Changes to be committed: 전처리 오류 수정, 학습 루프 구성

modified: DIVA_dataset.pt
modified: Models/Vector2MIDI.py
modified: preprocessing.ipynb
modified: train.ipynb

DIVA_dataset.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:51a657804e01360dbf4ae774d45e959d3955e3be0b7f9a84e467c5911d5f7cc3
-size 243341
+oid sha256:1f4db440c793f1db309541cace07cf4f2b83290173f9d5889ca31349fbde0377
+size 243790

Models/Vector2MIDI.py

@@ -1,45 +1,83 @@
-from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
-import torch.nn as nn
 import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
+
 
 class Vector2MIDI(nn.Module):
     def __init__(self, input_dim, hidden_dim, n_vocab, dropout=0.2):
-        super().__init__()  # 부모 클래스 생성자 호출
-        self.input_fc = nn.Linear(input_dim, hidden_dim) # 입력 차원에서 은닉 차원으로 변환
+        super().__init__()
+
+        self.n_vocab = n_vocab
+
+        self.init_hidden = nn.Linear(input_dim, hidden_dim)
+        self.init_cell = nn.Linear(input_dim, hidden_dim)
 
         # 과적합 방지 드롭아웃 LSTM
-        self.lstm = nn.LSTM(hidden_dim, hidden_dim, num_layers=2, batch_first=True, dropout=dropout)
+        self.lstm = nn.LSTM(n_vocab, hidden_dim, num_layers=2, batch_first=True, dropout=dropout)
 
         self.fc_mid = nn.Linear(hidden_dim, 256)
         self.fc_out = nn.Linear(256, n_vocab)
 
-    def forward(self, x, lengths, total_length=None):
-        print("input to forward:", x.shape)
-        B, feat_dim = x.size()
-        T = lengths.max()
-
-        # [B, 1, feat_dim] → [B, T, feat_dim]
-        x = x.unsqueeze(1).expand(B, T, feat_dim)
+    def forward(self, x, lengths, target_tokens):
+        """
+        x: (B, input_dim) - 입력 벡터
+        lengths: [B] - 시퀀스 길이
+        target_tokens: (B, T, n_vocab) - one-hot 또는 임베딩된 토큰 입력
+        """
+        B = x.size(0)
 
-        x = self.input_fc(x)
+        h0 = self.init_hidden(x).unsqueeze(0).repeat(2, 1, 1)  # (num_layers, B, H)
+        c0 = self.init_cell(x).unsqueeze(0).repeat(2, 1, 1)
 
-        packed_x = nn.utils.rnn.pack_padded_sequence(
-            x, lengths.cpu(), batch_first=True, enforce_sorted=False
-        )
-        packed_out, _ = self.lstm(packed_x)
-
-        out, _ = nn.utils.rnn.pad_packed_sequence(
-            packed_out, batch_first=True, total_length=total_length
-        )
+        packed_input = pack_padded_sequence(target_tokens, lengths.cpu(), batch_first=True, enforce_sorted=False)
+        packed_out, _ = self.lstm(packed_input, (h0, c0))
+        out, _ = pad_packed_sequence(packed_out, batch_first=True)
 
         out = self.fc_mid(out)
-        out = self.fc_out(out)  # [B, max_len, vocab_size]
+        out = self.fc_out(out)  # (B, T, vocab_size)
+
         return out
 
-    def generate(self, x, lengths, total_length=None):
-        out = self.forward(x, lengths, total_length)
+    def generate(self, vector, device, max_len=512, temperature=1.0, top_k=None, start_token_idx=0, end_token_idx=None):
+        """
+        스타일 벡터 하나로 시퀀스 생성 (확률적 샘플링 기반)
+        """
+        self.eval()
+
+        vector = vector.to(device)
+        h = self.init_hidden(vector).unsqueeze(0).repeat(2, 1, 1)  # (num_layers, 1, hidden)
+        c = self.init_cell(vector).unsqueeze(0).repeat(2, 1, 1)
+
+        # one-hot start token
+        x = F.one_hot(torch.tensor([start_token_idx], device=device), num_classes=self.n_vocab).float()
+        x = x.unsqueeze(1)  # (1, 1, n_vocab)
+
+        outputs = []
+
+        for _ in range(max_len):
+            out, (h, c) = self.lstm(x, (h, c))  # (1, 1, hidden)
+            out = self.fc_mid(out)
+            logits = self.fc_out(out).squeeze(0).squeeze(0)  # (n_vocab,)
+
+            # temperature scaling
+            logits = logits / temperature
+            probs = F.softmax(logits, dim=-1)
+
+            # top-k filtering
+            if top_k is not None:
+                top_vals, top_idx = torch.topk(probs, top_k)
+                probs = torch.zeros_like(probs).scatter_(0, top_idx, top_vals)
+                probs = probs / probs.sum()
+
+            pred_token = torch.multinomial(probs, 1).item()
+
+            if end_token_idx is not None and pred_token == end_token_idx:
+                break
+
+            outputs.append(pred_token)
+
+            # 다음 timestep의 입력으로 사용
+            x = F.one_hot(torch.tensor([pred_token], device=device), num_classes=self.n_vocab).float().unsqueeze(1)
 
-        preds = torch.argmax(out, dim=-1)  # [B, T], 가장 큰 점수 클래스 선택
-        external = preds - 2                  # 내부 표현 → 외부 표현
-        external[external == -2] = 0          # PAD 처리
-        return external
+        return outputs

preprocessing.ipynb

@@ -307,66 +307,10 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 3,
+   "execution_count": 7,
    "id": "f7b77c0c",
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "tensor([[[81,  3, 65,  ...,  3, 53,  3],\n",
-      "         [ 0,  1,  0,  ...,  1,  0,  1],\n",
-      "         [81,  2, 65,  ...,  2, 53,  2],\n",
-      "         ...,\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1],\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1],\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1]],\n",
-      "\n",
-      "        [[77,  2, 65,  ...,  2, 53,  2],\n",
-      "         [ 0,  2,  0,  ...,  2,  0,  2],\n",
-      "         [89,  1, 65,  ...,  1, 53,  1],\n",
-      "         ...,\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1],\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1],\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1]],\n",
-      "\n",
-      "        [[78,  2, 63,  ...,  2, 51,  2],\n",
-      "         [ 0,  2,  0,  ...,  2,  0,  2],\n",
-      "         [78,  1, 63,  ...,  1, 51,  2],\n",
-      "         ...,\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1],\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1],\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1]],\n",
-      "\n",
-      "        ...,\n",
-      "\n",
-      "        [[74,  2, 62,  ...,  2, 50,  2],\n",
-      "         [ 0,  2,  0,  ...,  2,  0,  2],\n",
-      "         [76,  2, 62,  ...,  2, 50,  2],\n",
-      "         ...,\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1],\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1],\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1]],\n",
-      "\n",
-      "        [[ 0,  4,  0,  ...,  4, 53,  4],\n",
-      "         [91,  2,  0,  ...,  2, 53,  2],\n",
-      "         [ 0,  2,  0,  ...,  2,  0,  2],\n",
-      "         ...,\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1],\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1],\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1]],\n",
-      "\n",
-      "        [[75,  2, 68,  ...,  2, 51,  2],\n",
-      "         [ 0,  2,  0,  ...,  2,  0,  2],\n",
-      "         [84,  2, 68,  ...,  2, 51,  2],\n",
-      "         ...,\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1],\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1],\n",
-      "         [-1, -1, -1,  ..., -1, -1, -1]]])\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
     "from sklearn.preprocessing import OneHotEncoder, MinMaxScaler\n",
     "from sklearn.compose import ColumnTransformer\n",
@@ -388,16 +332,15 @@
     "X_tensor = torch.tensor(X, dtype=torch.float32)\n",
     "Y_tensor = [torch.tensor(item['token'], dtype=torch.long) for item in tokenized_data]\n",
     "\n",
-    "seq_lengths = [len(seq) for seq in Y_tensor]\n",
+    "seq_lengths = torch.tensor([len(seq) for seq in Y_tensor])\n",
     "\n",
     "# 패딩 처리\n",
-    "padded_Y = pad_sequence(Y_tensor, batch_first=True, padding_value=-1)  # (batch_size, max_len, 7)\n",
-    "print(padded_Y)"
+    "padded_Y = pad_sequence(Y_tensor, batch_first=True, padding_value=-1)  # (batch_size, max_len, 7)"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 4,
+   "execution_count": 8,
    "id": "dd840788",
    "metadata": {},
    "outputs": [
@@ -406,18 +349,20 @@
      "output_type": "stream",
      "text": [
       "X shape: torch.Size([34, 25])\n",
-      "Y shape: torch.Size([34, 125, 7])\n"
+      "Y shape: torch.Size([34, 125, 7])\n",
+      "l shape: torch.Size([34])\n"
      ]
     }
    ],
    "source": [
     "print(\"X shape:\", X_tensor.shape)\n",
-    "print(\"Y shape:\", padded_Y.shape)"
+    "print(\"Y shape:\", padded_Y.shape)\n",
+    "print(\"l shape:\", seq_lengths.shape)"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 5,
+   "execution_count": 9,
    "id": "4f5f5dc1",
    "metadata": {},
    "outputs": [],

