatch detected: rewards {rewards.shape}, next_q_values {next_q_values.shape}")
|
||||
# Use the smaller size to prevent index errors
|
||||
min_size = min(rewards.shape[0], next_q_values.shape[0])
|
||||
rewards = rewards[:min_size]
|
||||
dones = dones[:min_size]
|
||||
next_q_values = next_q_values[:min_size]
|
||||
current_q_values = current_q_values[:min_size]
|
||||
|
||||
target_q_values = rewards + (1 - dones) * self.gamma * next_q_values
|
||||
|
||||
# Compute Q-value loss (primary task)
|
||||
q_loss = nn.MSELoss()(current_q_values, target_q_values)
|
||||
|
||||
# Initialize loss with q_loss
|
||||
loss = q_loss
|
||||
|
||||
# Try to extract price from current and next states
|
||||
try:
|
||||
# Extract price feature from sequence data (if available)
|
||||
if len(states.shape) == 3: # [batch, seq, features]
|
||||
current_prices = states[:, -1, -1] # Last timestep, last feature
|
||||
next_prices = next_states[:, -1, -1]
|
||||
else: # [batch, features]
|
||||
current_prices = states[:, -1] # Last feature
|
||||
next_prices = next_states[:, -1]
|
||||
|
||||
# Calculate price change for different timeframes
|
||||
immediate_changes = (next_prices - current_prices) / current_prices
|
||||
|
||||
# Get the actual batch size for this calculation
|
||||
actual_batch_size = states.shape[0]
|
||||
|
||||
# Create price direction labels - simplified for training
|
||||
# 0 = down, 1 = sideways, 2 = up
|
||||
immediate_labels = torch.ones(actual_batch_size, dtype=torch.long, device=self.device) * 1 # Default: sideways
|
||||
midterm_labels = torch.ones(actual_batch_size, dtype=torch.long, device=self.device) * 1
|
||||
longterm_labels = torch.ones(actual_batch_size, dtype=torch.long, device=self.device) * 1
|
||||
|
||||
# Immediate term direction (1s, 1m)
|
||||
immediate_up = (immediate_changes > 0.0005)
|
||||
immediate_down = (immediate_changes < -0.0005)
|
||||
immediate_labels[immediate_up] = 2 # Up
|
||||
immediate_labels[immediate_down] = 0 # Down
|
||||
|
||||
# For mid and long term, we can only approximate during training
|
||||
# In a real system, we'd need historical data to validate these
|
||||
# Here we'll use the immediate term with increasing thresholds as approximation
|
||||
|
||||
# Mid-term (1h) - use slightly higher threshold
|
||||
midterm_up = (immediate_changes > 0.001)
|
||||
midterm_down = (immediate_changes < -0.001)
|
||||
midterm_labels[midterm_up] = 2 # Up
|
||||
midterm_labels[midterm_down] = 0 # Down
|
||||
|
||||
# Long-term (1d) - use even higher threshold
|
||||
longterm_up = (immediate_changes > 0.002)
|
||||
longterm_down = (immediate_changes < -0.002)
|
||||
longterm_labels[longterm_up] = 2 # Up
|
||||
longterm_labels[longterm_down] = 0 # Down
|
||||
|
||||
# Generate target values for price change regression
|
||||
# For simplicity, we'll use the immediate change and scaled versions for longer timeframes
|
||||
price_value_targets = torch.zeros((actual_batch_size, 4), device=self.device)
|
||||
price_value_targets[:, 0] = immediate_changes
|
||||
price_value_targets[:, 1] = immediate_changes * 2.0 # Approximate 1h change
|
||||
price_value_targets[:, 2] = immediate_changes * 4.0 # Approximate 1d change
|
||||
price_value_targets[:, 3] = immediate_changes * 6.0 # Approximate 1w change
|
||||
|
||||
# Calculate loss for price direction prediction (classification)
|
||||
if len(current_price_pred['immediate'].shape) > 1 and current_price_pred['immediate'].shape[0] >= actual_batch_size:
|
||||
# Slice predictions to match the adjusted batch size
|
||||
immediate_pred = current_price_pred['immediate'][:actual_batch_size]
|
||||
midterm_pred = current_price_pred['midterm'][:actual_batch_size]
|
||||
longterm_pred = current_price_pred['longterm'][:actual_batch_size]
|
||||
price_values_pred = current_price_pred['values'][:actual_batch_size]
|
||||
# Get next Q values from target network
|
||||
with torch.no_grad():
|
||||
next_q_values, next_extrema_pred, next_price_pred, next_hidden_features, next_advanced_pred = self.target_net(next_states)
|
||||
next_q_values = next_q_values.max(1)[0]
|
||||
|
||||
# Compute losses for each task
|
||||
immediate_loss = nn.CrossEntropyLoss()(immediate_pred, immediate_labels)
|
||||
midterm_loss = nn.CrossEntropyLoss()(midterm_pred, midterm_labels)
|
||||
