Commit
·
ced6e93
0
Parent(s):
Initial commit: AutoEncoder model
Browse files- .gitignore +10 -0
- README.md +445 -0
- __init__.py +19 -0
- configuration_autoencoder.py +253 -0
- modeling_autoencoder.py +1099 -0
- register_autoencoder.py +62 -0
.gitignore
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Python-generated files
|
2 |
+
__pycache__/
|
3 |
+
*.py[oc]
|
4 |
+
build/
|
5 |
+
dist/
|
6 |
+
wheels/
|
7 |
+
*.egg-info
|
8 |
+
|
9 |
+
# Virtual environments
|
10 |
+
.venv
|
README.md
ADDED
@@ -0,0 +1,445 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Autoencoder Implementation for Hugging Face Transformers
|
2 |
+
|
3 |
+
A complete autoencoder implementation that integrates seamlessly with the Hugging Face Transformers ecosystem, providing all the standard functionality you expect from transformer models.
|
4 |
+
|
5 |
+
## 🚀 Features
|
6 |
+
|
7 |
+
- **Full Hugging Face Integration**: Compatible with `AutoModel`, `AutoConfig`, and `AutoTokenizer` patterns
|
8 |
+
- **Standard Training Workflows**: Works with `Trainer`, `TrainingArguments`, and all HF training utilities
|
9 |
+
- **Model Hub Compatible**: Save and share models on Hugging Face Hub with `push_to_hub()`
|
10 |
+
- **Flexible Architecture**: Configurable encoder-decoder architecture with various activation functions
|
11 |
+
- **Multiple Loss Functions**: Support for MSE, BCE, L1, Huber, Smooth L1, KL Divergence, Cosine, Focal, Dice, Tversky, SSIM, and Perceptual loss
|
12 |
+
- **Multiple Autoencoder Types (7)**: Classic, Variational (VAE), Beta-VAE, Denoising, Sparse, Contractive, and Recurrent autoencoders
|
13 |
+
- **Extended Activation Functions**: 18+ activation functions including ReLU, GELU, Swish, Mish, ELU, and more
|
14 |
+
- **Learnable Preprocessing**: Neural Scaler and Normalizing Flow preprocessors (2D and 3D tensors)
|
15 |
+
- **Extensible Design**: Easy to extend for new autoencoder variants and custom loss functions
|
16 |
+
- **Production Ready**: Proper serialization, checkpointing, and inference support
|
17 |
+
|
18 |
+
## 📦 Installation
|
19 |
+
|
20 |
+
```bash
|
21 |
+
uv sync # or: pip install -e .
|
22 |
+
```
|
23 |
+
|
24 |
+
Dependencies (see pyproject.toml):
|
25 |
+
- `torch>=2.8.0`
|
26 |
+
- `transformers>=4.55.2`
|
27 |
+
- `numpy>=2.3.2`
|
28 |
+
- `scikit-learn>=1.7.1`
|
29 |
+
- `datasets>=4.0.0`
|
30 |
+
- `accelerate>=1.10.0`
|
31 |
+
|
32 |
+
## 🏗️ Architecture
|
33 |
+
|
34 |
+
Note: This repository has been trimmed to essentials for easy reuse and distribution. Example scripts and tests were removed by request.
|
35 |
+
|
36 |
+
The implementation consists of three main components:
|
37 |
+
|
38 |
+
### 1. AutoencoderConfig
|
39 |
+
Configuration class that inherits from `PretrainedConfig`:
|
40 |
+
- Defines model architecture parameters
|
41 |
+
- Handles validation and serialization
|
42 |
+
- Enables `AutoConfig.from_pretrained()` functionality
|
43 |
+
|
44 |
+
### 2. AutoencoderModel
|
45 |
+
Base model class that inherits from `PreTrainedModel`:
|
46 |
+
- Implements encoder-decoder architecture
|
47 |
+
- Provides latent space representation
|
48 |
+
- Returns structured outputs with `AutoencoderOutput`
|
49 |
+
|
50 |
+
### 3. AutoencoderForReconstruction
|
51 |
+
Task-specific model for reconstruction:
|
52 |
+
- Adds reconstruction loss calculation
|
53 |
+
- Compatible with `Trainer` for easy training
|
54 |
+
- Returns `AutoencoderForReconstructionOutput` with loss
|
55 |
+
|
56 |
+
## 🔧 Quick Start
|
57 |
+
|
58 |
+
### Basic Usage
|
59 |
+
|
60 |
+
```python
|
61 |
+
from configuration_autoencoder import AutoencoderConfig
|
62 |
+
from modeling_autoencoder import AutoencoderForReconstruction
|
63 |
+
import torch
|
64 |
+
|
65 |
+
# Create configuration
|
66 |
+
config = AutoencoderConfig(
|
67 |
+
input_dim=784, # Input dimensionality (e.g., 28x28 images flattened)
|
68 |
+
hidden_dims=[512, 256], # Encoder hidden layers
|
69 |
+
latent_dim=64, # Latent space dimension
|
70 |
+
activation="gelu", # Activation function (18+ options available)
|
71 |
+
reconstruction_loss="mse", # Loss function (12+ options available)
|
72 |
+
autoencoder_type="classic", # Autoencoder type (7 types available)
|
73 |
+
# Optional learnable preprocessing
|
74 |
+
use_learnable_preprocessing=True,
|
75 |
+
preprocessing_type="neural_scaler", # or "normalizing_flow"
|
76 |
+
)
|
77 |
+
|
78 |
+
# Create model
|
79 |
+
model = AutoencoderForReconstruction(config)
|
80 |
+
|
81 |
+
# Forward pass
|
82 |
+
input_data = torch.randn(32, 784) # Batch of 32 samples
|
83 |
+
outputs = model(input_values=input_data)
|
84 |
+
|
85 |
+
print(f"Reconstruction loss: {outputs.loss}")
|
86 |
+
print(f"Latent shape: {outputs.last_hidden_state.shape}")
|
87 |
+
print(f"Reconstructed shape: {outputs.reconstructed.shape}")
|
88 |
+
```
|
89 |
+
|
90 |
+
### Training with Hugging Face Trainer
|
91 |
+
|
92 |
+
```python
|
93 |
+
from transformers import Trainer, TrainingArguments
|
94 |
+
from torch.utils.data import Dataset
|
95 |
+
|
96 |
+
class AutoencoderDataset(Dataset):
|
97 |
+
def __init__(self, data):
|
98 |
+
self.data = torch.FloatTensor(data)
|
99 |
+
|
100 |
+
def __len__(self):
|
101 |
+
return len(self.data)
|
102 |
+
|
103 |
+
def __getitem__(self, idx):
|
104 |
+
return {
|
105 |
+
"input_values": self.data[idx],
|
106 |
+
"labels": self.data[idx] # For autoencoder, input = target
|
107 |
+
}
|
108 |
+
|
109 |
+
# Prepare data
|
110 |
+
train_dataset = AutoencoderDataset(your_training_data)
|
111 |
+
val_dataset = AutoencoderDataset(your_validation_data)
|
112 |
+
|
113 |
+
# Training arguments
|
114 |
+
training_args = TrainingArguments(
|
115 |
+
output_dir="./autoencoder_output",
|
116 |
+
num_train_epochs=10,
|
117 |
+
per_device_train_batch_size=64,
|
118 |
+
per_device_eval_batch_size=64,
|
119 |
+
warmup_steps=500,
|
120 |
+
weight_decay=0.01,
|
121 |
+
logging_dir="./logs",
|
122 |
+
evaluation_strategy="steps",
|
123 |
+
eval_steps=500,
|
124 |
+
save_steps=1000,
|
125 |
+
load_best_model_at_end=True,
|
126 |
+
)
|
127 |
+
|
128 |
+
# Create trainer
|
129 |
+
trainer = Trainer(
|
130 |
+
model=model,
|
131 |
+
args=training_args,
|
132 |
+
train_dataset=train_dataset,
|
133 |
+
eval_dataset=val_dataset,
|
134 |
+
)
|
135 |
+
|
136 |
+
# Train
|
137 |
+
trainer.train()
|
138 |
+
|
139 |
+
# Save model
|
140 |
+
model.save_pretrained("./my_autoencoder")
|
141 |
+
config.save_pretrained("./my_autoencoder")
|
142 |
+
```
|
143 |
+
|
144 |
+
### Using AutoModel Framework
|
145 |
+
|
146 |
+
```python
|
147 |
+
from register_autoencoder import register_autoencoder_models
|
148 |
+
from transformers import AutoConfig, AutoModel
|
149 |
+
|
150 |
+
# Register models with AutoModel framework
|
151 |
+
register_autoencoder_models()
|
152 |
+
|
153 |
+
# Now you can use standard HF patterns
|
154 |
+
config = AutoConfig.from_pretrained("./my_autoencoder")
|
155 |
+
model = AutoModel.from_pretrained("./my_autoencoder")
|
156 |
+
|
157 |
+
# Use the model
|
158 |
+
outputs = model(input_values=your_data)
|
159 |
+
```
|
160 |
+
|
161 |
+
## ⚙️ Configuration Options
|
162 |
+
|
163 |
+
The `AutoencoderConfig` class supports extensive customization:
|
164 |
+
|
165 |
+
```python
|
166 |
+
config = AutoencoderConfig(
|
167 |
+
input_dim=784, # Input dimension
|
168 |
+
hidden_dims=[512, 256, 128], # Encoder hidden layers
|
169 |
+
latent_dim=64, # Latent space dimension
|
170 |
+
activation="gelu", # Activation function (see full list below)
|
171 |
+
dropout_rate=0.1, # Dropout rate (0.0 to 1.0)
|
172 |
+
use_batch_norm=True, # Use batch normalization
|
173 |
+
tie_weights=False, # Tie encoder/decoder weights
|
174 |
+
reconstruction_loss="mse", # Loss function (see full list below)
|
175 |
+
autoencoder_type="variational", # Autoencoder type (see types below)
|
176 |
+
beta=0.5, # Beta parameter for β-VAE
|
177 |
+
temperature=1.0, # Temperature for Gumbel softmax
|
178 |
+
noise_factor=0.1, # Noise factor for denoising AE
|
179 |
+
# Recurrent autoencoder parameters
|
180 |
+
rnn_type="lstm", # RNN type: "lstm", "gru", "rnn"
|
181 |
+
num_layers=2, # Number of RNN layers
|
182 |
+
bidirectional=True, # Bidirectional encoding
|
183 |
+
sequence_length=None, # Fixed sequence length (None for variable)
|
184 |
+
teacher_forcing_ratio=0.5, # Teacher forcing ratio during training
|
185 |
+
# Learnable preprocessing parameters
|
186 |
+
use_learnable_preprocessing=False, # Enable learnable preprocessing
|
187 |
+
preprocessing_type="none", # "none", "neural_scaler", "normalizing_flow"
|
188 |
+
preprocessing_hidden_dim=64, # Hidden dimension for preprocessing networks
|
189 |
+
preprocessing_num_layers=2, # Number of layers in preprocessing networks
|
190 |
+
learn_inverse_preprocessing=True, # Learn inverse transformation
|
191 |
+
flow_coupling_layers=4, # Number of coupling layers for flows
|
192 |
+
)
|
193 |
+
```
|
194 |
+
|
195 |
+
### 🎛️ Available Activation Functions
|
196 |
+
|
197 |
+
**Standard Activations:**
|
198 |
+
- `relu`, `leaky_relu`, `relu6`, `elu`, `prelu`
|
199 |
+
- `tanh`, `sigmoid`, `hardsigmoid`, `hardtanh`
|
200 |
+
- `gelu`, `swish`, `silu`, `hardswish`
|
201 |
+
- `mish`, `softplus`, `softsign`, `tanhshrink`, `threshold`
|
202 |
+
|
203 |
+
### 📊 Available Loss Functions
|
204 |
+
|
205 |
+
**Regression Losses:**
|
206 |
+
- `mse` - Mean Squared Error
|
207 |
+
- `l1` - L1/MAE Loss
|
208 |
+
- `huber` - Huber Loss
|
209 |
+
- `smooth_l1` - Smooth L1 Loss
|
210 |
+
|
211 |
+
**Classification/Probability Losses:**
|
212 |
+
- `bce` - Binary Cross Entropy
|
213 |
+
- `kl_div` - KL Divergence
|
214 |
+
- `focal` - Focal Loss
|
215 |
+
|
216 |
+
**Similarity Losses:**
|
217 |
+
- `cosine` - Cosine Similarity Loss
|
218 |
+
- `ssim` - Structural Similarity Loss
|
219 |
+
- `perceptual` - Perceptual Loss
|
220 |
+
|
221 |
+
**Segmentation Losses:**
|
222 |
+
- `dice` - Dice Loss
|
223 |
+
- `tversky` - Tversky Loss
|
224 |
+
|
225 |
