train_mobilenet.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "%load_ext autoreload\n",
    "%autoreload 2\n",
    "import torch\n",
    "from solver.solver import Solver\n",
    "from models.cnn_transformer import get_resnet_transformer\n",
    "from models.resnet101_3d_cnn import get_resnet101_3d\n",
    "from dataset.subset_loader import GestureSubset\n",
    "from torch.utils.data import DataLoader\n",
    "from models.utils import count_trainable_parameters\n",
    "import torch.nn as nn\n",
    "import torch.optim as optim\n",
    "from torchvision import transforms\n",
    "import os\n",
    "import numpy as np\n",
    "import json\n",
    "\n",
    "from edl_playground.edl.losses import (\n",
    "    LinearAnnealingFactor, ExpAnnealingFactor,\n",
    "    Type2MaximumLikelihoodLoss, BayesRiskForCrossEntropyLoss, BayesRiskForSSELoss,\n",
    "    KL_Divergence_RegularizationLoss, EUC_RegularizationLoss,\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "seed = 42\n",
    "np.random.seed(seed)\n",
    "torch.manual_seed(seed)\n",
    "torch.cuda.manual_seed(seed)\n",
    "torch.backends.cudnn.deterministic = True"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "annotations_train = \"Annot_TrainList.txt\"\n",
    "annotations_test = \"Annot_TestList.txt\"\n",
    "\n",
    "path_annotations_train = os.path.join(\".\", \"IPN_Hand\", \"annotations\", annotations_train)\n",
    "path_annotations_test = os.path.join(\".\", \"IPN_Hand\", \"annotations\", annotations_test)\n",
    "path_frames = os.path.join(\".\", \"IPN_Hand\", \"hand_gestures.h5\")\n",
    "\n",
    "path_to_save = \"cnn_trans_edl.pth\"\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "# transform = transforms.Compose([\n",
    "# transforms.ToTensor()\n",
    "# ])\n",
    "\n",
    "# SAMPLE_DURATION = 4\n",
    "\n",
    "# ds_train = GestureSubset(path_frames, path_annotations_train, transform, sample_duration=SAMPLE_DURATION)\n",
    "# ds_test = GestureSubset(path_frames, path_annotations_test, transform, sample_duration=SAMPLE_DURATION)\n",
    "\n",
    "# from collections import Counter\n",
    "\n",
    "# print(Counter([s[1] for s in ds_train]))  # Counter({0: 922, 1: 745, 2: 743, 3: 148, 4: 148, 5: 148, 6: 149, 7: 148, 8: 148, 9: 148, 10: 148, 11: 148, 12: 148, 13: 148})\n",
    "# print(Counter([s[1] for s in ds_test]))   # Counter({0: 509, 1: 265, 2: 264, 3: 52, 4: 52, 5: 52, 6: 52, 7: 52, 8: 52, 9: 52, 10: 52, 11: 52, 12: 52, 13: 52})\n",
    "# counter.most_common()\n",
    "# sorted(counter.items())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "BALANCED = False\n",
    "UNFREEZE_LAYERS = [\n",
    "    # [1, 5],\n",
    "    [1, 8],\n",
    "]\n",
    "# LOSS = \"focal\"\n",
    "# LOSS = \"focalcb\"\n",
    "# LOSS = \"sigmoidcb\"\n",
    "LOSS = \"softmaxcb\"\n",
    "# LOSS = \"weightedce\"\n",
    "# LOSS = \"ce\"\n",
    "\n",
    "assert(not (BALANCED and LOSS == \"weightedce\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "transform = transforms.Compose([\n",
    "transforms.ToTensor()\n",
    "])\n",
    "\n",
    "BATCH_SIZE = 8\n",
    "CLASS_IDS = list(range(5))\n",
    "no_of_classes = len(CLASS_IDS)\n",
    "if BALANCED:\n",
    "    N_SAMPLES_PER_CLASS_TRAIN = (BATCH_SIZE * 2 * np.array([3, 3, 3, 3, 3])).tolist()\n",
    "else:\n",
    "    N_SAMPLES_PER_CLASS_TRAIN = (BATCH_SIZE * 2 * np.array([6, 5, 5, 1, 1])).tolist()\n",
    "N_SAMPLES_PER_CLASS_TEST = (BATCH_SIZE * 2 * np.array([10, 5, 5, 1, 1])).tolist()\n",
    "SAMPLE_DURATION = 4\n",
    "\n",
    "ds_train = GestureSubset(path_frames, path_annotations_train, transform, sample_duration=SAMPLE_DURATION, class_ids=CLASS_IDS, n_samples_per_class=N_SAMPLES_PER_CLASS_TRAIN)\n",
    "ds_test = GestureSubset(path_frames, path_annotations_test, transform, sample_duration=SAMPLE_DURATION, class_ids=CLASS_IDS, n_samples_per_class=N_SAMPLES_PER_CLASS_TEST)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.31614906832298134"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(GestureSubset(path_frames, path_annotations_test, transform, sample_duration=4, class_ids=[0]))/len(GestureSubset(path_frames, path_annotations_test, transform, sample_duration=4))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train Dataset\n",
      "Batches: 288\n",
      "Elements per Sample: 3\n",
      "Image shape torch.Size([4, 3, 128, 128])\n"
     ]
    }
   ],
   "source": [
    "print(\"Train Dataset\")\n",
    "print(\"Batches:\", len(ds_train)) # Batch x np.array(frames), id, label\n",
    "print(\"Elements per Sample:\", len(ds_train[0]))\n",
    "print(\"Image shape\", ds_train[0][0].shape) # Sample duration, Channels, H, W\n",
    "# unique_ids = {s[1]: True for s in ds_train}\n",
    "# print(\"ID (single number) from possible values:\", sorted(list(unique_ids.keys())))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Test Dataset\n",
      "Batches: 352\n",
      "Elements per Sample: 3\n",
      "Image shape torch.Size([4, 3, 128, 128])\n"
     ]
    }
   ],
   "source": [
    "print(\"Test Dataset\")\n",
    "print(\"Batches:\", len(ds_test)) # Batch x np.array(frames), id, label\n",
    "print(\"Elements per Sample:\", len(ds_test[0]))\n",
    "print(\"Image shape\", ds_test[0][0].shape) # Sample duration, Channels, H, W\n",
    "# unique_ids = {s[1]: True for s in ds_test}\n",
    "# print(\"ID (single number) from possible values:\", sorted(list(unique_ids.keys())))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "import timm\n",
    "from torchsummary import summary\n",
    "feature_model = timm.create_model('tf_mobilenetv3_small_100.in1k', pretrained=True)#, num_classes=len(CLASS_IDS))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "======================================================================\n",
      "Layer (type:depth-idx)                        Param #\n",
      "======================================================================\n",
      "├─Conv2dSame: 1-1                             432\n",
      "├─BatchNormAct2d: 1-2                         --\n",
      "|    └─Identity: 2-1                          --\n",
      "|    └─Hardswish: 2-2                         --\n",
      "├─Sequential: 1-3                             --\n",
      "|    └─Sequential: 2-3                        --\n",
      "|    |    └─DepthwiseSeparableConv: 3-1       744\n",
      "|    └─Sequential: 2-4                        --\n",
      "|    |    └─InvertedResidual: 3-2             3,864\n",
      "|    |    └─InvertedResidual: 3-3             5,416\n",
      "|    └─Sequential: 2-5                        --\n",
      "|    |    └─InvertedResidual: 3-4             13,736\n",
      "|    |    └─InvertedResidual: 3-5             57,264\n",
      "|    |    └─InvertedResidual: 3-6             57,264\n",
      "|    └─Sequential: 2-6                        --\n",
