Merge pull request #14 from boostcampaitech4lv23nlp2/feat/dashboard

Feat/dashboard add dashboard to analyze result

.github/workflows/pr-reminder.yml

@@ -0,0 +1,23 @@
+name: PRs reviews reminder
+
+on:
+  schedule:
+    - cron: "0 1 * * *" #KST: 10:00
+    - cron: "0 3 * * *" #KST: 12:00
+    - cron: "0 4 * * *" #KST: 13:00
+    - cron: "0 7 * * *" #KST: 16:00
+    - cron: "0 9 * * *" #KST: 18:00
+
+jobs:
+  pr-reviews-reminder:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: davideviolante/[email protected]
+      env:
+        GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+      with:
+        webhook-url: 'https://hooks.slack.com/services/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' # Required
+        provider: 'slack' # Required (slack or msteams)
+        channel: '#ecl-free-talking' # Optional, eg: #general
+        github-provider-map: 'wbin0718:U041WE3RDMX,FacerAin:U041WE4P8GZ,ghlrobin:U041HN2FGMR,kyc3492:U041388FBM5,jinmyeongAN:U041HR962M8' # Optional, eg: DavideViolante:UEABCDEFG,foobar:UAABCDEFG
+        ignore-label: '' # Optional, eg: no-reminder

Makefile

@@ -3,6 +3,7 @@ quality: set-style-dep check-quality
 style: set-style-dep set-style
 setup: set-precommit set-style-dep set-test-dep set-git set-dev set-dataset set-directory
 test: set-test-dep set-test
+dashboard: set-dashboard
 
 
 ##### basic #####
@@ -58,4 +59,8 @@ set-directory:
 	mkdir -p ./src/prediction
 	mkdir -p ./src/logs
 	mkdir -p ./src/best_model
-	mkdir -p ./src/results
+	mkdir -p ./src/results
+
+
+set-dashboard:
+	streamlit run dashboard/app.py

dashboard/app.py

@@ -0,0 +1,46 @@
+import argparse
+import pickle as pickle
+
+import pandas as pd
+import streamlit as st
+
+from .utils import get_filtered_result, test
+
+
+def app(args):
+    """Run streamlit app"""
+    test_df = pd.read_csv(args.valid_data_path)
+
+    st.set_page_config(page_icon="❄️", page_title="Into the RE", layout="wide")
+
+    st.title("Into the Re")
+
+    result_df = test(args)
+    filtered_df = get_filtered_result(result_df, test_df)
+
+    st.dataframe(filtered_df)
+    st.text(f"전체 {len(test_df)} 중 {len(filtered_df)}개를 틀렸습니다.")
+    st.text("실제 정답 분포")
+    st.bar_chart(filtered_df["answer"].value_counts())
+    st.text("예측 라벨 분포")
+    st.bar_chart(filtered_df["pred_label"].value_counts())
+
+
+parser = argparse.ArgumentParser()
+
+parser.add_argument("--model_name", default="klue/bert-base", type=str)
+parser.add_argument(
+    "--model_dir",
+    default="src/best_model",
+    type=str,
+)
+parser.add_argument(
+    "--valid_data_path",
+    default="dataset/train/dev.csv",
+    type=str,
+)
+
+
+args = parser.parse_args()
+
+app(args)