+
### 🏗️ Available Autoencoder Types
|
226 |
+
|
227 |
+
**Classic Autoencoder (`classic`)**
|
228 |
+
- Standard encoder-decoder architecture
|
229 |
+
- Direct reconstruction loss minimization
|
230 |
+
|
231 |
+
**Variational Autoencoder (`variational`)**
|
232 |
+
- Probabilistic latent space with mean and variance
|
233 |
+
- KL divergence regularization
|
234 |
+
- Reparameterization trick for sampling
|
235 |
+
|
236 |
+
**Beta-VAE (`beta_vae`)**
|
237 |
+
- Variational autoencoder with adjustable β parameter
|
238 |
+
- Better disentanglement of latent factors
|
239 |
+
|
240 |
+
**Denoising Autoencoder (`denoising`)**
|
241 |
+
- Adds noise to input during training
|
242 |
+
- Learns robust representations
|
243 |
+
- Configurable noise factor
|
244 |
+
|
245 |
+
**Sparse Autoencoder (`sparse`)**
|
246 |
+
- Encourages sparse latent representations
|
247 |
+
- L1 regularization on latent activations
|
248 |
+
- Useful for feature selection
|
249 |
+
|
250 |
+
**Contractive Autoencoder (`contractive`)**
|
251 |
+
- Penalizes large gradients of latent w.r.t. input
|
252 |
+
- Learns smooth manifold representations
|
253 |
+
- Robust to small input perturbations
|
254 |
+
|
255 |
+
**Recurrent Autoencoder (`recurrent`)**
|
256 |
+
- LSTM/GRU/RNN encoder-decoder architecture
|
257 |
+
- Bidirectional encoding for better sequence representations
|
258 |
+
- Variable length sequence support with padding
|
259 |
+
- Teacher forcing during training for stable learning
|
260 |
+
- Sequence-to-sequence reconstruction
|
261 |
+
```
|
262 |
+
|
263 |
+
## 📊 Model Outputs
|
264 |
+
|
265 |
+
### AutoencoderOutput
|
266 |
+
```python
|
267 |
+
@dataclass
|
268 |
+
class AutoencoderOutput(ModelOutput):
|
269 |
+
last_hidden_state: torch.FloatTensor = None # Latent representation
|
270 |
+
reconstructed: torch.FloatTensor = None # Reconstructed input
|
271 |
+
hidden_states: Tuple[torch.FloatTensor] = None # Intermediate states
|
272 |
+
attentions: Tuple[torch.FloatTensor] = None # Not used
|
273 |
+
```
|
274 |
+
|
275 |
+
### AutoencoderForReconstructionOutput
|
276 |
+
```python
|
277 |
+
@dataclass
|
278 |
+
class AutoencoderForReconstructionOutput(ModelOutput):
|
279 |
+
loss: torch.FloatTensor = None # Reconstruction loss
|
280 |
+
reconstructed: torch.FloatTensor = None # Reconstructed input
|
281 |
+
last_hidden_state: torch.FloatTensor = None # Latent representation
|
282 |
+
hidden_states: Tuple[torch.FloatTensor] = None # Intermediate states
|
283 |
+
```
|
284 |
+
|
285 |
+
## 🔬 Advanced Usage
|
286 |
+
|
287 |
+
### Custom Loss Functions
|
288 |
+
|
289 |
+
You can easily extend the model with custom loss functions:
|
290 |
+
|
291 |
+
```python
|
292 |
+
class CustomAutoencoder(AutoencoderForReconstruction):
|
293 |
+
def _compute_reconstruction_loss(self, reconstructed, target):
|
294 |
+
# Custom loss implementation
|
295 |
+
return your_custom_loss(reconstructed, target)
|
296 |
+
```
|
297 |
+
|
298 |
+
### Recurrent Autoencoder for Sequences
|
299 |
+
|
300 |
+
Perfect for time series, text, and sequential data:
|
301 |
+
|
302 |
+
```python
|
303 |
+
config = AutoencoderConfig(
|
304 |
+
input_dim=50, # Feature dimension per timestep
|
305 |
+
latent_dim=32, # Compressed representation size
|
306 |
+
autoencoder_type="recurrent",
|
307 |
+
rnn_type="lstm", # or "gru", "rnn"
|
308 |
+
num_layers=2, # Number of RNN layers
|
309 |
+
bidirectional=True, # Bidirectional encoding
|
310 |
+
teacher_forcing_ratio=0.7, # Teacher forcing during training
|
311 |
+
sequence_length=None # Variable length sequences
|
312 |
+
)
|
313 |
+
|
314 |
+
# Usage with sequence data
|
315 |
+
model = AutoencoderForReconstruction(config)
|
316 |
+
sequence_data = torch.randn(batch_size, seq_len, input_dim)
|
317 |
+
outputs = model(input_values=sequence_data)
|
318 |
+
```
|
319 |
+
|
320 |
+
### Learnable Preprocessing
|
321 |
+
|
322 |
+
Deep learning-based data normalization that adapts to your data:
|
323 |
+
|
324 |
+
```python
|
325 |
+
# Neural Scaler - Learnable alternative to StandardScaler
|
326 |
+
config = AutoencoderConfig(
|
327 |
+
input_dim=20,
|
328 |
+
latent_dim=10,
|
329 |
+
use_learnable_preprocessing=True,
|
330 |
+
preprocessing_type="neural_scaler",
|
331 |
+
preprocessing_hidden_dim=64
|
332 |
+
)
|
333 |
+
|
334 |
+
# Normalizing Flow - Invertible transformations
|
335 |
+
config = AutoencoderConfig(
|
336 |
+
input_dim=20,
|
337 |
+
latent_dim=10,
|
338 |
+
use_learnable_preprocessing=True,
|
339 |
+
preprocessing_type="normalizing_flow",
|
340 |
+
flow_coupling_layers=4
|
341 |
+
)
|
342 |
+
|
343 |
+
# Works with all autoencoder types and sequence data
|
344 |
+
model = AutoencoderForReconstruction(config)
|
345 |
+
outputs = model(input_values=data)
|
346 |
+
print(f"Preprocessing loss: {outputs.preprocessing_loss}")
|
347 |
+
```
|
348 |
+
|
349 |
+
### Variational Autoencoder Extension
|
350 |
+
|
351 |
+
The configuration supports variational autoencoders:
|
352 |
+
|
353 |
+
```python
|
354 |
+
config = AutoencoderConfig(
|
355 |
+
autoencoder_type="variational",
|
356 |
+
beta=0.5, # β-VAE parameter
|
357 |
+
# ... other parameters
|
358 |
+
)
|
359 |
+
```
|
360 |
+
|
361 |
+
### Integration with Datasets Library
|
362 |
+
|
363 |
+
```python
|
364 |
+
from datasets import Dataset
|
365 |
+
|
366 |
+
# Convert your data to HF Dataset
|
367 |
+
dataset = Dataset.from_dict({
|
368 |
+
"input_values": your_data_list
|
369 |
+
})
|
370 |
+
|
371 |
+
# Use with Trainer
|
372 |
+
trainer = Trainer(
|
373 |
+
model=model,
|
374 |
+
train_dataset=dataset,
|
375 |
+
# ... other arguments
|
376 |
+
)
|
377 |
+
```
|
378 |
+
|
379 |
+
## 🧪 Testing
|
380 |
+
|
381 |
+
This repository has been trimmed to essential files. Example scripts and test files were removed by request. You can create your own quick checks using the Quick Start snippet above.
|
382 |
+
|
383 |
+
## 📁 Project Structure
|
384 |
+
|
385 |
+
```
|
386 |
+
autoencoder/
|
387 |
+
├── __init__.py # Package initialization
|
388 |
+
├── configuration_autoencoder.py # Configuration class
|
389 |
+
├── modeling_autoencoder.py # Model implementations
|
390 |
+
├── register_autoencoder.py # AutoModel registration
|
391 |
+
├── example_usage.py # Usage examples
|
392 |
+
├── test_save_load.py # Test suite
|
393 |
+
├── requirements.txt # Dependencies
|
394 |
+
└── README.md # This file
|
395 |
+
```
|
396 |
+
|
397 |
+
## 🤝 Contributing
|
398 |
+
|
399 |
+
This implementation follows Hugging Face conventions and can be easily extended:
|
400 |
+
|
401 |
+
1. **Adding new architectures**: Extend `AutoencoderModel` or create new model classes
|
402 |
+
2. **Custom configurations**: Add parameters to `AutoencoderConfig`
|
403 |
+
3. **Task-specific heads**: Create new classes like `AutoencoderForReconstruction`
|
404 |
+
4. **Integration**: Register new models with the AutoModel framework
|
405 |
+
|
406 |
+
## 📚 References
|
407 |
+
|
408 |
+
- [Hugging Face Transformers Documentation](https://huggingface.co/docs/transformers)
|
409 |
+
- [Custom Models Guide](https://huggingface.co/docs/transformers/custom_models)
|
410 |
+
- [AutoModel Documentation](https://huggingface.co/docs/transformers/model_doc/auto)
|
411 |
+
|
412 |
+
## 🎯 Use Cases
|
413 |
+
|
414 |
+
This autoencoder implementation is perfect for:
|
415 |
+
|
416 |
+
- **Dimensionality Reduction**: Compress high-dimensional data to lower dimensions
|
417 |
+
- **Anomaly Detection**: Identify outliers based on reconstruction error
|
418 |
+
- **Data Denoising**: Remove noise from corrupted data
|
419 |
+
- **Feature Learning**: Learn meaningful representations for downstream tasks
|
420 |
+
- **Data Generation**: Generate new samples similar to training data
|
421 |
+
- **Pretraining**: Initialize encoders for other tasks
|
422 |
+
|
423 |
+
## 🔍 Model Comparison
|
424 |
+
|
425 |
+
| Feature | Standard PyTorch | This Implementation |
|
426 |
+
|---------|------------------|-------------------|
|
427 |
+
| HF Integration | ❌ | ✅ |
|
428 |
+
| AutoModel Support | ❌ | ✅ |
|
429 |
+
| Trainer Compatible | ❌ | ✅ |
|
430 |
+
| Hub Integration | ❌ | ✅ |
|
431 |
+
| Config Management | Manual | ✅ Automatic |
|
432 |
+
| Serialization | Manual | ✅ Built-in |
|
433 |
+
| Checkpointing | Manual | ✅ Built-in |
|
434 |
+
|
435 |
+
## 🚀 Performance Tips
|
436 |
+
|
437 |
+
1. **Batch Size**: Use larger batch sizes for better GPU utilization
|
438 |
+
2. **Learning Rate**: Start with 1e-3 and adjust based on convergence
|
439 |
+
3. **Architecture**: Gradually decrease hidden dimensions for better compression
|
440 |
+
4. **Regularization**: Use dropout and batch normalization for better generalization
|
441 |
+
5. **Loss Function**: Choose appropriate loss based on your data type
|
442 |
+
|
443 |
+
## 📄 License
|
444 |
+
|
445 |
+
This implementation is provided as an example and follows the same license terms as Hugging Face Transformers.
|
__init__.py
ADDED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Autoencoder models for Hugging Face Transformers.
|
3 |
+
"""
|
4 |
+
|
5 |
+
from configuration_autoencoder import AutoencoderConfig
|
6 |
+
from modeling_autoencoder import (
|
7 |
+
AutoencoderModel,
|
8 |
+
AutoencoderForReconstruction,
|
9 |
+
AutoencoderOutput,
|
10 |
+
AutoencoderForReconstructionOutput,
|
11 |
+
)
|
12 |
+
|
13 |
+
__all__ = [
|
14 |
+
"AutoencoderConfig",
|
15 |
+
"AutoencoderModel",
|
16 |
+
"AutoencoderForReconstruction",
|
17 |
+
"AutoencoderOutput",
|
18 |
+
"AutoencoderForReconstructionOutput",
|
19 |
+
]
|
configuration_autoencoder.py
ADDED
@@ -0,0 +1,253 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Autoencoder configuration for Hugging Face Transformers.
|
3 |
+
"""
|
4 |
+
|
5 |
+
from transformers import PretrainedConfig
|
6 |
+
from typing import List, Optional
|
7 |
+
|
8 |
+
|
9 |
+
class AutoencoderConfig(PretrainedConfig):
|
10 |
+
"""
|
11 |
+
Configuration class for Autoencoder models.