      "|    |    └─InvertedResidual: 3-7             21,968\n",
      "|    |    └─InvertedResidual: 3-8             29,800\n",
      "|    └─Sequential: 2-7                        --\n",
      "|    |    └─InvertedResidual: 3-9             91,848\n",
      "|    |    └─InvertedResidual: 3-10            294,096\n",
      "|    |    └─InvertedResidual: 3-11            294,096\n",
      "|    └─Sequential: 2-8                        --\n",
      "|    |    └─ConvBnAct: 3-12                   56,448\n",
      "├─SelectAdaptivePool2d: 1-4                   --\n",
      "|    └─AdaptiveAvgPool2d: 2-9                 --\n",
      "|    └─Identity: 2-10                         --\n",
      "├─Conv2d: 1-5                                 590,848\n",
      "├─Hardswish: 1-6                              --\n",
      "├─Flatten: 1-7                                --\n",
      "├─Linear: 1-8                                 1,025,000\n",
      "======================================================================\n",
      "Total params: 2,542,824\n",
      "Trainable params: 2,542,824\n",
      "Non-trainable params: 0\n",
      "======================================================================\n"
     ]
    }
   ],
   "source": [
    "summary(feature_model);"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "from itertools import chain\n",
    "\n",
    "def get_max_depth(unfreeze_layers_depth_idx):\n",
    "    max_depth = 0\n",
    "    for depth_idx in unfreeze_layers_depth_idx:\n",
    "        depth = depth_idx[0]\n",
    "        if depth > max_depth:\n",
    "            max_depth = depth\n",
    "    return max_depth\n",
    "\n",
    "def call_children_bfs(queue, unfreeze_layers_depth_idx, depth, max_depth):\n",
    "    if depth > max_depth:\n",
    "        return\n",
    "    depth_idx_for_this_depth = [idx for idx in unfreeze_layers_depth_idx if idx[0] == depth]\n",
    "    new_queue = []\n",
    "    for i, child in enumerate(queue, start=1):\n",
    "        if [depth, i] in depth_idx_for_this_depth:\n",
    "            for param in child.parameters():\n",
    "                param.requires_grad = True\n",
    "        new_queue.append(child.children())\n",
    "    new_queue = chain.from_iterable(new_queue)\n",
    "    call_children_bfs(new_queue, unfreeze_layers_depth_idx, depth+1, max_depth)\n",
    "\n",
    "def unfreeze_layer(model, unfreeze_layers_depth_idx):\n",
    "    max_depth = get_max_depth(unfreeze_layers_depth_idx)\n",
    "    queue = model.children()\n",
    "    call_children_bfs(queue, unfreeze_layers_depth_idx, 1, max_depth)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "for param in feature_model.parameters():\n",
    "    param.requires_grad = False\n",
    "\n",
    "unfreeze_layer(feature_model, UNFREEZE_LAYERS)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "===============================================================================================\n",
      "Layer (type:depth-idx)                        Output Shape              Param #\n",
      "===============================================================================================\n",
      "├─Conv2dSame: 1-1                             [-1, 16, 64, 64]          (432)\n",
      "├─BatchNormAct2d: 1-2                         [-1, 16, 64, 64]          --\n",
      "|    └─Identity: 2-1                          [-1, 16, 64, 64]          --\n",
      "|    └─Hardswish: 2-2                         [-1, 16, 64, 64]          --\n",
      "├─Sequential: 1-3                             [-1, 576, 4, 4]           --\n",
      "|    └─Sequential: 2-3                        [-1, 16, 32, 32]          --\n",
      "|    |    └─DepthwiseSeparableConv: 3-1       [-1, 16, 32, 32]          (744)\n",
      "|    └─Sequential: 2-4                        [-1, 24, 16, 16]          --\n",
      "|    |    └─InvertedResidual: 3-2             [-1, 24, 16, 16]          (3,864)\n",
      "|    |    └─InvertedResidual: 3-3             [-1, 24, 16, 16]          (5,416)\n",
      "|    └─Sequential: 2-5                        [-1, 40, 8, 8]            --\n",
      "|    |    └─InvertedResidual: 3-4             [-1, 40, 8, 8]            (13,736)\n",
      "|    |    └─InvertedResidual: 3-5             [-1, 40, 8, 8]            (57,264)\n",
      "|    |    └─InvertedResidual: 3-6             [-1, 40, 8, 8]            (57,264)\n",
      "|    └─Sequential: 2-6                        [-1, 48, 8, 8]            --\n",
      "|    |    └─InvertedResidual: 3-7             [-1, 48, 8, 8]            (21,968)\n",
      "|    |    └─InvertedResidual: 3-8             [-1, 48, 8, 8]            (29,800)\n",
      "|    └─Sequential: 2-7                        [-1, 96, 4, 4]            --\n",
      "|    |    └─InvertedResidual: 3-9             [-1, 96, 4, 4]            (91,848)\n",
      "|    |    └─InvertedResidual: 3-10            [-1, 96, 4, 4]            (294,096)\n",
      "|    |    └─InvertedResidual: 3-11            [-1, 96, 4, 4]            (294,096)\n",
      "|    └─Sequential: 2-8                        [-1, 576, 4, 4]           --\n",
      "|    |    └─ConvBnAct: 3-12                   [-1, 576, 4, 4]           (56,448)\n",
      "├─SelectAdaptivePool2d: 1-4                   [-1, 576, 1, 1]           --\n",
      "|    └─AdaptiveAvgPool2d: 2-9                 [-1, 576, 1, 1]           --\n",
      "|    └─Identity: 2-10                         [-1, 576, 1, 1]           --\n",
      "├─Conv2d: 1-5                                 [-1, 1024, 1, 1]          (590,848)\n",
      "├─Hardswish: 1-6                              [-1, 1024, 1, 1]          --\n",
      "├─Flatten: 1-7                                [-1, 1024]                --\n",
      "├─Linear: 1-8                                 [-1, 1000]                1,025,000\n",
      "===============================================================================================\n",
      "Total params: 2,542,824\n",
      "Trainable params: 1,025,000\n",
      "Non-trainable params: 1,517,824\n",
      "Total mult-adds (M): 22.61\n",
      "===============================================================================================\n",
      "Input size (MB): 0.19\n",
      "Forward/backward pass size (MB): 3.53\n",
      "Params size (MB): 9.70\n",
      "Estimated Total Size (MB): 13.42\n",
      "===============================================================================================\n"
     ]
    }
   ],
   "source": [
    "summary(feature_model, (3, 128, 128));"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "class ApplyToSeq(nn.Module):\n",
    "    def __init__(self, module):\n",
    "        super().__init__()\n",
    "        self.module = module\n",
    "\n",
    "    def forward(self, x):\n",
    "        x = x.transpose(1, 2) # B, C, F, H, W > B, F, C, H, W\n",
    "        b, f = x.shape[:2]\n",
    "        y = x.reshape(b * f, *x.shape[2:])\n",