dashboard/utils.py

@@ -0,0 +1,208 @@
+import json
+import pickle as pickle
+
+import numpy as np
+import pandas as pd
+import torch
+import torch.nn.functional as F
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+from transformers import AutoModelForSequenceClassification, AutoTokenizer
+
+DICT_NUM_TO_LABEL_PATH = "dashboard/dict_num_to_label.pkl"
+with open(DICT_NUM_TO_LABEL_PATH, "rb") as f:
+    dict_num_to_label = pickle.load(f)
+
+
+def inference(model, tokenized_sent, device):
+    """
+    test dataset을 DataLoader로 만들어 준 후,
+    batch_size로 나눠 model이 예측 합니다.
+    """
+    dataloader = DataLoader(tokenized_sent, batch_size=16, shuffle=False)
+    model.eval()
+    output_pred = []
+    output_prob = []
+    for i, data in enumerate(tqdm(dataloader)):
+        with torch.no_grad():
+            outputs = model(
+                input_ids=data["input_ids"].to(device),
+                attention_mask=data["attention_mask"].to(device),
+                token_type_ids=data["token_type_ids"].to(device),
+            )
+        logits = outputs[0]
+        prob = F.softmax(logits, dim=-1).detach().cpu().numpy()
+        logits = logits.detach().cpu().numpy()
+        result = np.argmax(logits, axis=-1)
+
+        output_pred.append(result)
+        output_prob.append(prob)
+
+    return (
+        np.concatenate(output_pred).tolist(),
+        np.concatenate(output_prob, axis=0).tolist(),
+    )
+
+
+def tokenized_dataset(dataset, tokenizer):
+    """tokenizer에 따라 sentence를 tokenizing 합니다."""
+    concat_entity = []
+    for e01, e02 in zip(dataset["subject_entity"], dataset["object_entity"]):
+        temp = ""
+        temp = e01 + "[SEP]" + e02
+        concat_entity.append(temp)
+    tokenized_sentences = tokenizer(
+        concat_entity,
+        list(dataset["sentence"]),
+        return_tensors="pt",
+        padding=True,
+        truncation=True,
+        max_length=256,
+        add_special_tokens=True,
+    )
+    return tokenized_sentences
+
+
+def preprocessing_dataset(dataset):
+    """처음 불러온 csv 파일을 원하는 형태의 DataFrame으로 변경 시켜줍니다."""
+    subject_entity = []
+    object_entity = []
+    for i, j in zip(dataset["subject_entity"], dataset["object_entity"]):
+        i = i[1:-1].split(",")[0].split(":")[1]
+        j = j[1:-1].split(",")[0].split(":")[1]
+
+        subject_entity.append(i)
+        object_entity.append(j)
+    out_dataset = pd.DataFrame(
+        {
+            "id": dataset["id"],
+            "sentence": dataset["sentence"],
+            "subject_entity": subject_entity,
+            "object_entity": object_entity,
+            "label": dataset["label"],
+        }
+    )
+    return out_dataset
+
+
+def load_data(dataset_dir):
+    """csv 파일을 경로에 맡게 불러 옵니다."""
+    pd_dataset = pd.read_csv(dataset_dir)
+    dataset = preprocessing_dataset(pd_dataset)
+
+    return dataset
+
+
+def load_test_dataset(dataset_dir, tokenizer):
+    """
+    test dataset을 불러온 후,
+    tokenizing 합니다.
+    """
+    test_dataset = load_data(dataset_dir)
+    test_label = [100 for i in range(len(test_dataset))]
+    # tokenizing dataset
+    tokenized_test = tokenized_dataset(test_dataset, tokenizer)
+    return test_dataset["id"], tokenized_test, test_label
+
+
+def num_to_label(label):
+    """
+    숫자로 되어 있던 class를 원본 문자열 라벨로 변환 합니다.
+    """
+    origin_label = []
+    for v in label:
+        origin_label.append(dict_num_to_label[v])
+
+    return origin_label
+
+
+class RE_Dataset(torch.utils.data.Dataset):
+    """Dataset 구성을 위한 class."""
+
+    def __init__(self, pair_dataset, labels):
+        self.pair_dataset = pair_dataset
+        self.labels = labels
+
+    def __getitem__(self, idx):
+        item = {key: val[idx].clone().detach() for key, val in self.pair_dataset.items()}
+        item["labels"] = torch.tensor(self.labels[idx])
+        return item
+
+    def __len__(self):
+        return len(self.labels)
+
+
+def get_topn_probs(probs, n=3):
+    """_summary_
+
+    Args:
+        probs (_type_): _description_
+        n (int, optional): _description_. Defaults to 3.
+
+    Returns:
+        _type_: _description_
+    """
+    pairs = []
+    top_n_idxs = list(reversed(np.array(probs).argsort()))[:n]
+    for idx in top_n_idxs:
+        pairs.append((dict_num_to_label[idx], probs[idx]))
+    return pairs
+
+
+def get_entity_word(row):
+    """_summary_
+
+    Args:
+        row (_type_): _description_
+
+    Returns:
+        _type_: _description_
+    """
+    row = row.replace("'", '"')
+    return json.loads(row)["word"]
+
+
+def get_filtered_result(new_df, test_df):
+    """_summary_
+
+    Args:
+        new_df (_type_): _description_
+        test_df (_type_): _description_
+
+    Returns:
+        _type_: _description_
+    """
+    new_df["sentence"] = test_df["sentence"]
+    new_df["answer"] = test_df["label"]
+    new_df["subject"] = test_df["subject_entity"].apply(get_entity_word)
+    new_df["object"] = test_df["object_entity"].apply(get_entity_word)
+    new_df["probs"] = new_df["probs"].apply(get_topn_probs)
+    new_df = new_df.loc[new_df["pred_label"] != new_df["answer"]]
+    new_df = new_df[["sentence", "subject", "object", "pred_label", "answer", "probs"]]
+    return new_df
+
+
+def test(args):
+    """Perform a test using model of model_dir
+
+    Returns:
+        _type_: pd.DataFrame
+    """
+    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+    tokenizer = AutoTokenizer.from_pretrained(args.model_name)
+
+    model = AutoModelForSequenceClassification.from_pretrained(args.model_dir)
+    model.to(device)
+
+    test_id, test_dataset, test_label = load_test_dataset(args.valid_data_path, tokenizer)
+    Re_test_dataset = RE_Dataset(test_dataset, test_label)
+
+    pred_answer, output_prob = inference(model, Re_test_dataset, device)
+    pred_answer = num_to_label(pred_answer)
+    output = pd.DataFrame(
+        {
+            "pred_label": pred_answer,
+            "probs": output_prob,
+        }
+    )
+    return output

requirements.txt

@@ -3,4 +3,5 @@ scikit-learn~=0.24.1
 transformers==4.10.0
 pytorch-lightning==1.7.7
 pyyaml==6.0
-mlflow==2.0.1
+mlflow==2.0.1
+streamlit==1.14.1