|
12 |
+
|
13 |
+
This configuration class stores the configuration of an autoencoder model. It is used to instantiate
|
14 |
+
an autoencoder model according to the specified arguments, defining the model architecture.
|
15 |
+
|
16 |
+
Args:
|
17 |
+
input_dim (int, optional): Dimensionality of the input data. Defaults to 784.
|
18 |
+
hidden_dims (List[int], optional): List of hidden layer dimensions for the encoder.
|
19 |
+
The decoder will use the reverse of this list. Defaults to [512, 256, 128].
|
20 |
+
latent_dim (int, optional): Dimensionality of the latent space. Defaults to 64.
|
21 |
+
activation (str, optional): Activation function to use. Options: "relu", "tanh", "sigmoid",
|
22 |
+
"leaky_relu", "gelu", "swish", "silu", "elu", "prelu", "relu6", "hardtanh",
|
23 |
+
"hardsigmoid", "hardswish", "mish", "softplus", "softsign", "tanhshrink", "threshold".
|
24 |
+
Defaults to "relu".
|
25 |
+
dropout_rate (float, optional): Dropout rate for regularization. Defaults to 0.1.
|
26 |
+
use_batch_norm (bool, optional): Whether to use batch normalization. Defaults to True.
|
27 |
+
tie_weights (bool, optional): Whether to tie encoder and decoder weights. Defaults to False.
|
28 |
+
reconstruction_loss (str, optional): Type of reconstruction loss. Options: "mse", "bce", "l1",
|
29 |
+
"huber", "smooth_l1", "kl_div", "cosine", "focal", "dice", "tversky", "ssim", "perceptual".
|
30 |
+
Defaults to "mse".
|
31 |
+
autoencoder_type (str, optional): Type of autoencoder architecture. Options: "classic",
|
32 |
+
"variational", "beta_vae", "denoising", "sparse", "contractive", "recurrent". Defaults to "classic".
|
33 |
+
beta (float, optional): Beta parameter for beta-VAE. Defaults to 1.0.
|
34 |
+
temperature (float, optional): Temperature parameter for Gumbel softmax or other operations. Defaults to 1.0.
|
35 |
+
noise_factor (float, optional): Noise factor for denoising autoencoders. Defaults to 0.1.
|
36 |
+
rnn_type (str, optional): Type of RNN cell for recurrent autoencoders. Options: "lstm", "gru", "rnn".
|
37 |
+
Defaults to "lstm".
|
38 |
+
num_layers (int, optional): Number of RNN layers for recurrent autoencoders. Defaults to 2.
|
39 |
+
bidirectional (bool, optional): Whether to use bidirectional RNN for encoding. Defaults to True.
|
40 |
+
sequence_length (int, optional): Fixed sequence length. If None, supports variable length sequences.
|
41 |
+
Defaults to None.
|
42 |
+
teacher_forcing_ratio (float, optional): Ratio of teacher forcing during training for recurrent decoders.
|
43 |
+
Defaults to 0.5.
|
44 |
+
use_learnable_preprocessing (bool, optional): Whether to use learnable preprocessing. Defaults to False.
|
45 |
+
preprocessing_type (str, optional): Type of learnable preprocessing. Options: "none", "neural_scaler",
|
46 |
+
"normalizing_flow". Defaults to "none".
|
47 |
+
preprocessing_hidden_dim (int, optional): Hidden dimension for preprocessing networks. Defaults to 64.
|
48 |
+
preprocessing_num_layers (int, optional): Number of layers in preprocessing networks. Defaults to 2.
|
49 |
+
learn_inverse_preprocessing (bool, optional): Whether to learn inverse preprocessing for reconstruction.
|
50 |
+
Defaults to True.
|
51 |
+
flow_coupling_layers (int, optional): Number of coupling layers for normalizing flows. Defaults to 4.
|
52 |
+
**kwargs: Additional keyword arguments passed to the parent class.
|
53 |
+
"""
|
54 |
+
|
55 |
+
model_type = "autoencoder"
|
56 |
+
|
57 |
+
def __init__(
|
58 |
+
self,
|
59 |
+
input_dim: int = 784,
|
60 |
+
hidden_dims: List[int] = None,
|
61 |
+
latent_dim: int = 64,
|
62 |
+
activation: str = "relu",
|
63 |
+
dropout_rate: float = 0.1,
|
64 |
+
use_batch_norm: bool = True,
|
65 |
+
tie_weights: bool = False,
|
66 |
+
reconstruction_loss: str = "mse",
|
67 |
+
autoencoder_type: str = "classic",
|
68 |
+
beta: float = 1.0,
|
69 |
+
temperature: float = 1.0,
|
70 |
+
noise_factor: float = 0.1,
|
71 |
+
# Recurrent autoencoder parameters
|
72 |
+
rnn_type: str = "lstm",
|
73 |
+
num_layers: int = 2,
|
74 |
+
bidirectional: bool = True,
|
75 |
+
sequence_length: Optional[int] = None,
|
76 |
+
teacher_forcing_ratio: float = 0.5,
|
77 |
+
# Deep learning preprocessing parameters
|
78 |
+
use_learnable_preprocessing: bool = False,
|
79 |
+
preprocessing_type: str = "none",
|
80 |
+
preprocessing_hidden_dim: int = 64,
|
81 |
+
preprocessing_num_layers: int = 2,
|
82 |
+
learn_inverse_preprocessing: bool = True,
|
83 |
+
flow_coupling_layers: int = 4,
|
84 |
+
**kwargs,
|
85 |
+
):
|
86 |
+
# Validate parameters
|
87 |
+
if hidden_dims is None:
|
88 |
+
hidden_dims = [512, 256, 128]
|
89 |
+
|
90 |
+
# Extended activation functions
|
91 |
+
valid_activations = [
|
92 |
+
"relu", "tanh", "sigmoid", "leaky_relu", "gelu", "swish", "silu",
|
93 |
+
"elu", "prelu", "relu6", "hardtanh", "hardsigmoid", "hardswish",
|
94 |
+
"mish", "softplus", "softsign", "tanhshrink", "threshold"
|
95 |
+
]
|
96 |
+
if activation not in valid_activations:
|
97 |
+
raise ValueError(
|
98 |
+
f"`activation` must be one of {valid_activations}, got {activation}."
|
99 |
+
)
|
100 |
+
|
101 |
+
# Extended loss functions
|
102 |
+
valid_losses = [
|
103 |
+
"mse", "bce", "l1", "huber", "smooth_l1", "kl_div", "cosine",
|
104 |
+
"focal", "dice", "tversky", "ssim", "perceptual"
|
105 |
+
]
|
106 |
+
if reconstruction_loss not in valid_losses:
|
107 |
+
raise ValueError(
|
108 |
+
f"`reconstruction_loss` must be one of {valid_losses}, got {reconstruction_loss}."
|
109 |
+
)
|
110 |
+
|
111 |
+
# Autoencoder types
|
112 |
+
valid_types = ["classic", "variational", "beta_vae", "denoising", "sparse", "contractive", "recurrent"]
|
113 |
+
if autoencoder_type not in valid_types:
|
114 |
+
raise ValueError(
|
115 |
+
f"`autoencoder_type` must be one of {valid_types}, got {autoencoder_type}."
|
116 |
+
)
|
117 |
+
|
118 |
+
# RNN types for recurrent autoencoders
|
119 |
+
valid_rnn_types = ["lstm", "gru", "rnn"]
|
120 |
+
if rnn_type not in valid_rnn_types:
|
121 |
+
raise ValueError(
|
122 |
+
f"`rnn_type` must be one of {valid_rnn_types}, got {rnn_type}."
|
123 |
+
)
|
124 |
+
|
125 |
+
if not (0.0 <= dropout_rate <= 1.0):
|
126 |
+
raise ValueError(f"`dropout_rate` must be between 0.0 and 1.0, got {dropout_rate}.")
|
127 |
+
|
128 |
+
if input_dim <= 0:
|
129 |
+
raise ValueError(f"`input_dim` must be positive, got {input_dim}.")
|
130 |
+
|
131 |
+
if latent_dim <= 0:
|
132 |
+
raise ValueError(f"`latent_dim` must be positive, got {latent_dim}.")
|
133 |
+
|
134 |
+
if not all(dim > 0 for dim in hidden_dims):
|
135 |
+
raise ValueError("All dimensions in `hidden_dims` must be positive.")
|
136 |
+
|
137 |
+
if beta <= 0:
|
138 |
+
raise ValueError(f"`beta` must be positive, got {beta}.")
|
139 |
+
|
140 |
+
if num_layers <= 0:
|
141 |
+
raise ValueError(f"`num_layers` must be positive, got {num_layers}.")
|
142 |
+
|
143 |
+
if not (0.0 <= teacher_forcing_ratio <= 1.0):
|
144 |
+
raise ValueError(f"`teacher_forcing_ratio` must be between 0.0 and 1.0, got {teacher_forcing_ratio}.")
|
145 |
+
|
146 |
+
if sequence_length is not None and sequence_length <= 0:
|
147 |
+
raise ValueError(f"`sequence_length` must be positive when specified, got {sequence_length}.")
|
148 |
+
|
149 |
+
# Preprocessing validation
|
150 |
+
valid_preprocessing = ["none", "neural_scaler", "normalizing_flow"]
|
151 |
+
if preprocessing_type not in valid_preprocessing:
|
152 |
+
raise ValueError(
|
153 |
+
f"`preprocessing_type` must be one of {valid_preprocessing}, got {preprocessing_type}."
|
154 |
+
)
|
155 |
+
|
156 |
+
if preprocessing_hidden_dim <= 0:
|
157 |
+
raise ValueError(f"`preprocessing_hidden_dim` must be positive, got {preprocessing_hidden_dim}.")
|
158 |
+
|
159 |
+
if preprocessing_num_layers <= 0:
|
160 |
+
raise ValueError(f"`preprocessing_num_layers` must be positive, got {preprocessing_num_layers}.")
|
161 |
+
|
162 |
+
if flow_coupling_layers <= 0:
|
163 |
+
raise ValueError(f"`flow_coupling_layers` must be positive, got {flow_coupling_layers}.")
|
164 |
+
|
165 |
+
# Set configuration attributes
|
166 |
+
self.input_dim = input_dim
|
167 |
+
self.hidden_dims = hidden_dims
|
168 |
+
self.latent_dim = latent_dim
|
169 |
+
self.activation = activation
|
170 |
+
self.dropout_rate = dropout_rate
|
171 |
+
self.use_batch_norm = use_batch_norm
|
172 |
+
self.tie_weights = tie_weights
|
173 |
+
self.reconstruction_loss = reconstruction_loss
|
174 |
+
self.autoencoder_type = autoencoder_type
|
175 |
+
self.beta = beta
|
176 |
+
self.temperature = temperature
|
177 |
+
self.noise_factor = noise_factor
|
178 |
+
self.rnn_type = rnn_type
|
179 |
+
self.num_layers = num_layers
|
180 |
+
self.bidirectional = bidirectional
|
181 |
+
self.sequence_length = sequence_length
|
182 |
+
self.teacher_forcing_ratio = teacher_forcing_ratio
|
183 |
+
self.use_learnable_preprocessing = use_learnable_preprocessing
|
184 |
+
self.preprocessing_type = preprocessing_type
|
185 |
+
self.preprocessing_hidden_dim = preprocessing_hidden_dim
|
186 |
+
self.preprocessing_num_layers = preprocessing_num_layers
|
187 |
+
self.learn_inverse_preprocessing = learn_inverse_preprocessing
|
188 |
+
self.flow_coupling_layers = flow_coupling_layers
|
189 |
+
|
190 |
+
# Call parent constructor
|
191 |
+
super().__init__(**kwargs)
|
192 |
+
|
193 |
+
@property
|
194 |
+
def decoder_dims(self) -> List[int]:
|
195 |
+
"""Get decoder dimensions (reverse of encoder hidden dims)."""
|
196 |
+
return list(reversed(self.hidden_dims))
|
197 |
+
|
198 |
+
@property
|
199 |
+
def is_variational(self) -> bool:
|
200 |
+
"""Check if this is a variational autoencoder."""
|
201 |
+
return self.autoencoder_type in ["variational", "beta_vae"]
|
202 |
+
|
203 |
+
@property
|
204 |
+
def is_denoising(self) -> bool:
|
205 |
+
"""Check if this is a denoising autoencoder."""
