Upload folder using huggingface_hub

README.md

@@ -0,0 +1,21 @@
+
+---
+language: en
+tags:
+- ner
+- pytorch
+license: mit
+---
+
+# CNN for Named Entity Recognition
+
+This model is a CNN-based model for Named Entity Recognition (NER) built on top of a pre-trained transformer model.
+
+## Model description
+
+The model uses a pre-trained transformer as a base and adds convolutional layers on top for NER tasks.
+
+## Intended uses & limitations
+
+This model is intended for Named Entity Recognition tasks. It should be used on Yoruba text data.
+

config.json

@@ -0,0 +1 @@
+{"model_type": "CNNForYorubaNER", "num_classes": 9, "max_length": 128, "pretrained_model_name": "masakhane/afroxlmr-large-ner-masakhaner-1.0_2.0"}

custom_modeling.py

@@ -0,0 +1,12 @@
+
+from transformers import AutoConfig, AutoModel
+from modeling_cnn_ner import CNNForNER
+
+def get_model(pretrained_model_name_or_path):
+    config = AutoConfig.from_pretrained(pretrained_model_name_or_path)
+    model = CNNForNER(
+        pretrained_model_name=config.pretrained_model_name,
+        num_classes=config.num_classes,
+        max_length=config.max_length
+    )
+    return model

modeling_cnn_ner.py

@@ -0,0 +1,32 @@
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import AutoModelForTokenClassification
+
+class CNNForNER(nn.Module):
+    def __init__(self, pretrained_model_name, num_classes, max_length=128):
+        super(CNNForNER, self).__init__()
+        self.transformer = AutoModelForTokenClassification.from_pretrained(pretrained_model_name)
+        self.max_length = max_length
+
+        # Get the number of labels from the pretrained model
+        pretrained_num_labels = self.transformer.num_labels
+
+        self.conv1 = nn.Conv1d(in_channels=pretrained_num_labels, out_channels=256, kernel_size=3, padding=1)
+        self.conv2 = nn.Conv1d(in_channels=256, out_channels=128, kernel_size=3, padding=1)
+        self.dropout = nn.Dropout(0.3)
+        self.fc = nn.Linear(in_features=128, out_features=num_classes)
+
+    def forward(self, input_ids, attention_mask):
+        outputs = self.transformer(input_ids=input_ids, attention_mask=attention_mask)
+        logits = outputs.logits  # Shape: (batch_size, sequence_length, pretrained_num_labels)
+        
+        # Apply CNN layers
+        logits = logits.permute(0, 2, 1)  # Shape: (batch_size, pretrained_num_labels, sequence_length)
+        conv1_out = F.relu(self.conv1(logits))
+        conv2_out = F.relu(self.conv2(conv1_out))
+        conv2_out = self.dropout(conv2_out)
+        conv2_out = conv2_out.permute(0, 2, 1)  # Shape: (batch_size, sequence_length, 128)
+        final_logits = self.fc(conv2_out)  # Shape: (batch_size, sequence_length, num_classes)
+        return final_logits

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f6df5d161a93b1a7470e4e0cb4f1d8a9dbeb74a9fb56b72d318dad0731c9a379
+size 2236003390