python / model.py
Princess3's picture
Update model.py
d5730cd verified
raw
history blame
8.2 kB
import os
import xml.etree.ElementTree as ET
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import List, Dict, Any, Optional
from collections import defaultdict
from accelerate import Accelerator
from torch.utils.data import DataLoader, TensorDataset
from torch.cuda.amp import GradScaler, autocast
class DynamicModel(nn.Module):
def __init__(self, sections: Dict[str, List[Dict[str, Any]]]):
super(DynamicModel, self).__init__()
self.sections = nn.ModuleDict()
if not sections:
sections = {
'default': [{
'input_size': 128,
'output_size': 256,
'activation': 'relu',
'batch_norm': True,
'dropout': 0.1
}]
}
for section_name, layers in sections.items():
self.sections[section_name] = nn.ModuleList()
for layer_params in layers:
print(f"Creating layer in section '{section_name}' with params: {layer_params}")
self.sections[section_name].append(self.create_layer(layer_params))
def create_layer(self, layer_params: Dict[str, Any]) -> nn.Module:
layers = []
layers.append(nn.Linear(layer_params['input_size'], layer_params['output_size']))
if layer_params.get('batch_norm', False):
layers.append(nn.BatchNorm1d(layer_params['output_size']))
activation = layer_params.get('activation', 'relu')
if activation == 'relu':
layers.append(nn.ReLU(inplace=True))
elif activation == 'tanh':
layers.append(nn.Tanh())
elif activation == 'sigmoid':
layers.append(nn.Sigmoid())
elif activation == 'leaky_relu':
layers.append(nn.LeakyReLU(negative_slope=0.01, inplace=True))
elif activation == 'elu':
layers.append(nn.ELU(alpha=1.0, inplace=True))
elif activation is not None:
raise ValueError(f"Unsupported activation function: {activation}")
if dropout_rate := layer_params.get('dropout', 0.0):
layers.append(nn.Dropout(p=dropout_rate))
if hidden_layers := layer_params.get('hidden_layers', []):
for hidden_layer_params in hidden_layers:
layers.append(self.create_layer(hidden_layer_params))
if layer_params.get('memory_augmentation', True):
layers.append(MemoryAugmentationLayer(layer_params['output_size']))
if layer_params.get('hybrid_attention', True):
layers.append(HybridAttentionLayer(layer_params['output_size']))
if layer_params.get('dynamic_flash_attention', True):
layers.append(DynamicFlashAttentionLayer(layer_params['output_size']))
return nn.Sequential(*layers)
def forward(self, x: torch.Tensor, section_name: Optional[str] = None) -> torch.Tensor:
if section_name is not None:
if section_name not in self.sections:
raise KeyError(f"Section '{section_name}' not found in model")
for layer in self.sections[section_name]:
x = layer(x)
else:
for section_name, layers in self.sections.items():
for layer in layers:
x = layer(x)
return x
class MemoryAugmentationLayer(nn.Module):
def __init__(self, size: int):
super(MemoryAugmentationLayer, self).__init__()
self.memory = nn.Parameter(torch.randn(size))
def forward(self, x: torch.Tensor) -> torch.Tensor:
return x + self.memory
class HybridAttentionLayer(nn.Module):
def __init__(self, size: int):
super(HybridAttentionLayer, self).__init__()
self.attention = nn.MultiheadAttention(size, num_heads=8)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = x.unsqueeze(1) # Add sequence dimension
attn_output, _ = self.attention(x, x, x)
return attn_output.squeeze(1)
class DynamicFlashAttentionLayer(nn.Module):
def __init__(self, size: int):
super(DynamicFlashAttentionLayer, self).__init__()
self.attention = nn.MultiheadAttention(size, num_heads=8)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = x.unsqueeze(1) # Add sequence dimension
attn_output, _ = self.attention(x, x, x)
return attn_output.squeeze(1)
def parse_xml_file(file_path: str) -> List[Dict[str, Any]]:
tree = ET.parse(file_path)
root = tree.getroot()
layers = []
for layer in root.findall('.//layer'):
layer_params = {}
layer_params['input_size'] = int(layer.get('input_size', 128))
layer_params['output_size'] = int(layer.get('output_size', 256))
layer_params['activation'] = layer.get('activation', 'relu').lower()
if layer_params['activation'] not in ['relu', 'tanh', 'sigmoid', 'none']:
raise ValueError(f"Unsupported activation function: {layer_params['activation']}")
if layer_params['input_size'] <= 0 or layer_params['output_size'] <= 0:
raise ValueError("Layer dimensions must be positive integers")
layers.append(layer_params)
if not layers:
layers.append({
'input_size': 128,
'output_size': 256,
'activation': 'relu'
})
return layers
def create_model_from_folder(folder_path: str) -> DynamicModel:
sections = defaultdict(list)
if not os.path.exists(folder_path):
print(f"Warning: Folder {folder_path} does not exist. Creating model with default configuration.")
return DynamicModel({})
xml_files_found = False
for root, dirs, files in os.walk(folder_path):
for file in files:
if file.endswith('.xml'):
xml_files_found = True
file_path = os.path.join(root, file)
try:
layers = parse_xml_file(file_path)
section_name = os.path.basename(root).replace('.', '_')
sections[section_name].extend(layers)
except Exception as e:
print(f"Error processing {file_path}: {str(e)}")
if not xml_files_found:
print("Warning: No XML files found. Creating model with default configuration.")
return DynamicModel({})
return DynamicModel(dict(sections))
def main():
folder_path = 'data'
model = create_model_from_folder(folder_path)
print(f"Created dynamic PyTorch model with sections: {list(model.sections.keys())}")
# Print the model architecture
print(model)
first_section = next(iter(model.sections.keys()))
first_layer = model.sections[first_section][0]
input_features = first_layer[0].in_features
# Ensure the input tensor size matches the expected input size
sample_input = torch.randn(1, input_features)
output = model(sample_input)
print(f"Sample output shape: {output.shape}")
accelerator = Accelerator()
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
criterion = nn.CrossEntropyLoss()
num_epochs = 10
dataset = TensorDataset(
torch.randn(100, input_features),
torch.randint(0, 2, (100,))
)
train_dataloader = DataLoader(
dataset,
batch_size=8, # Reduced batch size
shuffle=True
)
model, optimizer, train_dataloader = accelerator.prepare(
model,
optimizer,
train_dataloader
)
scaler = GradScaler() # Mixed precision training
for epoch in range(num_epochs):
model.train()
total_loss = 0
for batch_idx, (inputs, labels) in enumerate(train_dataloader):
optimizer.zero_grad()
with autocast(): # Mixed precision training
outputs = model(inputs)
loss = criterion(outputs, labels)
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
total_loss += loss.item()
avg_loss = total_loss / len(train_dataloader)
print(f"Epoch {epoch+1}/{num_epochs}, Average Loss: {avg_loss:.4f}")
if __name__ == "__main__":
main()