minor - media (hierarchical)

interfaces/cap_minor_media.py

@@ -7,8 +7,10 @@ import pandas as pd
 from transformers import AutoModelForSequenceClassification
 from transformers import AutoTokenizer
 from huggingface_hub import HfApi
+from collections import defaultdict
 
-from label_dicts import CAP_MEDIA_NUM_DICT, CAP_MEDIA_LABEL_NAMES
+from label_dicts import (CAP_MEDIA_NUM_DICT, CAP_MEDIA_LABEL_NAMES,
+                        CAP_MIN_NUM_DICT, CAP_MIN_LABEL_NAMES)
 
 from .utils import is_disk_full
 
@@ -31,46 +33,83 @@ def check_huggingface_path(checkpoint_path: str):
         return False
 
 def build_huggingface_path(language: str, domain: str):
-    return "poltextlab/xlm-roberta-large-pooled-cap-media"
+    return ("poltextlab/xlm-roberta-large-pooled-cap-media", "poltextlab/xlm-roberta-large-pooled-cap-minor-v3")
 
-def predict(text, model_id, tokenizer_id):
+def predict(text, major_model_id, minor_model_id, tokenizer_id, HF_TOKEN=None):
     device = torch.device("cpu")
-    model = AutoModelForSequenceClassification.from_pretrained(model_id, low_cpu_mem_usage=True, device_map="auto", offload_folder="offload", token=HF_TOKEN)
+
+    # Load major and minor models + tokenizer
+    major_model = AutoModelForSequenceClassification.from_pretrained(
+        major_model_id,
+        low_cpu_mem_usage=True,
+        device_map="auto",
+        offload_folder="offload",
+        token=HF_TOKEN
+    ).to(device)
+
+    minor_model = AutoModelForSequenceClassification.from_pretrained(
+        minor_model_id,
+        low_cpu_mem_usage=True,
+        device_map="auto",
+        offload_folder="offload",
+        token=HF_TOKEN
+    ).to(device)
+
     tokenizer = AutoTokenizer.from_pretrained(tokenizer_id)
 
-    inputs = tokenizer(text,
-                       max_length=256,
-                       truncation=True,
-                       padding="do_not_pad",
-                       return_tensors="pt").to(device)
-    model.eval()
+    # Tokenize input
+    inputs = tokenizer(text, max_length=256, truncation=True, padding="do_not_pad", return_tensors="pt").to(device)
 
+    # Predict major topic
+    major_model.eval()
     with torch.no_grad():
-        logits = model(**inputs).logits
+        major_logits = major_model(**inputs).logits
+        major_probs = F.softmax(major_logits, dim=-1)
+    major_probs_np = major_probs.cpu().numpy().flatten()
+    top_major_index = int(np.argmax(major_probs_np))
+    top_major_id = major_index_to_id[top_major_index]
+
+    # Default: show major topic predictions
+    output_pred = {
+        f"[{major_index_to_id[i]}] {CAP_MEDIA_LABEL_NAMES[major_index_to_id[i]]}": float(major_probs_np[i])
+        for i in np.argsort(major_probs_np)[::-1]
+    }
+
+    # If eligible for minor prediction
+    if top_major_id in major_to_minor_map:
+        valid_minor_ids = major_to_minor_map[top_major_id]
+        minor_model.eval()
+        with torch.no_grad():
+            minor_logits = minor_model(**inputs).logits
+            minor_probs = F.softmax(minor_logits, dim=-1)
+
+        # Restrict to valid minor codes
+        valid_indices = [minor_id_to_index[mid] for mid in valid_minor_ids if mid in minor_id_to_index]
+        filtered_probs = {minor_index_to_id[i]: float(minor_probs[0][i]) for i in valid_indices}
+        output_pred = {
+            f"[{k}] {CAP_MIN_LABEL_NAMES[k]}": v
+            for k, v in sorted(filtered_probs.items(), key=lambda item: item[1], reverse=True)
+        }
+
+    output_info = f'<p style="text-align: center; display: block">Prediction used <a href="https://huggingface.co/{major_model_id}">{major_model_id}</a> and <a href="https://huggingface.co/{minor_model_id}">{minor_model_id}</a>.</p>'
 
-    probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten()
-    output_pred = {f"[{CAP_MEDIA_NUM_DICT[i]}] {CAP_MEDIA_LABEL_NAMES[CAP_MEDIA_NUM_DICT[i]]}": probs[i] for i in np.argsort(probs)[::-1]}
-    output_info = f'<p style="text-align: center; display: block">Prediction was made using the <a href="https://huggingface.co/{model_id}">{model_id}</a> model.</p>'
     return output_pred, output_info
 
 def predict_cap(text, language, domain):
     domain = domains[domain]
-    model_id = build_huggingface_path(language, domain)
+    major_model_id, minor_model_id = build_huggingface_path(language, domain)
     tokenizer_id = "xlm-roberta-large"
     
     if is_disk_full():
         os.system('rm -rf /data/models*')
         os.system('rm -r ~/.cache/huggingface/hub')
         
-    return predict(text, model_id, tokenizer_id)
-
-#demo = gr.Interface(
-#    title="CAP Media Topics Babel Demo",
-#    fn=predict_cap,
-#    inputs=[gr.Textbox(lines=6, label="Input"),
-#            gr.Dropdown(languages, label="Language"),
-#            gr.Dropdown(domains.keys(), label="Domain")],
-#    outputs=[gr.Label(num_top_classes=5, label="Output"), gr.Markdown()])
-with gr.Blocks() as demo:
-    gr.Markdown("""# CAP Media Topics Babel Demo
-    ## 🚧 Coming Soon 🚧""")
+    return predict(text, major_model_id, minor_model_id, tokenizer_id)
+
+demo = gr.Interface(
+    title="CAP Media Topics Babel Demo",
+    fn=predict_cap,
+    inputs=[gr.Textbox(lines=6, label="Input"),
+            gr.Dropdown(languages, label="Language"),
+            gr.Dropdown(domains.keys(), label="Domain")],
+    outputs=[gr.Label(num_top_classes=5, label="Output"), gr.Markdown()])