🎉 Add app

.gitignore

@@ -0,0 +1,5 @@
+.env
+
+.DS_Store
+
+__pycache__

app.py

@@ -0,0 +1,57 @@
+import streamlit as st
+import inference
+import pandas as pd
+import const
+
+st.title("Japanese to Emotion classification")
+st.write(
+    "I fine-tuned the BERT-based distillation model for classification of Japanese text."
+)
+
+if "input_text" not in st.session_state:
+    st.session_state.input_text = ""
+
+input_text = st.text_area(
+    "Japanese text", value=st.session_state.input_text, max_chars=512
+)
+
+suggestions = ["今日は愛犬と散歩した", "猫カフェ行きたい", "自転車盗まれた"]
+
+COLUMNS_NUM = len(suggestions)
+cols = st.columns(COLUMNS_NUM)
+
+for i, suggestion in enumerate(suggestions):
+    with cols[i]:
+        if st.button(suggestion, use_container_width=True):
+            st.session_state.input_text = suggestion
+            st.rerun()
+
+st.session_state.input_text = input_text
+
+if input_text:
+    probs_dict = inference.exec(input_text)
+
+    label_dict = {
+        const.EMOTIONS[0]: "😊 Joy",
+        const.EMOTIONS[1]: "😢 Sadness",
+        const.EMOTIONS[2]: "😮 Anticipation",
+        const.EMOTIONS[3]: "😲 Surprise",
+        const.EMOTIONS[4]: "😠 Anger",
+        const.EMOTIONS[5]: "😨 Fear",
+        const.EMOTIONS[6]: "😖 Disgust",
+        const.EMOTIONS[7]: "😉 Trust",
+    }
+
+    df = pd.DataFrame(
+        {
+            "Emotion": label_dict.values(),
+            "Probs": [probs_dict[emotion] for emotion in const.EMOTIONS],
+        }
+    )
+
+    st.bar_chart(df.set_index("Emotion"), horizontal=True)
+
+st.write('''
+- [GitHub](https://github.com/koshin01)
+- [Blog](https://zenn.dev/koshin)
+''')

const.py

@@ -0,0 +1,10 @@
+EMOTIONS = [
+    "Joy",
+    "Sadness",
+    "Anticipation",
+    "Surprise",
+    "Anger",
+    "Fear",
+    "Disgust",
+    "Trust",
+]

inference.py

@@ -0,0 +1,35 @@
+from transformers import AutoModelForSequenceClassification, AutoTokenizer
+import torch
+import torch.nn.functional as F
+import const
+
+
+def load_model():
+    return AutoModelForSequenceClassification.from_pretrained(
+        "koshin2001/Japanese-to-emotions"
+    ).eval()
+
+
+def load_tokenizer():
+    return AutoTokenizer.from_pretrained("koshin2001/Japanese-to-emotions")
+
+
+def exec(text):
+    model = load_model()
+    tokenizer = load_tokenizer()
+
+    inputs = tokenizer(
+        text,
+        return_tensors="pt",
+        truncation=True,
+        return_token_type_ids=False,
+        max_length=512,
+    )
+    output = model(**inputs)
+
+    output_logits = torch.tensor(output.logits).clone().detach().requires_grad_(True)
+    probs = F.softmax(output_logits, dim=-1).tolist()[0]
+
+    emotion_probs = dict(zip(const.EMOTIONS, probs))
+
+    return emotion_probs

requirements.txt

@@ -0,0 +1 @@
+transformers == 4.44.2