|
206 |
+
return self.autoencoder_type == "denoising"
|
207 |
+
|
208 |
+
@property
|
209 |
+
def is_sparse(self) -> bool:
|
210 |
+
"""Check if this is a sparse autoencoder."""
|
211 |
+
return self.autoencoder_type == "sparse"
|
212 |
+
|
213 |
+
@property
|
214 |
+
def is_contractive(self) -> bool:
|
215 |
+
"""Check if this is a contractive autoencoder."""
|
216 |
+
return self.autoencoder_type == "contractive"
|
217 |
+
|
218 |
+
@property
|
219 |
+
def is_recurrent(self) -> bool:
|
220 |
+
"""Check if this is a recurrent autoencoder."""
|
221 |
+
return self.autoencoder_type == "recurrent"
|
222 |
+
|
223 |
+
@property
|
224 |
+
def rnn_hidden_size(self) -> int:
|
225 |
+
"""Get the RNN hidden size (same as latent_dim for recurrent AE)."""
|
226 |
+
return self.latent_dim
|
227 |
+
|
228 |
+
@property
|
229 |
+
def rnn_output_size(self) -> int:
|
230 |
+
"""Get the RNN output size considering bidirectionality."""
|
231 |
+
return self.latent_dim * (2 if self.bidirectional else 1)
|
232 |
+
|
233 |
+
@property
|
234 |
+
def has_preprocessing(self) -> bool:
|
235 |
+
"""Check if learnable preprocessing is enabled."""
|
236 |
+
return self.use_learnable_preprocessing and self.preprocessing_type != "none"
|
237 |
+
|
238 |
+
@property
|
239 |
+
def is_neural_scaler(self) -> bool:
|
240 |
+
"""Check if using neural scaler preprocessing."""
|
241 |
+
return self.preprocessing_type == "neural_scaler"
|
242 |
+
|
243 |
+
@property
|
244 |
+
def is_normalizing_flow(self) -> bool:
|
245 |
+
"""Check if using normalizing flow preprocessing."""
|
246 |
+
return self.preprocessing_type == "normalizing_flow"
|
247 |
+
|
248 |
+
def to_dict(self):
|
249 |
+
"""
|
250 |
+
Serializes this instance to a Python dictionary.
|
251 |
+
"""
|
252 |
+
output = super().to_dict()
|
253 |
+
return output
|
modeling_autoencoder.py
ADDED
@@ -0,0 +1,1099 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
PyTorch Autoencoder model for Hugging Face Transformers.
|
3 |
+
"""
|
4 |
+
|
5 |
+
import torch
|
6 |
+
import torch.nn as nn
|
7 |
+
import torch.nn.functional as F
|
8 |
+
from typing import Optional, Tuple, Union, Dict, Any, List
|
9 |
+
from dataclasses import dataclass
|
10 |
+
import random
|
11 |
+
|
12 |
+
from transformers import PreTrainedModel
|
13 |
+
from transformers.modeling_outputs import BaseModelOutput
|
14 |
+
from transformers.utils import ModelOutput
|
15 |
+
|
16 |
+
from configuration_autoencoder import AutoencoderConfig
|
17 |
+
|
18 |
+
|
19 |
+
class NeuralScaler(nn.Module):
|
20 |
+
"""Learnable alternative to StandardScaler using neural networks."""
|
21 |
+
|
22 |
+
def __init__(self, config: AutoencoderConfig):
|
23 |
+
super().__init__()
|
24 |
+
self.config = config
|
25 |
+
input_dim = config.input_dim
|
26 |
+
hidden_dim = config.preprocessing_hidden_dim
|
27 |
+
|
28 |
+
# Networks to learn data-dependent statistics
|
29 |
+
self.mean_estimator = nn.Sequential(
|
30 |
+
nn.Linear(input_dim, hidden_dim),
|
31 |
+
nn.ReLU(),
|
32 |
+
nn.Linear(hidden_dim, hidden_dim),
|
33 |
+
nn.ReLU(),
|
34 |
+
nn.Linear(hidden_dim, input_dim)
|
35 |
+
)
|
36 |
+
|
37 |
+
self.std_estimator = nn.Sequential(
|
38 |
+
nn.Linear(input_dim, hidden_dim),
|
39 |
+
nn.ReLU(),
|
40 |
+
nn.Linear(hidden_dim, hidden_dim),
|
41 |
+
nn.ReLU(),
|
42 |
+
nn.Linear(hidden_dim, input_dim),
|
43 |
+
nn.Softplus() # Ensure positive standard deviation
|
44 |
+
)
|
45 |
+
|
46 |
+
# Learnable affine transformation parameters
|
47 |
+
self.weight = nn.Parameter(torch.ones(input_dim))
|
48 |
+
self.bias = nn.Parameter(torch.zeros(input_dim))
|
49 |
+
|
50 |
+
# Running statistics for inference (like BatchNorm)
|
51 |
+
self.register_buffer('running_mean', torch.zeros(input_dim))
|
52 |
+
self.register_buffer('running_std', torch.ones(input_dim))
|
53 |
+
self.register_buffer('num_batches_tracked', torch.tensor(0, dtype=torch.long))
|
54 |
+
|
55 |
+
# Momentum for running statistics
|
56 |
+
self.momentum = 0.1
|
57 |
+
|
58 |
+
def forward(self, x: torch.Tensor, inverse: bool = False) -> Tuple[torch.Tensor, torch.Tensor]:
|
59 |
+
"""
|
60 |
+
Forward pass through neural scaler.
|
61 |
+
|
62 |
+
Args:
|
63 |
+
x: Input tensor (2D or 3D)
|
64 |
+
inverse: Whether to apply inverse transformation
|
65 |
+
|
66 |
+
Returns:
|
67 |
+
Tuple of (transformed_tensor, regularization_loss)
|
68 |
+
"""
|
69 |
+
if inverse:
|
70 |
+
return self._inverse_transform(x)
|
71 |
+
|
72 |
+
# Handle both 2D and 3D tensors
|
73 |
+
original_shape = x.shape
|
74 |
+
if x.dim() == 3:
|
75 |
+
# Reshape (batch, seq, features) -> (batch*seq, features)
|
76 |
+
x = x.view(-1, x.size(-1))
|
77 |
+
|
78 |
+
if self.training:
|
79 |
+
# Training mode: learn statistics from current batch
|
80 |
+
batch_mean = x.mean(dim=0, keepdim=True)
|
81 |
+
batch_std = x.std(dim=0, keepdim=True)
|
82 |
+
|
83 |
+
# Learn data-dependent adjustments
|
84 |
+
learned_mean_adj = self.mean_estimator(batch_mean)
|
85 |
+
learned_std_adj = self.std_estimator(batch_std)
|
86 |
+
|
87 |
+
# Combine batch statistics with learned adjustments
|
88 |
+
effective_mean = batch_mean + learned_mean_adj
|
89 |
+
effective_std = batch_std + learned_std_adj + 1e-8
|
90 |
+
|
91 |
+
# Update running statistics
|
92 |
+
with torch.no_grad():
|
93 |
+
self.num_batches_tracked += 1
|
94 |
+
if self.num_batches_tracked == 1:
|
95 |
+
self.running_mean.copy_(batch_mean.squeeze())
|
96 |
+
self.running_std.copy_(batch_std.squeeze())
|
97 |
+
else:
|
98 |
+
self.running_mean.mul_(1 - self.momentum).add_(batch_mean.squeeze(), alpha=self.momentum)
|
99 |
+
self.running_std.mul_(1 - self.momentum).add_(batch_std.squeeze(), alpha=self.momentum)
|
100 |
+
else:
|
101 |
+
# Inference mode: use running statistics
|
102 |
+
effective_mean = self.running_mean.unsqueeze(0)
|
103 |
+
effective_std = self.running_std.unsqueeze(0) + 1e-8
|
104 |
+
|
105 |
+
# Normalize
|
106 |
+
normalized = (x - effective_mean) / effective_std
|
107 |
+
|
108 |
+
# Apply learnable affine transformation
|
109 |
+
transformed = normalized * self.weight + self.bias
|
110 |
+
|
111 |
+
# Reshape back to original shape if needed
|
112 |
+
if len(original_shape) == 3:
|
113 |
+
transformed = transformed.view(original_shape)
|
114 |
+
|
115 |
+
# Regularization loss to encourage meaningful learning
|
116 |
+
reg_loss = 0.01 * (self.weight.var() + self.bias.var())
|
117 |
+
|
118 |
+
return transformed, reg_loss
|
119 |
+
|
120 |
+
def _inverse_transform(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
|
121 |
+
"""Apply inverse transformation to get back original scale."""
|
122 |
+
if not self.config.learn_inverse_preprocessing:
|
123 |
+
return x, torch.tensor(0.0, device=x.device)
|
124 |
+
|
125 |
+
# Handle both 2D and 3D tensors
|
126 |
+
original_shape = x.shape
|
127 |
+
if x.dim() == 3:
|
128 |
+
# Reshape (batch, seq, features) -> (batch*seq, features)
|
129 |
+
x = x.view(-1, x.size(-1))
|
130 |
+
|
131 |
+
# Reverse affine transformation
|
132 |
+
x = (x - self.bias) / (self.weight + 1e-8)
|
133 |
+
|
134 |
+
# Reverse normalization using running statistics
|
135 |
+
effective_mean = self.running_mean.unsqueeze(0)
|
136 |
+
effective_std = self.running_std.unsqueeze(0) + 1e-8
|
137 |
+
x = x * effective_std + effective_mean
|
138 |
+
|
139 |
+
# Reshape back to original shape if needed
|
140 |
+
if len(original_shape) == 3:
|
141 |
+
x = x.view(original_shape)
|
142 |
+
|
143 |
+
return x, torch.tensor(0.0, device=x.device)
|
144 |
+
|
145 |
+
|
146 |
+
class CouplingLayer(nn.Module):
|
147 |
+
"""Coupling layer for normalizing flows."""
|
148 |
+
|
149 |
+
def __init__(self, input_dim: int, hidden_dim: int = 64, mask_type: str = "alternating"):
|
150 |
+
super().__init__()
|
151 |
+
self.input_dim = input_dim
|
152 |
+
self.hidden_dim = hidden_dim
|
153 |
+
|
154 |
+
# Create mask for coupling
|
155 |
+
if mask_type == "alternating":
|
156 |
+
self.register_buffer('mask', torch.arange(input_dim) % 2)
|
157 |
+
elif mask_type == "half":
|
158 |
+
mask = torch.zeros(input_dim)
|
159 |
+
mask[:input_dim // 2] = 1
|
160 |
+
self.register_buffer('mask', mask)
|
161 |
+
else:
|
162 |
+
raise ValueError(f"Unknown mask type: {mask_type}")
|
163 |
+
|
164 |
+
# Scale and translation networks
|
165 |
+
masked_dim = int(self.mask.sum().item())
|
166 |
+
unmasked_dim = input_dim - masked_dim
|
167 |
+
|
168 |
+
self.scale_net = nn.Sequential(
|
169 |
+
nn.Linear(masked_dim, hidden_dim),
|
170 |
+
nn.ReLU(),
|
171 |
+
nn.Linear(hidden_dim, hidden_dim),
|
172 |
+
nn.ReLU(),
|
173 |
+
nn.Linear(hidden_dim, unmasked_dim),
|
174 |
+
nn.Tanh() # Bounded output for stability
|
175 |
+
)
|
176 |
+
|
177 |
+
self.translate_net = nn.Sequential(
|
178 |
+
nn.Linear(masked_dim, hidden_dim),
|
179 |
+
nn.ReLU(),
|
180 |
+
nn.Linear(hidden_dim, hidden_dim),
|
181 |
+
nn.ReLU(),
|
182 |
+
nn.Linear(hidden_dim, unmasked_dim)
|
183 |
+
)
|
184 |
+
|
185 |
+
def forward(self, x: torch.Tensor, inverse: bool = False) -> Tuple[torch.Tensor, torch.Tensor]:
|
186 |
+
"""
|
187 |
+
Forward pass through coupling layer.