train.ipynb

@@ -32,21 +32,24 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 1,
+   "execution_count": null,
    "id": "630dd7ad",
    "metadata": {},
    "outputs": [],
    "source": [
     "from Models.Vector2MIDI import Vector2MIDI  # 클래스 정의가 필요\n",
     "import torch.optim as optim\n",
-    "import torch.nn as nn\n",
+    "from torch.nn import HuberLoss\n",
+    "from pysdtw import SoftDTW\n",
+    "#from utility.lossf import get_loss_function # 나중에 직접 해보자\n",
     "import torch\n",
     "\n",
     "device = torch.device(\"cuda\") # GPU 사용\n",
     "#device = torch.device(\"cpu\") # CPU 사용\n",
     "\n",
-    "model = Vector2MIDI(25, 128, 303).to(device)\n",
-    "criterion = nn.CrossEntropyLoss(ignore_index=0) # 손실함수 패딩(0) 무시\n",
+    "model = Vector2MIDI(25, 1024, 7).to(device)\n",
+    "sdtw = SoftDTW(0.6) # Soft Dynamic Time Warping (timestep 끼리 비교해 loss 계산 ->  gradient 가 흐르도록 함) https://judy-son.tistory.com/3\n",
+    "huber = HuberLoss(reduction='none', delta=1.0).to(device) # HuberLoss (reduction='none'로 개별 timestep loss 계산)\n",
     "optimizer = optim.Adam(model.parameters(), lr=1e-3)"
    ]
   },
@@ -61,7 +64,7 @@
      "output_type": "stream",
      "text": [
       "X_tensor shape: torch.Size([34, 25])\n",
-      "Y_tensor shape: torch.Size([34, 1185])\n",
+      "Y_tensor shape: torch.Size([34, 125, 7])\n",
       "lengths shape: torch.Size([34])\n"
      ]
     }
@@ -69,12 +72,12 @@
    "source": [
     "# 전처리 데이터 로드\n",
     "from torch.utils.data import DataLoader\n",
-    "from dataset import MIDIDataset\n",
+    "from utility.dataset import MIDIDataset\n",
     "import torch\n",
     "\n",
     "data = torch.load(\"DIVA_dataset.pt\")\n",
-    "X_tensor = data[\"X\"]\n",
-    "Y_tensor = data[\"Y\"]\n",
+    "X_tensor = data[\"X\"].float()\n",
+    "Y_tensor = data[\"Y\"].float()\n",
     "lengths = data[\"lengths\"]\n",
     "\n",
     "print(\"X_tensor shape:\", X_tensor.shape)\n",
@@ -124,521 +127,38 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 4,
+   "execution_count": 10,
    "id": "16a14b5f",
    "metadata": {},
    "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "c:\\Users\\rrayy\\anaconda3\\envs\\diva\\Lib\\site-packages\\numba\\cuda\\dispatcher.py:536: NumbaPerformanceWarning: \u001b[1mGrid size 8 will likely result in GPU under-utilization due to low occupancy.\u001b[0m\n",
+      "  warn(NumbaPerformanceWarning(msg))\n"
+     ]
+    },
     {
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "Epoch 1, Loss: 5.5722\n",
-      "Epoch 2, Loss: 4.4941\n",
-      "Epoch 3, Loss: 3.0559\n",
-      "Epoch 4, Loss: 2.9349\n",
-      "Epoch 5, Loss: 2.8127\n",
-      "Epoch 6, Loss: 2.8296\n",
-      "Epoch 7, Loss: 2.8393\n",
-      "Epoch 8, Loss: 2.7751\n",
-      "Epoch 9, Loss: 2.7641\n",
-      "Epoch 10, Loss: 2.7667\n",
-      "Epoch 11, Loss: 2.7747\n",
-      "Epoch 12, Loss: 2.7678\n",
-      "Epoch 13, Loss: 2.7582\n",
-      "Epoch 14, Loss: 2.7487\n",
-      "Epoch 15, Loss: 2.7586\n",
-      "Epoch 16, Loss: 2.7464\n",
-      "Epoch 17, Loss: 2.7516\n",
-      "Epoch 18, Loss: 2.7802\n",
-      "Epoch 19, Loss: 2.7607\n",
-      "Epoch 20, Loss: 2.7453\n",
-      "Epoch 21, Loss: 2.7715\n",
-      "Epoch 22, Loss: 2.7349\n",
-      "Epoch 23, Loss: 2.7714\n",
-      "Epoch 24, Loss: 2.7450\n",
-      "Epoch 25, Loss: 2.7568\n",
-      "Epoch 26, Loss: 2.7671\n",
-      "Epoch 27, Loss: 2.7577\n",
-      "Epoch 28, Loss: 2.7876\n",
-      "Epoch 29, Loss: 2.7568\n",
-      "Epoch 30, Loss: 2.7816\n",
-      "Epoch 31, Loss: 2.7863\n",
-      "Epoch 32, Loss: 2.7670\n",
-      "Epoch 33, Loss: 2.7259\n",
-      "Epoch 34, Loss: 2.6940\n",
-      "Epoch 35, Loss: 2.7501\n",
-      "Epoch 36, Loss: 2.7534\n",
-      "Epoch 37, Loss: 2.7127\n",
-      "Epoch 38, Loss: 2.7385\n",
-      "Epoch 39, Loss: 2.7298\n",
-      "Epoch 40, Loss: 2.7018\n",
-      "Epoch 41, Loss: 2.7540\n",
-      "Epoch 42, Loss: 2.7208\n",
-      "Epoch 43, Loss: 2.7112\n",
-      "Epoch 44, Loss: 2.6953\n",
-      "Epoch 45, Loss: 2.7329\n",
-      "Epoch 46, Loss: 2.7132\n",
-      "Epoch 47, Loss: 2.7179\n",
-      "Epoch 48, Loss: 2.6945\n",
-      "Epoch 49, Loss: 2.7133\n",
-      "Epoch 50, Loss: 2.7182\n",
-      "Epoch 51, Loss: 2.7321\n",
-      "Epoch 52, Loss: 2.7044\n",
-      "Epoch 53, Loss: 2.7128\n",
-      "Epoch 54, Loss: 2.7104\n",
-      "Epoch 55, Loss: 2.7089\n",
-      "Epoch 56, Loss: 2.7058\n",
-      "Epoch 57, Loss: 2.7132\n",
-      "Epoch 58, Loss: 2.7087\n",
-      "Epoch 59, Loss: 2.7084\n",
-      "Epoch 60, Loss: 2.7122\n",
-      "Epoch 61, Loss: 2.6939\n",
-      "Epoch 62, Loss: 2.6903\n",
-      "Epoch 63, Loss: 2.6926\n",
-      "Epoch 64, Loss: 2.6913\n",
-      "Epoch 65, Loss: 2.6882\n",
-      "Epoch 66, Loss: 2.7012\n",
-      "Epoch 67, Loss: 2.7040\n",
-      "Epoch 68, Loss: 2.7105\n",
-      "Epoch 69, Loss: 2.6827\n",
-      "Epoch 70, Loss: 2.7106\n",
-      "Epoch 71, Loss: 2.6837\n",
-      "Epoch 72, Loss: 2.6731\n",
-      "Epoch 73, Loss: 2.7332\n",
-      "Epoch 74, Loss: 2.6874\n",
-      "Epoch 75, Loss: 2.6647\n",
-      "Epoch 76, Loss: 2.6892\n",
-      "Epoch 77, Loss: 2.6797\n",
-      "Epoch 78, Loss: 2.6951\n",
-      "Epoch 79, Loss: 2.6704\n",
-      "Epoch 80, Loss: 2.6847\n",
-      "Epoch 81, Loss: 2.6603\n",
-      "Epoch 82, Loss: 2.6721\n",
-      "Epoch 83, Loss: 2.6744\n",
-      "Epoch 84, Loss: 2.6632\n",
-      "Epoch 85, Loss: 2.6685\n",
-      "Epoch 86, Loss: 2.6679\n",
-      "Epoch 87, Loss: 2.7077\n",
-      "Epoch 88, Loss: 2.6781\n",
-      "Epoch 89, Loss: 2.6045\n",
-      "Epoch 90, Loss: 2.6393\n",
-      "Epoch 91, Loss: 2.6398\n",
-      "Epoch 92, Loss: 2.6777\n",
