IDS training_d3-originalnoise-final.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "540b46d5",
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.nn.functional as F\n",
    "import random\n",
    "from torch.utils.data import Dataset, DataLoader\n",
    "import os\n",
    "os.environ[\"CUDA_VISIBLE_DEVICES\"] = \"2\"  # Select the GPU index\n",
    "import nltk\n",
    "from torch.optim.lr_scheduler import _LRScheduler\n",
    "import time\n",
    "import math\n",
    "import matplotlib.pyplot as plt\n",
    "import spacy\n",
    "from scipy.io import savemat\n",
    "import dill as pickle\n",
    "import thop"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "a7459a68",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "['ACGTAGCAGA', 'ACTACAGACG', 'AGACGATGCA', 'AGCGACTATC', 'ATAGCTCGTG', 'CACGAGCTAT', 'CGTCTGTAGT', 'CTATCAGCGA', 'GATAGTCGCT', 'GCAGACATCA', 'GTGCTCGATA', 'TATCGAGCAC', 'TCGCTGATAG', 'TCTGCTACGT', 'TGATGTCTGC', 'TGCATCGTCT']\n"
     ]
    }
   ],
   "source": [
    "# Dictionary\n",
    "dicno=3\n",
    "dictionaries =[[],\n",
    "               \n",
    "               [],\n",
    "               \n",
    "               [],\n",
    "               \n",
    "               ['ACGTAGCAGA', 'ACTACAGACG', 'AGACGATGCA', 'AGCGACTATC', 'ATAGCTCGTG', 'CACGAGCTAT','CGTCTGTAGT',\n",
    "                'CTATCAGCGA', 'GATAGTCGCT', 'GCAGACATCA', 'GTGCTCGATA', 'TATCGAGCAC', 'TCGCTGATAG', 'TCTGCTACGT',\n",
    "                'TGATGTCTGC', 'TGCATCGTCT'],\n",
    "               \n",
    "               ['ATCGATCGATCGAT','TACGTAGCTAGCTA', 'TAGCTACGTAGCTA', 'ATGCATGCATCGAT', 'ATGCTACGATGCTA', \n",
    "                'TAGCATGCTACGAT', 'TACGATCGTACGAT', 'ATCGTAGCATGCTA', 'TACGATGCTAGCAT', 'ATCGTACGATCGTA', \n",
    "                'ATGCTAGCATCGTA', 'TAGCATCGTAGCAT', 'TAGCTAGCTACGTA', 'ATGCATCGATGCAT', 'ATCGATGCATGCAT', \n",
    "                'TACGTACGTACGTA'],\n",
    "              \n",
    "               []\n",
    "              \n",
    "               ]\n",
    "dicts=dictionaries[dicno]\n",
    "print(dicts)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "8751ef87",
   "metadata": {},
   "outputs": [],
   "source": [
    "batch_size=2000\n",
    "device = 'cuda' if torch.cuda.is_available() else 'cpu'\n",
    "lr=0.001\n",
    "num_epochs=500\n",
    "final_num_epochs=200\n",
    "patience=400\n",
    "train_num_samples=30000\n",
    "test_num_samples=10000\n",
    "final_train_num_samples=500000\n",
    "final_test_num_samples=10000\n",
    "num_errors=4\n",
    "K=len(dicts[0])\n",
    "hidden_size=32"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "c8e3d5d0",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'finalmodels/BiLSTM/Dictionary_3/batch_size_2000/num_errors_4/initial_lr_0.001/epochs200_p400_tr500000ts10000K10hs32/'"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "path_directory=f\"finalmodels/BiLSTM/Dictionary_{dicno}/batch_size_{batch_size}/num_errors_{num_errors}/initial_lr_{lr}/epochs{final_num_epochs}_p{patience}_tr{final_train_num_samples}ts{final_test_num_samples}K{K}hs{hidden_size}/\"\n",
    "path_directory"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "9689e71a",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Check if \"models\" folder exists, create it if it doesn't\n",
    "if not os.path.exists(path_directory):\n",
    "    os.makedirs(path_directory)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "e0ed2ba0",
   "metadata": {},
   "outputs": [],
   "source": [
    "class BiLSTMDeletionModel(nn.Module):\n",
    "\n",
    "    def __init__(self, hidden_size, dropout):\n",
    "        super().__init__()\n",
    "        \n",
    "        self.lstm1 = nn.LSTM(\n",
    "            input_size=5, \n",
    "            hidden_size=hidden_size,\n",
    "            num_layers=2,\n",
    "            batch_first=True,\n",
    "            bidirectional=True,\n",
    "            dropout=dropout\n",
    "        )\n",
    "        self.fc = nn.Linear(2*hidden_size, 4)\n",
    "        self.softmax = nn.Softmax(dim=2)\n",
    "    def forward(self, x):\n",
    "\n",
    "        # BiLSTM 1\n",
    "        out, _ = self.lstm1(x)\n",
    "        out = self.softmax(self.fc(out))\n",
    "        return out"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "fe39060e",
   "metadata": {},
   "outputs": [],
   "source": [
    "deletion_model = BiLSTMDeletionModel(hidden_size,0.1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "687c5cba",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'deletionmodels/BiLSTM/Dictionary_3/batch_size_2000/num_errors_4/initial_lr_0.001/epochs500_p400_tr30000ts10000K10hs32/'"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "deletion_path_directory=f\"deletionmodels/BiLSTM/Dictionary_{dicno}/batch_size_{batch_size}/num_errors_{num_errors}/initial_lr_{lr}/epochs{num_epochs}_p{patience}_tr{train_num_samples}ts{test_num_samples}K{K}hs{hidden_size}/\"\n",
    "deletion_path_directory"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "7c4b4c52",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "_IncompatibleKeys(missing_keys=[], unexpected_keys=['total_ops', 'total_params'])"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "deletion_model.load_state_dict(torch.load(deletion_path_directory+\"model.pth\"), strict=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "be948844",
   "metadata": {},
   "outputs": [],
   "source": [
    "class BiLSTMsubstitutionModel(nn.Module):\n",
    "\n",
    "    def __init__(self, hidden_size, dropout):\n",
    "        super().__init__()\n",
    "        \n",
    "        self.lstm1 = nn.LSTM(\n",
    "            input_size=4, \n",
    "            hidden_size=hidden_size,\n",
    "            num_layers=2,\n",
    "            batch_first=True,\n",
    "            bidirectional=True,\n",
    "            dropout=dropout\n",
    "        )\n",
    "        self.fc = nn.Linear(2*hidden_size, 4)\n",
    "        self.softmax = nn.Softmax(dim=2)\n",
    "    def forward(self, x):\n",
    "\n",
    "        # BiLSTM 1\n",
    "        out, _ = self.lstm1(x)\n",
    "        out = self.softmax(self.fc(out))\n",
    "        return out"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "066e764a",
   "metadata": {},
   "outputs": [],
   "source": [
    "substitution_model = BiLSTMsubstitutionModel(hidden_size,0.1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "70bcdf1b",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'substitutionmodels/BiLSTM/Dictionary_3/batch_size_2000/num_errors_4/initial_lr_0.001/epochs500_p400_tr30000ts10000K10hs32/'"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "substitution_path_directory=f\"substitutionmodels/BiLSTM/Dictionary_{dicno}/batch_size_{batch_size}/num_errors_{num_errors}/initial_lr_{lr}/epochs{num_epochs}_p{patience}_tr{train_num_samples}ts{test_num_samples}K{K}hs{hidden_size}/\"\n",
    "substitution_path_directory"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "fd4d87ce",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "_IncompatibleKeys(missing_keys=[], unexpected_keys=['total_ops', 'total_params'])"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "substitution_model.load_state_dict(torch.load(substitution_path_directory+\"model.pth\"), strict=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "aae1fe9c",