    "        y = self.module(y)\n",
    "        x = y.view(b, f, *y.shape[1:]) # B, F, C_out\n",
    "        # x = x.transpose(1, 2) # B, F, C_out > B, C_out, F\n",
    "        return x"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "class SelectFinalState(nn.Module):\n",
    "    def forward(self, x):\n",
    "        x = x[1]\n",
    "        x = x.view(x.shape[1:])\n",
    "        return x"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "class PrintShape(nn.Module):\n",
    "    def forward(self, x):\n",
    "        print(\"Shape\", x.shape)\n",
    "        return x"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = nn.Sequential(\n",
    "    # PrintShape(),\n",
    "    ApplyToSeq(feature_model),\n",
    "    # PrintShape(),\n",
    "    nn.GRU(1000, 256, batch_first=True),\n",
    "    SelectFinalState(),\n",
    "    # PrintShape(),\n",
    "    nn.Linear(256, 64),\n",
    "    # PrintShape(),\n",
    "    nn.Dropout(),\n",
    "    nn.Linear(64, len(CLASS_IDS)),\n",
    "    # PrintShape(),\n",
    "    nn.Softmax(dim=-1)\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "======================================================================\n",
      "Layer (type:depth-idx)                        Param #\n",
      "======================================================================\n",
      "├─ApplyToSeq: 1-1                             --\n",
      "|    └─MobileNetV3: 2-1                       --\n",
      "|    |    └─Conv2dSame: 3-1                   (432)\n",
      "|    |    └─BatchNormAct2d: 3-2               (32)\n",
      "|    |    └─Sequential: 3-3                   (926,544)\n",
      "|    |    └─SelectAdaptivePool2d: 3-4         --\n",
      "|    |    └─Conv2d: 3-5                       (590,848)\n",
      "|    |    └─Hardswish: 3-6                    --\n",
      "|    |    └─Flatten: 3-7                      --\n",
      "|    |    └─Linear: 3-8                       1,025,000\n",
      "├─GRU: 1-2                                    966,144\n",
      "├─SelectFinalState: 1-3                       --\n",
      "├─Linear: 1-4                                 16,448\n",
      "├─Dropout: 1-5                                --\n",
      "├─Linear: 1-6                                 325\n",
      "├─Softmax: 1-7                                --\n",
      "======================================================================\n",
      "Total params: 3,525,773\n",
      "Trainable params: 2,007,917\n",
      "Non-trainable params: 1,517,856\n",
      "======================================================================\n"
     ]
    }
   ],
   "source": [
    "summary(model);"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [],
   "source": [
    "device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "from torch import Tensor\n",
    "\n",
    "class FocalLoss(nn.Module):\n",
    "    # Multiclass version from https://arxiv.org/pdf/1708.02002.pdf\n",
    "    def __init__(self, alpha: float=0.25, gamma: float=2): #, reduction: str='mean'\n",
    "        super().__init__()\n",
    "        self.alpha = alpha\n",
    "        self.gamma = gamma\n",
    "        # self.reduction = reduction\n",
    "\n",
    "    def forward(self, input: Tensor, target: Tensor) -> Tensor:\n",
    "        ce_loss = torch.nn.functional.cross_entropy(input, target, reduction=\"none\")\n",
    "        p_t = torch.exp(-ce_loss)\n",
    "        focal_loss = (self.alpha * (1-p_t)**self.gamma * ce_loss).mean()\n",
    "        return focal_loss"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch.nn.functional as F\n",
    "\n",
    "def focal_loss(labels, logits, alpha, gamma):\n",
    "    \"\"\"Compute the focal loss between `logits` and the ground truth `labels`.\n",
    "\n",
    "    Focal loss = -alpha_t * (1-pt)^gamma * log(pt)\n",
    "    where pt is the probability of being classified to the true class.\n",
    "    pt = p (if true class), otherwise pt = 1 - p. p = sigmoid(logit).\n",
    "\n",
    "    Args:\n",
    "      labels: A float tensor of size [batch, num_classes].\n",
    "      logits: A float tensor of size [batch, num_classes].\n",
    "      alpha: A float tensor of size [batch_size]\n",
    "        specifying per-example weight for balanced cross entropy.\n",
    "      gamma: A float scalar modulating loss from hard and easy examples.\n",
    "\n",
    "    Returns:\n",
    "      focal_loss: A float32 scalar representing normalized total loss.\n",
    "    \"\"\"    \n",
    "    BCLoss = F.binary_cross_entropy_with_logits(input = logits, target = labels,reduction = \"none\")\n",
    "\n",
    "    if gamma == 0.0:\n",
    "        modulator = 1.0\n",
    "    else:\n",
    "        modulator = torch.exp(-gamma * labels * logits - gamma * torch.log(1 + \n",
    "            torch.exp(-1.0 * logits)))\n",
    "\n",
    "    loss = modulator * BCLoss\n",
    "\n",
    "    weighted_loss = alpha * loss\n",
    "    focal_loss = torch.sum(weighted_loss)\n",
    "\n",
    "    focal_loss /= torch.sum(labels)\n",
    "    return focal_loss\n",
    "\n",
    "def CB_loss(labels, logits, samples_per_cls, no_of_classes, loss_type, beta, gamma, device):\n",
    "    \"\"\"Compute the Class Balanced Loss between `logits` and the ground truth `labels`.\n",
    "\n",
    "    Based on: https://arxiv.org/pdf/1901.05555.pdf from\n",
    "    https://github.com/vandit15/Class-balanced-loss-pytorch/blob/master/class_balanced_loss.py\n",
    "\n",
    "    Class Balanced Loss: ((1-beta)/(1-beta^n))*Loss(labels, logits)\n",
    "    where Loss is one of the standard losses used for Neural Networks.\n",
    "\n",
    "    Args:\n",
    "      labels: A int tensor of size [batch].\n",
    "      logits: A float tensor of size [batch, no_of_classes].\n",
    "      samples_per_cls: A python list of size [no_of_classes].\n",
    "      no_of_classes: total number of classes. int\n",
    "      loss_type: string. One of \"sigmoid\", \"focal\", \"softmax\".\n",
    "      beta: float. Hyperparameter for Class balanced loss.\n",
    "      gamma: float. Hyperparameter for Focal loss.\n",
    "\n",
    "    Returns:\n",
    "      cb_loss: A float tensor representing class balanced loss\n",
    "    \"\"\"\n",
    "    effective_num = 1.0 - np.power(beta, samples_per_cls)\n",
    "    weights = (1.0 - beta) / np.array(effective_num)\n",
    "    weights = weights / np.sum(weights) * no_of_classes\n",
    "\n",
    "    labels_one_hot = F.one_hot(labels, no_of_classes).float()\n",
    "\n",
    "    weights = torch.tensor(weights).float()\n",
    "    weights = weights.unsqueeze(0)\n",
    "    weights = weights.repeat(labels_one_hot.shape[0],1).to(device) * labels_one_hot\n",
    "    weights = weights.sum(1)\n",
    "    weights = weights.unsqueeze(1)\n",
    "    weights = weights.repeat(1,no_of_classes)\n",
    "\n",