-      "Epoch 93, Loss: 2.6694\n",
-      "Epoch 94, Loss: 2.5960\n",
-      "Epoch 95, Loss: 2.6379\n",
-      "Epoch 96, Loss: 2.6252\n",
-      "Epoch 97, Loss: 2.6191\n",
-      "Epoch 98, Loss: 2.5861\n",
-      "Epoch 99, Loss: 2.6027\n",
-      "Epoch 100, Loss: 2.5682\n",
-      "Epoch 101, Loss: 2.5900\n",
-      "Epoch 102, Loss: 2.5685\n",
-      "Epoch 103, Loss: 2.5704\n",
-      "Epoch 104, Loss: 2.5715\n",
-      "Epoch 105, Loss: 2.5142\n",
-      "Epoch 106, Loss: 2.5458\n",
-      "Epoch 107, Loss: 2.5558\n",
-      "Epoch 108, Loss: 2.5480\n",
-      "Epoch 109, Loss: 2.4956\n",
-      "Epoch 110, Loss: 2.4933\n",
-      "Epoch 111, Loss: 2.5003\n",
-      "Epoch 112, Loss: 2.5570\n",
-      "Epoch 113, Loss: 2.4918\n",
-      "Epoch 114, Loss: 2.4801\n",
-      "Epoch 115, Loss: 2.4920\n",
-      "Epoch 116, Loss: 2.4286\n",
-      "Epoch 117, Loss: 2.4576\n",
-      "Epoch 118, Loss: 2.4352\n",
-      "Epoch 119, Loss: 2.4678\n",
-      "Epoch 120, Loss: 2.4387\n",
-      "Epoch 121, Loss: 2.3994\n",
-      "Epoch 122, Loss: 2.4908\n",
-      "Epoch 123, Loss: 2.5221\n",
-      "Epoch 124, Loss: 2.4906\n",
-      "Epoch 125, Loss: 2.4710\n",
-      "Epoch 126, Loss: 2.4122\n",
-      "Epoch 127, Loss: 2.4305\n",
-      "Epoch 128, Loss: 2.4538\n",
-      "Epoch 129, Loss: 2.4856\n",
-      "Epoch 130, Loss: 2.4429\n",
-      "Epoch 131, Loss: 2.4655\n",
-      "Epoch 132, Loss: 2.4415\n",
-      "Epoch 133, Loss: 2.3814\n",
-      "Epoch 134, Loss: 2.3545\n",
-      "Epoch 135, Loss: 2.3763\n",
-      "Epoch 136, Loss: 2.3961\n",
-      "Epoch 137, Loss: 2.3468\n",
-      "Epoch 138, Loss: 2.3336\n",
-      "Epoch 139, Loss: 2.4034\n",
-      "Epoch 140, Loss: 2.3725\n",
-      "Epoch 141, Loss: 2.4021\n",
-      "Epoch 142, Loss: 2.3808\n",
-      "Epoch 143, Loss: 2.3289\n",
-      "Epoch 144, Loss: 2.3159\n",
-      "Epoch 145, Loss: 2.3318\n",
-      "Epoch 146, Loss: 2.3034\n",
-      "Epoch 147, Loss: 2.3529\n",
-      "Epoch 148, Loss: 2.3036\n",
-      "Epoch 149, Loss: 2.3199\n",
-      "Epoch 150, Loss: 2.3196\n",
-      "Epoch 151, Loss: 2.3636\n",
-      "Epoch 152, Loss: 2.3570\n",
-      "Epoch 153, Loss: 2.3199\n",
-      "Epoch 154, Loss: 2.3474\n",
-      "Epoch 155, Loss: 2.3376\n",
-      "Epoch 156, Loss: 2.3279\n",
-      "Epoch 157, Loss: 2.2390\n",
-      "Epoch 158, Loss: 2.2388\n",
-      "Epoch 159, Loss: 2.2676\n",
-      "Epoch 160, Loss: 2.2972\n",
-      "Epoch 161, Loss: 2.2829\n",
-      "Epoch 162, Loss: 2.2616\n",
-      "Epoch 163, Loss: 2.2766\n",
-      "Epoch 164, Loss: 2.2627\n",
-      "Epoch 165, Loss: 2.2495\n",
-      "Epoch 166, Loss: 2.2004\n",
-      "Epoch 167, Loss: 2.2778\n",
-      "Epoch 168, Loss: 2.2962\n",
-      "Epoch 169, Loss: 2.2827\n",
-      "Epoch 170, Loss: 2.2971\n",
-      "Epoch 171, Loss: 2.3118\n",
-      "Epoch 172, Loss: 2.3182\n",
-      "Epoch 173, Loss: 2.2547\n",
-      "Epoch 174, Loss: 2.2558\n",
-      "Epoch 175, Loss: 2.2282\n",
-      "Epoch 176, Loss: 2.2383\n",
-      "Epoch 177, Loss: 2.2618\n",
-      "Epoch 178, Loss: 2.3070\n",
-      "Epoch 179, Loss: 2.2940\n",
-      "Epoch 180, Loss: 2.2900\n",
-      "Epoch 181, Loss: 2.2707\n",
-      "Epoch 182, Loss: 2.2771\n",
-      "Epoch 183, Loss: 2.2522\n",
-      "Epoch 184, Loss: 2.2848\n",
-      "Epoch 185, Loss: 2.2608\n",
-      "Epoch 186, Loss: 2.2561\n",
-      "Epoch 187, Loss: 2.2359\n",
-      "Epoch 188, Loss: 2.2281\n",
-      "Epoch 189, Loss: 2.2654\n",
-      "Epoch 190, Loss: 2.2352\n",
-      "Epoch 191, Loss: 2.2345\n",
-      "Epoch 192, Loss: 2.2051\n",
-      "Epoch 193, Loss: 2.2366\n",
-      "Epoch 194, Loss: 2.2277\n",
-      "Epoch 195, Loss: 2.2157\n",
-      "Epoch 196, Loss: 2.2035\n",
-      "Epoch 197, Loss: 2.2165\n",
-      "Epoch 198, Loss: 2.1960\n",
-      "Epoch 199, Loss: 2.1935\n",
-      "Epoch 200, Loss: 2.1733\n",
-      "Epoch 201, Loss: 2.1952\n",
-      "Epoch 202, Loss: 2.2147\n",
-      "Epoch 203, Loss: 2.2009\n",
-      "Epoch 204, Loss: 2.2007\n",
-      "Epoch 205, Loss: 2.2009\n",
-      "Epoch 206, Loss: 2.1869\n",
-      "Epoch 207, Loss: 2.2003\n",
-      "Epoch 208, Loss: 2.1820\n",
-      "Epoch 209, Loss: 2.1792\n",
-      "Epoch 210, Loss: 2.1342\n",
-      "Epoch 211, Loss: 2.1702\n",
-      "Epoch 212, Loss: 2.1725\n",
-      "Epoch 213, Loss: 2.1858\n",
-      "Epoch 214, Loss: 2.2000\n",
-      "Epoch 215, Loss: 2.2105\n",
-      "Epoch 216, Loss: 2.2001\n",
-      "Epoch 217, Loss: 2.2053\n",
-      "Epoch 218, Loss: 2.1930\n",
-      "Epoch 219, Loss: 2.1952\n",
-      "Epoch 220, Loss: 2.1801\n",
-      "Epoch 221, Loss: 2.1974\n",
-      "Epoch 222, Loss: 2.1729\n",
-      "Epoch 223, Loss: 2.1888\n",
-      "Epoch 224, Loss: 2.1726\n",
-      "Epoch 225, Loss: 2.1784\n",
-      "Epoch 226, Loss: 2.1842\n",
-      "Epoch 227, Loss: 2.1666\n",
-      "Epoch 228, Loss: 2.1716\n",
-      "Epoch 229, Loss: 2.1842\n",
-      "Epoch 230, Loss: 2.2009\n",
-      "Epoch 231, Loss: 2.1572\n",
-      "Epoch 232, Loss: 2.1893\n",
-      "Epoch 233, Loss: 2.1743\n",
-      "Epoch 234, Loss: 2.1879\n",
-      "Epoch 235, Loss: 2.2010\n",
-      "Epoch 236, Loss: 2.1647\n",
-      "Epoch 237, Loss: 2.1824\n",
-      "Epoch 238, Loss: 2.1587\n",
-      "Epoch 239, Loss: 2.1600\n",
-      "Epoch 240, Loss: 2.1812\n",
-      "Epoch 241, Loss: 2.1634\n",
-      "Epoch 242, Loss: 2.1967\n",
-      "Epoch 243, Loss: 2.1885\n",
-      "Epoch 244, Loss: 2.2712\n",
-      "Epoch 245, Loss: 2.2293\n",
-      "Epoch 246, Loss: 2.2223\n",
-      "Epoch 247, Loss: 2.2170\n",
-      "Epoch 248, Loss: 2.1715\n",
-      "Epoch 249, Loss: 2.1775\n",
-      "Epoch 250, Loss: 2.1647\n",
-      "Epoch 251, Loss: 2.2222\n",