|
188 |
+
|
189 |
+
Args:
|
190 |
+
x: Input tensor
|
191 |
+
inverse: Whether to apply inverse transformation
|
192 |
+
|
193 |
+
Returns:
|
194 |
+
Tuple of (transformed_tensor, log_determinant)
|
195 |
+
"""
|
196 |
+
mask = self.mask.bool()
|
197 |
+
x_masked = x[:, mask]
|
198 |
+
x_unmasked = x[:, ~mask]
|
199 |
+
|
200 |
+
# Compute scale and translation
|
201 |
+
s = self.scale_net(x_masked)
|
202 |
+
t = self.translate_net(x_masked)
|
203 |
+
|
204 |
+
if not inverse:
|
205 |
+
# Forward transformation
|
206 |
+
y_unmasked = x_unmasked * torch.exp(s) + t
|
207 |
+
log_det = s.sum(dim=1)
|
208 |
+
else:
|
209 |
+
# Inverse transformation
|
210 |
+
y_unmasked = (x_unmasked - t) * torch.exp(-s)
|
211 |
+
log_det = -s.sum(dim=1)
|
212 |
+
|
213 |
+
# Reconstruct output
|
214 |
+
y = torch.zeros_like(x)
|
215 |
+
y[:, mask] = x_masked
|
216 |
+
y[:, ~mask] = y_unmasked
|
217 |
+
|
218 |
+
return y, log_det
|
219 |
+
|
220 |
+
|
221 |
+
class NormalizingFlowPreprocessor(nn.Module):
|
222 |
+
"""Normalizing flow for learnable data preprocessing."""
|
223 |
+
|
224 |
+
def __init__(self, config: AutoencoderConfig):
|
225 |
+
super().__init__()
|
226 |
+
self.config = config
|
227 |
+
input_dim = config.input_dim
|
228 |
+
hidden_dim = config.preprocessing_hidden_dim
|
229 |
+
num_layers = config.flow_coupling_layers
|
230 |
+
|
231 |
+
# Create coupling layers with alternating masks
|
232 |
+
self.layers = nn.ModuleList()
|
233 |
+
for i in range(num_layers):
|
234 |
+
mask_type = "alternating" if i % 2 == 0 else "half"
|
235 |
+
self.layers.append(CouplingLayer(input_dim, hidden_dim, mask_type))
|
236 |
+
|
237 |
+
# Optional: Add batch normalization between layers
|
238 |
+
if config.use_batch_norm:
|
239 |
+
self.batch_norms = nn.ModuleList([
|
240 |
+
nn.BatchNorm1d(input_dim) for _ in range(num_layers - 1)
|
241 |
+
])
|
242 |
+
else:
|
243 |
+
self.batch_norms = None
|
244 |
+
|
245 |
+
def forward(self, x: torch.Tensor, inverse: bool = False) -> Tuple[torch.Tensor, torch.Tensor]:
|
246 |
+
"""
|
247 |
+
Forward pass through normalizing flow.
|
248 |
+
|
249 |
+
Args:
|
250 |
+
x: Input tensor (2D or 3D)
|
251 |
+
inverse: Whether to apply inverse transformation
|
252 |
+
|
253 |
+
Returns:
|
254 |
+
Tuple of (transformed_tensor, total_log_determinant)
|
255 |
+
"""
|
256 |
+
# Handle both 2D and 3D tensors
|
257 |
+
original_shape = x.shape
|
258 |
+
if x.dim() == 3:
|
259 |
+
# Reshape (batch, seq, features) -> (batch*seq, features)
|
260 |
+
x = x.view(-1, x.size(-1))
|
261 |
+
|
262 |
+
log_det_total = torch.zeros(x.size(0), device=x.device)
|
263 |
+
|
264 |
+
if not inverse:
|
265 |
+
# Forward pass
|
266 |
+
for i, layer in enumerate(self.layers):
|
267 |
+
x, log_det = layer(x, inverse=False)
|
268 |
+
log_det_total += log_det
|
269 |
+
|
270 |
+
# Apply batch normalization (except for last layer)
|
271 |
+
if self.batch_norms and i < len(self.layers) - 1:
|
272 |
+
x = self.batch_norms[i](x)
|
273 |
+
else:
|
274 |
+
# Inverse pass
|
275 |
+
for i, layer in enumerate(reversed(self.layers)):
|
276 |
+
# Reverse batch normalization (except for first layer in reverse)
|
277 |
+
if self.batch_norms and i > 0:
|
278 |
+
# Note: This is approximate inverse of batch norm
|
279 |
+
bn_idx = len(self.layers) - 1 - i
|
280 |
+
x = self.batch_norms[bn_idx](x)
|
281 |
+
|
282 |
+
x, log_det = layer(x, inverse=True)
|
283 |
+
log_det_total += log_det
|
284 |
+
|
285 |
+
# Reshape back to original shape if needed
|
286 |
+
if len(original_shape) == 3:
|
287 |
+
x = x.view(original_shape)
|
288 |
+
|
289 |
+
# Convert log determinant to regularization loss
|
290 |
+
# Encourage the flow to preserve information (log_det close to 0)
|
291 |
+
reg_loss = 0.01 * log_det_total.abs().mean()
|
292 |
+
|
293 |
+
return x, reg_loss
|
294 |
+
|
295 |
+
|
296 |
+
class LearnablePreprocessor(nn.Module):
|
297 |
+
"""Unified interface for learnable preprocessing methods."""
|
298 |
+
|
299 |
+
def __init__(self, config: AutoencoderConfig):
|
300 |
+
super().__init__()
|
301 |
+
self.config = config
|
302 |
+
|
303 |
+
if not config.has_preprocessing:
|
304 |
+
self.preprocessor = nn.Identity()
|
305 |
+
elif config.is_neural_scaler:
|
306 |
+
self.preprocessor = NeuralScaler(config)
|
307 |
+
elif config.is_normalizing_flow:
|
308 |
+
self.preprocessor = NormalizingFlowPreprocessor(config)
|
309 |
+
else:
|
310 |
+
raise ValueError(f"Unknown preprocessing type: {config.preprocessing_type}")
|
311 |
+
|
312 |
+
def forward(self, x: torch.Tensor, inverse: bool = False) -> Tuple[torch.Tensor, torch.Tensor]:
|
313 |
+
"""
|
314 |
+
Apply preprocessing transformation.
|
315 |
+
|
316 |
+
Args:
|
317 |
+
x: Input tensor
|
318 |
+
inverse: Whether to apply inverse transformation
|
319 |
+
|
320 |
+
Returns:
|
321 |
+
Tuple of (transformed_tensor, regularization_loss)
|
322 |
+
"""
|
323 |
+
if isinstance(self.preprocessor, nn.Identity):
|
324 |
+
return x, torch.tensor(0.0, device=x.device)
|
325 |
+
|
326 |
+
return self.preprocessor(x, inverse=inverse)
|
327 |
+
|
328 |
+
|
329 |
+
@dataclass
|
330 |
+
class AutoencoderOutput(ModelOutput):
|
331 |
+
"""
|
332 |
+
Output type of AutoencoderModel.
|
333 |
+
|
334 |
+
Args:
|
335 |
+
last_hidden_state (torch.FloatTensor): The latent representation of the input.
|
336 |
+
reconstructed (torch.FloatTensor, optional): The reconstructed input.
|
337 |
+
hidden_states (tuple(torch.FloatTensor), optional): Hidden states of the encoder layers.
|
338 |
+
attentions (tuple(torch.FloatTensor), optional): Not used in basic autoencoder.
|
339 |
+
preprocessing_loss (torch.FloatTensor, optional): Loss from learnable preprocessing.
|
340 |
+
"""
|
341 |
+
|
342 |
+
last_hidden_state: torch.FloatTensor = None
|
343 |
+
reconstructed: Optional[torch.FloatTensor] = None
|
344 |
+
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
|
345 |
+
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
346 |
+
preprocessing_loss: Optional[torch.FloatTensor] = None
|
347 |
+
|
348 |
+
|
349 |
+
@dataclass
|
350 |
+
class AutoencoderForReconstructionOutput(ModelOutput):
|
351 |
+
"""
|
352 |
+
Output type of AutoencoderForReconstruction.
|
353 |
+
|
354 |
+
Args:
|
355 |
+
loss (torch.FloatTensor, optional): The reconstruction loss.
|
356 |
+
reconstructed (torch.FloatTensor): The reconstructed input.
|
357 |
+
last_hidden_state (torch.FloatTensor): The latent representation.
|
358 |
+
hidden_states (tuple(torch.FloatTensor), optional): Hidden states of the encoder layers.
|
359 |
+
preprocessing_loss (torch.FloatTensor, optional): Loss from learnable preprocessing.
|
360 |
+
"""
|
361 |
+
|
362 |
+
loss: Optional[torch.FloatTensor] = None
|
363 |
+
reconstructed: torch.FloatTensor = None
|
364 |
+
last_hidden_state: torch.FloatTensor = None
|
365 |
+
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
|
366 |
+
preprocessing_loss: Optional[torch.FloatTensor] = None
|
367 |
+
|
368 |
+
|
369 |
+
class AutoencoderEncoder(nn.Module):
|
370 |
+
"""Encoder part of the autoencoder."""
|
371 |
+
|
372 |
+
def __init__(self, config: AutoencoderConfig):
|
373 |
+
super().__init__()
|
374 |
+
self.config = config
|
375 |
+
|
376 |
+
# Build encoder layers
|
377 |
+
layers = []
|
378 |
+
input_dim = config.input_dim
|
379 |
+
|
380 |
+
for hidden_dim in config.hidden_dims:
|
381 |
+
layers.append(nn.Linear(input_dim, hidden_dim))
|
382 |
+
|
383 |
+
if config.use_batch_norm:
|
384 |
+
layers.append(nn.BatchNorm1d(hidden_dim))
|
385 |
+
|
386 |
+
layers.append(self._get_activation(config.activation))
|
387 |
+
|
388 |
+
if config.dropout_rate > 0:
|
389 |
+
layers.append(nn.Dropout(config.dropout_rate))
|
390 |
+
|
391 |
+
input_dim = hidden_dim
|
392 |
+
|
393 |
+
self.encoder = nn.Sequential(*layers)
|
394 |
+
|
395 |
+
# For variational autoencoders, we need separate layers for mean and log variance
|
396 |
+
if config.is_variational:
|
397 |
+
self.fc_mu = nn.Linear(input_dim, config.latent_dim)
|
398 |
+
self.fc_logvar = nn.Linear(input_dim, config.latent_dim)
|
399 |
+
else:
|
400 |
+
# Standard encoder output
|
401 |
+
self.fc_out = nn.Linear(input_dim, config.latent_dim)
|
402 |
+
|
403 |
+
def _get_activation(self, activation: str) -> nn.Module:
|
404 |
+
"""Get activation function by name."""
|
405 |
+
activations = {
|
406 |
+
"relu": nn.ReLU(),
|
407 |
+
"tanh": nn.Tanh(),
|
408 |
+
"sigmoid": nn.Sigmoid(),
|
409 |
+
"leaky_relu": nn.LeakyReLU(),
|
410 |
+
"gelu": nn.GELU(),
|
411 |
+
"swish": nn.SiLU(),
|
412 |
+
"silu": nn.SiLU(),
|
413 |
+
"elu": nn.ELU(),
|
414 |
+
"prelu": nn.PReLU(),
|
415 |
+
"relu6": nn.ReLU6(),
|
416 |
+
"hardtanh": nn.Hardtanh(),
|
417 |
+
"hardsigmoid": nn.Hardsigmoid(),
|
418 |
+
"hardswish": nn.Hardswish(),
|
419 |
+
"mish": nn.Mish(),
|
420 |
+
"softplus": nn.Softplus(),
|
421 |
+
"softsign": nn.Softsign(),
|
422 |
+
"tanhshrink": nn.Tanhshrink(),
|
423 |
+
"threshold": nn.Threshold(threshold=0.1, value=0),
|
424 |
+
}
|
425 |
+
return activations[activation]
|
426 |
+
|
427 |
+
def forward(self, x: torch.Tensor) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor, torch.Tensor]]:
|
428 |
+
"""Forward pass through encoder."""
|
429 |
+
# Add noise for denoising autoencoders
|
430 |
+
if self.config.is_denoising and self.training:
|
431 |
+
noise = torch.randn_like(x) * self.config.noise_factor
|
432 |
+
x = x + noise
|
433 |
+
|
434 |
+
encoded = self.encoder(x)
|
435 |
+
|
436 |
+
if self.config.is_variational:
|
437 |
+
# Variational autoencoder: return mean, log variance, and sampled latent
|
438 |
+
mu = self.fc_mu(encoded)
|
439 |
+
logvar = self.fc_logvar(encoded)
|
440 |
+
|
441 |
+
# Reparameterization trick
|
442 |
+
if self.training:
|
443 |
+
std = torch.exp(0.5 * logvar)
|
444 |
+
eps = torch.randn_like(std)
|
445 |
+
z = mu + eps * std
|
446 |
+
else:
|
447 |
+
z = mu # Use mean during inference
|
448 |
+
|
449 |
+
return z, mu, logvar
|
450 |
+
else:
|
451 |
+
# Standard autoencoder
|
452 |
+
latent = self.fc_out(encoded)
|
453 |
+
|
454 |
+
# Add sparsity constraint for sparse autoencoders
|
455 |
+
if self.config.is_sparse and self.training:
|
456 |
+
# Apply L1 regularization to encourage sparsity
|
457 |
+
latent = F.relu(latent) # Ensure non-negative activations
|
458 |
+
|
459 |
+
return latent
|
460 |
+
|
461 |
+
|
462 |
+
class AutoencoderDecoder(nn.Module):
|
463 |
+
"""Decoder part of the autoencoder."""