    "    if loss_type == \"focal\":\n",
    "        cb_loss = focal_loss(labels_one_hot, logits, weights, gamma)\n",
    "    elif loss_type == \"sigmoid\":\n",
    "        cb_loss = F.binary_cross_entropy_with_logits(input = logits,target = labels_one_hot, weight = weights)\n",
    "    elif loss_type == \"softmax\":\n",
    "        pred = logits.softmax(dim = 1)\n",
    "        cb_loss = F.binary_cross_entropy(input = pred, target = labels_one_hot, weight = weights)\n",
    "    return cb_loss\n",
    "\n",
    "class _CBLoss(nn.Module):\n",
    "    def __init__(self, samples_per_cls, no_of_classes, beta: float=0.9999, gamma: float=2.0, device=None):\n",
    "        super().__init__()\n",
    "        self.samples_per_cls = samples_per_cls\n",
    "        self.no_of_classes = no_of_classes\n",
    "        self.beta = beta\n",
    "        self.gamma = gamma\n",
    "        self.device = device\n",
    "\n",
    "class FocalCBLoss(_CBLoss):\n",
    "    def forward(self, input: Tensor, target: Tensor) -> Tensor:\n",
    "        loss = CB_loss(target, input, self.samples_per_cls, self.no_of_classes, loss_type=\"focal\", beta=self.beta, gamma=self.gamma, device=self.device)\n",
    "        return loss\n",
    "\n",
    "class SigmoidCBLoss(_CBLoss):\n",
    "    def forward(self, input: Tensor, target: Tensor) -> Tensor:\n",
    "        loss = CB_loss(target, input, self.samples_per_cls, self.no_of_classes, loss_type=\"sigmoid\", beta=self.beta, gamma=self.gamma, device=self.device)\n",
    "        return loss\n",
    "\n",
    "class SoftmaxCBLoss(_CBLoss):\n",
    "    def forward(self, input: Tensor, target: Tensor) -> Tensor:\n",
    "        loss = CB_loss(target, input, self.samples_per_cls, self.no_of_classes, loss_type=\"softmax\", beta=self.beta, gamma=self.gamma, device=self.device)\n",
    "        return loss"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [],
   "source": [
    "if LOSS == \"focal\":\n",
    "    criterion = FocalLoss()\n",
    "elif LOSS == \"focalcb\":\n",
    "    criterion = FocalCBLoss(N_SAMPLES_PER_CLASS_TRAIN, no_of_classes, device=device)\n",
    "elif LOSS == \"sigmoidcb\":\n",
    "    criterion = SigmoidCBLoss(N_SAMPLES_PER_CLASS_TRAIN, no_of_classes, device=device)\n",
    "elif LOSS == \"softmaxcb\":\n",
    "    criterion = SoftmaxCBLoss(N_SAMPLES_PER_CLASS_TRAIN, no_of_classes, device=device)\n",
    "elif LOSS == \"weightedce\":\n",
    "    criterion = nn.CrossEntropyLoss(weight=torch.tensor([6, 5, 5, 1, 1], dtype=torch.float).to(device))\n",
    "elif LOSS == \"ce\":\n",
    "    criterion = nn.CrossEntropyLoss()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = optim.Adam(model.parameters(), lr=0.001)\n",
    "scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=30, gamma=0.1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "solver = Solver(\n",
    "    model,\n",
    "    ds_train,\n",
    "    ds_test,\n",
    "    criterion,\n",
    "    optimizer,\n",
    "    scheduler,\n",
    "    device,\n",
    "    batch_size=BATCH_SIZE,\n",
    "    num_classes=len(CLASS_IDS),\n",
    "    save_every=None,\n",
    "    path_to_save=None\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'loss': [0.2818208672106266,\n",
       "  0.2621240150183439,\n",
       "  0.26177744334563613,\n",
       "  0.24733486492186785,\n",
       "  0.241692706476897,\n",
       "  0.23505750391632318,\n",
       "  0.22735050227493048,\n",
       "  0.22488300129771233,\n",
       "  0.21820963500067592,\n",
       "  0.21645781164988875,\n",
       "  0.20285547198727727,\n",
       "  0.20425537624396384,\n",
       "  0.20168026420287788,\n",
       "  0.1989373799879104,\n",
       "  0.2010414833202958,\n",
       "  0.18811332620680332,\n",
       "  0.19126420421525836,\n",
       "  0.19797184481285512,\n",
       "  0.18722586776129901,\n",
       "  0.18821436818689108,\n",
       "  0.20225538220256567,\n",
       "  0.20623054122552276,\n",
       "  0.20597467268817127,\n",
       "  0.19790399144403636,\n",
       "  0.19810148351825774,\n",
       "  0.19358408777043223,\n",
       "  0.1972361346706748,\n",
       "  0.18818847695365548,\n",
       "  0.1831229105591774,\n",
       "  0.18498404417186975,\n",
       "  0.18154895189218223,\n",
       "  0.18109870632179081,\n",
       "  0.18079688609577715,\n",
       "  0.18464053864590824,\n",
       "  0.1803507290314883,\n",
       "  0.18079466512426734,\n",
       "  0.17867135349661112,\n",
       "  0.17792699253186584,\n",
       "  0.18248162884265184,\n",
       "  0.17858957848511636],\n",
       " 'train_accuracy': [0.5328185558319092,\n",
       "  0.5984556078910828,\n",
       "  0.6447876691818237,\n",
       "  0.6061776280403137,\n",
       "  0.7065637111663818,\n",
       "  0.7027027010917664,\n",
       "  0.6911196708679199,\n",
       "  0.7528957724571228,\n",
       "  0.8494208455085754,\n",
       "  0.8571428656578064,\n",
       "  0.9111968874931335,\n",
       "  0.8957529067993164,\n",
       "  0.8571428656578064,\n",
       "  0.837837815284729,\n",
       "  0.8494208455085754,\n",
       "  0.9111968874931335,\n",
       "  0.9111968874931335,\n",
       "  0.876447856426239,\n",
       "  0.915057897567749,\n",
       "  0.8996139168739319,\n",
       "  0.9382239580154419,\n",
       "  0.8687258958816528,\n",
       "  0.9034749269485474,\n",
       "  0.92277991771698,\n",
       "  0.9343629479408264,\n",
       "  0.915057897567749,\n",
       "  0.9305019378662109,\n",
       "  0.9575289487838745,\n",
       "  0.9420849680900574,\n",
       "  0.96138995885849,\n",
       "  0.9768339991569519,\n",
       "  0.9768339991569519,\n",
       "  0.9884169697761536,\n",
       "  0.9768339991569519,\n",
       "  0.9806950092315674,\n",
       "  0.969111979007721,\n",
       "  0.969111979007721,\n",
       "  0.9806950092315674,\n",
       "  0.9845559597015381,\n",
       "  0.9845559597015381],\n",
       " 'train_precision': [0.5328185558319092,\n",
       "  0.5984556078910828,\n",
       "  0.6447876691818237,\n",
       "  0.6061776280403137,\n",
       "  0.7065637111663818,\n",
       "  0.7027027010917664,\n",
       "  0.6911196708679199,\n",
       "  0.7528957724571228,\n",
       "  0.8494208455085754,\n",
       "  0.8571428656578064,\n",
       "  0.9111968874931335,\n",
       "  0.8957529067993164,\n",
       "  0.8571428656578064,\n",
       "  0.837837815284729,\n",
       "  0.8494208455085754,\n",
       "  0.9111968874931335,\n",
       "  0.9111968874931335,\n",
       "  0.876447856426239,\n",
       "  0.915057897567749,\n",
       "  0.8996139168739319,\n",
       "  0.9382239580154419,\n",
       "  0.8687258958816528,\n",
       "  0.9034749269485474,\n",
       "  0.92277991771698,\n",
       "  0.9343629479408264,\n",
       "  0.915057897567749,\n",
       "  0.9305019378662109,\n",