longterm_loss = nn.CrossEntropyLoss()(longterm_pred, longterm_labels)
|
||||
# Check for dimension mismatch and fix it
|
||||
if rewards.shape[0] != next_q_values.shape[0]:
|
||||
# Log the shape mismatch for debugging
|
||||
logger.warning(f"Shape mismatch detected: rewards {rewards.shape}, next_q_values {next_q_values.shape}")
|
||||
# Use the smaller size to prevent index errors
|
||||
min_size = min(rewards.shape[0], next_q_values.shape[0])
|
||||
rewards = rewards[:min_size]
|
||||
dones = dones[:min_size]
|
||||
next_q_values = next_q_values[:min_size]
|
||||
current_q_values = current_q_values[:min_size]
|
||||
|
||||
# MSE loss for price value regression
|
||||
price_value_loss = nn.MSELoss()(price_values_pred, price_value_targets)
|
||||
|
||||
# Combine all price prediction losses
|
||||
price_loss = immediate_loss + 0.7 * midterm_loss + 0.5 * longterm_loss + 0.3 * price_value_loss
|
||||
|
||||
# Create extrema labels (same as before)
|
||||
extrema_labels = torch.ones(actual_batch_size, dtype=torch.long, device=self.device) * 2 # Default: neither
|
||||
|
||||
# Identify potential bottoms (significant negative change)
|
||||
bottoms = (immediate_changes < -0.003)
|
||||
extrema_labels[bottoms] = 0
|
||||
|
||||
# Identify potential tops (significant positive change)
|
||||
tops = (immediate_changes > 0.003)
|
||||
extrema_labels[tops] = 1
|
||||
|
||||
# Calculate extrema prediction loss
|
||||
if len(current_extrema_pred.shape) > 1 and current_extrema_pred.shape[0] >= actual_batch_size:
|
||||
current_extrema_pred = current_extrema_pred[:actual_batch_size]
|
||||
extrema_loss = nn.CrossEntropyLoss()(current_extrema_pred, extrema_labels)
|
||||
|
||||
# Combined loss with all components
|
||||
# Primary task: Q-value learning (RL objective)
|
||||
# Secondary tasks: extrema detection and price prediction (supervised objectives)
|
||||
loss = q_loss + 0.3 * extrema_loss + 0.3 * price_loss
|
||||
|
||||
# Log loss components occasionally
|
||||
if random.random() < 0.01: # Log 1% of the time
|
||||
logger.info(
|
||||
f"Mixed precision losses: Q-loss={q_loss.item():.4f}, "
|
||||
f"Extrema-loss={extrema_loss.item():.4f}, "
|
||||
f"Price-loss={price_loss.item():.4f}"
|
||||
)
|
||||
except Exception as e:
|
||||
# Fallback if price extraction fails
|
||||
logger.warning(f"Failed to calculate price prediction loss: {str(e)}. Using only Q-value loss.")
|
||||
# Just use Q-value loss
|
||||
target_q_values = rewards + (1 - dones) * self.gamma * next_q_values
|
||||
|
||||
# Compute Q-value loss (primary task)
|
||||
q_loss = nn.MSELoss()(current_q_values, target_q_values)
|
||||
|
||||
# Initialize loss with q_loss
|
||||
loss = q_loss
|
||||
|
||||
# Backward pass with scaled gradients
|
||||
self.scaler.scale(loss).backward()
|
||||
|
||||
# Gradient clipping on scaled gradients
|
||||
self.scaler.unscale_(self.optimizer)
|
||||
torch.nn.utils.clip_grad_norm_(self.policy_net.parameters(), 1.0)
|
||||
|
||||
# Update with scaler
|
||||
self.scaler.step(self.optimizer)
|
||||
self.scaler.update()
|
||||
|
||||
# Update target network if needed
|
||||
self.update_count += 1
|
||||
if self.update_count % self.target_update == 0:
|
||||
self.target_net.load_state_dict(self.policy_net.state_dict())
|
||||
|
||||
# Track and decay epsilon
|
||||
self.epsilon = max(self.epsilon_min, self.epsilon * self.epsilon_decay)
|
||||
|
||||
return loss.item()
|
||||
|
||||
|
||||
# Try to extract price from current and next states
|
||||
try:
|
||||
# Extract price feature from sequence data (if available)
|
||||
if len(states.shape) == 3: # [batch, seq, features]
|
||||
current_prices = states[:, -1, -1] # Last timestep, last feature
|
||||
next_prices = next_states[:, -1, -1]
|
||||
else: # [batch, features]
|
||||
current_prices = states[:, -1] # Last feature
|
||||
next_prices = next_states[:, -1]
|
||||
|
||||
# Calculate price change for different timeframes
|
||||
immediate_changes = (next_prices - current_prices) / current_prices
|
||||
|
||||
# Get the actual batch size for this calculation
|
||||
actual_batch_size = states.shape[0]
|
||||
|
||||
# Create price direction labels - simplified for training
|
||||
# 0 = down, 1 = sideways, 2 = up
|
||||
immediate_labels = torch.ones(actual_batch_size, dtype=torch.long, device=self.device) * 1 # Default: sideways
|
||||
