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

README.md

@@ -18,13 +18,12 @@ It includes both an encoder (compresses weights into latent representations) and
 - **Architecture**: Transformer encoder-decoder
 - **Training Dataset**: maximuspowers/muat-fourier-5
 - **Input Mode**: signature
-- **Latent Dimension**: 128
+- **Latent Dimension**: 256
 
 ## Tokenization
 
-- **Chunk Size**: 1 weight values per token
+- **Granularity**: neuron
 - **Max Tokens**: 64
-- **Metadata**: True
 
 ## Training Config
 
@@ -35,8 +34,8 @@ It includes both an encoder (compresses weights into latent representations) and
 
 ## Performance Metrics (Test Set)
 
-- **MSE**: 0.105820
-- **MAE**: 0.208260
-- **RMSE**: 0.325300
-- **Cosine Similarity**: 0.9560
-- **R² Score**: 0.9830
+- **MSE**: 0.125011
+- **MAE**: 0.259796
+- **RMSE**: 0.353570
+- **Cosine Similarity**: 0.0348
+- **R² Score**: -0.0097

config.yaml

@@ -1,22 +1,27 @@
 architecture:
-  latent_dim: 128
+  arch_encoder:
+    embed_dim: 64
+    max_layers: 6
+    max_neurons: 8
+  latent_dim: 256
   transformer:
     decoder:
       activation: gelu
       d_model: 512
       dim_feedforward: 2048
-      dropout: 0.1
+      dropout: 0.025
       num_heads: 8
       num_layers: 6
     encoder:
       activation: gelu
       d_model: 512
       dim_feedforward: 2048
-      dropout: 0.1
+      dropout: 0.025
       num_heads: 8
-      num_layers: 6
+      num_layers: 10
       pooling: mean
       positional_encoding: learned
+      use_positional_encoding: true
   type: transformer
 dataloader:
   num_workers: 0
@@ -75,26 +80,26 @@ loss:
   contrastive:
     enabled: true
     projection_head:
-      hidden_dim: 64
-      input_dim: 128
-      output_dim: 32
+      hidden_dim: 128
+      input_dim: 256
+      output_dim: 64
     temperature: 0.1
-    weight: 0.4
+    weight: 0.05
   functional:
-    benchmark_path: /configs/autoencoder/benchmark_dataset.json
+    benchmark_path: /Users/max/Desktop/muat/model_zoo/configs/autoencoder/benchmark_dataset.json
     enabled: true
     test_samples: null
-    weight: 0.4
+    weight: 0.5
   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
+    weight: 0.6
+run_dir: /Users/max/Desktop/muat/model_zoo/runs/train-encoder-decoder_config_2025-12-18_22-23-46
 run_log_cleanup: false
 tokenization:
   chunk_size: 1
   granularity: neuron
-  include_metadata: true
+  include_metadata: false
   max_tokens: 64
 training:
   batch_size: 32
@@ -102,8 +107,8 @@ training:
     enabled: true
     mode: min
     monitor: val_loss
-    patience: 15
-  epochs: 250
+    patience: 50
+  epochs: 1000
   gradient_accumulation_steps: 4
   learning_rate: 0.0001
   lr_scheduler:

decoder.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1a7e1b2bed452a4562d4f0e6fb7e47a75e917bfbf6a68f660bdfc3194fabfdca
-size 101365774
+oid sha256:0c8bdb28e21e756364067eaee646097f7e24a78aa907a48d4c34a08c1c7df45f
+size 103255652

encoder.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4ebcb2592d5bb6ef3f7806da61037cc769ad5f29534c6dbdb683228624a2db38
-size 76106790
+oid sha256:394e0021fc9d3862b3307c018dbce6118ea0c2b6664766d1729081a3904ca1d8
+size 126816306

tokenizer_config.json

@@ -1,9 +1,7 @@
 {
   "chunk_size": 1,
   "max_tokens": 64,
-  "include_metadata": true,
-  "metadata_features": 5,
-  "token_dim": 14,
+  "token_dim": 9,
   "granularity": "neuron",
   "max_neuron_data_size": 9
 }