       "  0.9575289487838745,\n",
       "  0.9420849680900574,\n",
       "  0.96138995885849,\n",
       "  0.9768339991569519,\n",
       "  0.9768339991569519,\n",
       "  0.9884169697761536,\n",
       "  0.9768339991569519,\n",
       "  0.9806950092315674,\n",
       "  0.969111979007721,\n",
       "  0.969111979007721,\n",
       "  0.9806950092315674,\n",
       "  0.9845559597015381,\n",
       "  0.9845559597015381],\n",
       " 'train_recall': [0.5328185558319092,\n",
       "  0.5984556078910828,\n",
       "  0.6447876691818237,\n",
       "  0.6061776280403137,\n",
       "  0.7065637111663818,\n",
       "  0.7027027010917664,\n",
       "  0.6911196708679199,\n",
       "  0.7528957724571228,\n",
       "  0.8494208455085754,\n",
       "  0.8571428656578064,\n",
       "  0.9111968874931335,\n",
       "  0.8957529067993164,\n",
       "  0.8571428656578064,\n",
       "  0.837837815284729,\n",
       "  0.8494208455085754,\n",
       "  0.9111968874931335,\n",
       "  0.9111968874931335,\n",
       "  0.876447856426239,\n",
       "  0.915057897567749,\n",
       "  0.8996139168739319,\n",
       "  0.9382239580154419,\n",
       "  0.8687258958816528,\n",
       "  0.9034749269485474,\n",
       "  0.92277991771698,\n",
       "  0.9343629479408264,\n",
       "  0.915057897567749,\n",
       "  0.9305019378662109,\n",
       "  0.9575289487838745,\n",
       "  0.9420849680900574,\n",
       "  0.96138995885849,\n",
       "  0.9768339991569519,\n",
       "  0.9768339991569519,\n",
       "  0.9884169697761536,\n",
       "  0.9768339991569519,\n",
       "  0.9806950092315674,\n",
       "  0.969111979007721,\n",
       "  0.969111979007721,\n",
       "  0.9806950092315674,\n",
       "  0.9845559597015381,\n",
       "  0.9845559597015381],\n",
       " 'train_f_beta_score': [0.5328185558319092,\n",
       "  0.5984556078910828,\n",
       "  0.6447876691818237,\n",
       "  0.6061776280403137,\n",
       "  0.7065637111663818,\n",
       "  0.7027027010917664,\n",
       "  0.6911196708679199,\n",
       "  0.7528957724571228,\n",
       "  0.8494208455085754,\n",
       "  0.8571428656578064,\n",
       "  0.9111968874931335,\n",
       "  0.8957529067993164,\n",
       "  0.8571428656578064,\n",
       "  0.837837815284729,\n",
       "  0.8494208455085754,\n",
       "  0.9111968874931335,\n",
       "  0.9111968874931335,\n",
       "  0.876447856426239,\n",
       "  0.915057897567749,\n",
       "  0.8996139168739319,\n",
       "  0.9382239580154419,\n",
       "  0.8687258958816528,\n",
       "  0.9034749269485474,\n",
       "  0.92277991771698,\n",
       "  0.9343629479408264,\n",
       "  0.915057897567749,\n",
       "  0.9305019378662109,\n",
       "  0.9575289487838745,\n",
       "  0.9420849680900574,\n",
       "  0.96138995885849,\n",
       "  0.9768339991569519,\n",
       "  0.9768339991569519,\n",
       "  0.9884169697761536,\n",
       "  0.9768339991569519,\n",
       "  0.9806950092315674,\n",
       "  0.969111979007721,\n",
       "  0.969111979007721,\n",
       "  0.9806950092315674,\n",
       "  0.9845559597015381,\n",
       "  0.9845559597015381],\n",
       " 'train_confusion_matrix': [tensor([[63,  1, 21,  0,  1],\n",
       "          [ 6,  4, 61,  0,  2],\n",
       "          [ 3,  0, 67,  0,  2],\n",
       "          [ 2,  1,  8,  0,  2],\n",
       "          [ 2,  0,  9,  0,  4]]),\n",
       "  tensor([[74,  8,  4,  0,  0],\n",
       "          [ 6, 46, 17,  0,  4],\n",
       "          [ 3, 36, 27,  0,  6],\n",
       "          [ 1,  6,  0,  0,  6],\n",
       "          [ 1,  5,  1,  0,  8]]),\n",
       "  tensor([[72,  0,  4,  9,  1],\n",
       "          [ 2, 37, 18,  7,  9],\n",
       "          [ 3, 18, 38,  3, 10],\n",
       "          [ 0,  2,  0,  9,  2],\n",
       "          [ 0,  1,  1,  2, 11]]),\n",
       "  tensor([[67,  2,  1, 12,  4],\n",
       "          [ 0, 33, 10, 25,  5],\n",
       "          [ 2,  9, 34, 21,  6],\n",
       "          [ 0,  0,  1, 12,  0],\n",
       "          [ 0,  1,  0,  3, 11]]),\n",
       "  tensor([[76,  0,  2,  0,  8],\n",
       "          [13, 35, 23,  0,  2],\n",
       "          [ 8,  9, 51,  0,  4],\n",
       "          [ 0,  0,  0,  7,  6],\n",
       "          [ 1,  0,  0,  0, 14]]),\n",
       "  tensor([[81,  0,  1,  0,  4],\n",
       "          [ 4, 36, 22,  9,  2],\n",
       "          [ 2,  5, 42, 10, 13],\n",
       "          [ 2,  0,  0, 11,  0],\n",
       "          [ 1,  0,  0,  2, 12]]),\n",
       "  tensor([[77,  3,  0,  4,  2],\n",
       "          [ 1, 43, 13, 15,  1],\n",
       "          [ 3, 11, 36, 17,  5],\n",
       "          [ 1,  0,  0, 12,  0],\n",
       "          [ 0,  0,  1,  3, 11]]),\n",
       "  tensor([[71,  2,  0,  4,  9],\n",
       "          [ 1, 57,  4,  9,  2],\n",
       "          [ 0, 19, 41,  8,  4],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  2, 13]]),\n",
       "  tensor([[79,  1,  0,  3,  3],\n",
       "          [ 2, 56, 11,  2,  2],\n",
       "          [ 1,  8, 59,  1,  3],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 1,  0,  1,  0, 13]]),\n",
       "  tensor([[82,  1,  1,  0,  2],\n",
       "          [ 0, 52, 21,  0,  0],\n",
       "          [ 1,  6, 63,  0,  2],\n",
       "          [ 0,  1,  1, 10,  1],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[84,  0,  1,  1,  0],\n",
       "          [ 0, 68,  5,  0,  0],\n",
       "          [ 1, 11, 58,  0,  2],\n",
       "          [ 1,  0,  0, 11,  1],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[81,  0,  2,  2,  1],\n",
       "          [ 0, 64,  6,  1,  2],\n",
       "          [ 0,  7, 61,  0,  4],\n",
       "          [ 1,  0,  0, 11,  1],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[79,  3,  0,  4,  0],\n",
       "          [ 0, 72,  0,  1,  0],\n",
       "          [ 1, 16, 47,  3,  5],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  4, 11]]),\n",
       "  tensor([[68,  3,  0,  0, 15],\n",
       "          [ 0, 69,  2,  0,  2],\n",
       "          [ 0, 10, 55,  0,  7],\n",
       "          [ 0,  0,  0, 10,  3],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[82,  0,  0,  1,  3],\n",
       "          [ 3, 62,  4,  2,  2],\n",
       "          [ 1, 16, 50,  0,  5],\n",
       "          [ 0,  0,  0, 11,  2],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[83,  0,  2,  1,  0],\n",
       "          [ 0, 69,  4,  0,  0],\n",
       "          [ 1, 11, 58,  1,  1],\n",
       "          [ 0,  0,  1, 12,  0],\n",
       "          [ 0,  1,  0,  0, 14]]),\n",
       "  tensor([[84,  0,  1,  0,  1],\n",
       "          [ 0, 68,  2,  1,  2],\n",
       "          [ 1,  7, 56,  2,  6],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[84,  1,  1,  0,  0],\n",
       "          [ 1, 62,  5,  5,  0],\n",