midterm_labels = torch.ones(actual_batch_size, dtype=torch.long, device=self.device) * 1
|
||||
longterm_labels = torch.ones(actual_batch_size, dtype=torch.long, device=self.device) * 1
|
||||
|
||||
# Immediate term direction (1s, 1m)
|
||||
immediate_up = (immediate_changes > 0.0005)
|
||||
immediate_down = (immediate_changes < -0.0005)
|
||||
immediate_labels[immediate_up] = 2 # Up
|
||||
immediate_labels[immediate_down] = 0 # Down
|
||||
|
||||
# For mid and long term, we can only approximate during training
|
||||
# In a real system, we'd need historical data to validate these
|
||||
# Here we'll use the immediate term with increasing thresholds as approximation
|
||||
|
||||
# Mid-term (1h) - use slightly higher threshold
|
||||
midterm_up = (immediate_changes > 0.001)
|
||||
midterm_down = (immediate_changes < -0.001)
|
||||
midterm_labels[midterm_up] = 2 # Up
|
||||
midterm_labels[midterm_down] = 0 # Down
|
||||
|
||||
# Long-term (1d) - use even higher threshold
|
||||
longterm_up = (immediate_changes > 0.002)
|
||||
longterm_down = (immediate_changes < -0.002)
|
||||
longterm_labels[longterm_up] = 2 # Up
|
||||
longterm_labels[longterm_down] = 0 # Down
|
||||
|
||||
# Generate target values for price change regression
|
||||
# For simplicity, we'll use the immediate change and scaled versions for longer timeframes
|
||||
price_value_targets = torch.zeros((actual_batch_size, 4), device=self.device)
|
||||
price_value_targets[:, 0] = immediate_changes
|
||||
price_value_targets[:, 1] = immediate_changes * 2.0 # Approximate 1h change
|
||||
price_value_targets[:, 2] = immediate_changes * 4.0 # Approximate 1d change
|
||||
price_value_targets[:, 3] = immediate_changes * 6.0 # Approximate 1w change
|
||||
|
||||
# Calculate loss for price direction prediction (classification)
|
||||
if len(current_price_pred['immediate'].shape) > 1 and current_price_pred['immediate'].shape[0] >= actual_batch_size:
|
||||
# Slice predictions to match the adjusted batch size
|
||||
immediate_pred = current_price_pred['immediate'][:actual_batch_size]
|
||||
midterm_pred = current_price_pred['midterm'][:actual_batch_size]
|
||||
longterm_pred = current_price_pred['longterm'][:actual_batch_size]
|
||||
price_values_pred = current_price_pred['values'][:actual_batch_size]
|
||||
|
||||
# Compute losses for each task
|
||||
immediate_loss = nn.CrossEntropyLoss()(immediate_pred, immediate_labels)
|
||||
midterm_loss = nn.CrossEntropyLoss()(midterm_pred, midterm_labels)
|
||||
longterm_loss = nn.CrossEntropyLoss()(longterm_pred, longterm_labels)
|
||||
|
||||
# MSE loss for price value regression
|
||||
price_value_loss = nn.MSELoss()(price_values_pred, price_value_targets)
|
||||
|
||||
# Combine all price prediction losses
|
||||
price_loss = immediate_loss + 0.7 * midterm_loss + 0.5 * longterm_loss + 0.3 * price_value_loss
|
||||
|
||||
# Create extrema labels (same as before)
|
||||
extrema_labels = torch.ones(actual_batch_size, dtype=torch.long, device=self.device) * 2 # Default: neither
|
||||
|
||||
# Identify potential bottoms (significant negative change)
|
||||
bottoms = (immediate_changes < -0.003)
|
||||
extrema_labels[bottoms] = 0
|
||||
|
||||
# Identify potential tops (significant positive change)
|
||||
tops = (immediate_changes > 0.003)
|
||||
extrema_labels[tops] = 1
|
||||
|
||||
# Calculate extrema prediction loss
|
||||
if len(current_extrema_pred.shape) > 1 and current_extrema_pred.shape[0] >= actual_batch_size:
|
||||
current_extrema_pred = current_extrema_pred[:actual_batch_size]
|
||||
extrema_loss = nn.CrossEntropyLoss()(current_extrema_pred, extrema_labels)
|
||||
|
||||
# Combined loss with all components
|
||||
# Primary task: Q-value learning (RL objective)
|
||||
# Secondary tasks: extrema detection and price prediction (supervised objectives)
|
||||
loss = q_loss + 0.3 * extrema_loss + 0.3 * price_loss
|
||||
|
||||
# Log loss components occasionally
|
||||
if random.random() < 0.01: # Log 1% of the time
|
||||
logger.info(
|
||||
f"Mixed precision losses: Q-loss={q_loss.item():.4f}, "
|
||||
f"Extrema-loss={extrema_loss.item():.4f}, "
|
||||
f"Price-loss={price_loss.item():.4f}"
|
||||
)
|
||||
except Exception as e:
|
||||
# Fallback if price extraction fails
|
||||
logger.warning(f"Failed to calculate price prediction loss: {str(e)}. Using only Q-value loss.")
