\n \n
Table of Contents:\n\n
\n
\n - 1. Install and Import Dependencies
\n - \n 2. Image Segmentation\n
\n - \n 3. Implement VGG16 from Scratch\n
\n - \n 4. Data Prepration and Training VGG16\n
\n - 5. Evaluating VGG16 and Predicting Labels
\n - 6. Preprocessing and Training ResNet50
\n - 7. Predicting Labels using ResNet50
\n - 8. Conclusion
\n
\n","metadata":{}},{"cell_type":"markdown","source":"
\n
\n 1. Install and Import Dependencies\n ¶\n
\n
","metadata":{}},{"cell_type":"code","source":"import numpy as np\nimport matplotlib.pyplot as plt\nimport pandas as pd\nfrom PIL import Image\nimport seaborn as sns\nsns.set_style('darkgrid')\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.metrics import classification_report, confusion_matrix\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\nfrom torch.utils.checkpoint import checkpoint\nfrom torch.utils.data import Dataset\nfrom torchvision import transforms, datasets, models\nimport torchvision\nfrom fastai.vision.all import * \nimport os\nfrom tempfile import TemporaryDirectory\nimport time\nimport pickle ","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:13.109128Z","iopub.execute_input":"2024-04-15T07:34:13.109601Z","iopub.status.idle":"2024-04-15T07:34:13.121353Z","shell.execute_reply.started":"2024-04-15T07:34:13.109566Z","shell.execute_reply":"2024-04-15T07:34:13.120176Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"paths = {\n 'source_train' : os.path.join('/kaggle', 'input', 'cifar-10', 'train.7z'),\n 'source_test' : os.path.join('/kaggle', 'input', 'cifar-10', 'test.7z'),\n 'dest' : os.path.join('/kaggle', 'working', 'cifar-10'),\n 'train' : os.path.join('/kaggle', 'working', 'cifar-10', 'train'),\n 'test' : os.path.join('/kaggle', 'working', 'cifar-10', 'test'),\n 'trainLabels' : os.path.join('/kaggle', 'input', 'cifar-10', 'trainLabels.csv')\n}","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:13.126947Z","iopub.execute_input":"2024-04-15T07:34:13.127231Z","iopub.status.idle":"2024-04-15T07:34:13.136481Z","shell.execute_reply.started":"2024-04-15T07:34:13.127206Z","shell.execute_reply":"2024-04-15T07:34:13.135676Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"!pip install py7zr","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:13.138027Z","iopub.execute_input":"2024-04-15T07:34:13.138368Z","iopub.status.idle":"2024-04-15T07:34:14.982515Z","shell.execute_reply.started":"2024-04-15T07:34:13.138337Z","shell.execute_reply":"2024-04-15T07:34:14.981225Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import py7zr","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:14.984634Z","iopub.execute_input":"2024-04-15T07:34:14.985652Z","iopub.status.idle":"2024-04-15T07:34:14.990332Z","shell.execute_reply.started":"2024-04-15T07:34:14.985603Z","shell.execute_reply":"2024-04-15T07:34:14.989325Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def extract_7z_file(archive_path, destination_path):\n with py7zr.SevenZipFile(archive_path, mode='r') as z:\n z.extractall(path=destination_path)","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:14.993045Z","iopub.execute_input":"2024-04-15T07:34:14.993355Z","iopub.status.idle":"2024-04-15T07:34:15.000166Z","shell.execute_reply.started":"2024-04-15T07:34:14.993324Z","shell.execute_reply":"2024-04-15T07:34:14.999279Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"extract_7z_file(paths['source_train'], paths['dest'])\nextract_7z_file(paths['source_test'], paths['dest'])","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:31.351913Z","iopub.execute_input":"2024-04-15T07:34:31.352813Z","iopub.status.idle":"2024-04-15T07:38:15.521771Z","shell.execute_reply.started":"2024-04-15T07:34:31.352768Z","shell.execute_reply":"2024-04-15T07:38:15.520699Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"
\n
\n 2. Image Segmentation\n ¶\n
\n
","metadata":{}},{"cell_type":"code","source":"train_images = os.listdir(paths['train'])\ntrain_images[:10]","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:39:23.558594Z","iopub.execute_input":"2024-04-15T07:39:23.559736Z","iopub.status.idle":"2024-04-15T07:39:23.595227Z","shell.execute_reply.started":"2024-04-15T07:39:23.559693Z","shell.execute_reply":"2024-04-15T07:39:23.594350Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"y_train = pd.read_csv(paths['trainLabels'])\ny_train.head()","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:39:26.634488Z","iopub.execute_input":"2024-04-15T07:39:26.634869Z","iopub.status.idle":"2024-04-15T07:39:26.662409Z","shell.execute_reply.started":"2024-04-15T07:39:26.634838Z","shell.execute_reply":"2024-04-15T07:39:26.661409Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"labels = y_train['label'].unique()\nlabels","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:39:33.669756Z","iopub.execute_input":"2024-04-15T07:39:33.670130Z","iopub.status.idle":"2024-04-15T07:39:33.680400Z","shell.execute_reply.started":"2024-04-15T07:39:33.670101Z","shell.execute_reply":"2024-04-15T07:39:33.679375Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"for