   "metadata": {},
   "outputs": [],
   "source": [
    "class BiLSTMinsertionModel(nn.Module):\n",
    "\n",
    "    def __init__(self, hidden_size, dropout):\n",
    "        super().__init__()\n",
    "        \n",
    "        self.lstm1 = nn.LSTM(\n",
    "            input_size=5, \n",
    "            hidden_size=hidden_size,\n",
    "            num_layers=2,\n",
    "            batch_first=True,\n",
    "            bidirectional=True,\n",
    "            dropout=dropout\n",
    "        )\n",
    "        self.fc = nn.Linear(2*hidden_size, 5)\n",
    "        self.softmax = nn.Softmax(dim=2)\n",
    "    def forward(self, x):\n",
    "\n",
    "        # BiLSTM 1\n",
    "        out, _ = self.lstm1(x)\n",
    "        out = self.softmax(self.fc(out))\n",
    "        return out"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "0a49b7fd",
   "metadata": {},
   "outputs": [],
   "source": [
    "insertion_model = BiLSTMinsertionModel(hidden_size,0.1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "58840053",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'insertionmodels/BiLSTM/Dictionary_3/batch_size_2000/num_errors_4/initial_lr_0.001/epochs500_p400_tr30000ts10000K10hs32/'"
      ]
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "insertion_path_directory=f\"insertionmodels/BiLSTM/Dictionary_{dicno}/batch_size_{batch_size}/num_errors_{num_errors}/initial_lr_{lr}/epochs{num_epochs}_p{patience}_tr{train_num_samples}ts{test_num_samples}K{K}hs{hidden_size}/\"\n",
    "insertion_path_directory"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "a9520aed",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "_IncompatibleKeys(missing_keys=[], unexpected_keys=['total_ops', 'total_params'])"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "insertion_model.load_state_dict(torch.load(insertion_path_directory+\"model.pth\"), strict=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "id": "b7f6d6ab",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "BiLSTMDeletionModel(\n",
       "  (lstm1): LSTM(5, 32, num_layers=2, batch_first=True, dropout=0.1, bidirectional=True)\n",
       "  (fc): Linear(in_features=64, out_features=4, bias=True)\n",
       "  (softmax): Softmax(dim=2)\n",
       ")"
      ]
     },
     "execution_count": 18,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "insertion_model.to(device)\n",
    "substitution_model.to(device)\n",
    "deletion_model.to(device)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "id": "f532a47a",
   "metadata": {},
   "outputs": [],
   "source": [
    "class BiLSTMcombinerModel(nn.Module):\n",
    "\n",
    "    def __init__(self, hidden_size, dropout):\n",
    "        super().__init__()\n",
    "        self.lstm1 = nn.LSTM(\n",
    "            input_size=4, \n",
    "            hidden_size=5*hidden_size,\n",
    "            num_layers=2,\n",
    "            batch_first=True,\n",
    "            bidirectional=True,\n",
    "            dropout=dropout\n",
    "        )\n",
    "        self.fc = nn.Linear(10*hidden_size, 5)\n",
    "        self.softmax = nn.Softmax(dim=2)\n",
    "    def forward(self, x):\n",
    "        # BiLSTM 1\n",
    "        out, _ = self.lstm1(x)\n",
    "        out = self.softmax(self.fc(out))\n",
    "        return out"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "id": "42d5258d",
   "metadata": {},
   "outputs": [],
   "source": [
    "combiner_model=BiLSTMcombinerModel(hidden_size,0.1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "6eaf70ed",
   "metadata": {},
   "outputs": [],
   "source": [
    "class ParallelEnsemble(nn.Module):\n",
    "    def __init__(self, insertion_model, deletion_model, substitution_model, combiner_model):\n",
    "        super().__init__()\n",
    "        self.insert_model = insertion_model \n",
    "        self.delete_model = deletion_model\n",
    "        self.subst_model = substitution_model\n",
    "        self.combiner_model = combiner_model\n",
    "        \n",
    "        # Freeze insertion, deletion, substitution models\n",
    "        for param in self.insert_model.parameters():\n",
    "            param.requires_grad = False\n",
    "\n",
    "        for param in self.delete_model.parameters():\n",
    "            param.requires_grad = False\n",
    "\n",
    "        for param in self.subst_model.parameters():\n",
    "            param.requires_grad = False\n",
    "\n",
    "        # Ensure combiner model is trainable\n",
    "        for param in self.combiner_model.parameters():\n",
    "            param.requires_grad = True\n",
    "\n",
    "    def forward(self, x):\n",
    "        # Get predictions in parallel \n",
    "        insert_in = one_hot_encode(x.to(device), 5)\n",
    "        insert_out = self.insert_model(insert_in)  \n",
    "        delete_in = one_hot_encode(x[:,:K].to(device), 5)\n",
    "        delete_out = self.delete_model(delete_in)\n",
    "        subst_in = one_hot_encode(x[:,:K].to(device), 5)\n",
    "        subst_out = self.subst_model(subst_in[:,:,:4])\n",
    "        # Combine parallel outputs  \n",
    "        combined_out = torch.cat([insert_out[:,:K,:4], delete_out[:,:K,:4], subst_out[:,:K,:4]], dim=1)\n",
    "        # Final correction  \n",
    "        y = self.combiner_model(combined_out)\n",
    "        return y\n",
    "\n",
    "ensemble = ParallelEnsemble(insertion_model, deletion_model, substitution_model, combiner_model).to(device)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "bb5db291",
   "metadata": {},
   "outputs": [],
   "source": [
    "def one_hot_encode(sequences, n_unique):\n",
    "\n",
    "    # Get size of batch and max sequence length\n",
    "    batch_size,seq_len=sequences.shape\n",
    "    #len(sequences)\n",
    "    #seq_len=len(sequences[0])\n",
    "    \n",
    "    # Initialize encoded tensor\n",
    "    oneencoding = torch.zeros(batch_size, seq_len, n_unique, device = device)\n",
    "    # Set 1 at index of each character\n",
    "    for i in range(batch_size):\n",
    "        for j in range(seq_len):\n",
    "            index = sequences[i][j]\n",
    "            oneencoding[i, j, index] = 1          \n",
    "    return oneencoding"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "id": "392d5923",
   "metadata": {},
   "outputs": [],
   "source": [
    "def one_hot_decode(encoded_seq):\n",
    "    \"\"\"Converts a batch of one-hot encoded sequences to indices\"\"\"\n",
    "    \n",
    "    # Dimensions will be (batch_size, seq_len, num_classes)\n",
    "    batch_size, seq_len, num_classes = encoded_seq.shape\n",
    "    \n",
    "    # Flatten to treat as a list of one-hot vectors\n",
    "    encoded_seq = encoded_seq.reshape(-1, num_classes)\n",
    "    \n",
    "    # Argmax over classes to get class index\n",
    "    indices = encoded_seq.argmax(dim=1).to(device)\n",
    "    \n",
    "    # Reshape back to original dimensions\n",
    "    indices = indices.reshape(batch_size, seq_len)\n",
    "    \n",