-      "Epoch 252, Loss: 2.2520\n",
-      "Epoch 253, Loss: 2.2165\n",
-      "Epoch 254, Loss: 2.2536\n",
-      "Epoch 255, Loss: 2.2271\n",
-      "Epoch 256, Loss: 2.2170\n",
-      "Epoch 257, Loss: 2.1954\n",
-      "Epoch 258, Loss: 2.2052\n",
-      "Epoch 259, Loss: 2.1957\n",
-      "Epoch 260, Loss: 2.1667\n",
-      "Epoch 261, Loss: 2.1562\n",
-      "Epoch 262, Loss: 2.1628\n",
-      "Epoch 263, Loss: 2.1560\n",
-      "Epoch 264, Loss: 2.1331\n",
-      "Epoch 265, Loss: 2.1412\n",
-      "Epoch 266, Loss: 2.1485\n",
-      "Epoch 267, Loss: 2.1643\n",
-      "Epoch 268, Loss: 2.1463\n",
-      "Epoch 269, Loss: 2.1454\n",
-      "Epoch 270, Loss: 2.1164\n",
-      "Epoch 271, Loss: 2.1524\n",
-      "Epoch 272, Loss: 2.1413\n",
-      "Epoch 273, Loss: 2.1966\n",
-      "Epoch 274, Loss: 2.1998\n",
-      "Epoch 275, Loss: 2.1584\n",
-      "Epoch 276, Loss: 2.2067\n",
-      "Epoch 277, Loss: 2.1642\n",
-      "Epoch 278, Loss: 2.1317\n",
-      "Epoch 279, Loss: 2.1466\n",
-      "Epoch 280, Loss: 2.1366\n",
-      "Epoch 281, Loss: 2.1396\n",
-      "Epoch 282, Loss: 2.0961\n",
-      "Epoch 283, Loss: 2.1695\n",
-      "Epoch 284, Loss: 2.1404\n",
-      "Epoch 285, Loss: 2.1688\n",
-      "Epoch 286, Loss: 2.1803\n",
-      "Epoch 287, Loss: 2.1537\n",
-      "Epoch 288, Loss: 2.1549\n",
-      "Epoch 289, Loss: 2.1642\n",
-      "Epoch 290, Loss: 2.1418\n",
-      "Epoch 291, Loss: 2.1355\n",
-      "Epoch 292, Loss: 2.1252\n",
-      "Epoch 293, Loss: 2.1335\n",
-      "Epoch 294, Loss: 2.1274\n",
-      "Epoch 295, Loss: 2.0980\n",
-      "Epoch 296, Loss: 2.1283\n",
-      "Epoch 297, Loss: 2.1466\n",
-      "Epoch 298, Loss: 2.1427\n",
-      "Epoch 299, Loss: 2.1472\n",
-      "Epoch 300, Loss: 2.1436\n",
-      "Epoch 301, Loss: 2.1546\n",
-      "Epoch 302, Loss: 2.1311\n",
-      "Epoch 303, Loss: 2.1920\n",
-      "Epoch 304, Loss: 2.1233\n",
-      "Epoch 305, Loss: 2.1415\n",
-      "Epoch 306, Loss: 2.1336\n",
-      "Epoch 307, Loss: 2.1153\n",
-      "Epoch 308, Loss: 2.1141\n",
-      "Epoch 309, Loss: 2.1147\n",
-      "Epoch 310, Loss: 2.1086\n",
-      "Epoch 311, Loss: 2.0999\n",
-      "Epoch 312, Loss: 2.0766\n",
-      "Epoch 313, Loss: 2.1061\n",
-      "Epoch 314, Loss: 2.1038\n",
-      "Epoch 315, Loss: 2.1097\n",
-      "Epoch 316, Loss: 2.0944\n",
-      "Epoch 317, Loss: 2.1001\n",
-      "Epoch 318, Loss: 2.0994\n",
-      "Epoch 319, Loss: 2.0951\n",
-      "Epoch 320, Loss: 2.1278\n",
-      "Epoch 321, Loss: 2.1183\n",
-      "Epoch 322, Loss: 2.1236\n",
-      "Epoch 323, Loss: 2.1069\n",
-      "Epoch 324, Loss: 2.1431\n",
-      "Epoch 325, Loss: 2.1437\n",
-      "Epoch 326, Loss: 2.1081\n",
-      "Epoch 327, Loss: 2.1248\n",
-      "Epoch 328, Loss: 2.1266\n",
-      "Epoch 329, Loss: 2.1096\n",
-      "Epoch 330, Loss: 2.0736\n",
-      "Epoch 331, Loss: 2.0968\n",
-      "Epoch 332, Loss: 2.1103\n",
-      "Epoch 333, Loss: 2.1250\n",
-      "Epoch 334, Loss: 2.0644\n",
-      "Epoch 335, Loss: 2.0949\n",
-      "Epoch 336, Loss: 2.1160\n",
-      "Epoch 337, Loss: 2.0806\n",
-      "Epoch 338, Loss: 2.1123\n",
-      "Epoch 339, Loss: 2.1143\n",
-      "Epoch 340, Loss: 2.0953\n",
-      "Epoch 341, Loss: 2.0875\n",
-      "Epoch 342, Loss: 2.1337\n",
-      "Epoch 343, Loss: 2.1420\n",
-      "Epoch 344, Loss: 2.1249\n",
-      "Epoch 345, Loss: 2.1215\n",
-      "Epoch 346, Loss: 2.1090\n",
-      "Epoch 347, Loss: 2.0963\n",
-      "Epoch 348, Loss: 2.0921\n",
-      "Epoch 349, Loss: 2.0933\n",
-      "Epoch 350, Loss: 2.0794\n",
-      "Epoch 351, Loss: 2.0959\n",
-      "Epoch 352, Loss: 2.0767\n",
-      "Epoch 353, Loss: 2.0906\n",
-      "Epoch 354, Loss: 2.1021\n",
-      "Epoch 355, Loss: 2.0927\n",
-      "Epoch 356, Loss: 2.1038\n",
-      "Epoch 357, Loss: 2.0741\n",
-      "Epoch 358, Loss: 2.0727\n",
-      "Epoch 359, Loss: 2.0753\n",
-      "Epoch 360, Loss: 2.0548\n",
-      "Epoch 361, Loss: 2.0923\n",
-      "Epoch 362, Loss: 2.0861\n",
-      "Epoch 363, Loss: 2.0771\n",
-      "Epoch 364, Loss: 2.0960\n",
-      "Epoch 365, Loss: 2.0745\n",
-      "Epoch 366, Loss: 2.0788\n",
-      "Epoch 367, Loss: 2.0733\n",
-      "Epoch 368, Loss: 2.0839\n",
-      "Epoch 369, Loss: 2.0971\n",
-      "Epoch 370, Loss: 2.0800\n",
-      "Epoch 371, Loss: 2.1154\n",
-      "Epoch 372, Loss: 2.0617\n",
-      "Epoch 373, Loss: 2.0934\n",
-      "Epoch 374, Loss: 2.0934\n",
-      "Epoch 375, Loss: 2.1069\n",
-      "Epoch 376, Loss: 2.0890\n",
-      "Epoch 377, Loss: 2.0881\n",
-      "Epoch 378, Loss: 2.1018\n",
-      "Epoch 379, Loss: 2.0697\n",
-      "Epoch 380, Loss: 2.0837\n",
-      "Epoch 381, Loss: 2.0858\n",
-      "Epoch 382, Loss: 2.0811\n",
-      "Epoch 383, Loss: 2.0630\n",
-      "Epoch 384, Loss: 2.0845\n",
-      "Epoch 385, Loss: 2.0732\n",
-      "Epoch 386, Loss: 2.0704\n",
-      "Epoch 387, Loss: 2.0790\n",
-      "Epoch 388, Loss: 2.0865\n",
-      "Epoch 389, Loss: 2.1035\n",
-      "Epoch 390, Loss: 2.0938\n",
-      "Epoch 391, Loss: 2.1012\n",
-      "Epoch 392, Loss: 2.0946\n",
-      "Epoch 393, Loss: 2.0570\n",
-      "Epoch 394, Loss: 2.0578\n",
-      "Epoch 395, Loss: 2.0493\n",
-      "Epoch 396, Loss: 2.0494\n",
-      "Epoch 397, Loss: 2.0473\n",
-      "Epoch 398, Loss: 2.0564\n",
-      "Epoch 399, Loss: 2.0497\n",
-      "Epoch 400, Loss: 2.0462\n",
-      "Epoch 401, Loss: 2.0484\n",
-      "Epoch 402, Loss: 2.0652\n",
-      "Epoch 403, Loss: 2.0719\n",
-      "Epoch 404, Loss: 2.1264\n",
-      "Epoch 405, Loss: 2.0922\n",
-      "Epoch 406, Loss: 2.0889\n",
-      "Epoch 407, Loss: 2.0744\n",
-      "Epoch 408, Loss: 2.0803\n",
-      "Epoch 409, Loss: 2.0559\n",
