Upload folder using huggingface_hub

CustomBertForMaskedLM.py

@@ -0,0 +1,37 @@
+from transformers import BertForMaskedLM
+import torch.nn as nn
+from RBFLayer import RBFLayer  # Assuming RBFLayer is your custom RBF implementation
+
+class CustomBertForMaskedLM(BertForMaskedLM):
+    def __init__(self, config):
+        super().__init__(config)
+
+        # Replace the feedforward MLP layers with RBF layers in BERT's encoder
+        for i, layer in enumerate(self.bert.encoder.layer):
+            in_features = 768
+            intermediate_features = 3072
+            
+            # Replace the intermediate dense layer (768 -> 3072) with RBF
+            layer.intermediate.dense = RBFLayer(
+                in_features_dim=in_features,
+                num_kernels=2,  # Number of kernels in the RBF layer
+                out_features_dim=intermediate_features,
+                radial_function=gaussian_rbf,
+                norm_function=euclidean_norm
+            )
+            
+            # Replace the output dense layer (3072 -> 768) with RBF
+            layer.output.dense = RBFLayer(
+                in_features_dim=intermediate_features,
+                num_kernels=2,
+                out_features_dim=in_features,
+                radial_function=gaussian_rbf,
+                norm_function=euclidean_norm
+            )
+
+# Define radial basis and norm functions
+def gaussian_rbf(x):
+    return torch.exp(-x**2)
+
+def euclidean_norm(x):
+    return torch.norm(x, p=2, dim=-1)

CustomRBFBert.py

@@ -0,0 +1,28 @@
+import torch
+import torch.nn as nn
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from RBFLayer import RBFLayer
+
+def l_norm(x, p=2):
+    return torch.norm(x, p=p, dim=-1)
+
+
+# Gaussian RBF
+def rbf_gaussian(x):
+    return (-x.pow(2)).exp()
+
+class CustomRBFFeedForward(nn.Module):
+    def __init__(self, in_features, out_features, num_kernels):
+        super(CustomRBFFeedForward, self).__init__()
+        # RBFLayer from the given implementation
+        self.rbf_layer = RBFLayer(
+            in_features_dim=in_features,  # Input size (e.g., 896)
+            num_kernels=num_kernels,  # Number of kernels in the RBF layer (can be tuned)
+            out_features_dim=out_features,  # Output size (e.g., 4864)
+            radial_function=rbf_gaussian,  # Use the Gaussian RBF
+            norm_function=l_norm  # Use Euclidean norm
+        )
+
+    def forward(self, x):
+        # Apply the RBF layer to the input x
+        return self.rbf_layer(x)