|
464 |
+
|
465 |
+
def __init__(self, config: AutoencoderConfig):
|
466 |
+
super().__init__()
|
467 |
+
self.config = config
|
468 |
+
|
469 |
+
# Build decoder layers (reverse of encoder)
|
470 |
+
layers = []
|
471 |
+
input_dim = config.latent_dim
|
472 |
+
decoder_dims = config.decoder_dims + [config.input_dim]
|
473 |
+
|
474 |
+
for i, hidden_dim in enumerate(decoder_dims):
|
475 |
+
layers.append(nn.Linear(input_dim, hidden_dim))
|
476 |
+
|
477 |
+
# Don't add batch norm, activation, or dropout to the final layer
|
478 |
+
if i < len(decoder_dims) - 1:
|
479 |
+
if config.use_batch_norm:
|
480 |
+
layers.append(nn.BatchNorm1d(hidden_dim))
|
481 |
+
|
482 |
+
layers.append(self._get_activation(config.activation))
|
483 |
+
|
484 |
+
if config.dropout_rate > 0:
|
485 |
+
layers.append(nn.Dropout(config.dropout_rate))
|
486 |
+
else:
|
487 |
+
# Final layer - add appropriate activation based on reconstruction loss
|
488 |
+
if config.reconstruction_loss == "bce":
|
489 |
+
layers.append(nn.Sigmoid())
|
490 |
+
|
491 |
+
input_dim = hidden_dim
|
492 |
+
|
493 |
+
self.decoder = nn.Sequential(*layers)
|
494 |
+
|
495 |
+
def _get_activation(self, activation: str) -> nn.Module:
|
496 |
+
"""Get activation function by name."""
|
497 |
+
activations = {
|
498 |
+
"relu": nn.ReLU(),
|
499 |
+
"tanh": nn.Tanh(),
|
500 |
+
"sigmoid": nn.Sigmoid(),
|
501 |
+
"leaky_relu": nn.LeakyReLU(),
|
502 |
+
"gelu": nn.GELU(),
|
503 |
+
"swish": nn.SiLU(),
|
504 |
+
"silu": nn.SiLU(),
|
505 |
+
"elu": nn.ELU(),
|
506 |
+
"prelu": nn.PReLU(),
|
507 |
+
"relu6": nn.ReLU6(),
|
508 |
+
"hardtanh": nn.Hardtanh(),
|
509 |
+
"hardsigmoid": nn.Hardsigmoid(),
|
510 |
+
"hardswish": nn.Hardswish(),
|
511 |
+
"mish": nn.Mish(),
|
512 |
+
"softplus": nn.Softplus(),
|
513 |
+
"softsign": nn.Softsign(),
|
514 |
+
"tanhshrink": nn.Tanhshrink(),
|
515 |
+
"threshold": nn.Threshold(threshold=0.1, value=0),
|
516 |
+
}
|
517 |
+
return activations[activation]
|
518 |
+
|
519 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
520 |
+
"""Forward pass through decoder."""
|
521 |
+
return self.decoder(x)
|
522 |
+
|
523 |
+
|
524 |
+
class RecurrentEncoder(nn.Module):
|
525 |
+
"""Recurrent encoder for sequence data."""
|
526 |
+
|
527 |
+
def __init__(self, config: AutoencoderConfig):
|
528 |
+
super().__init__()
|
529 |
+
self.config = config
|
530 |
+
|
531 |
+
# Get RNN class
|
532 |
+
if config.rnn_type == "lstm":
|
533 |
+
rnn_class = nn.LSTM
|
534 |
+
elif config.rnn_type == "gru":
|
535 |
+
rnn_class = nn.GRU
|
536 |
+
elif config.rnn_type == "rnn":
|
537 |
+
rnn_class = nn.RNN
|
538 |
+
else:
|
539 |
+
raise ValueError(f"Unknown RNN type: {config.rnn_type}")
|
540 |
+
|
541 |
+
# Create RNN layers
|
542 |
+
self.rnn = rnn_class(
|
543 |
+
input_size=config.input_dim,
|
544 |
+
hidden_size=config.latent_dim,
|
545 |
+
num_layers=config.num_layers,
|
546 |
+
batch_first=True,
|
547 |
+
dropout=config.dropout_rate if config.num_layers > 1 else 0,
|
548 |
+
bidirectional=config.bidirectional
|
549 |
+
)
|
550 |
+
|
551 |
+
# Projection layer for bidirectional RNN
|
552 |
+
if config.bidirectional:
|
553 |
+
self.projection = nn.Linear(config.latent_dim * 2, config.latent_dim)
|
554 |
+
else:
|
555 |
+
self.projection = None
|
556 |
+
|
557 |
+
# Batch normalization
|
558 |
+
if config.use_batch_norm:
|
559 |
+
self.batch_norm = nn.BatchNorm1d(config.latent_dim)
|
560 |
+
else:
|
561 |
+
self.batch_norm = None
|
562 |
+
|
563 |
+
# Dropout
|
564 |
+
if config.dropout_rate > 0:
|
565 |
+
self.dropout = nn.Dropout(config.dropout_rate)
|
566 |
+
else:
|
567 |
+
self.dropout = None
|
568 |
+
|
569 |
+
def forward(self, x: torch.Tensor, lengths: Optional[torch.Tensor] = None) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor, torch.Tensor]]:
|
570 |
+
"""
|
571 |
+
Forward pass through recurrent encoder.
|
572 |
+
|
573 |
+
Args:
|
574 |
+
x: Input tensor of shape (batch_size, seq_len, input_dim)
|
575 |
+
lengths: Sequence lengths for packed sequences (optional)
|
576 |
+
|
577 |
+
Returns:
|
578 |
+
Encoded representation or tuple for VAE
|
579 |
+
"""
|
580 |
+
batch_size, seq_len, _ = x.shape
|
581 |
+
|
582 |
+
# Add noise for denoising autoencoders
|
583 |
+
if self.config.is_denoising and self.training:
|
584 |
+
noise = torch.randn_like(x) * self.config.noise_factor
|
585 |
+
x = x + noise
|
586 |
+
|
587 |
+
# Pack sequences if lengths provided
|
588 |
+
if lengths is not None:
|
589 |
+
x = nn.utils.rnn.pack_padded_sequence(x, lengths, batch_first=True, enforce_sorted=False)
|
590 |
+
|
591 |
+
# RNN forward pass
|
592 |
+
if self.config.rnn_type == "lstm":
|
593 |
+
output, (hidden, cell) = self.rnn(x)
|
594 |
+
else:
|
595 |
+
output, hidden = self.rnn(x)
|
596 |
+
cell = None
|
597 |
+
|
598 |
+
# Unpack if necessary
|
599 |
+
if lengths is not None:
|
600 |
+
output, _ = nn.utils.rnn.pad_packed_sequence(output, batch_first=True)
|
601 |
+
|
602 |
+
# Use last hidden state as encoding
|
603 |
+
if self.config.bidirectional:
|
604 |
+
# Concatenate forward and backward hidden states
|
605 |
+
hidden = hidden.view(self.config.num_layers, 2, batch_size, self.config.latent_dim)
|
606 |
+
hidden = hidden[-1] # Take last layer
|
607 |
+
hidden = hidden.transpose(0, 1).contiguous().view(batch_size, -1) # Concatenate directions
|
608 |
+
|
609 |
+
# Project to latent dimension
|
610 |
+
if self.projection:
|
611 |
+
hidden = self.projection(hidden)
|
612 |
+
else:
|
613 |
+
hidden = hidden[-1] # Take last layer
|
614 |
+
|
615 |
+
# Apply batch normalization
|
616 |
+
if self.batch_norm:
|
617 |
+
hidden = self.batch_norm(hidden)
|
618 |
+
|
619 |
+
# Apply dropout
|
620 |
+
if self.dropout and self.training:
|
621 |
+
hidden = self.dropout(hidden)
|
622 |
+
|
623 |
+
# Handle variational encoding
|
624 |
+
if self.config.is_variational:
|
625 |
+
# Split hidden into mean and log variance
|
626 |
+
mu = hidden[:, :self.config.latent_dim // 2]
|
627 |
+
logvar = hidden[:, self.config.latent_dim // 2:]
|
628 |
+
|
629 |
+
# Reparameterization trick
|
630 |
+
if self.training:
|
631 |
+
std = torch.exp(0.5 * logvar)
|
632 |
+
eps = torch.randn_like(std)
|
633 |
+
z = mu + eps * std
|
634 |
+
else:
|
635 |
+
z = mu
|
636 |
+
|
637 |
+
return z, mu, logvar
|
638 |
+
else:
|
639 |
+
return hidden
|
640 |
+
|
641 |
+
|
642 |
+
class RecurrentDecoder(nn.Module):
|
643 |
+
"""Recurrent decoder for sequence data."""
|
644 |
+
|
645 |
+
def __init__(self, config: AutoencoderConfig):
|
646 |
+
super().__init__()
|
647 |
+
self.config = config
|
648 |
+
|
649 |
+
# Get RNN class
|
650 |
+
if config.rnn_type == "lstm":
|
651 |
+
rnn_class = nn.LSTM
|
652 |
+
elif config.rnn_type == "gru":
|
653 |
+
rnn_class = nn.GRU
|
654 |
+
elif config.rnn_type == "rnn":
|
655 |
+
rnn_class = nn.RNN
|
656 |
+
else:
|
657 |
+
raise ValueError(f"Unknown RNN type: {config.rnn_type}")
|
658 |
+
|
659 |
+
# Create RNN layers
|
660 |
+
self.rnn = rnn_class(
|
661 |
+
input_size=config.latent_dim,
|
662 |
+
hidden_size=config.latent_dim,
|
663 |
+
num_layers=config.num_layers,
|
664 |
+
batch_first=True,
|
665 |
+
dropout=config.dropout_rate if config.num_layers > 1 else 0,
|
666 |
+
bidirectional=False # Decoder is always unidirectional
|
667 |
+
)
|
668 |
+
|
669 |
+
# Output projection
|
670 |
+
self.output_projection = nn.Linear(config.latent_dim, config.input_dim)
|
671 |
+
|
672 |
+
# Batch normalization
|
673 |
+
if config.use_batch_norm:
|
674 |
+
self.batch_norm = nn.BatchNorm1d(config.latent_dim)
|
675 |
+
else:
|
676 |
+
self.batch_norm = None
|
677 |
+
|
678 |
+
# Dropout
|
679 |
+
if config.dropout_rate > 0:
|
680 |
+
self.dropout = nn.Dropout(config.dropout_rate)
|
681 |
+
else:
|
682 |
+
self.dropout = None
|
683 |
+
|
684 |
+
def forward(self, z: torch.Tensor, target_length: int, target_sequence: Optional[torch.Tensor] = None) -> torch.Tensor:
|
685 |
+
"""
|
686 |
+
Forward pass through recurrent decoder.
