Upload artifacts/inference_pathora.py with huggingface_hub

artifacts/inference_pathora.py

@@ -0,0 +1,51 @@
+import numpy as np, re, joblib, torch
+import tensorflow as tf
+from tensorflow import keras
+from tensorflow.keras import backend as K
+from transformers import BertTokenizer, BertModel
+from sentence_transformers import SentenceTransformer
+
+class FeatureAttention(keras.layers.Layer):
+    def __init__(self, **kw): super().__init__(**kw)
+    def build(self, s): self.W = self.add_weight(shape=(s[-1],), initializer="ones", trainable=True)
+    def call(self, x): return x * tf.nn.softmax(self.W)
+
+@keras.saving.register_keras_serializable()
+def focal_loss(g=2.0, a=0.5):
+    def fn(y, p):
+        y = tf.squeeze(tf.cast(y, tf.int32))
+        p = K.clip(p, K.epsilon(), 1-K.epsilon())
+        ce = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y, logits=tf.math.log(p+K.epsilon()))
+        idx = tf.stack([tf.range(tf.shape(p)[0], dtype=tf.int32), y], axis=-1)
+        return K.mean(a * K.pow(1-tf.gather_nd(p,idx), g) * ce)
+    return fn
+
+class PathOraPredictor:
+    def __init__(self, model_path="pathora_model.keras", le_path="label_encoder.joblib"):
+        self.model = keras.models.load_model(model_path, custom_objects={"FeatureAttention":FeatureAttention,"loss_fn":focal_loss()})
+        self.le = joblib.load(le_path)
+        self.device = torch.device("cpu")
+        self.tokenizer = BertTokenizer.from_pretrained("bert-resume-classifier-final")
+        self.bert = BertModel.from_pretrained("bert-resume-classifier-final").to(self.device)
+        self.bert.eval()
+        self.st = SentenceTransformer("all-MiniLM-L6-v2")
+
+    def predict(self, text, top_n=5):
+        text = re.sub(r"<[^>]+>", " ", re.sub(r"\s+", " ", text)).strip()
+        enc = self.tokenizer(text, padding=True, truncation=True, max_length=128, return_tensors="pt")
+        with torch.no_grad():
+            out = self.bert(**enc)
+        bert_emb = out.last_hidden_state[:, 0, :].numpy()
+        st_emb = self.st.encode([text], convert_to_numpy=True)
+        emb = np.concatenate([bert_emb, st_emb], axis=1)
+        probs = self.model.predict(emb, verbose=0)[0]
+        top = np.argsort(probs)[::-1][:top_n]
+        return [(self.le.classes_[i], float(probs[i])) for i in top]
+
+if __name__ == "__main__":
+    p = PathOraPredictor()
+    for s in ["Python developer, TensorFlow, AWS.", "Financial analyst, CFA, Excel.", "Nurse, ICU, BLS."]:
+        print("Text:", s)
+        for cat, conf in p.predict(s):
+            print("  {}: {:.1%}".format(cat, conf))
+        print()