RBFLayer.py

@@ -0,0 +1,101 @@
+import torch
+import torch.nn as nn
+from typing import Callable
+
+
+class RBFLayer(nn.Module):
+    def __init__(self,
+                 in_features_dim: int,
+                 num_kernels: int,
+                 out_features_dim: int,
+                 radial_function: Callable[[torch.Tensor], torch.Tensor],
+                 norm_function: Callable[[torch.Tensor], torch.Tensor],
+                 normalization: bool = True,
+                 initial_shape_parameter: torch.Tensor = None,
+                 initial_centers_parameter: torch.Tensor = None,
+                 initial_weights_parameters: torch.Tensor = None,
+                 constant_shape_parameter: bool = False,
+                 constant_centers_parameter: bool = False,
+                 constant_weights_parameters: bool = False):
+        super(RBFLayer, self).__init__()
+
+        self.in_features_dim = in_features_dim
+        self.num_kernels = num_kernels
+        self.out_features_dim = out_features_dim
+        self.radial_function = radial_function
+        self.norm_function = norm_function
+        self.normalization = normalization
+
+        self.initial_shape_parameter = initial_shape_parameter
+        self.constant_shape_parameter = constant_shape_parameter
+
+        self.initial_centers_parameter = initial_centers_parameter
+        self.constant_centers_parameter = constant_centers_parameter
+
+        self.initial_weights_parameters = initial_weights_parameters
+        self.constant_weights_parameters = constant_weights_parameters
+
+        self._make_parameters()
+
+    def _make_parameters(self) -> None:
+        # Initialize linear combination weights
+        if self.constant_weights_parameters:
+            self.weights = nn.Parameter(self.initial_weights_parameters, requires_grad=False)
+        else:
+            self.weights = nn.Parameter(torch.zeros(self.out_features_dim, self.num_kernels, dtype=torch.float32))
+
+        # Initialize kernels' centers
+        if self.constant_centers_parameter:
+            self.kernels_centers = nn.Parameter(self.initial_centers_parameter, requires_grad=False)
+        else:
+            self.kernels_centers = nn.Parameter(torch.zeros(self.num_kernels, self.in_features_dim, dtype=torch.float32))
+
+        # Initialize shape parameter
+        if self.constant_shape_parameter:
+            self.log_shapes = nn.Parameter(self.initial_shape_parameter, requires_grad=False)
+        else:
+            self.log_shapes = nn.Parameter(torch.zeros(self.num_kernels, dtype=torch.float32))
+
+        self.reset()
+
+    def reset(self, upper_bound_kernels: float = 1.0, std_shapes: float = 0.1, gain_weights: float = 1.0) -> None:
+        if self.initial_centers_parameter is None:
+            nn.init.uniform_(self.kernels_centers, a=-upper_bound_kernels, b=upper_bound_kernels)
+
+        if self.initial_shape_parameter is None:
+            nn.init.normal_(self.log_shapes, mean=0.0, std=std_shapes)
+
+        if self.initial_weights_parameters is None:
+            nn.init.xavier_uniform_(self.weights, gain=gain_weights)
+
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        """
+        Computes the output of the RBF layer given an input tensor.
+        Input has size [batch_size, sequence_length, in_features].
+        """
+
+        batch_size = input.size(0)
+        sequence_length = input.size(1)
+
+        # Expand centers to match the batch and sequence length
+        c = self.kernels_centers.expand(batch_size, sequence_length, self.num_kernels, self.in_features_dim)
+
+        # Compute differences between input and centers
+        diff = input.unsqueeze(2) - c  # Shape: [batch_size, sequence_length, num_kernels, in_features_dim]
+
+        # Apply norm function to get distances
+        r = self.norm_function(diff)  # Shape: [batch_size, sequence_length, num_kernels]
+
+        # Apply shape parameters (log_shapes) to the distances
+        eps_r = self.log_shapes.exp().unsqueeze(0).unsqueeze(0) * r
+
+        # Apply radial basis function (e.g., Gaussian)
+        rbfs = self.radial_function(eps_r)
+
+        if self.normalization:
+            rbfs = rbfs / (1e-9 + rbfs.sum(dim=-1, keepdim=True))
+
+        # Combine RBF outputs using the weights
+        out = (self.weights.unsqueeze(0).unsqueeze(0) * rbfs.unsqueeze(2)).sum(dim=-1)
+
+        return out

config.json

@@ -0,0 +1,32 @@
+{
+  "_name_or_path": "google-bert/bert-base-uncased",
+  "architectures": [
+    "CustomBertForMaskedLM" 
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "classifier_dropout": null,
+  "gradient_checkpointing": false,
+  "hidden_act": "gelu",  
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,  
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,  
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,  
+  "pad_token_id": 0,
+  "position_embedding_type": "absolute",
+  "torch_dtype": "float32",
+  "transformers_version": "4.41.1",
+  "type_vocab_size": 2,
+  "use_cache": true,
+  "vocab_size": 30522,
+  
+  "custom_layers": {
+    "use_rbf": true,  
+    "rbf_num_kernels": 2,  
+    "rbf_intermediate_size": 3072 
+  }
+}

generation_config.json

@@ -0,0 +1,5 @@
+{
+  "_from_model_config": true,
+  "pad_token_id": 0,
+  "transformers_version": "4.41.1"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc47217a75b4c60c273e782fa8855e06a91fb8dcebf301378e149221462519cd
+size 212145392

special_tokens_map.json

@@ -0,0 +1,7 @@
+{
+  "cls_token": "[CLS]",
+  "mask_token": "[MASK]",
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "unk_token": "[UNK]"
+}

tokenizer_config.json

@@ -0,0 +1,55 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "100": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "101": {
+      "content": "[CLS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "102": {
+      "content": "[SEP]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "103": {
+      "content": "[MASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "clean_up_tokenization_spaces": true,
+  "cls_token": "[CLS]",
+  "do_lower_case": true,
+  "mask_token": "[MASK]",
+  "model_max_length": 512,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "strip_accents": null,
+  "tokenize_chinese_chars": true,
+  "tokenizer_class": "BertTokenizer",
+  "unk_token": "[UNK]"
+}