-      "Epoch 410, Loss: 2.0484\n",
-      "Epoch 411, Loss: 2.0358\n",
-      "Epoch 412, Loss: 2.0422\n",
-      "Epoch 413, Loss: 2.0323\n",
-      "Epoch 414, Loss: 2.0358\n",
-      "Epoch 415, Loss: 2.0284\n",
-      "Epoch 416, Loss: 2.0365\n",
-      "Epoch 417, Loss: 2.0580\n",
-      "Epoch 418, Loss: 2.0814\n",
-      "Epoch 419, Loss: 2.0985\n",
-      "Epoch 420, Loss: 2.0845\n",
-      "Epoch 421, Loss: 2.1305\n",
-      "Epoch 422, Loss: 2.1280\n",
-      "Epoch 423, Loss: 2.0703\n",
-      "Epoch 424, Loss: 2.0926\n",
-      "Epoch 425, Loss: 2.0963\n",
-      "Epoch 426, Loss: 2.0651\n",
-      "Epoch 427, Loss: 2.0548\n",
-      "Epoch 428, Loss: 2.0529\n",
-      "Epoch 429, Loss: 2.0274\n",
-      "Epoch 430, Loss: 2.0400\n",
-      "Epoch 431, Loss: 2.0409\n",
-      "Epoch 432, Loss: 2.0379\n",
-      "Epoch 433, Loss: 2.0234\n",
-      "Epoch 434, Loss: 2.0314\n",
-      "Epoch 435, Loss: 1.9965\n",
-      "Epoch 436, Loss: 2.0345\n",
-      "Epoch 437, Loss: 2.0361\n",
-      "Epoch 438, Loss: 2.0215\n",
-      "Epoch 439, Loss: 2.0387\n",
-      "Epoch 440, Loss: 2.0397\n",
-      "Epoch 441, Loss: 2.0126\n",
-      "Epoch 442, Loss: 2.0365\n",
-      "Epoch 443, Loss: 2.0224\n",
-      "Epoch 444, Loss: 2.0329\n",
-      "Epoch 445, Loss: 2.0341\n",
-      "Epoch 446, Loss: 2.0324\n",
-      "Epoch 447, Loss: 2.0453\n",
-      "Epoch 448, Loss: 2.0491\n",
-      "Epoch 449, Loss: 2.0387\n",
-      "Epoch 450, Loss: 2.0504\n",
-      "Epoch 451, Loss: 2.0397\n",
-      "Epoch 452, Loss: 2.0357\n",
-      "Epoch 453, Loss: 2.0398\n",
-      "Epoch 454, Loss: 2.0317\n",
-      "Epoch 455, Loss: 2.0258\n",
-      "Epoch 456, Loss: 2.0260\n",
-      "Epoch 457, Loss: 2.0194\n",
-      "Epoch 458, Loss: 2.0161\n",
-      "Epoch 459, Loss: 2.0133\n",
-      "Epoch 460, Loss: 2.0416\n",
-      "Epoch 461, Loss: 2.0170\n",
-      "Epoch 462, Loss: 2.0286\n",
-      "Epoch 463, Loss: 2.0244\n",
-      "Epoch 464, Loss: 2.0286\n",
-      "Epoch 465, Loss: 1.9974\n",
-      "Epoch 466, Loss: 2.0162\n",
-      "Epoch 467, Loss: 2.0040\n",
-      "Epoch 468, Loss: 2.0190\n",
-      "Epoch 469, Loss: 2.0180\n",
-      "Epoch 470, Loss: 1.9842\n",
-      "Epoch 471, Loss: 2.0325\n",
-      "Epoch 472, Loss: 2.0165\n",
-      "Epoch 473, Loss: 2.0149\n",
-      "Epoch 474, Loss: 2.0333\n",
-      "Epoch 475, Loss: 2.0147\n",
-      "Epoch 476, Loss: 2.0180\n",
-      "Epoch 477, Loss: 2.0313\n",
-      "Epoch 478, Loss: 2.0278\n",
-      "Epoch 479, Loss: 2.0228\n",
-      "Epoch 480, Loss: 2.0036\n",
-      "Epoch 481, Loss: 2.0114\n",
-      "Epoch 482, Loss: 2.0111\n",
-      "Epoch 483, Loss: 2.0239\n",
-      "Epoch 484, Loss: 2.0085\n",
-      "Epoch 485, Loss: 2.0084\n",
-      "Epoch 486, Loss: 2.0402\n",
-      "Epoch 487, Loss: 2.0372\n",
-      "Epoch 488, Loss: 2.0807\n",
-      "Epoch 489, Loss: 2.0684\n",
-      "Epoch 490, Loss: 2.0992\n",
-      "Epoch 491, Loss: 2.0516\n",
-      "Epoch 492, Loss: 2.1279\n",
-      "Epoch 493, Loss: 2.1087\n",
-      "Epoch 494, Loss: 2.0793\n",
-      "Epoch 495, Loss: 2.0580\n",
-      "Epoch 496, Loss: 2.0744\n",
-      "Epoch 497, Loss: 2.0852\n",
-      "Epoch 498, Loss: 2.0631\n",
-      "Epoch 499, Loss: 2.0341\n",
-      "Epoch 500, Loss: 2.0277\n"
+      "Epoch 1, Loss: 123961.2219\n"
+     ]
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "c:\\Users\\rrayy\\anaconda3\\envs\\diva\\Lib\\site-packages\\numba\\cuda\\dispatcher.py:536: NumbaPerformanceWarning: \u001b[1mGrid size 2 will likely result in GPU under-utilization due to low occupancy.\u001b[0m\n",
+      "  warn(NumbaPerformanceWarning(msg))\n"
      ]
     }
    ],
    "source": [
     "# 학습 루프\n",
     "\n",
-    "EPOCH = 500\n",
+    "EPOCH = 1\n",
     "\n",
     "for i in range(EPOCH):\n",
     "    total_loss = 0\n",
@@ -648,17 +168,24 @@
     "        lengths_batch = lengths_batch.to(device)\n",
     "\n",
     "        optimizer.zero_grad()\n",
-    "        outputs = model.forward(X_batch, lengths_batch, total_length=Y_batch.size(1))\n",
+    "        outputs = model(X_batch, lengths_batch, Y_batch)\n",
+    "\n",
+    "        min_len = min(outputs.size(1), Y_batch.size(1))\n",
+    "\n",
+    "        loss_HL = huber(outputs[:, :min_len, :], Y_batch[:, :min_len, :]) # 슬라이싱을 이용해 output과 target(Y) 길이가 달라도 loss 측정 가능\n",
+    "        loss_HL = loss_HL.mean(dim=2)  # (B, T), 7차원 평균\n",
     "\n",
-    "        # Loss 계산: (B*T, vocab) vs (B*T)\n",
-    "        outputs = outputs.view(-1, outputs.size(-1))\n",
-    "        targets = Y_batch.view(-1)\n",
+    "        max_len = Y_batch.size(1)\n",
+    "        mask = torch.arange(max_len, device=device).unsqueeze(0) < lengths_batch.unsqueeze(1)  # (B, T)\n",
+    "        loss_HL = (loss_HL * mask[:, :min_len]).sum() / mask[:, :min_len].sum()  # huber만 padding 제외 (sdtw랑 shape가 달라서)\n",
     "\n",
-    "        loss_f = criterion(outputs, targets)\n",
-    "        loss_f.backward()\n",
+    "        loss_sdtw = sdtw(outputs, Y_batch).mean()  # 스칼라\n",
+    "        loss = 0.7*loss_HL+0.3*loss_sdtw # 가중합(다른 loss function 동시에 사용 가능)\n",
+    "\n",
+    "        loss.backward()\n",
     "        optimizer.step()\n",
     "\n",
-    "        total_loss += loss_f.item()\n",
+    "        total_loss += loss.item()\n",
     "\n",
     "    print(f\"Epoch {i+1}, Loss: {total_loss/len(dataloader):.4f}\")"
    ]
@@ -691,10 +218,22 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 5,
+   "execution_count": 2,
    "id": "da89b45a",
    "metadata": {},
-   "outputs": [],
+   "outputs": [
+    {
+     "ename": "NameError",
+     "evalue": "name 'model' is not defined",
+     "output_type": "error",
+     "traceback": [