|
687 |
+
|
688 |
+
Args:
|
689 |
+
z: Latent representation of shape (batch_size, latent_dim)
|
690 |
+
target_length: Length of sequence to generate
|
691 |
+
target_sequence: Target sequence for teacher forcing (optional)
|
692 |
+
|
693 |
+
Returns:
|
694 |
+
Decoded sequence of shape (batch_size, seq_len, input_dim)
|
695 |
+
"""
|
696 |
+
batch_size = z.size(0)
|
697 |
+
device = z.device
|
698 |
+
|
699 |
+
# Initialize hidden state with latent representation
|
700 |
+
if self.config.rnn_type == "lstm":
|
701 |
+
h_0 = z.unsqueeze(0).repeat(self.config.num_layers, 1, 1)
|
702 |
+
c_0 = torch.zeros_like(h_0)
|
703 |
+
hidden = (h_0, c_0)
|
704 |
+
else:
|
705 |
+
hidden = z.unsqueeze(0).repeat(self.config.num_layers, 1, 1)
|
706 |
+
|
707 |
+
outputs = []
|
708 |
+
|
709 |
+
# Initialize input (can be learned or zero)
|
710 |
+
current_input = torch.zeros(batch_size, 1, self.config.latent_dim, device=device)
|
711 |
+
|
712 |
+
for t in range(target_length):
|
713 |
+
# Teacher forcing decision
|
714 |
+
use_teacher_forcing = (target_sequence is not None and
|
715 |
+
self.training and
|
716 |
+
random.random() < self.config.teacher_forcing_ratio)
|
717 |
+
|
718 |
+
if use_teacher_forcing and t > 0:
|
719 |
+
# Use previous target as input
|
720 |
+
current_input = target_sequence[:, t-1:t, :]
|
721 |
+
# Project to latent dimension if needed
|
722 |
+
if current_input.size(-1) != self.config.latent_dim:
|
723 |
+
current_input = torch.zeros(batch_size, 1, self.config.latent_dim, device=device)
|
724 |
+
|
725 |
+
# RNN forward step
|
726 |
+
if self.config.rnn_type == "lstm":
|
727 |
+
output, hidden = self.rnn(current_input, hidden)
|
728 |
+
else:
|
729 |
+
output, hidden = self.rnn(current_input, hidden)
|
730 |
+
|
731 |
+
# Apply batch normalization and dropout
|
732 |
+
output_flat = output.squeeze(1) # Remove sequence dimension
|
733 |
+
|
734 |
+
if self.batch_norm:
|
735 |
+
output_flat = self.batch_norm(output_flat)
|
736 |
+
|
737 |
+
if self.dropout and self.training:
|
738 |
+
output_flat = self.dropout(output_flat)
|
739 |
+
|
740 |
+
# Project to output dimension
|
741 |
+
step_output = self.output_projection(output_flat)
|
742 |
+
outputs.append(step_output.unsqueeze(1))
|
743 |
+
|
744 |
+
# Use output as next input (for non-teacher forcing)
|
745 |
+
if not use_teacher_forcing:
|
746 |
+
# Project output back to latent dimension for next step
|
747 |
+
current_input = torch.zeros(batch_size, 1, self.config.latent_dim, device=device)
|
748 |
+
|
749 |
+
# Concatenate all outputs
|
750 |
+
return torch.cat(outputs, dim=1)
|
751 |
+
|
752 |
+
|
753 |
+
class AutoencoderModel(PreTrainedModel):
|
754 |
+
"""
|
755 |
+
The bare Autoencoder Model transformer outputting raw hidden-states without any specific head on top.
|
756 |
+
|
757 |
+
This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the
|
758 |
+
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
759 |
+
etc.)
|
760 |
+
|
761 |
+
This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the
|
762 |
+
PyTorch documentation for all matter related to general usage and behavior.
|
763 |
+
"""
|
764 |
+
|
765 |
+
config_class = AutoencoderConfig
|
766 |
+
base_model_prefix = "autoencoder"
|
767 |
+
supports_gradient_checkpointing = False
|
768 |
+
|
769 |
+
def __init__(self, config: AutoencoderConfig):
|
770 |
+
super().__init__(config)
|
771 |
+
self.config = config
|
772 |
+
|
773 |
+
# Initialize learnable preprocessing
|
774 |
+
if config.has_preprocessing:
|
775 |
+
self.preprocessor = LearnablePreprocessor(config)
|
776 |
+
else:
|
777 |
+
self.preprocessor = None
|
778 |
+
|
779 |
+
# Initialize encoder and decoder based on type
|
780 |
+
if config.is_recurrent:
|
781 |
+
self.encoder = RecurrentEncoder(config)
|
782 |
+
self.decoder = RecurrentDecoder(config)
|
783 |
+
else:
|
784 |
+
self.encoder = AutoencoderEncoder(config)
|
785 |
+
self.decoder = AutoencoderDecoder(config)
|
786 |
+
|
787 |
+
# Tie weights if specified
|
788 |
+
if config.tie_weights:
|
789 |
+
self._tie_weights()
|
790 |
+
|
791 |
+
# Initialize weights
|
792 |
+
self.post_init()
|
793 |
+
|
794 |
+
def _tie_weights(self):
|
795 |
+
"""Tie encoder and decoder weights (transpose relationship)."""
|
796 |
+
# This is a simplified weight tying - in practice, you might want more sophisticated tying
|
797 |
+
pass
|
798 |
+
|
799 |
+
def get_input_embeddings(self):
|
800 |
+
"""Get input embeddings (not applicable for basic autoencoder)."""
|
801 |
+
return None
|
802 |
+
|
803 |
+
def set_input_embeddings(self, value):
|
804 |
+
"""Set input embeddings (not applicable for basic autoencoder)."""
|
805 |
+
pass
|
806 |
+
|
807 |
+
def forward(
|
808 |
+
self,
|
809 |
+
input_values: torch.Tensor,
|
810 |
+
sequence_lengths: Optional[torch.Tensor] = None,
|
811 |
+
target_length: Optional[int] = None,
|
812 |
+
output_hidden_states: Optional[bool] = None,
|
813 |
+
return_dict: Optional[bool] = None,
|
814 |
+
) -> Union[Tuple[torch.Tensor], AutoencoderOutput]:
|
815 |
+
"""
|
816 |
+
Forward pass through the autoencoder.
|
817 |
+
|
818 |
+
Args:
|
819 |
+
input_values (torch.Tensor): Input tensor. Shape depends on autoencoder type:
|
820 |
+
- Standard: (batch_size, input_dim)
|
821 |
+
- Recurrent: (batch_size, seq_len, input_dim)
|
822 |
+
sequence_lengths (torch.Tensor, optional): Sequence lengths for recurrent AE.
|
823 |
+
target_length (int, optional): Target sequence length for recurrent decoder.
|
824 |
+
output_hidden_states (bool, optional): Whether to return hidden states.
|
825 |
+
return_dict (bool, optional): Whether to return a ModelOutput instead of a plain tuple.
|
826 |
+
|
827 |
+
Returns:
|
828 |
+
AutoencoderOutput or tuple: The model outputs.
|
829 |
+
"""
|
830 |
+
output_hidden_states = (
|
831 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
832 |
+
)
|
833 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
834 |
+
|
835 |
+
# Apply learnable preprocessing
|
836 |
+
preprocessing_loss = torch.tensor(0.0, device=input_values.device)
|
837 |
+
if self.preprocessor is not None:
|
838 |
+
input_values, preprocessing_loss = self.preprocessor(input_values, inverse=False)
|
839 |
+
|
840 |
+
# Handle different autoencoder types
|
841 |
+
if self.config.is_recurrent:
|
842 |
+
# Recurrent autoencoder
|
843 |
+
if sequence_lengths is not None:
|
844 |
+
encoder_output = self.encoder(input_values, sequence_lengths)
|
845 |
+
else:
|
846 |
+
encoder_output = self.encoder(input_values)
|
847 |
+
|
848 |
+
if self.config.is_variational:
|
849 |
+
latent, mu, logvar = encoder_output
|
850 |
+
self._mu = mu
|
851 |
+
self._logvar = logvar
|
852 |
+
else:
|
853 |
+
latent = encoder_output
|
854 |
+
self._mu = None
|
855 |
+
self._logvar = None
|
856 |
+
|
857 |
+
# Determine target length for decoder
|
858 |
+
if target_length is None:
|
859 |
+
if self.config.sequence_length is not None:
|
860 |
+
target_length = self.config.sequence_length
|
861 |
+
else:
|
862 |
+
target_length = input_values.size(1) # Use input sequence length
|
863 |
+
|
864 |
+
# Decode latent back to sequence space
|
865 |
+
reconstructed = self.decoder(latent, target_length, input_values if self.training else None)
|
866 |
+
else:
|
867 |
+
# Standard autoencoder
|
868 |
+
encoder_output = self.encoder(input_values)
|
869 |
+
|
870 |
+
if self.config.is_variational:
|
871 |
+
latent, mu, logvar = encoder_output
|
872 |
+
self._mu = mu
|
873 |
+
self._logvar = logvar
|
874 |
+
else:
|
875 |
+
latent = encoder_output
|
876 |
+
self._mu = None
|
877 |
+
self._logvar = None
|
878 |
+
|
879 |
+
# Decode latent back to input space
|
880 |
+
reconstructed = self.decoder(latent)
|
881 |
+
|
882 |
+
# Apply inverse preprocessing to reconstruction
|
883 |
+
if self.preprocessor is not None and self.config.learn_inverse_preprocessing:
|
884 |
+
reconstructed, inverse_loss = self.preprocessor(reconstructed, inverse=True)
|
885 |
+
preprocessing_loss += inverse_loss
|
886 |
+
|
887 |
+
hidden_states = None
|
888 |
+
if output_hidden_states:
|
889 |
+
if self.config.is_variational:
|
890 |
+
hidden_states = (latent, mu, logvar)
|
891 |
+
else:
|
892 |
+
hidden_states = (latent,)
|
893 |
+
|
894 |
+
if not return_dict:
|
895 |
+
return tuple(v for v in [latent, reconstructed, hidden_states] if v is not None)
|
896 |
+
|
897 |
+
return AutoencoderOutput(
|
898 |
+
last_hidden_state=latent,
|
899 |
+
reconstructed=reconstructed,
|
900 |
+
hidden_states=hidden_states,
|
901 |
+
preprocessing_loss=preprocessing_loss,
|
902 |
+
)
|
903 |
+
|
904 |
+
|
905 |
+
class AutoencoderForReconstruction(PreTrainedModel):
|
906 |
+
"""
|
907 |
+
Autoencoder Model with a reconstruction head on top for reconstruction tasks.
|
908 |
+
|
909 |
+
This model inherits from PreTrainedModel and adds a reconstruction loss calculation.
|
910 |
+
"""
|
911 |
+
|
912 |
+
config_class = AutoencoderConfig
|
913 |
+
base_model_prefix = "autoencoder"
|
914 |
+
|
915 |
+
def __init__(self, config: AutoencoderConfig):
|
916 |
+
super().__init__(config)
|
917 |
+
self.config = config
|
918 |
+
|
919 |
+
# Initialize the base autoencoder model
|
920 |
+
self.autoencoder = AutoencoderModel(config)
|
921 |
+
|
922 |
+
# Initialize weights
|
923 |
+
self.post_init()
|
924 |
+
|
925 |
+
def get_input_embeddings(self):
|
926 |
+
"""Get input embeddings."""
|
927 |
+
return self.autoencoder.get_input_embeddings()
|
928 |
+
|
929 |
+
def set_input_embeddings(self, value):
|
930 |
+
"""Set input embeddings."""
|
931 |
+
self.autoencoder.set_input_embeddings(value)
|
932 |
+
|
933 |
+
def _compute_reconstruction_loss(
|
934 |
+
self,
|
935 |
+
reconstructed: torch.Tensor,
|
936 |
+
target: torch.Tensor
|
937 |
+
) -> torch.Tensor:
|
938 |
+
"""Compute reconstruction loss based on the configured loss type."""
|
939 |
+
if self.config.reconstruction_loss == "mse":
|
940 |
+
return F.mse_loss(reconstructed, target, reduction="mean")
|
941 |
+
elif self.config.reconstruction_loss == "bce":
|
942 |
+
return F.binary_cross_entropy_with_logits(reconstructed, target, reduction="mean")
|
943 |
+
elif self.config.reconstruction_loss == "l1":
|
944 |
+
return F.l1_loss(reconstructed, target, reduction="mean")
|
945 |
+
elif self.config.reconstruction_loss == "huber":
|
946 |
+
return F.huber_loss(reconstructed, target, reduction="mean")
|
947 |
+
elif self.config.reconstruction_loss == "smooth_l1":
|
948 |
+
return F.smooth_l1_loss(reconstructed, target, reduction="mean")
|
949 |
+
elif self.config.reconstruction_loss == "kl_div":
|
950 |
+
return F.kl_div(F.log_softmax(reconstructed, dim=-1), F.softmax(target, dim=-1), reduction="mean")
|
951 |
+
elif self.config.reconstruction_loss == "cosine":
|
952 |
+
return 1 - F.cosine_similarity(reconstructed, target, dim=-1).mean()
|
953 |
+
elif self.config.reconstruction_loss == "focal":
|
954 |
+
return self._focal_loss(reconstructed, target)
|
955 |
+
elif self.config.reconstruction_loss == "dice":
|
956 |
+
return self._dice_loss(reconstructed, target)
|
957 |
+
elif self.config.reconstruction_loss == "tversky":
|
958 |
+
return self._tversky_loss(reconstructed, target)
|
959 |
+
elif self.config.reconstruction_loss == "ssim":
|
960 |
+
return self._ssim_loss(reconstructed, target)
|
961 |
+
elif self.config.reconstruction_loss == "perceptual":
|
962 |
+
return self._perceptual_loss(reconstructed, target)
|
963 |
+
else:
|
964 |
+
raise ValueError(f"Unknown reconstruction loss: {self.config.reconstruction_loss}")
|
965 |
+
|
966 |
+
def _focal_loss(self, pred: torch.Tensor, target: torch.Tensor, alpha: float = 1.0, gamma: float = 2.0) -> torch.Tensor:
|
967 |
+
"""Compute focal loss for handling class imbalance."""