       "          [ 6,  7, 56,  2,  1],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  3, 12]]),\n",
       "  tensor([[85,  0,  1,  0,  0],\n",
       "          [ 0, 63,  9,  0,  1],\n",
       "          [ 1,  3, 62,  0,  6],\n",
       "          [ 0,  0,  0, 12,  1],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[84,  0,  0,  0,  2],\n",
       "          [ 1, 63,  3,  3,  3],\n",
       "          [ 1, 11, 58,  0,  2],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[82,  3,  0,  0,  1],\n",
       "          [ 0, 72,  1,  0,  0],\n",
       "          [ 1,  7, 61,  0,  3],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[67,  3,  0,  0, 16],\n",
       "          [ 3, 68,  2,  0,  0],\n",
       "          [ 0,  3, 65,  1,  3],\n",
       "          [ 0,  0,  0, 11,  2],\n",
       "          [ 0,  0,  0,  1, 14]]),\n",
       "  tensor([[78,  3,  0,  0,  5],\n",
       "          [ 0, 73,  0,  0,  0],\n",
       "          [ 2, 11, 57,  1,  1],\n",
       "          [ 0,  0,  0, 12,  1],\n",
       "          [ 0,  1,  0,  0, 14]]),\n",
       "  tensor([[83,  3,  0,  0,  0],\n",
       "          [ 0, 68,  5,  0,  0],\n",
       "          [ 2,  3, 62,  4,  1],\n",
       "          [ 0,  0,  0, 11,  2],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[82,  2,  1,  1,  0],\n",
       "          [ 0, 66,  7,  0,  0],\n",
       "          [ 0,  2, 68,  0,  2],\n",
       "          [ 0,  0,  0, 12,  1],\n",
       "          [ 0,  0,  0,  1, 14]]),\n",
       "  tensor([[81,  2,  0,  1,  2],\n",
       "          [ 0, 68,  5,  0,  0],\n",
       "          [ 0,  6, 61,  1,  4],\n",
       "          [ 0,  0,  0, 12,  1],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[84,  2,  0,  0,  0],\n",
       "          [ 0, 65,  6,  1,  1],\n",
       "          [ 1,  3, 65,  0,  3],\n",
       "          [ 0,  0,  0, 12,  1],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[83,  1,  0,  0,  2],\n",
       "          [ 0, 72,  1,  0,  0],\n",
       "          [ 0,  5, 65,  0,  2],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[81,  0,  1,  0,  4],\n",
       "          [ 0, 73,  0,  0,  0],\n",
       "          [ 1,  5, 63,  0,  3],\n",
       "          [ 0,  0,  0, 12,  1],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[86,  0,  0,  0,  0],\n",
       "          [ 0, 73,  0,  0,  0],\n",
       "          [ 1,  8, 62,  0,  1],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[86,  0,  0,  0,  0],\n",
       "          [ 0, 73,  0,  0,  0],\n",
       "          [ 0,  3, 66,  0,  3],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[86,  0,  0,  0,  0],\n",
       "          [ 0, 72,  1,  0,  0],\n",
       "          [ 0,  3, 67,  0,  2],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[86,  0,  0,  0,  0],\n",
       "          [ 0, 73,  0,  0,  0],\n",
       "          [ 0,  1, 69,  0,  2],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[85,  1,  0,  0,  0],\n",
       "          [ 0, 73,  0,  0,  0],\n",
       "          [ 0,  3, 67,  0,  2],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[86,  0,  0,  0,  0],\n",
       "          [ 0, 73,  0,  0,  0],\n",
       "          [ 1,  3, 67,  0,  1],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[84,  1,  0,  1,  0],\n",
       "          [ 0, 72,  0,  1,  0],\n",
       "          [ 1,  2, 67,  1,  1],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[85,  0,  0,  1,  0],\n",
       "          [ 0, 72,  0,  1,  0],\n",
       "          [ 0,  3, 66,  0,  3],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[86,  0,  0,  0,  0],\n",
       "          [ 0, 73,  0,  0,  0],\n",
       "          [ 1,  2, 67,  0,  2],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[86,  0,  0,  0,  0],\n",
       "          [ 0, 73,  0,  0,  0],\n",
       "          [ 0,  2, 68,  0,  2],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  0, 15]]),\n",
       "  tensor([[86,  0,  0,  0,  0],\n",
       "          [ 0, 73,  0,  0,  0],\n",
       "          [ 0,  2, 68,  0,  2],\n",
       "          [ 0,  0,  0, 13,  0],\n",
       "          [ 0,  0,  0,  0, 15]])],\n",
       " 'val_accuracy': [0.4137931168079376,\n",
       "  0.3103448152542114,\n",
       "  0.3448275923728943,\n",
       "  0.24137930572032928,\n",
       "  0.37931033968925476,\n",
       "  0.37931033968925476,\n",
       "  0.4482758641242981,\n",
       "  0.48275861144065857,\n",
       "  0.37931033968925476,\n",
       "  0.48275861144065857,\n",
       "  0.5517241358757019,\n",
       "  0.37931033968925476,\n",
       "  0.4482758641242981,\n",
       "  0.3103448152542114,\n",
       "  0.4137931168079376,\n",
       "  0.5517241358757019,\n",
       "  0.5517241358757019,\n",
       "  0.5517241358757019,\n",
       "  0.48275861144065857,\n",
       "  0.48275861144065857,\n",
       "  0.4482758641242981,\n",
       "  0.37931033968925476,\n",
       "  0.5517241358757019,\n",
       "  0.6896551847457886,\n",
       "  0.517241358757019,\n",
       "  0.4482758641242981,\n",
       "  0.4482758641242981,\n",
       "  0.5517241358757019,\n",
       "  0.6206896305084229,\n",
       "  0.6206896305084229,\n",
       "  0.5517241358757019,\n",
       "  0.5517241358757019,\n",
       "  0.6206896305084229,\n",
       "  0.5862069129943848,\n",
       "  0.5862069129943848,\n",
       "  0.6206896305084229,\n",
       "  0.5517241358757019,\n",
       "  0.5862069129943848,\n",
       "  0.6206896305084229,\n",
       "  0.6206896305084229],\n",
       " 'val_precision': [0.4137931168079376,\n",
       "  0.3103448152542114,\n",
       "  0.3448275923728943,\n",
       "  0.24137930572032928,\n",
       "  0.37931033968925476,\n",
       "  0.37931033968925476,\n",
       "  0.4482758641242981,\n",
       "  0.48275861144065857,\n",
       "  0.37931033968925476,\n",
       "  0.48275861144065857,\n",
       "  0.5517241358757019,\n",
       "  0.37931033968925476,\n",
       "  0.4482758641242981,\n",
       "  0.3103448152542114,\n",
       "  0.4137931168079376,\n",
       "  0.5517241358757019,\n",
       "  0.5517241358757019,\n",
       "  0.5517241358757019,\n",
       "  0.48275861144065857,\n",
       "  0.48275861144065857,\n",
       "  0.4482758641242981,\n",
       "  0.37931033968925476,\n",
       "  0.5517241358757019,\n",
       "  0.6896551847457886,\n",
       "  0.517241358757019,\n",
       "  0.4482758641242981,\n",
       "  0.4482758641242981,\n",
       "  0.5517241358757019,\n",
       "  0.6206896305084229,\n",
       "  0.6206896305084229,\n",
       "  0.5517241358757019,\n",
       "  0.5517241358757019,\n",
       "  0.6206896305084229,\n",
       "  0.5862069129943848,\n",
       "  0.5862069129943848,\n",
       "  0.6206896305084229,\n",
       "  0.5517241358757019,\n",
       "  0.5862069129943848,\n",
       "  0.6206896305084229,\n",
       "  0.6206896305084229],\n",