|
||||
# Just use Q-value loss
|
||||
loss = q_loss
|
||||
|
||||
# Backward pass with scaled gradients
|
||||
self.scaler.scale(loss).backward()
|
||||
|
||||
# Gradient clipping on scaled gradients
|
||||
self.scaler.unscale_(self.optimizer)
|
||||
torch.nn.utils.clip_grad_norm_(self.policy_net.parameters(), 1.0)
|
||||
|
||||
# Update with scaler
|
||||
self.scaler.step(self.optimizer)
|
||||
self.scaler.update()
|
||||
|
||||
# Update target network if needed
|
||||
self.update_count += 1
|
||||
if self.update_count % self.target_update == 0:
|
||||
self.target_net.load_state_dict(self.policy_net.state_dict())
|
||||
|
||||
# Track and decay epsilon
|
||||
self.epsilon = max(self.epsilon_min, self.epsilon * self.epsilon_decay)
|
||||
|
||||
return loss.item()
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Error in mixed precision training: {str(e)}")
|
||||
logger.warning("Falling back to standard precision training")
|
||||
@ -1565,6 +1579,14 @@ class DQNAgent:
|
||||
try:
|
||||
# If state is already a numpy array, return it
|
||||
if isinstance(state, np.ndarray):
|
||||
# Check for empty array
|
||||
if state.size == 0:
|
||||
logger.warning("Received empty numpy array state. Using fallback dimensions.")
|
||||
expected_size = getattr(self, 'state_size', getattr(self, 'state_dim', 403))
|
||||
if isinstance(expected_size, tuple):
|
||||
expected_size = np.prod(expected_size)
|
||||
return np.zeros(int(expected_size), dtype=np.float32)
|
||||
|
||||
# Check for non-numeric data and handle it
|
||||
if state.dtype == object:
|
||||
# Convert object array to float array
|
||||
@ -1581,6 +1603,14 @@ class DQNAgent:
|
||||
|
||||
# If state is a list or tuple, convert to array
|
||||
elif isinstance(state, (list, tuple)):
|
||||
# Check for empty list/tuple
|
||||
if len(state) == 0:
|
||||
logger.warning("Received empty list/tuple state. Using fallback dimensions.")
|
||||
expected_size = getattr(self, 'state_size', getattr(self, 'state_dim', 403))
|
||||
if isinstance(expected_size, tuple):
|
||||
expected_size = np.prod(expected_size)
|
||||
return np.zeros(int(expected_size), dtype=np.float32)
|
||||
|
||||
# Recursively sanitize each element
|
||||
sanitized = []
|
||||
for item in state:
|
||||
@ -1591,7 +1621,18 @@ class DQNAgent:
|
||||
sanitized.append(sanitized_row)
|
||||
else:
|
||||
sanitized.append(self._extract_numeric_value(item))
|
||||
return np.array(sanitized, dtype=np.float32)
|
||||
|
||||
result = np.array(sanitized, dtype=np.float32)
|
||||
|
||||
# Check if result is empty and provide fallback
|
||||
if result.size == 0:
|
||||
logger.warning("Sanitized state resulted in empty array. Using fallback dimensions.")
|
||||
expected_size = getattr(self, 'state_size', getattr(self, 'state_dim', 403))
|
||||
if isinstance(expected_size, tuple):
|
||||
expected_size = np.prod(expected_size)
|
||||
return np.zeros(int(expected_size), dtype=np.float32)
|
||||
|
||||
return result
|
||||
|
||||
# If state is a dict, try to extract values
|
||||
elif isinstance(state, dict):
|
||||
|
||||
Reference in New Issue
Block a user