minilm_keyword_pred.py

import torch
import torch.nn as nn
import pickle
from transformers import AutoTokenizer, AutoModel
from tqdm import tqdm
import numpy as np
   
OFFLINE_MODEL_PATH = "all-MiniLM-L6-v2" 

# ==============================================================================
#  STEP 1: DEFINE THE MODEL ARCHITECTURE
#  This MUST be the exact same class definition you used for training.
# ==============================================================================
class ImprovedMultiTaskClassifier(nn.Module):
    def __init__(self, model_name, num_keywords, num_groups, dropout_rate=0.1):
        super(ImprovedMultiTaskClassifier, self).__init__()
        self.transformer = AutoModel.from_pretrained(model_name)
        hidden_size = self.transformer.config.hidden_size

        self.keyword_classifier = nn.Sequential(
            nn.Linear(hidden_size, hidden_size),
            nn.LayerNorm(hidden_size), nn.ReLU(), nn.Dropout(dropout_rate),
            nn.Linear(hidden_size, hidden_size // 2),
            nn.LayerNorm(hidden_size // 2), nn.ReLU(), nn.Dropout(dropout_rate),
            nn.Linear(hidden_size // 2, num_keywords)
        )
        self.group_classifier = nn.Sequential(
            nn.Linear(hidden_size, hidden_size),
            nn.LayerNorm(hidden_size), nn.ReLU(), nn.Dropout(dropout_rate),
            nn.Linear(hidden_size, hidden_size // 2),
            nn.LayerNorm(hidden_size // 2), nn.ReLU(), nn.Dropout(dropout_rate),
            nn.Linear(hidden_size // 2, num_groups)
        )

    def forward(self, input_ids, attention_mask):
        outputs = self.transformer(input_ids=input_ids, attention_mask=attention_mask)
        token_embeddings = outputs.last_hidden_state
        attention_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
        sum_embeddings = torch.sum(token_embeddings * attention_mask_expanded, 1)
        sum_mask = torch.clamp(attention_mask_expanded.sum(1), min=1e-9)
        pooled_output = sum_embeddings / sum_mask
        keyword_logits = self.keyword_classifier(pooled_output)
        group_logits = self.group_classifier(pooled_output)
        return keyword_logits, group_logits


# ==============================================================================
#  STEP 2: LOAD ALL SAVED COMPONENTS
# ==============================================================================
print("Loading all components for inference...")

# Set device
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"Using device: {device}")

# Load config
with open('minilm_keyword_classifier_gemini/inference_config.pkl', 'rb') as f:
    config = pickle.load(f)

# *** IMPORTANT: Override the model_name to use the local path ***
config['model_name'] = OFFLINE_MODEL_PATH
# Load tokenizer from the same offline path
tokenizer = AutoTokenizer.from_pretrained(OFFLINE_MODEL_PATH)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained('minilm_keyword_classifier_gemini/inference_tokenizer')

# Load label encoders
with open('minilm_keyword_classifier_gemini/inference_mlb_keywords.pkl', 'rb') as f:
    mlb_keywords = pickle.load(f)
with open('minilm_keyword_classifier_gemini/inference_mlb_groups.pkl', 'rb') as f:
    mlb_groups = pickle.load(f)

# Instantiate the model architecture
num_keywords = len(mlb_keywords.classes_)
num_groups = len(mlb_groups.classes_)
model = ImprovedMultiTaskClassifier(config['model_name'], num_keywords, num_groups).to(device)

# Load the trained weights
model.load_state_dict(torch.load('minilm_keyword_classifier_gemini/inference_model.pth', map_location=device))

# Set model to evaluation mode (very important!)
model.eval()

print("✅ All components loaded and model is ready for inference.")


# ==============================================================================
#  STEP 3: CREATE THE PREDICTION FUNCTION (MODIFIED TO INCLUDE SCORES)
# ==============================================================================
def predict_on_text(text: str):
    """
    Takes a string of text and returns the predicted keywords and groups
    along with their confidence scores.
    """
    with torch.no_grad():
        encoding = tokenizer(
            text,
            truncation=True,
            padding='max_length',
            max_length=512,
            return_tensors='pt'
        )
        input_ids = encoding['input_ids'].to(device)
        attention_mask = encoding['attention_mask'].to(device)

        keyword_logits, group_logits = model(input_ids, attention_mask)

        keyword_probs = torch.sigmoid(keyword_logits).cpu().numpy()[0]
        group_probs = torch.sigmoid(group_logits).cpu().numpy()[0]

        kw_threshold = config['optimal_keyword_threshold']
        gr_threshold = config['optimal_group_threshold']

        # --- MODIFICATION START ---
        # Get keywords that are above the threshold
        kw_indices = np.where(keyword_probs > kw_threshold)[0]
        predicted_keywords_with_scores = [
            (mlb_keywords.classes_[i], keyword_probs[i]) for i in kw_indices
        ]

        # Get groups that are above the threshold
        gr_indices = np.where(group_probs > gr_threshold)[0]
        predicted_groups_with_scores = [
            (mlb_groups.classes_[i], group_probs[i]) for i in gr_indices
        ]
        
        # Sort predictions by score in descending order
        predicted_keywords_with_scores.sort(key=lambda x: x[1], reverse=True)
        predicted_groups_with_scores.sort(key=lambda x: x[1], reverse=True)
        # --- MODIFICATION END ---
        
        return {
            'predicted_keywords_with_scores': predicted_keywords_with_scores,
            'predicted_groups_with_scores': predicted_groups_with_scores,
        }



# list through all csv files in automarked\todo folder. Read the content column and loop through all the content there as text
# for file in glob.glob('automarked\\todo\\*.csv'):
#     with open(file, 'r', newline='', encoding='utf-8', errors='ignore') as f:
#         reader = csv.DictReader(f)
#         for row in reader:
#             text = row['content']

text = """I want you to understand, people think there are many problems in the world. There are no many problems in the world. There's only one problem in the world – human being. What other problem, I'm asking"""

dpred = predict_on_text(text) 
for d in dpred['predicted_groups_with_scores']:
    print(d[0], d[1], d[1] > 0.5)