    "    return indices"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "id": "6aac4a8f",
   "metadata": {},
   "outputs": [],
   "source": [
    "class LanguageDataset1(Dataset):\n",
    "\n",
    "    def __init__(self, dictionary, num_samples, num_errors):\n",
    "        self.dictionary = dictionary\n",
    "        self.num_samples = num_samples \n",
    "        self.num_errors = num_errors\n",
    "\n",
    "    def __len__(self):\n",
    "        return self.num_samples\n",
    "\n",
    "    def __getitem__(self, idx):\n",
    "        source_idx = idx % len(self.dictionary)\n",
    "        sequence = self.dictionary[source_idx]\n",
    "        choice = torch.rand(1).item()\n",
    "        if self.num_errors > 0:\n",
    "            if choice >= 0 and choice< 0.10646:\n",
    "                sequence = self.introduce_substitution_errors(sequence, self.num_errors)\n",
    "                for _ in range(self.num_errors):\n",
    "                    sequence.append('N')  \n",
    "                sequence=''.join(sequence)\n",
    "                #sequence= torch.cat([sequence, 'NNNN'], dim=1)\n",
    "            elif choice >=0.10646 and choice < 0.14218:\n",
    "                sequence = self.introduce_deletion_errors(sequence, self.num_errors)\n",
    "                for _ in range(self.num_errors):\n",
    "                    sequence.append('N') \n",
    "                sequence=''.join(sequence)\n",
    "                #sequence= torch.cat([sequence, 'NNNN'], dim=1)\n",
    "            elif choice >=0.14218 and choice < 0.15163:\n",
    "                sequence = self.introduce_insertion_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "            elif choice >=0.15163 and choice < 0.156:\n",
    "                sequence = self.introduce_substitution_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "                sequence = self.introduce_deletion_errors(sequence, self.num_errors)\n",
    "                for _ in range(self.num_errors):\n",
    "                    sequence.append('N') \n",
    "                sequence=''.join(sequence)\n",
    "                #sequence= torch.cat([sequence, 'NNNN'], dim=1)\n",
    "            elif choice >=0.156 and choice < 0.15701:\n",
    "                sequence = self.introduce_substitution_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "                sequence = self.introduce_insertion_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "                \n",
    "            elif choice >=0.15701 and choice < 0.1574:\n",
    "                sequence = self.introduce_deletion_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "                sequence = self.introduce_insertion_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "            elif choice >=0.1574 and choice < 0.15747:\n",
    "                sequence = self.introduce_substitution_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "                sequence = self.introduce_deletion_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "                sequence = self.introduce_insertion_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)  \n",
    "            else:\n",
    "                sequence = list(sequence)\n",
    "                for _ in range(self.num_errors):\n",
    "                    sequence.append('N') \n",
    "                sequence=''.join(sequence) \n",
    "            \n",
    "        x = torch.tensor([self.char_to_idx(c) for c in sequence], dtype=torch.long)\n",
    "        y = torch.tensor([self.char_to_idx(c) for c in self.dictionary[source_idx]], dtype=torch.long)\n",
    "        return x.to(device), y.to(device)\n",
    "\n",
    "    def introduce_deletion_errors(self, sequence, num_errors):\n",
    "        ne = random.randint(1, num_errors)\n",
    "        sequence = list(sequence)\n",
    "        \n",
    "        for _ in range(ne):\n",
    "            i = random.randint(0, len(sequence)-1)\n",
    "            del sequence[i]\n",
    "            sequence.append('N') \n",
    "        \n",
    "        return sequence\n",
    "    \n",
    "    def introduce_substitution_errors(self, sequence, num_errors):\n",
    "        ne = random.randint(1, num_errors)\n",
    "        sequence = list(sequence)\n",
    "        for _ in range(ne):\n",
    "            i = random.randint(0, len(sequence)-1)\n",
    "            if sequence[i] == 'A':\n",
    "                sequence[i] = random.choice(['T', 'G', 'C'])\n",
    "            elif sequence[i] == 'T':\n",
    "                sequence[i] = random.choice(['A', 'G', 'C'])\n",
    "            elif sequence[i] == 'G':\n",
    "                sequence[i] = random.choice(['A', 'T', 'C'])\n",
    "            elif sequence[i] == 'C':\n",
    "                sequence[i] = random.choice(['A', 'T', 'G'])\n",
    "        \n",
    "        return sequence\n",
    "    \n",
    "    def introduce_insertion_errors(self, sequence, num_errors):\n",
    "        ne = random.randint(1, num_errors)\n",
    "        sequence = list(sequence)\n",
    "        for _ in range(num_errors-ne):\n",
    "            sequence.append('N')\n",
    "        for _ in range(ne):\n",
    "            i = random.randint(0, len(sequence))\n",
    "            random_char = random.choice(['A','C','T','G']) \n",
    "            sequence.insert(i, random_char)\n",
    "        return sequence\n",
    "\n",
    "    def char_to_idx(self, char):\n",
    "        return {'A': 0, 'C': 1, 'T': 2, 'G': 3, 'N': 4}[char]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "id": "8b0352f4",
   "metadata": {},
   "outputs": [],
   "source": [
    "train_data = LanguageDataset1(dicts, num_samples=final_train_num_samples, num_errors=num_errors)\n",
    "test_data = LanguageDataset1(dicts, num_samples=final_test_num_samples, num_errors=num_errors)\n",
    "\n",
    "train_loader = DataLoader(train_data, batch_size=batch_size, shuffle=True)\n",
    "test_loader = DataLoader(test_data, batch_size=batch_size)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "id": "f720cb12",
   "metadata": {},
   "outputs": [],
   "source": [
    "criterion= nn.MSELoss().to(device)\n",
    "optimizer = torch.optim.Adam(ensemble.parameters(), lr=lr, betas=(0.9, 0.98), eps=1e-9)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "id": "cbe2519c",
   "metadata": {},
   "outputs": [],
   "source": [
    "class WarmUpCosineAnnealingLR(_LRScheduler):\n",
    "    def __init__(self, optimizer, T_max, T_warmup, eta_min=0, last_epoch=-1):\n",
    "        self.T_max = T_max\n",
    "        self.T_warmup = T_warmup\n",
    "        self.eta_min = eta_min\n",
    "        super(WarmUpCosineAnnealingLR, self).__init__(optimizer, last_epoch)\n",
    "\n",
    "    def get_lr(self):\n",
    "        if self.last_epoch < self.T_warmup:\n",
    "            return [base_lr * self.last_epoch / self.T_warmup for base_lr in self.base_lrs]\n",
    "        else:\n",
    "            k = 1 + math.cos(math.pi * (self.last_epoch - self.T_warmup) / (self.T_max - self.T_warmup))\n",
    "            return [self.eta_min + (base_lr - self.eta_min) * k / 2 for base_lr in self.base_lrs]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "id": "b6017282",