+      "\u001b[31m---------------------------------------------------------------------------\u001b[39m",
+      "\u001b[31mNameError\u001b[39m                                 Traceback (most recent call last)",
+      "\u001b[36mCell\u001b[39m\u001b[36m \u001b[39m\u001b[32mIn[2]\u001b[39m\u001b[32m, line 3\u001b[39m\n\u001b[32m      1\u001b[39m \u001b[38;5;28;01mimport\u001b[39;00m\u001b[38;5;250m \u001b[39m\u001b[34;01mtorch\u001b[39;00m\n\u001b[32m----> \u001b[39m\u001b[32m3\u001b[39m torch.save(\u001b[43mmodel\u001b[49m.state_dict(), \u001b[33m'\u001b[39m\u001b[33mDIVA_Model_dict.pt\u001b[39m\u001b[33m'\u001b[39m) \u001b[38;5;66;03m# 모델 가중치, 매개변수 저장\u001b[39;00m\n\u001b[32m      4\u001b[39m torch.save(model, \u001b[33m'\u001b[39m\u001b[33mDIVA_Model_full.pt\u001b[39m\u001b[33m'\u001b[39m) \u001b[38;5;66;03m# 모델 전체 저장\u001b[39;00m\n",
+      "\u001b[31mNameError\u001b[39m: name 'model' is not defined"
+     ]
+    }
+   ],
    "source": [
     "import torch\n",
     "\n",
@@ -704,536 +243,87 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 14,
+   "execution_count": 11,
    "id": "75530554",
    "metadata": {},
    "outputs": [
     {
      "data": {
       "text/plain": [
-       "[296,\n",
-       " 12,\n",
-       " 73,\n",
-       " 4,\n",
-       " 12,\n",
-       " 84,\n",
-       " 22,\n",
-       " 76,\n",
-       " 4,\n",
-       " 12,\n",
-       " 22,\n",
-       " 4,\n",
-       " 12,\n",
-       " 12,\n",
-       " 2,\n",
-       " 4,\n",
-       " 12,\n",
-       " 48,\n",
-       " 2,\n",
-       " 4,\n",
-       " 12,\n",
-       " 4,\n",
-       " 4,\n",
-       " 3,\n",
-       " 3,\n",
-       " 4,\n",
-       " 73,\n",
-       " 22,\n",
-       " 3,\n",
-       " 3,\n",
-       " 12,\n",
-       " 83,\n",
-       " 12,\n",
-       " 12,\n",
-       " 4,\n",
-       " 83,\n",
-       " 93,\n",
-       " 4,\n",
-       " 22,\n",
-       " 4,\n",
-       " 10,\n",
-       " 86,\n",
-       " 12,\n",
-       " 3,\n",
-       " 86,\n",
-       " 63,\n",
-       " 12,\n",
-       " 12,\n",
-       " 3,\n",
-       " 3,\n",
-       " 1,\n",
-       " 50,\n",
-       " 12,\n",
-       " 1,\n",
-       " 4,\n",
-       " 12,\n",
-       " 84,\n",
-       " 12,\n",
-       " 3,\n",
-       " 12,\n",
-       " 78,\n",
-       " 4,\n",
-       " 12,\n",
-       " 1,\n",
-       " 53,\n",
-       " 4,\n",
-       " 67,\n",
-       " 4,\n",
-       " 4,\n",
-       " 12,\n",
-       " 4,\n",
-       " 4,\n",
-       " 12,\n",
-       " 12,\n",
-       " 1,\n",
-       " 67,\n",
-       " 3,\n",
-       " 22,\n",
-       " 79,\n",
-       " 4,\n",
-       " 4,\n",
-       " 1,\n",
-       " 3,\n",
-       " 8,\n",
-       " 4,\n",
-       " 57,\n",
-       " 86,\n",
-       " 45,\n",
-       " 53,\n",
-       " 3,\n",
-       " 4,\n",
-       " 12,\n",
-       " 73,\n",
-       " 3,\n",
-       " 302,\n",
-       " 12,\n",
-       " 22,\n",
-       " 62,\n",
-       " 3,\n",
-       " 71,\n",
-       " 1,\n",
-       " 51,\n",
-       " 4,\n",
-       " 12,\n",
-       " 71,\n",
-       " 45,\n",
-       " 1,\n",
-       " 1,\n",
-       " 80,\n",
-       " 1,\n",
-       " 1,\n",
-       " 4,\n",
-       " 302,\n",
-       " 64,\n",
-       " 8,\n",
-       " 1,\n",
-       " 12,\n",
-       " 3,\n",
-       " 4,\n",
-       " 12,\n",
-       " 12,\n",
-       " 1,\n",
-       " 1,\n",
-       " 22,\n",
-       " 3,\n",
-       " 5,\n",
-       " 76,\n",
-       " 66,\n",
-       " 1,\n",
-       " 22,\n",
-       " 56,\n",
-       " 4,\n",
-       " 22,\n",
-       " 4,\n",
-       " 77,\n",
-       " 12,\n",
-       " 22,\n",
-       " 52,\n",
-       " 12,\n",
-       " 3,\n",
-       " 12,\n",
-       " 80,\n",
-       " 4,\n",
-       " 12,\n",
-       " 22,\n",
-       " 12,\n",
-       " 50,\n",
-       " 4,\n",
-       " 86,\n",
-       " 4,\n",
-       " 22,\n",
-       " 5,\n",
-       " 4,\n",
-       " 43,\n",
-       " 4,\n",
-       " 3,\n",
-       " 4,\n",
-       " 64,\n",
-       " 3,\n",
-       " 12,\n",
-       " 5,\n",
-       " 12,\n",
-       " 85,\n",
-       " 4,\n",
-       " 12,\n",
-       " 22,\n",
-       " 6,\n",
-       " 6,\n",
-       " 3,\n",
-       " 53,\n",
-       " 1,\n",
-       " 12,\n",
-       " 12,\n",
-       " 12,\n",
-       " 68,\n",
-       " 4,\n",
-       " 63,\n",
-       " 3,\n",
-       " 86,\n",
-       " 3,\n",
-       " 12,\n",
-       " 22,\n",
-       " 22,\n",
-       " 130,\n",
-       " 90,\n",
-       " 69,\n",
-       " 4,\n",
-       " 4,\n",
-       " 1,\n",
-       " 4,\n",
-       " 3,\n",
-       " 12,\n",
-       " 1,\n",
-       " 3,\n",
-       " 1,\n",
-       " 1,\n",
-       " 4,\n",
-       " 1,\n",
-       " 3,\n",
-       " 5,\n",
-       " 49,\n",
-       " 65,\n",
-       " 4,\n",
-       " 1,\n",
-       " 6,\n",
-       " 202,\n",
-       " 1,\n",
-       " 81,\n",
-       " 67,\n",
-       " 52,\n",
-       " 12,\n",
-       " 7,\n",
-       " 12,\n",
-       " 3,\n",
-       " 3,\n",
-       " 5,\n",
-       " 4,\n",
-       " 12,\n",
-       " 4,\n",
-       " 3,\n",
-       " 12,\n",
-       " 12,\n",
-       " 4,\n",
-       " 5,\n",
-       " 77,\n",
-       " 4,\n",
-       " 3,\n",
-       " 3,\n",
-       " 84,\n",
-       " 78,\n",
-       " 63,\n",
-       " 83,\n",
-       " 4,\n",
-       " 46,\n",
-       " 22,\n",
-       " 22,\n",
-       " 7,\n",
-       " 12,\n",
-       " 177,\n",
-       " 62,\n",
-       " 4,\n",
-       " 70,\n",
-       " 66,\n",
-       " 7,\n",
-       " 1,\n",
-       " 79,\n",
-       " 82,\n",
-       " 4,\n",
-       " 6,\n",
-       " 22,\n",
-       " 12,\n",
-       " 4,\n",
-       " 53,\n",
-       " 4,\n",
-       " 4,\n",
-       " 12,\n",
-       " 1,\n",
-       " 22,\n",