|
968 |
+
ce_loss = F.mse_loss(pred, target, reduction="none")
|
969 |
+
pt = torch.exp(-ce_loss)
|
970 |
+
focal_loss = alpha * (1 - pt) ** gamma * ce_loss
|
971 |
+
return focal_loss.mean()
|
972 |
+
|
973 |
+
def _dice_loss(self, pred: torch.Tensor, target: torch.Tensor, smooth: float = 1e-6) -> torch.Tensor:
|
974 |
+
"""Compute Dice loss for segmentation-like tasks."""
|
975 |
+
pred_flat = pred.view(-1)
|
976 |
+
target_flat = target.view(-1)
|
977 |
+
intersection = (pred_flat * target_flat).sum()
|
978 |
+
dice = (2.0 * intersection + smooth) / (pred_flat.sum() + target_flat.sum() + smooth)
|
979 |
+
return 1 - dice
|
980 |
+
|
981 |
+
def _tversky_loss(self, pred: torch.Tensor, target: torch.Tensor, alpha: float = 0.7, beta: float = 0.3, smooth: float = 1e-6) -> torch.Tensor:
|
982 |
+
"""Compute Tversky loss, a generalization of Dice loss."""
|
983 |
+
pred_flat = pred.view(-1)
|
984 |
+
target_flat = target.view(-1)
|
985 |
+
true_pos = (pred_flat * target_flat).sum()
|
986 |
+
false_neg = (target_flat * (1 - pred_flat)).sum()
|
987 |
+
false_pos = ((1 - target_flat) * pred_flat).sum()
|
988 |
+
tversky = (true_pos + smooth) / (true_pos + alpha * false_neg + beta * false_pos + smooth)
|
989 |
+
return 1 - tversky
|
990 |
+
|
991 |
+
def _ssim_loss(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
|
992 |
+
"""Compute SSIM-based loss (simplified version)."""
|
993 |
+
# Simplified SSIM for 1D data
|
994 |
+
mu1 = pred.mean(dim=-1, keepdim=True)
|
995 |
+
mu2 = target.mean(dim=-1, keepdim=True)
|
996 |
+
sigma1_sq = ((pred - mu1) ** 2).mean(dim=-1, keepdim=True)
|
997 |
+
sigma2_sq = ((target - mu2) ** 2).mean(dim=-1, keepdim=True)
|
998 |
+
sigma12 = ((pred - mu1) * (target - mu2)).mean(dim=-1, keepdim=True)
|
999 |
+
|
1000 |
+
c1, c2 = 0.01, 0.03
|
1001 |
+
ssim = ((2 * mu1 * mu2 + c1) * (2 * sigma12 + c2)) / ((mu1**2 + mu2**2 + c1) * (sigma1_sq + sigma2_sq + c2))
|
1002 |
+
return 1 - ssim.mean()
|
1003 |
+
|
1004 |
+
def _perceptual_loss(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
|
1005 |
+
"""Compute perceptual loss (simplified version using feature differences)."""
|
1006 |
+
# For simplicity, use L2 loss on normalized features
|
1007 |
+
pred_norm = F.normalize(pred, p=2, dim=-1)
|
1008 |
+
target_norm = F.normalize(target, p=2, dim=-1)
|
1009 |
+
return F.mse_loss(pred_norm, target_norm)
|
1010 |
+
|
1011 |
+
def forward(
|
1012 |
+
self,
|
1013 |
+
input_values: torch.Tensor,
|
1014 |
+
labels: Optional[torch.Tensor] = None,
|
1015 |
+
sequence_lengths: Optional[torch.Tensor] = None,
|
1016 |
+
target_length: Optional[int] = None,
|
1017 |
+
output_hidden_states: Optional[bool] = None,
|
1018 |
+
return_dict: Optional[bool] = None,
|
1019 |
+
) -> Union[Tuple[torch.Tensor], AutoencoderForReconstructionOutput]:
|
1020 |
+
"""
|
1021 |
+
Forward pass with reconstruction loss calculation.
|
1022 |
+
|
1023 |
+
Args:
|
1024 |
+
input_values (torch.Tensor): Input tensor. Shape depends on autoencoder type:
|
1025 |
+
- Standard: (batch_size, input_dim)
|
1026 |
+
- Recurrent: (batch_size, seq_len, input_dim)
|
1027 |
+
labels (torch.Tensor, optional): Target tensor for reconstruction. If None, uses input_values.
|
1028 |
+
sequence_lengths (torch.Tensor, optional): Sequence lengths for recurrent AE.
|
1029 |
+
target_length (int, optional): Target sequence length for recurrent decoder.
|
1030 |
+
output_hidden_states (bool, optional): Whether to return hidden states.
|
1031 |
+
return_dict (bool, optional): Whether to return a ModelOutput instead of a plain tuple.
|
1032 |
+
|
1033 |
+
Returns:
|
1034 |
+
AutoencoderForReconstructionOutput or tuple: The model outputs including loss.
|
1035 |
+
"""
|
1036 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1037 |
+
|
1038 |
+
# If no labels provided, use input as target (standard autoencoder)
|
1039 |
+
if labels is None:
|
1040 |
+
labels = input_values
|
1041 |
+
|
1042 |
+
# Forward pass through autoencoder
|
1043 |
+
outputs = self.autoencoder(
|
1044 |
+
input_values=input_values,
|
1045 |
+
sequence_lengths=sequence_lengths,
|
1046 |
+
target_length=target_length,
|
1047 |
+
output_hidden_states=output_hidden_states,
|
1048 |
+
return_dict=True,
|
1049 |
+
)
|
1050 |
+
|
1051 |
+
reconstructed = outputs.reconstructed
|
1052 |
+
latent = outputs.last_hidden_state
|
1053 |
+
hidden_states = outputs.hidden_states
|
1054 |
+
|
1055 |
+
# Compute reconstruction loss
|
1056 |
+
recon_loss = self._compute_reconstruction_loss(reconstructed, labels)
|
1057 |
+
|
1058 |
+
# Add regularization losses based on autoencoder type
|
1059 |
+
total_loss = recon_loss
|
1060 |
+
|
1061 |
+
# Add preprocessing loss if available
|
1062 |
+
if hasattr(outputs, 'preprocessing_loss') and outputs.preprocessing_loss is not None:
|
1063 |
+
total_loss += outputs.preprocessing_loss
|
1064 |
+
|
1065 |
+
if self.config.is_variational and hasattr(self.autoencoder, '_mu') and self.autoencoder._mu is not None:
|
1066 |
+
# KL divergence loss for variational autoencoders
|
1067 |
+
kl_loss = -0.5 * torch.sum(1 + self.autoencoder._logvar - self.autoencoder._mu.pow(2) - self.autoencoder._logvar.exp())
|
1068 |
+
kl_loss = kl_loss / (self.autoencoder._mu.size(0) * self.autoencoder._mu.size(1)) # Normalize by batch size and latent dim
|
1069 |
+
total_loss = recon_loss + self.config.beta * kl_loss
|
1070 |
+
|
1071 |
+
elif self.config.is_sparse:
|
1072 |
+
# Sparsity loss for sparse autoencoders
|
1073 |
+
latent = outputs.last_hidden_state
|
1074 |
+
sparsity_loss = torch.mean(torch.abs(latent)) # L1 sparsity
|
1075 |
+
total_loss = recon_loss + 0.1 * sparsity_loss # Sparsity weight
|
1076 |
+
|
1077 |
+
elif self.config.is_contractive:
|
1078 |
+
# Contractive loss - penalize large gradients of hidden representation w.r.t. input
|
1079 |
+
latent = outputs.last_hidden_state
|
1080 |
+
latent.retain_grad()
|
1081 |
+
if latent.grad is not None:
|
1082 |
+
contractive_loss = torch.sum(latent.grad ** 2)
|
1083 |
+
total_loss = recon_loss + 0.1 * contractive_loss
|
1084 |
+
|
1085 |
+
loss = total_loss
|
1086 |
+
|
1087 |
+
if not return_dict:
|
1088 |
+
output = (reconstructed, latent)
|
1089 |
+
if hidden_states is not None:
|
1090 |
+
output = output + (hidden_states,)
|
1091 |
+
return ((loss,) + output) if loss is not None else output
|
1092 |
+
|
1093 |
+
return AutoencoderForReconstructionOutput(
|
1094 |
+
loss=loss,
|
1095 |
+
reconstructed=reconstructed,
|
1096 |
+
last_hidden_state=latent,
|
1097 |
+
hidden_states=hidden_states,
|
1098 |
+
preprocessing_loss=outputs.preprocessing_loss if hasattr(outputs, 'preprocessing_loss') else None,
|
1099 |
+
)
|
register_autoencoder.py
ADDED
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Registration script for Autoencoder models with Hugging Face AutoModel framework.
|
3 |
+
"""
|
4 |
+
|
5 |
+
from transformers import AutoConfig, AutoModel
|
6 |
+
from configuration_autoencoder import AutoencoderConfig
|
7 |
+
from modeling_autoencoder import AutoencoderModel, AutoencoderForReconstruction
|
8 |
+
|
9 |
+
|
10 |
+
def register_autoencoder_models():
|
11 |
+
"""
|
12 |
+
Register the autoencoder models with the Hugging Face AutoModel framework.
|
13 |
+
|
14 |
+
This function registers:
|
15 |
+
- AutoencoderConfig with AutoConfig
|
16 |
+
- AutoencoderModel with AutoModel
|
17 |
+
- AutoencoderForReconstruction with AutoModel (for reconstruction tasks)
|
18 |
+
|
19 |
+
After calling this function, you can use:
|
20 |
+
- AutoConfig.from_pretrained() to load autoencoder configs
|
21 |
+
- AutoModel.from_pretrained() to load autoencoder models
|
22 |
+
"""
|
23 |
+
|
24 |
+
# Register configuration
|
25 |
+
AutoConfig.register("autoencoder", AutoencoderConfig)
|
26 |
+
|
27 |
+
# Register base model
|
28 |
+
AutoModel.register(AutoencoderConfig, AutoencoderModel)
|
29 |
+
|
30 |
+
# Note: For task-specific models like AutoencoderForReconstruction,
|
31 |
+
# we would typically create a custom AutoModelForReconstruction class
|
32 |
+
# and register it separately. For now, users can import directly.
|
33 |
+
|
34 |
+
print("✅ Autoencoder models registered with Hugging Face AutoModel framework!")
|
35 |
+
print("You can now use:")
|
36 |
+
print(" - AutoConfig.from_pretrained() for configs")
|
37 |
+
print(" - AutoModel.from_pretrained() for models")
|
38 |
+
print(" - Direct imports for task-specific models")
|
39 |
+
|
40 |
+
|
41 |
+
def register_for_auto_class():
|
42 |
+
"""
|
43 |
+
Register models for auto class functionality when saving/loading.
|
44 |
+
|
45 |
+
This enables the models to be automatically discovered when using
|
46 |
+
save_pretrained() and from_pretrained() methods.
|
47 |
+
"""
|
48 |
+
|
49 |
+
# Register config for auto class
|
50 |
+
AutoencoderConfig.register_for_auto_class()
|
51 |
+
|
52 |
+
# Register models for auto class
|
53 |
+
AutoencoderModel.register_for_auto_class("AutoModel")
|
54 |
+
AutoencoderForReconstruction.register_for_auto_class("AutoModel")
|
55 |
+
|
56 |
+
print("✅ Models registered for auto class functionality!")
|
57 |
+
|
58 |
+
|
59 |
+
if __name__ == "__main__":
|
60 |
+
# Register models when script is run directly
|
61 |
+
register_autoencoder_models()
|
62 |
+
register_for_auto_class()
|