   "metadata": {},
   "outputs": [],
   "source": [
    "scheduler = WarmUpCosineAnnealingLR(optimizer=optimizer,\n",
    "                                            T_max=final_num_epochs * len(train_loader),\n",
    "                                            T_warmup=50 * len(train_loader),\n",
    "                                            eta_min=1e-6)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "id": "5b353dc4",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "epoch:1/200 average TL:0.15592927 average VL:0.14370334 epoch time:1134 seconds, lr:2.00e-05\n",
      "epoch:2/200 average TL:0.09058636 average VL:0.06203678 epoch time:1122 seconds, lr:4.00e-05\n",
      "epoch:3/200 average TL:0.05297061 average VL:0.04881890 epoch time:1121 seconds, lr:6.00e-05\n",
      "epoch:4/200 average TL:0.04558529 average VL:0.03864570 epoch time:1151 seconds, lr:8.00e-05\n",
      "epoch:5/200 average TL:0.01873799 average VL:0.00501806 epoch time:1124 seconds, lr:1.00e-04\n",
      "epoch:6/200 average TL:0.00474837 average VL:0.00420403 epoch time:1128 seconds, lr:1.20e-04\n",
      "epoch:7/200 average TL:0.00419213 average VL:0.00366787 epoch time:1121 seconds, lr:1.40e-04\n",
      "epoch:8/200 average TL:0.00345084 average VL:0.00284482 epoch time:1144 seconds, lr:1.60e-04\n",
      "epoch:9/200 average TL:0.00270564 average VL:0.00221378 epoch time:1127 seconds, lr:1.80e-04\n",
      "epoch:10/200 average TL:0.00211922 average VL:0.00178604 epoch time:1121 seconds, lr:2.00e-04\n",
      "epoch:11/200 average TL:0.00173508 average VL:0.00158271 epoch time:1152 seconds, lr:2.20e-04\n",
      "epoch:12/200 average TL:0.00152830 average VL:0.00141087 epoch time:1122 seconds, lr:2.40e-04\n",
      "epoch:13/200 average TL:0.00140898 average VL:0.00130543 epoch time:1152 seconds, lr:2.60e-04\n",
      "epoch:14/200 average TL:0.00132382 average VL:0.00105895 epoch time:1128 seconds, lr:2.80e-04\n",
      "epoch:15/200 average TL:0.00124456 average VL:0.00118853 epoch time:1126 seconds, lr:3.00e-04\n",
      "epoch:16/200 average TL:0.00121053 average VL:0.00120716 epoch time:1129 seconds, lr:3.20e-04\n",
      "epoch:17/200 average TL:0.00117899 average VL:0.00121522 epoch time:1122 seconds, lr:3.40e-04\n",
      "epoch:18/200 average TL:0.00113512 average VL:0.00096504 epoch time:1125 seconds, lr:3.60e-04\n",
      "epoch:19/200 average TL:0.00115872 average VL:0.00107973 epoch time:1122 seconds, lr:3.80e-04\n",
      "epoch:20/200 average TL:0.00114042 average VL:0.00103038 epoch time:1124 seconds, lr:4.00e-04\n",
      "epoch:21/200 average TL:0.00111686 average VL:0.00120114 epoch time:1124 seconds, lr:4.20e-04\n",
      "epoch:22/200 average TL:0.00115883 average VL:0.00136030 epoch time:1127 seconds, lr:4.40e-04\n",
      "epoch:23/200 average TL:0.00106391 average VL:0.00086355 epoch time:1173 seconds, lr:4.60e-04\n",
      "epoch:24/200 average TL:0.00104279 average VL:0.00102997 epoch time:1145 seconds, lr:4.80e-04\n",
      "epoch:25/200 average TL:0.00102520 average VL:0.00097728 epoch time:1140 seconds, lr:5.00e-04\n",
      "epoch:26/200 average TL:0.00102808 average VL:0.00105752 epoch time:1137 seconds, lr:5.20e-04\n",
      "epoch:27/200 average TL:0.00101856 average VL:0.00110042 epoch time:1140 seconds, lr:5.40e-04\n",
      "epoch:28/200 average TL:0.00099290 average VL:0.00095386 epoch time:1124 seconds, lr:5.60e-04\n",
      "epoch:29/200 average TL:0.00095879 average VL:0.00095197 epoch time:1132 seconds, lr:5.80e-04\n",
      "epoch:30/200 average TL:0.00098142 average VL:0.00086822 epoch time:1120 seconds, lr:6.00e-04\n",
      "epoch:31/200 average TL:0.00094956 average VL:0.00084838 epoch time:1139 seconds, lr:6.20e-04\n",
      "epoch:32/200 average TL:0.00092854 average VL:0.00091676 epoch time:1121 seconds, lr:6.40e-04\n",
      "epoch:33/200 average TL:0.00092695 average VL:0.00088483 epoch time:1129 seconds, lr:6.60e-04\n",
      "epoch:34/200 average TL:0.00091708 average VL:0.00090793 epoch time:1122 seconds, lr:6.80e-04\n",
      "epoch:35/200 average TL:0.00091359 average VL:0.00076348 epoch time:1123 seconds, lr:7.00e-04\n",
      "epoch:36/200 average TL:0.00090720 average VL:0.00090714 epoch time:1120 seconds, lr:7.20e-04\n",
      "epoch:37/200 average TL:0.00088160 average VL:0.00080865 epoch time:1160 seconds, lr:7.40e-04\n",
      "epoch:38/200 average TL:0.00085872 average VL:0.00079121 epoch time:1119 seconds, lr:7.60e-04\n",
      "epoch:39/200 average TL:0.00085246 average VL:0.00070570 epoch time:1126 seconds, lr:7.80e-04\n",
      "epoch:40/200 average TL:0.00084088 average VL:0.00081569 epoch time:1148 seconds, lr:8.00e-04\n",
      "epoch:41/200 average TL:0.00084833 average VL:0.00084128 epoch time:1184 seconds, lr:8.20e-04\n",
      "epoch:42/200 average TL:0.00080671 average VL:0.00076897 epoch time:1119 seconds, lr:8.40e-04\n",
      "epoch:43/200 average TL:0.00081056 average VL:0.00064532 epoch time:1198 seconds, lr:8.60e-04\n",
      "epoch:44/200 average TL:0.00080706 average VL:0.00083982 epoch time:1173 seconds, lr:8.80e-04\n",
      "epoch:45/200 average TL:0.00079468 average VL:0.00069288 epoch time:1122 seconds, lr:9.00e-04\n",
      "epoch:46/200 average TL:0.00077051 average VL:0.00073019 epoch time:1136 seconds, lr:9.20e-04\n",
      "epoch:47/200 average TL:0.00077462 average VL:0.00088546 epoch time:1121 seconds, lr:9.40e-04\n",
      "epoch:48/200 average TL:0.00076243 average VL:0.00071173 epoch time:1120 seconds, lr:9.60e-04\n",
      "epoch:49/200 average TL:0.00076067 average VL:0.00063579 epoch time:1132 seconds, lr:9.80e-04\n",
      "epoch:50/200 average TL:0.00077209 average VL:0.00066146 epoch time:1143 seconds, lr:1.00e-03\n",
      "epoch:51/200 average TL:0.00075964 average VL:0.00061390 epoch time:1128 seconds, lr:1.00e-03\n",
      "epoch:52/200 average TL:0.00074001 average VL:0.00073872 epoch time:1120 seconds, lr:1.00e-03\n",
      "epoch:53/200 average TL:0.00074570 average VL:0.00066249 epoch time:1122 seconds, lr:9.99e-04\n",
      "epoch:54/200 average TL:0.00073301 average VL:0.00069401 epoch time:1120 seconds, lr:9.98e-04\n",
      "epoch:55/200 average TL:0.00070841 average VL:0.00083154 epoch time:1152 seconds, lr:9.97e-04\n",
      "epoch:56/200 average TL:0.00071324 average VL:0.00056817 epoch time:1126 seconds, lr:9.96e-04\n",
      "epoch:57/200 average TL:0.00071251 average VL:0.00064402 epoch time:1143 seconds, lr:9.95e-04\n",
      "epoch:58/200 average TL:0.00070249 average VL:0.00072744 epoch time:1130 seconds, lr:9.93e-04\n",
      "epoch:59/200 average TL:0.00071311 average VL:0.00061042 epoch time:1122 seconds, lr:9.91e-04\n",
      "epoch:60/200 average TL:0.00069654 average VL:0.00066567 epoch time:1135 seconds, lr:9.89e-04\n",
      "epoch:61/200 average TL:0.00072318 average VL:0.00073426 epoch time:1149 seconds, lr:9.87e-04\n",
      "epoch:62/200 average TL:0.00070775 average VL:0.00060861 epoch time:1120 seconds, lr:9.84e-04\n",
      "epoch:63/200 average TL:0.00070326 average VL:0.00071113 epoch time:1140 seconds, lr:9.82e-04\n",
      "epoch:64/200 average TL:0.00067885 average VL:0.00054428 epoch time:1129 seconds, lr:9.79e-04\n",