       " 'val_recall': [0.4137931168079376,\n",
       "  0.3103448152542114,\n",
       "  0.3448275923728943,\n",
       "  0.24137930572032928,\n",
       "  0.37931033968925476,\n",
       "  0.37931033968925476,\n",
       "  0.4482758641242981,\n",
       "  0.48275861144065857,\n",
       "  0.37931033968925476,\n",
       "  0.48275861144065857,\n",
       "  0.5517241358757019,\n",
       "  0.37931033968925476,\n",
       "  0.4482758641242981,\n",
       "  0.3103448152542114,\n",
       "  0.4137931168079376,\n",
       "  0.5517241358757019,\n",
       "  0.5517241358757019,\n",
       "  0.5517241358757019,\n",
       "  0.48275861144065857,\n",
       "  0.48275861144065857,\n",
       "  0.4482758641242981,\n",
       "  0.37931033968925476,\n",
       "  0.5517241358757019,\n",
       "  0.6896551847457886,\n",
       "  0.517241358757019,\n",
       "  0.4482758641242981,\n",
       "  0.4482758641242981,\n",
       "  0.5517241358757019,\n",
       "  0.6206896305084229,\n",
       "  0.6206896305084229,\n",
       "  0.5517241358757019,\n",
       "  0.5517241358757019,\n",
       "  0.6206896305084229,\n",
       "  0.5862069129943848,\n",
       "  0.5862069129943848,\n",
       "  0.6206896305084229,\n",
       "  0.5517241358757019,\n",
       "  0.5862069129943848,\n",
       "  0.6206896305084229,\n",
       "  0.6206896305084229],\n",
       " 'val_f_beta_score': [0.4137931168079376,\n",
       "  0.3103448152542114,\n",
       "  0.3448275923728943,\n",
       "  0.24137930572032928,\n",
       "  0.37931033968925476,\n",
       "  0.37931033968925476,\n",
       "  0.4482758641242981,\n",
       "  0.48275861144065857,\n",
       "  0.37931033968925476,\n",
       "  0.48275861144065857,\n",
       "  0.5517241358757019,\n",
       "  0.37931033968925476,\n",
       "  0.4482758641242981,\n",
       "  0.3103448152542114,\n",
       "  0.4137931168079376,\n",
       "  0.5517241358757019,\n",
       "  0.5517241358757019,\n",
       "  0.5517241358757019,\n",
       "  0.48275861144065857,\n",
       "  0.48275861144065857,\n",
       "  0.4482758641242981,\n",
       "  0.37931033968925476,\n",
       "  0.5517241358757019,\n",
       "  0.6896551847457886,\n",
       "  0.517241358757019,\n",
       "  0.4482758641242981,\n",
       "  0.4482758641242981,\n",
       "  0.5517241358757019,\n",
       "  0.6206896305084229,\n",
       "  0.6206896305084229,\n",
       "  0.5517241358757019,\n",
       "  0.5517241358757019,\n",
       "  0.6206896305084229,\n",
       "  0.5862069129943848,\n",
       "  0.5862069129943848,\n",
       "  0.6206896305084229,\n",
       "  0.5517241358757019,\n",
       "  0.5862069129943848,\n",
       "  0.6206896305084229,\n",
       "  0.6206896305084229],\n",
       " 'val_confusion_matrix': [tensor([[4, 0, 5, 0, 1],\n",
       "          [0, 0, 7, 0, 0],\n",
       "          [0, 0, 8, 0, 0],\n",
       "          [1, 0, 2, 0, 0],\n",
       "          [0, 0, 1, 0, 0]]),\n",
       "  tensor([[4, 4, 1, 0, 1],\n",
       "          [0, 3, 2, 0, 2],\n",
       "          [1, 3, 2, 0, 2],\n",
       "          [0, 2, 0, 0, 1],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[5, 2, 1, 2, 0],\n",
       "          [0, 2, 2, 1, 2],\n",
       "          [0, 3, 2, 1, 2],\n",
       "          [0, 0, 0, 1, 2],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[4, 0, 1, 4, 1],\n",
       "          [0, 1, 1, 4, 1],\n",
       "          [0, 1, 1, 5, 1],\n",
       "          [0, 0, 1, 1, 1],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[5, 0, 0, 0, 5],\n",
       "          [1, 1, 5, 0, 0],\n",
       "          [2, 0, 4, 0, 2],\n",
       "          [0, 0, 0, 1, 2],\n",
       "          [0, 0, 1, 0, 0]]),\n",
       "  tensor([[7, 0, 1, 1, 1],\n",
       "          [0, 1, 3, 2, 1],\n",
       "          [0, 2, 2, 1, 3],\n",
       "          [0, 0, 0, 1, 2],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[7, 0, 1, 1, 1],\n",
       "          [0, 2, 2, 2, 1],\n",
       "          [1, 2, 3, 2, 0],\n",
       "          [1, 0, 0, 1, 1],\n",
       "          [0, 0, 0, 1, 0]]),\n",
       "  tensor([[3, 1, 1, 1, 4],\n",
       "          [0, 6, 0, 1, 0],\n",
       "          [1, 1, 4, 1, 1],\n",
       "          [0, 0, 1, 1, 1],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[4, 0, 1, 3, 2],\n",
       "          [1, 2, 3, 0, 1],\n",
       "          [1, 1, 3, 2, 1],\n",
       "          [0, 1, 0, 2, 0],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[6, 0, 1, 0, 3],\n",
       "          [0, 4, 3, 0, 0],\n",
       "          [0, 2, 4, 1, 1],\n",
       "          [0, 1, 0, 0, 2],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[6, 2, 0, 1, 1],\n",
       "          [0, 6, 1, 0, 0],\n",
       "          [0, 3, 4, 1, 0],\n",
       "          [0, 0, 1, 0, 2],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[6, 1, 1, 0, 2],\n",
       "          [0, 2, 3, 1, 1],\n",
       "          [0, 2, 3, 1, 2],\n",
       "          [0, 0, 1, 0, 2],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[6, 2, 0, 2, 0],\n",
       "          [0, 5, 1, 0, 1],\n",
       "          [0, 3, 1, 2, 2],\n",
       "          [0, 1, 0, 1, 1],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[2, 2, 0, 0, 6],\n",
       "          [0, 5, 1, 1, 0],\n",
       "          [0, 3, 2, 1, 2],\n",
       "          [0, 1, 0, 0, 2],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[3, 1, 0, 1, 5],\n",
       "          [0, 4, 2, 1, 0],\n",
       "          [0, 2, 5, 0, 1],\n",
       "          [0, 0, 1, 0, 2],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[7, 1, 0, 1, 1],\n",
       "          [0, 5, 2, 0, 0],\n",
       "          [0, 4, 4, 0, 0],\n",
       "          [1, 1, 0, 0, 1],\n",
       "          [0, 0, 1, 0, 0]]),\n",
       "  tensor([[5, 1, 1, 1, 2],\n",
       "          [0, 6, 0, 1, 0],\n",
       "          [0, 2, 5, 1, 0],\n",
       "          [1, 0, 1, 0, 1],\n",
       "          [0, 0, 1, 0, 0]]),\n",
       "  tensor([[6, 1, 0, 3, 0],\n",
       "          [0, 4, 1, 2, 0],\n",
       "          [1, 1, 4, 2, 0],\n",
       "          [0, 1, 0, 2, 0],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[6, 2, 0, 0, 2],\n",
       "          [1, 4, 2, 0, 0],\n",
       "          [1, 2, 3, 0, 2],\n",
       "          [0, 1, 0, 1, 1],\n",
       "          [0, 0, 1, 0, 0]]),\n",
       "  tensor([[4, 2, 0, 3, 1],\n",
       "          [0, 5, 2, 0, 0],\n",
       "          [0, 3, 4, 0, 1],\n",
       "          [0, 1, 0, 1, 1],\n",
       "          [0, 0, 0, 1, 0]]),\n",
       "  tensor([[4, 2, 0, 2, 2],\n",
       "          [0, 5, 2, 0, 0],\n",
       "          [0, 5, 2, 1, 0],\n",
       "          [0, 1, 0, 2, 0],\n",
       "          [0, 0, 1, 0, 0]]),\n",
       "  tensor([[2, 2, 0, 1, 5],\n",
       "          [0, 5, 1, 1, 0],\n",
       "          [1, 2, 3, 1, 1],\n",