-       " 12,\n",
-       " 3,\n",
-       " 48,\n",
-       " 12,\n",
-       " 4,\n",
-       " 12,\n",
-       " 4,\n",
-       " 53,\n",
-       " 12,\n",
-       " 12,\n",
-       " 3,\n",
-       " 4,\n",
-       " 1,\n",
-       " 12,\n",
-       " 3,\n",
-       " 22,\n",
-       " 12,\n",
-       " 12,\n",
-       " 76,\n",
-       " 12,\n",
-       " 78,\n",
-       " 22,\n",
-       " 22,\n",
-       " 4,\n",
-       " 78,\n",
-       " 4,\n",
-       " 3,\n",
-       " 1,\n",
-       " 4,\n",
-       " 6,\n",
-       " 5,\n",
-       " 64,\n",
-       " 4,\n",
-       " 4,\n",
-       " 47,\n",
-       " 22,\n",
-       " 22,\n",
-       " 1,\n",
-       " 12,\n",
-       " 3,\n",
-       " 3,\n",
-       " 68,\n",
-       " 4,\n",
-       " 1,\n",
-       " 22,\n",
-       " 12,\n",
-       " 22,\n",
-       " 3,\n",
-       " 12,\n",
-       " 12,\n",
-       " 4,\n",
-       " 1,\n",
-       " 3,\n",
-       " 3,\n",
-       " 1,\n",
-       " 7,\n",
-       " 4,\n",
-       " 3,\n",
-       " 12,\n",
-       " 81,\n",
-       " 3,\n",
-       " 49,\n",
-       " 4,\n",
-       " 12,\n",
-       " 1,\n",
-       " 88,\n",
-       " 4,\n",
-       " 4,\n",
-       " 66,\n",
-       " 22,\n",
-       " 1,\n",
-       " 12,\n",
-       " 45,\n",
-       " 78,\n",
-       " 78,\n",
-       " 22,\n",
-       " 12,\n",
-       " 6,\n",
-       " 12,\n",
-       " 52,\n",
-       " 47,\n",
-       " 4,\n",
-       " 12,\n",
-       " 76,\n",
-       " 5,\n",
-       " 12,\n",
-       " 64,\n",
-       " 52,\n",
-       " 12,\n",
-       " 4,\n",
-       " 22,\n",
-       " 4,\n",
-       " 4,\n",
-       " 202,\n",
-       " 1,\n",
-       " 22,\n",
-       " 22,\n",
-       " 73,\n",
-       " 65,\n",
-       " 4,\n",
-       " 1,\n",
-       " 1,\n",
-       " 3,\n",
-       " 22,\n",
-       " 6,\n",
-       " 3,\n",
-       " 12,\n",
-       " 12,\n",
-       " 69,\n",
-       " 58,\n",
-       " 84,\n",
-       " 5,\n",
-       " 4,\n",
-       " 12,\n",
-       " 1,\n",
-       " 12,\n",
-       " 22,\n",
-       " 12,\n",
-       " 51,\n",
-       " 1,\n",
-       " 1,\n",
-       " 22,\n",
-       " 1,\n",
-       " 12,\n",
-       " 4,\n",
-       " 4,\n",
-       " 4,\n",
-       " 4,\n",
-       " 3,\n",
-       " 3,\n",
-       " 7,\n",
-       " 4,\n",
-       " 84,\n",
-       " 22,\n",
-       " 12,\n",
-       " 4,\n",
-       " 3,\n",
-       " 66,\n",
-       " 51,\n",
-       " 22,\n",
-       " 49,\n",
-       " 4,\n",
-       " 4,\n",
-       " 64,\n",
-       " 1,\n",
-       " 12,\n",
-       " 56,\n",
-       " 12,\n",
-       " 54,\n",
-       " 3,\n",
-       " 77,\n",
-       " 4,\n",
-       " 4,\n",
-       " 71,\n",
-       " 4,\n",
-       " 12,\n",
-       " 3,\n",
-       " 22,\n",
-       " 76,\n",
-       " 45,\n",
-       " 12,\n",
-       " 4,\n",
-       " 82,\n",
-       " 4,\n",
-       " 22,\n",
-       " 1,\n",
-       " 12,\n",
-       " 49,\n",
-       " 4,\n",
-       " 12,\n",
-       " 1,\n",
-       " 12,\n",
-       " 22,\n",
-       " 4,\n",
-       " 22,\n",
-       " 12,\n",
-       " 45,\n",
-       " 73,\n",
-       " 12,\n",
-       " 22,\n",
-       " 12,\n",
-       " 4,\n",
-       " 4,\n",
-       " 12,\n",
-       " 72,\n",
-       " 4,\n",
-       " 3,\n",
-       " 1,\n",
-       " 6,\n",
-       " 1,\n",
-       " 50,\n",
-       " 3,\n",
-       " 1,\n",
-       " 4,\n",
-       " 12,\n",
-       " 22,\n",
-       " 47,\n",
-       " 4,\n",
-       " 1,\n",
-       " 1,\n",
-       " 3,\n",
-       " 50,\n",
-       " 80,\n",
-       " 4,\n",
-       " 4,\n",
-       " 1,\n",
-       " 4,\n",
-       " 49,\n",
-       " 4,\n",
-       " 4,\n",
-       " 71,\n",
-       " 77,\n",
-       " 3,\n",
-       " 3,\n",
-       " 22,\n",
-       " 1,\n",
-       " 12,\n",
-       " 78,\n",
-       " 4,\n",
-       " 4,\n",
-       " 66,\n",
-       " 22,\n",
-       " 22,\n",
-       " 4,\n",
-       " 3,\n",
-       " 3,\n",
-       " 12,\n",
-       " 73,\n",
-       " 1,\n",
-       " 3,\n",
-       " 12,\n",
-       " 22,\n",
-       " 4,\n",
-       " 3,\n",
-       " 12,\n",
-       " 5,\n",
-       " 4,\n",
-       " 12,\n",
-       " 3,\n",
-       " 22,\n",
-       " 12,\n",
-       " 12,\n",
-       " 12,\n",
-       " 12,\n",
-       " 3,\n",
-       " 12,\n",
-       " 7,\n",
-       " 11,\n",
-       " 12,\n",
-       " 4,\n",
-       " 22,\n",
-       " 66,\n",
-       " 12,\n",
-       " 12]"
+       "<All keys matched successfully>"
       ]
      },
-     "execution_count": 14,
+     "execution_count": 11,
      "metadata": {},
      "output_type": "execute_result"
     }
    ],
    "source": [
-    "model.load_state_dict(torch.load('DIVA_Model_dict.pt')) # 모델 가중치, 매개변수 불러오기\n",
+    "model.load_state_dict(torch.load('DIVA_Model_dict.pt')) # 모델 가중치, 매개변수 불러오기"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 14,