      "epoch:65/200 average TL:0.00069572 average VL:0.00059209 epoch time:1122 seconds, lr:9.76e-04\n",
      "epoch:66/200 average TL:0.00068077 average VL:0.00058569 epoch time:1136 seconds, lr:9.72e-04\n",
      "epoch:67/200 average TL:0.00069046 average VL:0.00063811 epoch time:1140 seconds, lr:9.69e-04\n",
      "epoch:68/200 average TL:0.00068648 average VL:0.00071056 epoch time:1142 seconds, lr:9.65e-04\n",
      "epoch:69/200 average TL:0.00067103 average VL:0.00073547 epoch time:1127 seconds, lr:9.61e-04\n",
      "epoch:70/200 average TL:0.00067951 average VL:0.00064285 epoch time:1123 seconds, lr:9.57e-04\n",
      "epoch:71/200 average TL:0.00067042 average VL:0.00060675 epoch time:1126 seconds, lr:9.52e-04\n",
      "epoch:72/200 average TL:0.00066343 average VL:0.00071368 epoch time:1133 seconds, lr:9.48e-04\n",
      "epoch:73/200 average TL:0.00065169 average VL:0.00061477 epoch time:1125 seconds, lr:9.43e-04\n",
      "epoch:74/200 average TL:0.00064322 average VL:0.00066877 epoch time:1199 seconds, lr:9.38e-04\n",
      "epoch:75/200 average TL:0.00065884 average VL:0.00067680 epoch time:1151 seconds, lr:9.33e-04\n",
      "epoch:76/200 average TL:0.00064478 average VL:0.00052619 epoch time:1124 seconds, lr:9.28e-04\n",
      "epoch:77/200 average TL:0.00064392 average VL:0.00064105 epoch time:1164 seconds, lr:9.22e-04\n",
      "epoch:78/200 average TL:0.00066106 average VL:0.00062171 epoch time:1152 seconds, lr:9.17e-04\n",
      "epoch:79/200 average TL:0.00064599 average VL:0.00070218 epoch time:1130 seconds, lr:9.11e-04\n",
      "epoch:80/200 average TL:0.00065870 average VL:0.00061299 epoch time:1136 seconds, lr:9.05e-04\n",
      "epoch:81/200 average TL:0.00064374 average VL:0.00056236 epoch time:1132 seconds, lr:8.98e-04\n",
      "epoch:82/200 average TL:0.00061804 average VL:0.00066799 epoch time:1152 seconds, lr:8.92e-04\n",
      "epoch:83/200 average TL:0.00064682 average VL:0.00073904 epoch time:1120 seconds, lr:8.85e-04\n",
      "epoch:84/200 average TL:0.00063596 average VL:0.00063792 epoch time:1202 seconds, lr:8.79e-04\n",
      "epoch:85/200 average TL:0.00063784 average VL:0.00048672 epoch time:1141 seconds, lr:8.72e-04\n",
      "epoch:86/200 average TL:0.00063475 average VL:0.00051970 epoch time:1122 seconds, lr:8.65e-04\n",
      "epoch:87/200 average TL:0.00062041 average VL:0.00058418 epoch time:1122 seconds, lr:8.57e-04\n",
      "epoch:88/200 average TL:0.00063310 average VL:0.00052125 epoch time:1120 seconds, lr:8.50e-04\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "epoch:89/200 average TL:0.00063131 average VL:0.00062439 epoch time:1122 seconds, lr:8.42e-04\n",
      "epoch:90/200 average TL:0.00062779 average VL:0.00062117 epoch time:1121 seconds, lr:8.35e-04\n",
      "epoch:91/200 average TL:0.00062870 average VL:0.00065006 epoch time:1125 seconds, lr:8.27e-04\n",
      "epoch:92/200 average TL:0.00060519 average VL:0.00051790 epoch time:1122 seconds, lr:8.19e-04\n",
      "epoch:93/200 average TL:0.00062576 average VL:0.00066642 epoch time:1135 seconds, lr:8.11e-04\n",
      "epoch:94/200 average TL:0.00062534 average VL:0.00046815 epoch time:1122 seconds, lr:8.02e-04\n",
      "epoch:95/200 average TL:0.00062332 average VL:0.00056984 epoch time:1127 seconds, lr:7.94e-04\n",
      "epoch:96/200 average TL:0.00060836 average VL:0.00063531 epoch time:1125 seconds, lr:7.86e-04\n",
      "epoch:97/200 average TL:0.00061028 average VL:0.00053676 epoch time:1121 seconds, lr:7.77e-04\n",
      "epoch:98/200 average TL:0.00059965 average VL:0.00062902 epoch time:1121 seconds, lr:7.68e-04\n",
      "epoch:99/200 average TL:0.00059592 average VL:0.00066299 epoch time:1133 seconds, lr:7.59e-04\n",
      "epoch:100/200 average TL:0.00061985 average VL:0.00053588 epoch time:1121 seconds, lr:7.50e-04\n",
      "epoch:101/200 average TL:0.00060258 average VL:0.00059439 epoch time:1130 seconds, lr:7.41e-04\n",
      "epoch:102/200 average TL:0.00059614 average VL:0.00055450 epoch time:1123 seconds, lr:7.32e-04\n",
      "epoch:103/200 average TL:0.00060052 average VL:0.00059955 epoch time:1123 seconds, lr:7.23e-04\n",
      "epoch:104/200 average TL:0.00060313 average VL:0.00053887 epoch time:1139 seconds, lr:7.13e-04\n",
      "epoch:105/200 average TL:0.00059988 average VL:0.00061950 epoch time:1122 seconds, lr:7.04e-04\n",
      "epoch:106/200 average TL:0.00061866 average VL:0.00065835 epoch time:1142 seconds, lr:6.94e-04\n",
      "epoch:107/200 average TL:0.00059698 average VL:0.00052022 epoch time:1122 seconds, lr:6.84e-04\n",
      "epoch:108/200 average TL:0.00058560 average VL:0.00061357 epoch time:1154 seconds, lr:6.75e-04\n",
      "epoch:109/200 average TL:0.00059967 average VL:0.00050974 epoch time:1129 seconds, lr:6.65e-04\n",
      "epoch:110/200 average TL:0.00059602 average VL:0.00051552 epoch time:1120 seconds, lr:6.55e-04\n",
      "epoch:111/200 average TL:0.00058655 average VL:0.00056903 epoch time:1122 seconds, lr:6.45e-04\n",
      "epoch:112/200 average TL:0.00058789 average VL:0.00060645 epoch time:1127 seconds, lr:6.35e-04\n",
      "epoch:113/200 average TL:0.00057821 average VL:0.00069740 epoch time:1143 seconds, lr:6.25e-04\n",
      "epoch:114/200 average TL:0.00059623 average VL:0.00048697 epoch time:1122 seconds, lr:6.15e-04\n",
      "epoch:115/200 average TL:0.00059243 average VL:0.00060945 epoch time:1125 seconds, lr:6.04e-04\n",
      "epoch:116/200 average TL:0.00057367 average VL:0.00044835 epoch time:1124 seconds, lr:5.94e-04\n",
      "epoch:117/200 average TL:0.00058660 average VL:0.00049836 epoch time:1148 seconds, lr:5.84e-04\n",
      "epoch:118/200 average TL:0.00057522 average VL:0.00053506 epoch time:1123 seconds, lr:5.73e-04\n",
      "epoch:119/200 average TL:0.00058676 average VL:0.00046697 epoch time:1123 seconds, lr:5.63e-04\n",
      "epoch:120/200 average TL:0.00057692 average VL:0.00048688 epoch time:1123 seconds, lr:5.53e-04\n",
      "epoch:121/200 average TL:0.00056761 average VL:0.00050673 epoch time:1128 seconds, lr:5.42e-04\n",
      "epoch:122/200 average TL:0.00057386 average VL:0.00056092 epoch time:1130 seconds, lr:5.32e-04\n",
      "epoch:123/200 average TL:0.00058236 average VL:0.00059367 epoch time:1124 seconds, lr:5.21e-04\n",
      "epoch:124/200 average TL:0.00056390 average VL:0.00055222 epoch time:1127 seconds, lr:5.11e-04\n",
      "epoch:125/200 average TL:0.00058109 average VL:0.00056059 epoch time:1129 seconds, lr:5.01e-04\n",
      "epoch:126/200 average TL:0.00057755 average VL:0.00059075 epoch time:1145 seconds, lr:4.90e-04\n",
      "epoch:127/200 average TL:0.00056258 average VL:0.00046768 epoch time:1121 seconds, lr:4.80e-04\n",
      "epoch:128/200 average TL:0.00055704 average VL:0.00046125 epoch time:1120 seconds, lr:4.69e-04\n",
      "epoch:129/200 average TL:0.00057207 average VL:0.00046427 epoch time:1136 seconds, lr:4.59e-04\n",
      "epoch:130/200 average TL:0.00054465 average VL:0.00055366 epoch time:1148 seconds, lr:4.48e-04\n",
      "epoch:131/200 average TL:0.00055899 average VL:0.00052072 epoch time:1125 seconds, lr:4.38e-04\n",
      "epoch:132/200 average TL:0.00055294 average VL:0.00057654 epoch time:1123 seconds, lr:4.28e-04\n",
