Upload weight-space autoencoder (encoder + decoder) and configuration

README.md

@@ -0,0 +1,42 @@
+---
+tags:
+- weight-space-learning
+- neural-network-autoencoder
+- autoencoder
+- transformer
+datasets:
+- maximuspowers/muat-fourier-5
+---
+
+# Weight-Space Autoencoder (TRANSFORMER)
+
+This model is a weight-space autoencoder trained on neural network activation weights/signatures.
+It includes both an encoder (compresses weights into latent representations) and a decoder (reconstructs weights from latent codes).
+
+## Model Description
+
+- **Architecture**: Transformer encoder-decoder
+- **Training Dataset**: maximuspowers/muat-fourier-5
+- **Input Mode**: signature
+- **Latent Dimension**: 128
+
+## Tokenization
+
+- **Chunk Size**: 1 weight values per token
+- **Max Tokens**: 64
+- **Metadata**: True
+
+## Training Config
+
+- **Loss Functions**: reconstruction, contrastive, functional
+- **Optimizer**: adamw
+- **Learning Rate**: 0.0001
+- **Batch Size**: 32
+
+## Performance Metrics (Test Set)
+
+- **MSE**: 0.105820
+- **MAE**: 0.208260
+- **RMSE**: 0.325300
+- **Cosine Similarity**: 0.9560
+- **R² Score**: 0.9830

config.yaml

@@ -0,0 +1,116 @@
+architecture:
+  latent_dim: 128
+  transformer:
+    decoder:
+      activation: gelu
+      d_model: 512
+      dim_feedforward: 2048
+      dropout: 0.1
+      num_heads: 8
+      num_layers: 6
+    encoder:
+      activation: gelu
+      d_model: 512
+      dim_feedforward: 2048
+      dropout: 0.1
+      num_heads: 8
+      num_layers: 6
+      pooling: mean
+      positional_encoding: learned
+  type: transformer
+dataloader:
+  num_workers: 0
+  pin_memory: true
+dataset:
+  hf_dataset: maximuspowers/muat-fourier-5
+  input_mode: signature
+  max_dimensions:
+    max_hidden_layers: 6
+    max_neurons_per_layer: 8
+    max_sequence_length: 5
+  neuron_profile:
+    features_per_neuron: 5
+    methods:
+    - fourier
+  random_seed: 42
+  test_split: 0.1
+  train_split: 0.8
+  val_split: 0.1
+device:
+  type: auto
+evaluation:
+  metrics:
+  - mse
+  - mae
+  - rmse
+  - cosine_similarity
+  - relative_error
+  - r2_score
+  per_layer_metrics: false
+hub:
+  enabled: true
+  private: false
+  push_logs: true
+  push_metrics: true
+  push_model: true
+  repo_id: maximuspowers/weight-autoencoder-mlp-v1
+  token: <REDACTED>
+logging:
+  checkpoint:
+    enabled: true
+    mode: min
+    monitor: val_loss
+    save_best_only: true
+  tensorboard:
+    auto_launch: true
+    enabled: true
+    log_interval: 10
+    port: 6006
+    visualizations:
+      enabled: true
+      log_interval: 1
+      num_image_samples: 4
+  verbose: true
+loss:
+  contrastive:
+    enabled: true
+    projection_head:
+      hidden_dim: 64
+      input_dim: 128
+      output_dim: 32
+    temperature: 0.1
+    weight: 0.4
+  functional:
+    benchmark_path: /configs/autoencoder/benchmark_dataset.json
+    enabled: true
+    test_samples: null
+    weight: 0.4
+  reconstruction:
+    enabled: true
+    type: mse
+    weight: 0.2
+run_dir: /Users/max/Desktop/muat/model_zoo/runs/train-encoder-decoder_config_2025-12-17_19-33-32
+run_log_cleanup: false
+tokenization:
+  chunk_size: 1
+  granularity: neuron
+  include_metadata: true
+  max_tokens: 64
+training:
+  batch_size: 32
+  early_stopping:
+    enabled: true
+    mode: min
+    monitor: val_loss
+    patience: 15
+  epochs: 250
+  gradient_accumulation_steps: 4
+  learning_rate: 0.0001
+  lr_scheduler:
+    enabled: true
+    factor: 0.5
+    min_lr: 1.0e-06
+    patience: 5
+  max_grad_norm: 1.0
+  optimizer: adamw
+  weight_decay: 0.0001

decoder.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a7e1b2bed452a4562d4f0e6fb7e47a75e917bfbf6a68f660bdfc3194fabfdca
+size 101365774

encoder.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ebcb2592d5bb6ef3f7806da61037cc769ad5f29534c6dbdb683228624a2db38
+size 76106790

tokenizer_config.json

@@ -0,0 +1,9 @@
+{
+  "chunk_size": 1,
+  "max_tokens": 64,
+  "include_metadata": true,
+  "metadata_features": 5,
+  "token_dim": 14,
+  "granularity": "neuron",
+  "max_neuron_data_size": 9
+}