       "          [0, 1, 0, 1, 1],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[5, 2, 0, 0, 3],\n",
       "          [0, 7, 0, 0, 0],\n",
       "          [1, 3, 3, 1, 0],\n",
       "          [1, 1, 0, 1, 0],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[7, 1, 0, 0, 2],\n",
       "          [0, 6, 1, 0, 0],\n",
       "          [0, 1, 7, 0, 0],\n",
       "          [1, 0, 1, 0, 1],\n",
       "          [0, 0, 0, 1, 0]]),\n",
       "  tensor([[6, 0, 1, 1, 2],\n",
       "          [0, 3, 4, 0, 0],\n",
       "          [1, 1, 6, 0, 0],\n",
       "          [2, 0, 1, 0, 0],\n",
       "          [0, 0, 0, 1, 0]]),\n",
       "  tensor([[6, 0, 1, 1, 2],\n",
       "          [1, 5, 1, 0, 0],\n",
       "          [1, 3, 2, 0, 2],\n",
       "          [1, 0, 0, 0, 2],\n",
       "          [0, 0, 0, 1, 0]]),\n",
       "  tensor([[5, 1, 1, 1, 2],\n",
       "          [0, 5, 2, 0, 0],\n",
       "          [0, 2, 3, 1, 2],\n",
       "          [1, 1, 0, 0, 1],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[5, 1, 1, 1, 2],\n",
       "          [0, 6, 0, 1, 0],\n",
       "          [1, 2, 4, 1, 0],\n",
       "          [0, 1, 0, 1, 1],\n",
       "          [0, 0, 0, 1, 0]]),\n",
       "  tensor([[7, 1, 0, 0, 2],\n",
       "          [0, 6, 1, 0, 0],\n",
       "          [0, 2, 5, 1, 0],\n",
       "          [1, 0, 1, 0, 1],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[6, 1, 0, 1, 2],\n",
       "          [0, 6, 1, 0, 0],\n",
       "          [0, 2, 5, 1, 0],\n",
       "          [1, 1, 0, 1, 0],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[6, 0, 1, 1, 2],\n",
       "          [0, 6, 1, 0, 0],\n",
       "          [0, 3, 4, 1, 0],\n",
       "          [1, 1, 0, 0, 1],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[6, 0, 1, 1, 2],\n",
       "          [0, 6, 1, 0, 0],\n",
       "          [0, 3, 4, 1, 0],\n",
       "          [1, 1, 0, 0, 1],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[6, 0, 1, 1, 2],\n",
       "          [0, 7, 0, 0, 0],\n",
       "          [0, 3, 4, 1, 0],\n",
       "          [1, 1, 0, 1, 0],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[6, 0, 1, 1, 2],\n",
       "          [0, 6, 1, 0, 0],\n",
       "          [0, 3, 4, 1, 0],\n",
       "          [0, 1, 0, 1, 1],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[6, 0, 1, 1, 2],\n",
       "          [0, 7, 0, 0, 0],\n",
       "          [1, 2, 4, 1, 0],\n",
       "          [1, 1, 0, 0, 1],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[6, 0, 1, 1, 2],\n",
       "          [0, 7, 0, 0, 0],\n",
       "          [0, 3, 4, 1, 0],\n",
       "          [1, 1, 0, 1, 0],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[7, 0, 1, 1, 1],\n",
       "          [0, 5, 2, 0, 0],\n",
       "          [1, 2, 4, 1, 0],\n",
       "          [1, 1, 0, 0, 1],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[6, 0, 1, 1, 2],\n",
       "          [0, 7, 0, 0, 0],\n",
       "          [1, 2, 4, 1, 0],\n",
       "          [1, 1, 0, 0, 1],\n",
       "          [0, 1, 0, 0, 0]]),\n",
       "  tensor([[6, 0, 1, 1, 2],\n",
       "          [0, 7, 0, 0, 0],\n",
       "          [0, 3, 4, 1, 0],\n",
       "          [1, 1, 0, 1, 0],\n",
       "          [0, 0, 0, 1, 0]]),\n",
       "  tensor([[6, 0, 1, 1, 2],\n",
       "          [0, 7, 0, 0, 0],\n",
       "          [0, 3, 4, 1, 0],\n",
       "          [1, 1, 0, 1, 0],\n",
       "          [0, 0, 0, 1, 0]])]}"
      ]
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "num_epochs = 40\n",
    "results = solver.train(num_epochs)\n",
    "results"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.legend.Legend at 0x170142165e0>"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "start = 0\n",
    "\n",
    "plt.plot(results[\"loss\"][start:], label='Loss')\n",
    "plt.legend()\n",
    "\n",
    "plt.figure()\n",
    "plt.plot(results[\"train_accuracy\"][start:], label='Train Acc')\n",
    "plt.plot(results[\"val_accuracy\"][start:], label='Val Acc')\n",
    "plt.legend()\n",
    "\n",
    "# plt.figure()\n",
    "# plt.plot(results[\"train_recall\"][start:], label='Train Recall')\n",
    "# plt.plot(results[\"val_recall\"][start:], label='Val Recall')\n",
    "# plt.legend()\n",
    "\n",
    "# plt.figure()\n",
    "# plt.plot(results[\"train_precision\"][start:], label='Train Precision')\n",
    "# plt.plot(results[\"val_precision\"][start:], label='Val Precision')\n",
    "# plt.legend()\n",
    "\n",
    "# plt.figure()\n",
    "# plt.plot(results[\"train_f_beta_score\"][start:], label='Train F-Beta Score')\n",
    "# plt.plot(results[\"val_f_beta_score\"][start:], label='Val F-Beta Score')\n",
    "# plt.legend()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<Axes: >"
      ]
     },
     "execution_count": 28,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "import pandas as pd\n",
    "import seaborn as sn\n",
    "\n",
    "confusion_matrix = results[\"val_confusion_matrix\"][-1]\n",
    "\n",
    "df_cm = pd.DataFrame(confusion_matrix)\n",
    "sn.heatmap(df_cm, annot=True, fmt=\"g\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [],
   "source": [
    "name = []\n",
    "[name.append(\"-\".join(map(str, layer))) for layer in UNFREEZE_LAYERS]\n",
    "name.append(LOSS)\n",
    "if BALANCED:\n",
    "    name.append(\"balanced\")\n",
    "name = \",\".join(name)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "| 1-8,softmaxcb | \n",
      "| ------ | \n",
      "| 0.4828 | \n",
      "| 0.5517 | \n",
      "| 0.6897 | \n",
      "| 0.6897 | \n",
      "\n"
     ]
    }
   ],
   "source": [
    "accs = results[\"val_accuracy\"]\n",
    "step = 10\n",
    "print(f\"| {name} | \")\n",
    "print(\"| ------ | \")\n",
    "[print(f\"| {max(accs[:(i+1)*step]):.4f} | \") for i in range(len(accs)//step)]\n",
    "print();"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Best Val Accs after N Epochs\n",
    "\n",
    "1-8 = Unfreeze only Last Linear Layer of pretrained model\\\n",
    "weightedce = Weighted ce loss according to train dataset distribution\\\n",
    "bal = Balanced Train dataset\n",
    "\n",
    "| ep | 1-5,1-8,ce | 1-8,ce | ce | 1-8,weightedce | 1-8,ce,bal | 1-8,focal | 1-8,focalcb | 1-8,sigmoidcb | 1-8,softmaxcb | \n",
    "| -- | ------ | ------ | ------ | ------ | ------ | ------ | ------ | ------ | ------ | \n",
    "| 10 | 0.5517 | 0.6207 | 0.6207 | 0.6552 | 0.4583 | 0.6552 | 0.5172 | 0.4828 | 0.4828 | \n",
    "| 20 | 0.5517 | 0.6207 | 0.6207 | 0.6552 | 0.5417 | 0.6552 | 0.5862 | 0.6207 | 0.5517 | \n",
    "| 30 | 0.5517 | 0.6897 | 0.6897 | 0.6552 | 0.5417 | 0.6552 | 0.6207 | 0.6207 | 0.6897 | \n",
    "| 40 | 0.5517 | 0.6897 | 0.6897 | 0.6552 | 0.5417 | 0.6552 | 0.6207 | 0.6207 | 0.6897 |"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.9.18"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}