      "epoch:133/200 average TL:0.00055073 average VL:0.00050834 epoch time:1152 seconds, lr:4.17e-04\n",
      "epoch:134/200 average TL:0.00056575 average VL:0.00054734 epoch time:1136 seconds, lr:4.07e-04\n",
      "epoch:135/200 average TL:0.00054991 average VL:0.00055409 epoch time:1150 seconds, lr:3.97e-04\n",
      "epoch:136/200 average TL:0.00054888 average VL:0.00051464 epoch time:1161 seconds, lr:3.86e-04\n",
      "epoch:137/200 average TL:0.00057094 average VL:0.00049313 epoch time:1166 seconds, lr:3.76e-04\n",
      "epoch:138/200 average TL:0.00055569 average VL:0.00054596 epoch time:1136 seconds, lr:3.66e-04\n",
      "epoch:139/200 average TL:0.00056493 average VL:0.00046756 epoch time:1141 seconds, lr:3.56e-04\n",
      "epoch:140/200 average TL:0.00054913 average VL:0.00050217 epoch time:1166 seconds, lr:3.46e-04\n",
      "epoch:141/200 average TL:0.00055865 average VL:0.00050881 epoch time:1146 seconds, lr:3.36e-04\n",
      "epoch:142/200 average TL:0.00055391 average VL:0.00051599 epoch time:1135 seconds, lr:3.26e-04\n",
      "epoch:143/200 average TL:0.00053845 average VL:0.00050900 epoch time:1135 seconds, lr:3.17e-04\n",
      "epoch:144/200 average TL:0.00054911 average VL:0.00049289 epoch time:1133 seconds, lr:3.07e-04\n",
      "epoch:145/200 average TL:0.00055715 average VL:0.00044847 epoch time:1149 seconds, lr:2.97e-04\n",
      "epoch:146/200 average TL:0.00055521 average VL:0.00052167 epoch time:1124 seconds, lr:2.88e-04\n",
      "epoch:147/200 average TL:0.00054458 average VL:0.00061373 epoch time:1125 seconds, lr:2.78e-04\n",
      "epoch:148/200 average TL:0.00055216 average VL:0.00042275 epoch time:1125 seconds, lr:2.69e-04\n",
      "epoch:149/200 average TL:0.00055099 average VL:0.00053275 epoch time:1138 seconds, lr:2.60e-04\n",
      "epoch:150/200 average TL:0.00053107 average VL:0.00048005 epoch time:1119 seconds, lr:2.51e-04\n",
      "epoch:151/200 average TL:0.00054083 average VL:0.00058928 epoch time:1123 seconds, lr:2.42e-04\n",
      "epoch:152/200 average TL:0.00054925 average VL:0.00049256 epoch time:1177 seconds, lr:2.33e-04\n",
      "epoch:153/200 average TL:0.00053445 average VL:0.00052247 epoch time:1131 seconds, lr:2.24e-04\n",
      "epoch:154/200 average TL:0.00054940 average VL:0.00058743 epoch time:1120 seconds, lr:2.15e-04\n",
      "epoch:155/200 average TL:0.00055192 average VL:0.00041659 epoch time:1146 seconds, lr:2.07e-04\n",
      "epoch:156/200 average TL:0.00055226 average VL:0.00044853 epoch time:1133 seconds, lr:1.99e-04\n",
      "epoch:157/200 average TL:0.00054616 average VL:0.00051742 epoch time:1120 seconds, lr:1.90e-04\n",
      "epoch:158/200 average TL:0.00054246 average VL:0.00047440 epoch time:1124 seconds, lr:1.82e-04\n",
      "epoch:159/200 average TL:0.00055557 average VL:0.00068125 epoch time:1121 seconds, lr:1.74e-04\n",
      "epoch:160/200 average TL:0.00054225 average VL:0.00044669 epoch time:1124 seconds, lr:1.66e-04\n",
      "epoch:161/200 average TL:0.00054377 average VL:0.00047566 epoch time:1137 seconds, lr:1.59e-04\n",
      "epoch:162/200 average TL:0.00053313 average VL:0.00051402 epoch time:1122 seconds, lr:1.51e-04\n",
      "epoch:163/200 average TL:0.00054396 average VL:0.00056994 epoch time:1149 seconds, lr:1.44e-04\n",
      "epoch:164/200 average TL:0.00054744 average VL:0.00045450 epoch time:1125 seconds, lr:1.36e-04\n",
      "epoch:165/200 average TL:0.00055292 average VL:0.00048513 epoch time:1128 seconds, lr:1.29e-04\n",
      "epoch:166/200 average TL:0.00052136 average VL:0.00058150 epoch time:1135 seconds, lr:1.22e-04\n",
      "epoch:167/200 average TL:0.00053336 average VL:0.00054082 epoch time:1126 seconds, lr:1.16e-04\n",
      "epoch:168/200 average TL:0.00052345 average VL:0.00048114 epoch time:1123 seconds, lr:1.09e-04\n",
      "epoch:169/200 average TL:0.00053357 average VL:0.00052512 epoch time:1149 seconds, lr:1.03e-04\n",
      "epoch:170/200 average TL:0.00053154 average VL:0.00045444 epoch time:1141 seconds, lr:9.64e-05\n",
      "epoch:171/200 average TL:0.00053292 average VL:0.00051755 epoch time:1147 seconds, lr:9.03e-05\n",
      "epoch:172/200 average TL:0.00054005 average VL:0.00048892 epoch time:1125 seconds, lr:8.45e-05\n",
      "epoch:173/200 average TL:0.00054650 average VL:0.00052781 epoch time:1125 seconds, lr:7.88e-05\n",
      "epoch:174/200 average TL:0.00053262 average VL:0.00041618 epoch time:1145 seconds, lr:7.32e-05\n",
      "epoch:175/200 average TL:0.00053334 average VL:0.00044559 epoch time:1138 seconds, lr:6.79e-05\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "epoch:176/200 average TL:0.00052627 average VL:0.00044513 epoch time:1127 seconds, lr:6.28e-05\n",
      "epoch:177/200 average TL:0.00052810 average VL:0.00045932 epoch time:1177 seconds, lr:5.78e-05\n",
      "epoch:178/200 average TL:0.00053555 average VL:0.00047685 epoch time:1122 seconds, lr:5.31e-05\n",
      "epoch:179/200 average TL:0.00053619 average VL:0.00057519 epoch time:1237 seconds, lr:4.85e-05\n",
      "epoch:180/200 average TL:0.00051405 average VL:0.00044181 epoch time:1124 seconds, lr:4.42e-05\n",
      "epoch:181/200 average TL:0.00053028 average VL:0.00050321 epoch time:1118 seconds, lr:4.00e-05\n",
      "epoch:182/200 average TL:0.00052683 average VL:0.00047185 epoch time:1126 seconds, lr:3.61e-05\n",
      "epoch:183/200 average TL:0.00052223 average VL:0.00046943 epoch time:1121 seconds, lr:3.23e-05\n",
      "epoch:184/200 average TL:0.00054029 average VL:0.00056459 epoch time:1121 seconds, lr:2.88e-05\n",
      "epoch:185/200 average TL:0.00052344 average VL:0.00054966 epoch time:1118 seconds, lr:2.54e-05\n",
      "epoch:186/200 average TL:0.00052868 average VL:0.00053174 epoch time:1125 seconds, lr:2.23e-05\n",
      "epoch:187/200 average TL:0.00051332 average VL:0.00052141 epoch time:1122 seconds, lr:1.94e-05\n",
      "epoch:188/200 average TL:0.00053540 average VL:0.00061183 epoch time:1120 seconds, lr:1.67e-05\n",
      "epoch:189/200 average TL:0.00053935 average VL:0.00049823 epoch time:1130 seconds, lr:1.42e-05\n",
      "epoch:190/200 average TL:0.00052234 average VL:0.00048683 epoch time:1125 seconds, lr:1.19e-05\n",
      "epoch:191/200 average TL:0.00052667 average VL:0.00054797 epoch time:1128 seconds, lr:9.85e-06\n",
      "epoch:192/200 average TL:0.00052176 average VL:0.00047502 epoch time:1120 seconds, lr:7.99e-06\n",
      "epoch:193/200 average TL:0.00052012 average VL:0.00050340 epoch time:1127 seconds, lr:6.36e-06\n",
      "epoch:194/200 average TL:0.00053528 average VL:0.00053315 epoch time:1165 seconds, lr:4.94e-06\n",
      "epoch:195/200 average TL:0.00051562 average VL:0.00041009 epoch time:1126 seconds, lr:3.74e-06\n",
      "epoch:196/200 average TL:0.00053298 average VL:0.00047237 epoch time:1123 seconds, lr:2.75e-06\n",
      "epoch:197/200 average TL:0.00052497 average VL:0.00044786 epoch time:1216 seconds, lr:1.99e-06\n",
      "epoch:198/200 average TL:0.00053373 average VL:0.00054342 epoch time:1160 seconds, lr:1.44e-06\n",
      "epoch:199/200 average TL:0.00052414 average VL:0.00051760 epoch time:1120 seconds, lr:1.11e-06\n",
      "epoch:200/200 average TL:0.00052914 average VL:0.00046012 epoch time:1139 seconds, lr:1.00e-06\n",
