Add fashion-mnist(code)

docker-compose.yaml

@@ -27,3 +27,12 @@ services:
     volumes:
       - .:/src
       - ~/.amplify:/src/src/token
+  run-fashion-mnist:
+    command:
+      - "python"
+      - "-u"
+      - "fashion-mnist.py"
+    build: .
+    volumes:
+      - .:/src
+      - ~/.amplify:/src/src/token

src/fashion-mnist.py

@@ -0,0 +1,122 @@
+import pandas as pd
+import matplotlib.pyplot as plt
+from sklearn.metrics import roc_auc_score, accuracy_score
+from omegaconf import OmegaConf
+from torch.utils.data import Dataset, DataLoader
+
+from utils import *
+
+conf = """
+base:
+  api_key_path: "token/token.json"
+  train_path: "../data/fashion_train_400.csv"
+  seed: 67
+  
+dataset:
+  img_size: 20 
+  features: 400  # 特徴量の数
+  val_size: 300  # validationに使うデータ数
+  target: 5  # ラベル
+  
+model:
+  timeout: 3000  # ms  計算時間
+  # 訓練に使うデータは(batch_size * n_iter)個
+  batch_size: 20  # バッチサイズ
+  n_iter: 5  # ループ数
+  l: 2  # 正則化項
+  each_weight: 1  # 重み係数
+  length_weight: 3  # 重みの層の数
+  multiprocessing: true
+"""
+
+cfg = OmegaConf.create(conf)
+
+init_client(cfg)
+
+
+# PyTorch形式
+class MyDataset(Dataset):
+    def __init__(self, data, label):
+        self.data = (data / 255).astype(np.float16)
+        self.label = label
+
+        self.label = np.where(self.label == cfg.dataset.target, 1, 0)
+        self.data[self.data == 0] = -1
+
+    def __len__(self):
+        return self.data.shape[0]
+
+    def __getitem__(self, idx):
+        return self.data[idx, :], self.label[idx]
+
+
+def get_ds(n, seed):
+    true_ = dataset.loc[dataset.iloc[:, 0] == cfg.dataset.target, :].sample(n // 2)
+    false_ = dataset.loc[dataset.iloc[:, 0] != cfg.dataset.target, :].sample(n // 2)
+    return pd.concat(
+        [
+            true_,
+            false_
+        ]
+    ).sample(
+        frac=1,
+        random_state=seed
+    ).values
+
+
+dataset = pd.read_csv(cfg.base.train_path)  # .sample(100).values
+for i in range(5, 10):
+    init_client(cfg)
+    cfg.dataset.target = i
+    train = get_ds(
+        int(
+            cfg.model.n_iter * cfg.model.batch_size
+        ),
+        cfg.base.seed
+    )
+
+    val = get_ds(
+        cfg.dataset.val_size,
+        cfg.base.seed + 1
+    )
+
+    train_ds = MyDataset(train[:, 1:], train[:, 0])
+    valid_ds = MyDataset(val[:, 1:], val[:, 0])
+    train_dl = DataLoader(
+        train_ds,
+        batch_size=cfg.model.batch_size,
+        shuffle=True
+    )
+    valid_dl = DataLoader(
+        valid_ds,
+        batch_size=64,
+        shuffle=False
+    )
+    weight = np.zeros(
+        (
+            cfg.dataset.features,
+            cfg.model.length_weight
+        )
+    )
+
+    weight = run_model(cfg, train_dl, weight, cfg.model.multiprocessing)
+    pred, label = eval_model(cfg, valid_dl, weight)
+
+    print("=" * 20, i, "=" * 20)
+    print("AUC:", roc_auc_score(label, pred))
+    print("ACC:", accuracy_score(label, np.round(pred)))
+    print("=" * 43)
+    weight = weight.sum(axis=1) * cfg.model.each_weight
+
+    plt.imshow(
+        weight.reshape(
+            cfg.dataset.img_size,
+            cfg.dataset.img_size
+        )
+    )
+    plt.axis('tight')
+    plt.axis('off')
+    plt.subplots_adjust(left=0, right=1, bottom=0, top=1)
+    plt.savefig(
+        f"../img/Fashion-MNIST_{cfg.dataset.target}_weight.png"
+    )