+   "id": "6c7f2aa0",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "[128, 100, 10, -1, 2, 10, 3, -1, 10, 10, 81, 10, -1, 10, 84, 1, 81, 10, -1, 79, 10, 10, 10, -1, 1, 2, 10, 1, 2, 1, 1, 79, 10, 10, 1, 10, 84, 2, 1, 86, 1, 84, 1, 84, 83, 10, 10, 1, 84, 83, -1, 10, 10, 10, -1, 10, -1, 2, 10, 10, -1, 10, 10, 81, 1, 83, 1, 10, 1, 10, 1, 81, 1, 10, 2, 10, 10, 10, 84, 10, -1, 1, 84, 10, -1, 10, 10, 1, 10, 10, 84, 1, 10, -1, 1, 2, 2, 10, 2, 83, 3, 10, 84, 10, 10, -1, 84, 83, 81, 2, 2, 10, 10, 10, 10, -1, 10, 81, 79, 2, 2, 1, 84, 10, 10, 10, 1, 1, 10, 10, 10, 1, 3, 81, 10, 10, 1, 2, 2, 10, 84, 2, 79, 1, 10, 91, -1, 86, 81, 84, -1, 84, 10, -1, 10, 2, 10, 10, 10, 84, 83, 79, 10, -1, -1, 1, 3, 10, 84, 1, 84, 2, 10, 1, -1, 10, 10, 81, 84, 10, 83, 2, 84, 1, 10, 2, 10, 3, -1, 10, 2, 1, 84, 1, 84, 84, 10, 10, 86, 10, 84, 83, 2, 3, -1, 10, 3, 91, 84, 10, 84, 2, 10, 10, 10, 83, 84, 84, 2, 10, 10, 10, 10, 91, 84, 10, 2, 2, 2, 1, 2, 60, 2, 65, 20, 1, 200, 1, 1, 2, 2, 2, 2, 3, 20, 1, 1, 20, 1, 3, 1, 1, 1, 20, -1, 2, 1, 2, 1, 20, 60, 1, 1, 20, 20, 1, 20, 8, 1, 8, 20, 8, 20, 1, 1, 20, -1, 20, 1, -1, 60, 1, 8, 60, 2, 2, -1, 8, 20, -1, 60, -1, -1, 60, 1, 8, 20, 60, 20, 60, 20, 65, 8, 20, 5, -1, 60, 20, 67, 60, 69, -1, 8, 1, -1, 69, 65, 69, 20, 20, 69, 2, 5, -1, 20, 5, 8, 2, 69, 2, 1, 5, -1, 67, 20, -1, 5, 69, 20, 2, 20, 2, 2, 1, 69, -1, 2, 2, 1, 8, 8, 67, 5, 2, 20, 65, 1, 3, 20, 1, 2, 20, 67, 1, -1, -1, 20, 2, 65, -1, 67, 300, 20, 1, 2, 1, -1, 1, 20, 1, 20, 67, 60, 20, 65, 20, -1, 20, -1, 67, -1, 20, 1, 5, 2, 20, 20, -1, 20, -1, 69, 20, 20, 20, -1, -1, -1, 2, -1, 1, 1, -1, -1, 2, 2, 1, 2, 2, 20, 1, 1, -1, 20, 2, 3, 67, 2, 20, 60, -1, 2, 1, 60, -1, 1, 20, 20, 20, 20, 1, 69, -1, 1, 2, -1, 60, 20, 2, 60, 65, -1, 8, -1, 20, -1, 20, 4, 2, 20, -1, 20, 20, 8, 65, 5, 1, 8, -1, 69, 1, 20, 1, 69, -1, 1, 67, 69, 20, 8, -1, 2, 1, 2, -1, 1, -1, 2, -1, 1, 67, 65, 20, 1, 67, 20, 65, 20, 5, 2, 2, 1, 2, -1, -1, -1, 67, -1, 2, 65, -1, 1, 67]\n"
+     ]
+    },
+    {
+     "ename": "ValueError",
+     "evalue": "invalid literal for int() with base 10: ''",
+     "output_type": "error",
+     "traceback": [
+      "\u001b[31m---------------------------------------------------------------------------\u001b[39m",
+      "\u001b[31mValueError\u001b[39m                                Traceback (most recent call last)",
+      "\u001b[36mCell\u001b[39m\u001b[36m \u001b[39m\u001b[32mIn[14]\u001b[39m\u001b[32m, line 8\u001b[39m\n\u001b[32m      5\u001b[39m \u001b[38;5;28mprint\u001b[39m(token)\n\u001b[32m      7\u001b[39m MIDI = Tokenizer()\n\u001b[32m----> \u001b[39m\u001b[32m8\u001b[39m \u001b[43mMIDI\u001b[49m\u001b[43m.\u001b[49m\u001b[43mset_id\u001b[49m\u001b[43m(\u001b[49m\u001b[43mtoken\u001b[49m\u001b[43m)\u001b[49m\n\u001b[32m     10\u001b[39m midi= MIDI.to_midi() \u001b[38;5;66;03m# This should generate MIDI from the stored melody and chords\u001b[39;00m\n\u001b[32m     11\u001b[39m midi.write(\u001b[33m'\u001b[39m\u001b[33mmidi\u001b[39m\u001b[33m'\u001b[39m, fp=\u001b[33m'\u001b[39m\u001b[33mtest_output.mid\u001b[39m\u001b[33m'\u001b[39m)  \u001b[38;5;66;03m# Save the generated MIDI to a file\u001b[39;00m\n",
+      "\u001b[36mFile \u001b[39m\u001b[32mc:\\Users\\rrayy\\anaconda3\\envs\\diva\\Lib\\site-packages\\HarmonyMIDIToken\\tokenizer.py:189\u001b[39m, in \u001b[36mHarmonyMIDIToken.set_id\u001b[39m\u001b[34m(self, token_id)\u001b[39m\n\u001b[32m    186\u001b[39m bass_tokens = token_id[token_id.index(\u001b[32m300\u001b[39m)+\u001b[32m1\u001b[39m:]\n\u001b[32m    188\u001b[39m \u001b[38;5;28mself\u001b[39m.melody = \u001b[38;5;28mself\u001b[39m._detokenize_note(melody_tokens)\n\u001b[32m--> \u001b[39m\u001b[32m189\u001b[39m \u001b[38;5;28mself\u001b[39m.chords = \u001b[38;5;28;43mself\u001b[39;49m\u001b[43m.\u001b[49m\u001b[43m_detokenize_chord\u001b[49m\u001b[43m(\u001b[49m\u001b[43mchords_tokens\u001b[49m\u001b[43m)\u001b[49m\n\u001b[32m    190\u001b[39m \u001b[38;5;28mself\u001b[39m.bass = \u001b[38;5;28mself\u001b[39m._detokenize_note(bass_tokens)\n",
+      "\u001b[36mFile \u001b[39m\u001b[32mc:\\Users\\rrayy\\anaconda3\\envs\\diva\\Lib\\site-packages\\HarmonyMIDIToken\\tokenizer.py:166\u001b[39m, in \u001b[36mHarmonyMIDIToken._detokenize_chord\u001b[39m\u001b[34m(self, token)\u001b[39m\n\u001b[32m    164\u001b[39m         output.append({\u001b[33m\"\u001b[39m\u001b[33mchord\u001b[39m\u001b[33m\"\u001b[39m: \u001b[33m\"\u001b[39m\u001b[33m\"\u001b[39m, \u001b[33m\"\u001b[39m\u001b[33mduration\u001b[39m\u001b[33m\"\u001b[39m: \u001b[38;5;28mfloat\u001b[39m(chord_list[-\u001b[32m2\u001b[39m])/\u001b[32m4\u001b[39m})\n\u001b[32m    165\u001b[39m     \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[32m--> \u001b[39m\u001b[32m166\u001b[39m         output.append({\u001b[33m\"\u001b[39m\u001b[33mchord\u001b[39m\u001b[33m\"\u001b[39m:\u001b[38;5;28mself\u001b[39m._intpitch_to_note_name(\u001b[38;5;28;43mint\u001b[39;49m\u001b[43m(\u001b[49m\u001b[43mchord_list\u001b[49m\u001b[43m[\u001b[49m\u001b[32;43m1\u001b[39;49m\u001b[43m]\u001b[49m\u001b[43m)\u001b[49m)[:-\u001b[32m1\u001b[39m]+inverse_quality_map[\u001b[38;5;28mint\u001b[39m(chord_list[\u001b[32m2\u001b[39m])], \u001b[33m\"\u001b[39m\u001b[33mduration\u001b[39m\u001b[33m\"\u001b[39m: \u001b[38;5;28mfloat\u001b[39m(chord_list[-\u001b[32m2\u001b[39m])/\u001b[32m4\u001b[39m})\n\u001b[32m    168\u001b[39m \u001b[38;5;28;01mreturn\u001b[39;00m output\n",
+      "\u001b[31mValueError\u001b[39m: invalid literal for int() with base 10: ''"
+     ]
+    }
+   ],
+   "source": [
+    "from HarmonyMIDIToken import HarmonyMIDIToken as Tokenizer\n",
+    "\n",
+    "Y = model.generate(X_tensor[0], device=device) # 스타일 벡터 하나로 시퀀스 생성\n",
+    "token = [i-2 for i in Y]\n",
+    "print(token)\n",
+    "\n",
+    "MIDI = Tokenizer()\n",
+    "MIDI.set_id(token)\n",
+    "\n",
+    "midi= MIDI.to_midi() # This should generate MIDI from the stored melody and chords\n",
+    "midi.write('midi', fp='test_output.mid')  # Save the generated MIDI to a file"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "id": "b2a75eeb",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "X shape: torch.Size([10, 5])\n",
+      "Y shape: torch.Size([10, 9])\n"
+     ]
+    }
+   ],
+   "source": [
+    "X = torch.rand((10, 5, 7), device=device, requires_grad=True)\n",
+    "Y = torch.rand((10, 9, 7), device=device)\n",
     "\n",
-    "model.generate(X_tensor[0], device=device) # 스타일 벡터 하나로 시퀀스 생성"
+    "print(\"X shape:\", X.shape[:2])\n",
+    "print(\"Y shape:\", Y.shape[:2])"
    ]
   }
  ],