      "\n",
      "Duration: 226830 seconds\n"
     ]
    }
   ],
   "source": [
    "start_time = time.time()\n",
    "num_train_batches=len(train_loader)\n",
    "num_test_batches=len(test_loader)\n",
    "train_losses = []\n",
    "val_losses = []\n",
    "patience_counter = 0\n",
    "best_val_loss = float('inf')\n",
    "\n",
    "for i in range(final_num_epochs):\n",
    "    loss1 = 0\n",
    "    epoch_time = time.time()\n",
    "    ensemble.train()\n",
    "    # Run the training batches\n",
    "    for b, (X_train, y_train) in enumerate(train_loader):  \n",
    "        optimizer.zero_grad()\n",
    "        # Apply the model\n",
    "        #xt=one_hot_encode(X_train.to(device), 5)\n",
    "        y_pred=ensemble(X_train)\n",
    "        #print(y_pred.shape)\n",
    "        y_train = torch.cat([y_train, 4*torch.ones(batch_size, 2*K, dtype=torch.int).to(device)], dim=1)\n",
    "        yt=one_hot_encode(y_train, 5).to(device)\n",
    "        loss = criterion(y_pred, yt) \n",
    "        # Update parameters\n",
    "        loss.backward()\n",
    "        optimizer.step()\n",
    "        scheduler.step()\n",
    "        loss1=loss1+loss\n",
    "        #if (b+1)%50==1:\n",
    "        #    print(f'epoch: {i+1}/{num_epochs} batch:{b+1}/{num_train_batches} loss: {loss.item():10.8f}')      \n",
    "    train_loss=loss1/num_train_batches  \n",
    "    train_losses.append(train_loss.item())\n",
    "    # Run the testing batches\n",
    "    ensemble.eval()\n",
    "    with torch.no_grad():\n",
    "        loss1=0\n",
    "        for b, (X_test, y_test) in enumerate(test_loader):\n",
    "            \n",
    "            # Apply the model\n",
    "            #xt=one_hot_encode(X_test.to(device), 5)\n",
    "            y_pred=ensemble(X_test)\n",
    "            y_test = torch.cat([y_test, 4*torch.ones(batch_size, 2*K, dtype=torch.int).to(device)], dim=1)\n",
    "            yt=one_hot_encode(y_test, 5).to(device)\n",
    "            loss = criterion(y_pred, yt) \n",
    "            loss1=loss1+loss         \n",
    "        val_loss=loss1/num_test_batches  \n",
    "        val_losses.append(val_loss.item())\n",
    "    \n",
    "    print(f'epoch:{i+1}/{final_num_epochs} average TL:{train_loss.item():10.8f} average VL:{val_loss.item():10.8f} epoch time:{time.time() - epoch_time:.0f} seconds, lr:{optimizer.param_groups[0][\"lr\"]:.2e}')               \n",
    "     # Early stopping\n",
    "    if val_loss < best_val_loss:\n",
    "        best_val_loss = val_loss\n",
    "        torch.save(ensemble.state_dict(), path_directory+\"model.pth\")\n",
    "        \n",
    "print(f'\\nDuration: {time.time() - start_time:.0f} seconds') # print the time elapsed            "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "id": "8fb0da02",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model Flops: 52639880000.0\n",
      "Model Params Num: 936850.0\n",
      "\n"
     ]
    }
   ],
   "source": [
    "t1=iter(train_loader)\n",
    "X,y=next(t1)\n",
    "flops, params = thop.profile(ensemble, inputs=(X,), verbose=False)\n",
    "print(f'Model Flops: {flops}')\n",
    "print(f'Model Params Num: {params}\\n')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "id": "0b3a903e",
   "metadata": {},
   "outputs": [],
   "source": [
    "dict={'train_losses':train_losses, 'val_losses':val_losses, 'flops':flops, 'parameters':params }\n",
    "savemat(path_directory+'losses.mat', dict)\n",
    "with open(path_directory+'losses.pkl', 'wb') as file:\n",
    "    \n",
    "    # A new file will be created\n",
    "    pickle.dump(dict, file)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "id": "51fd549e",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Plotting\n",
    "plt.plot(train_losses, label='Train Loss')\n",
    "plt.plot(val_losses, label='Validation Loss')\n",
    "plt.xlabel('Epoch')\n",
    "plt.ylabel('Loss')\n",
    "plt.legend()\n",
    "plt.grid()\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7401de54",
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "id": "b3bc825d",
   "metadata": {},
   "outputs": [],
   "source": [
    "class LanguageDataset(Dataset):\n",
    "\n",
    "    def __init__(self, dictionary, num_samples, num_errors):\n",
    "        self.dictionary = dictionary\n",
    "        self.num_samples = num_samples \n",
    "        self.num_errors = num_errors\n",
    "\n",
    "    def __len__(self):\n",
    "        return self.num_samples\n",
    "\n",
    "    def __getitem__(self, idx):\n",
    "        source_idx = idx % len(self.dictionary)\n",
    "        sequence = self.dictionary[source_idx]\n",
    "        choice = torch.rand(1).item()\n",
    "        if self.num_errors > 0:\n",
    "            if choice >= 0 and choice< 0.10646:\n",
    "                sequence = self.introduce_substitution_errors(sequence, self.num_errors)\n",
    "                for _ in range(self.num_errors):\n",
    "                    sequence.append('N')  \n",
    "                sequence=''.join(sequence)\n",
    "                #sequence= torch.cat([sequence, 'NNNN'], dim=1)\n",
    "            elif choice >=0.10646 and choice < 0.14218:\n",
    "                sequence = self.introduce_deletion_errors(sequence, self.num_errors)\n",
    "                for _ in range(self.num_errors):\n",
    "                    sequence.append('N') \n",
    "                sequence=''.join(sequence)\n",
    "                #sequence= torch.cat([sequence, 'NNNN'], dim=1)\n",
    "            elif choice >=0.14218 and choice < 0.15163:\n",
    "                sequence = self.introduce_insertion_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "            elif choice >=0.15163 and choice < 0.156:\n",
    "                sequence = self.introduce_substitution_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "                sequence = self.introduce_deletion_errors(sequence, self.num_errors)\n",
    "                for _ in range(self.num_errors):\n",
    "                    sequence.append('N') \n",
    "                sequence=''.join(sequence)\n",
    "                #sequence= torch.cat([sequence, 'NNNN'], dim=1)\n",
    "            elif choice >=0.156 and choice < 0.15701:\n",
    "                sequence = self.introduce_substitution_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "                sequence = self.introduce_insertion_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "                \n",
    "            elif choice >=0.15701 and choice < 0.1574:\n",
    "                sequence = self.introduce_deletion_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "                sequence = self.introduce_insertion_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "            elif choice >=0.1574 and choice < 0.15747:\n",
    "                sequence = self.introduce_substitution_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "                sequence = self.introduce_deletion_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)\n",
    "                sequence = self.introduce_insertion_errors(sequence, self.num_errors)\n",
    "                sequence=''.join(sequence)  \n",
    "            else:\n",
    "                sequence = list(sequence)\n",
    "                for _ in range(self.num_errors):\n",
    "                    sequence.append('N') \n",
    "                sequence=''.join(sequence)   \n",
    "            \n",
    "        x = torch.tensor([self.char_to_idx(c) for c in sequence], dtype=torch.long)\n",
    "        y = torch.tensor([self.char_to_idx(c) for c in self.dictionary[source_idx]], dtype=torch.long)\n",
    "        return x.to(device), y.to(device)\n",
    "\n",
    "    def introduce_deletion_errors(self, sequence, num_errors):\n",
    "        #ne = random.randint(1, num_errors)\n",
    "        sequence = list(sequence)\n",
    "        \n",
    "        for _ in range(num_errors):\n",
    "            i = random.randint(0, len(sequence)-1)\n",
    "            del sequence[i]\n",
    "            sequence.append('N') \n",
    "        \n",
    "        return sequence\n",
    "    \n",
    "    def introduce_substitution_errors(self, sequence, num_errors):\n",
    "        #ne = random.randint(1, num_errors)\n",
    "        sequence = list(sequence)\n",
    "        for _ in range(num_errors):\n",
    "            i = random.randint(0, len(sequence)-1)\n",
    "            if sequence[i] == 'A':\n",
    "                sequence[i] = random.choice(['T', 'G', 'C'])\n",
    "            elif sequence[i] == 'T':\n",
    "                sequence[i] = random.choice(['A', 'G', 'C'])\n",
    "            elif sequence[i] == 'G':\n",
    "                sequence[i] = random.choice(['A', 'T', 'C'])\n",
    "            elif sequence[i] == 'C':\n",
    "                sequence[i] = random.choice(['A', 'T', 'G'])\n",
    "        \n",
    "        return sequence\n",
    "    \n",
    "    def introduce_insertion_errors(self, sequence, num_errors):\n",
    "        #ne = random.randint(1, num_errors)\n",
    "        sequence = list(sequence)\n",
    "        #for _ in range(num_errors-ne):\n",
    "        #    sequence.append('N')\n",
    "        for _ in range(num_errors):\n",
    "            i = random.randint(0, len(sequence))\n",
    "            random_char = random.choice(['A','C','T','G']) \n",
    "            sequence.insert(i, random_char)\n",
    "        return sequence\n",
    "\n",
    "    def char_to_idx(self, char):\n",
    "        return {'A': 0, 'C': 1, 'T': 2, 'G': 3, 'N': 4}[char]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "id": "47e6dab4",
   "metadata": {},
   "outputs": [],
   "source": [
    "start_test=1\n",
    "stop_test=4\n",
    "num_test=200000\n",
    "test_error_points=torch.arange(start_test, stop_test+1)\n",
    "#nucleotide error rate\n",
    "ner=torch.zeros((stop_test-start_test+1)).to(device)\n",
    "#sequence error rate\n",
    "ser=torch.zeros((stop_test-start_test+1)).to(device)\n",
    "time_tensor =torch.zeros((stop_test-start_test+1)).to(device)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "id": "973c5d20",
   "metadata": {},
   "outputs": [],
   "source": [
    "def count_different_rows(matrix1, matrix2):\n",
    "    \"\"\"\n",
    "    Calculates the number of rows that are not equal between two 2D PyTorch matrices.\n",
    "\n",
    "    Args:\n",
    "        matrix1 (torch.Tensor): A 2D PyTorch tensor of shape (m, n).\n",
    "        matrix2 (torch.Tensor): A 2D PyTorch tensor of shape (m, n).\n",
    "\n",
    "    Returns:\n",
    "        int: The number of rows that are not equal between the two matrices.\n",
    "    \"\"\"\n",
    "    assert matrix1.shape == matrix2.shape, \"The input matrices must have the same shape.\"\n",
    "    equal_rows = torch.all(matrix1 == matrix2, dim=1)\n",
    "    return matrix1.shape[0] - torch.sum(equal_rows).item()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "id": "c53a8c64",
   "metadata": {},
   "outputs": [],
   "source": [
    "i1=0\n",
    "for i in range(start_test, stop_test+1):\n",
    "    test_data1 = LanguageDataset(dicts, num_samples=num_test, num_errors=i)\n",
    "    test_loader1 = DataLoader(test_data1, batch_size=batch_size, shuffle=True)\n",
    "    n_tbatch=len(test_loader1)\n",
    "    ne=0.0\n",
    "    count=0.0\n",
    "    decoding_time =0\n",
    "    # Run the testing batches\n",
    "    ensemble.eval()\n",
    "    with torch.no_grad():\n",
    "        for b, (X_test, y_test) in enumerate(test_loader1):\n",
    "            # Apply the model\n",
    "            start_time = time.time()\n",
    "            y_pred=ensemble(X_test)\n",
    "            decoding_time = decoding_time + time.time()-start_time\n",
    "            y_pred_d=one_hot_decode(y_pred)\n",
    "            \n",
    "            y_pred_dk = y_pred_d[:,:K]\n",
    "            ne=ne+torch.sum(y_pred_dk!=y_test).item()\n",
    "            count=count+count_different_rows(y_test,y_pred_dk)\n",
    "        time_tensor[i1] = decoding_time/(n_tbatch*batch_size)\n",
    "        ner[i1]=ne/(n_tbatch*batch_size*K)\n",
    "        ser[i1]=count/(n_tbatch*batch_size)\n",
    "    i1=i1+1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "id": "e6f272cc",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor([0.0010, 0.0010, 0.0011, 0.0011], device='cuda:0')"
      ]
     },
     "execution_count": 37,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "time_tensor"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "id": "a0c8cf5b",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor([0.0029, 0.0050, 0.0060, 0.0128], device='cuda:0')"
      ]
     },
     "execution_count": 38,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "ner"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "id": "5cc695a1",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor([0.0054, 0.0092, 0.0115, 0.0221], device='cuda:0')"
      ]
     },
     "execution_count": 39,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "ser"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "id": "f5da72f3",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.plot(torch.Tensor.cpu(test_error_points).numpy(),torch.Tensor.cpu(ner).numpy())\n",
    "plt.yscale('log')\n",
    "plt.xlabel('Number of errors')\n",
    "plt.ylabel('NER')    \n",
    "plt.grid(True)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "id": "aa18ab62",
   "metadata": {},
   "outputs": [],
   "source": [
    "dict1={'ner':torch.Tensor.cpu(ner).numpy(), 'ser':torch.Tensor.cpu(ser).numpy()}\n",
    "savemat(path_directory+'metric.mat', dict1)\n",
    "with open(path_directory+'metric.pkl', 'wb') as file:\n",
    "    \n",
    "    # A new file will be created\n",
    "    pickle.dump(dict1, file)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "id": "9dc7e09e",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.plot(torch.Tensor.cpu(test_error_points).numpy(),torch.Tensor.cpu(ser).numpy())\n",
    "plt.yscale('log')\n",
    "plt.xlabel('Number of errors')\n",
    "plt.ylabel('SER')    \n",
    "plt.grid(True)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "6e546b61",
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "39846d7e",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.10.9"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}