state in ['train', 'eval']:\n for label in labels:\n if os.path.exists(os.path.join(paths['dest'], 'train', state)) == False:\n os.mkdir(os.path.join(paths['dest'], 'train', state,))\n \n if os.path.exists(os.path.join(paths['dest'], 'train', state, label)) == False:\n os.mkdir(os.path.join(paths['dest'], 'train', state, label))","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:39:36.303687Z","iopub.execute_input":"2024-04-15T07:39:36.304059Z","iopub.status.idle":"2024-04-15T07:39:36.311623Z","shell.execute_reply.started":"2024-04-15T07:39:36.304030Z","shell.execute_reply":"2024-04-15T07:39:36.310664Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"li = []\nfor item in os.listdir(paths['train']):\n if '.png' not in item:\n li.append(item)\nli","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:39:39.589104Z","iopub.execute_input":"2024-04-15T07:39:39.589767Z","iopub.status.idle":"2024-04-15T07:39:39.634215Z","shell.execute_reply.started":"2024-04-15T07:39:39.589734Z","shell.execute_reply":"2024-04-15T07:39:39.633266Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"png_files = [item for item in os.listdir(paths['train']) if '.png' in item]\nnumbers = [int(os.path.splitext(png)[0]) for png in png_files if type(int(os.path.splitext(png)[0])) == int]\nnumbers.sort()","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:39:42.068410Z","iopub.execute_input":"2024-04-15T07:39:42.069349Z","iopub.status.idle":"2024-04-15T07:39:42.322322Z","shell.execute_reply.started":"2024-04-15T07:39:42.069312Z","shell.execute_reply":"2024-04-15T07:39:42.321181Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"X = y_train.copy()\ny = X['label'].values\nX.drop(['id', 'label'], axis=1, inplace=True)\nX['image name'] = [str(num) + '.' + 'png' for num in numbers]\nX.shape, y.shape","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:39:44.839586Z","iopub.execute_input":"2024-04-15T07:39:44.840480Z","iopub.status.idle":"2024-04-15T07:39:44.874984Z","shell.execute_reply.started":"2024-04-15T07:39:44.840446Z","shell.execute_reply":"2024-04-15T07:39:44.874121Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def make_image_grid(labels:np.ndarray, images:list):\n plt.figure(figsize=(10, 10))\n for i in range(32):\n plt.subplot(4, 8, i + 1)\n labeli = labels[i]\n imgi_arr = np.asarray(Image.open(os.path.join(paths['train'], images[i])))\n if imgi_arr.shape == (32, 32, 3):\n pass\n elif imgi_arr.shape == (3, 32, 32):\n imgi_arr = np.transpose(imgi_arr, (1,2,0))\n plt.imshow(imgi_arr)\n plt.title(labeli, size=9, alpha=.8)\n plt.axis('off')","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:39:48.163536Z","iopub.execute_input":"2024-04-15T07:39:48.164224Z","iopub.status.idle":"2024-04-15T07:39:48.171381Z","shell.execute_reply.started":"2024-04-15T07:39:48.164191Z","shell.execute_reply":"2024-04-15T07:39:48.170396Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"make_image_grid(y, X['image name'].values.tolist())","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:39:51.397439Z","iopub.execute_input":"2024-04-15T07:39:51.398081Z","iopub.status.idle":"2024-04-15T07:39:55.801665Z","shell.execute_reply.started":"2024-04-15T07:39:51.398045Z","shell.execute_reply":"2024-04-15T07:39:55.800710Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"X_train, X_eval, y_train, y_eval = train_test_split(X, y, random_state=42, train_size=.8)\nprint(X_train.shape, y_train.shape)\nprint(X_eval.shape, y_eval.shape)","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:39:59.814822Z","iopub.execute_input":"2024-04-15T07:39:59.815717Z","iopub.status.idle":"2024-04-15T07:39:59.832224Z","shell.execute_reply.started":"2024-04-15T07:39:59.815683Z","shell.execute_reply":"2024-04-15T07:39:59.831157Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"make_image_grid(y_train, X_train['image name'].values.tolist())","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:40:02.831323Z","iopub.execute_input":"2024-04-15T07:40:02.831681Z","iopub.status.idle":"2024-04-15T07:40:07.184618Z","shell.execute_reply.started":"2024-04-15T07:40:02.831654Z","shell.execute_reply":"2024-04-15T07:40:07.183689Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"make_image_grid(y_eval, X_eval['image name'].values.tolist())","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:40:10.996809Z","iopub.execute_input":"2024-04-15T07:40:10.997474Z","iopub.status.idle":"2024-04-15T07:40:15.368363Z","shell.execute_reply.started":"2024-04-15T07:40:10.997440Z","shell.execute_reply":"2024-04-15T07:40:15.367429Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def segment_images(source, state):\n states = {\n 'train' : pd.DataFrame({\n 'image name' : X_train['image name'].values.tolist(),\n 'label' : y_train\n }),\n 'eval' : pd.DataFrame({\n 'image name' : X_eval['image name'].values.tolist(),\n 'label' : y_eval\n })\n }\n \n for _, row in states[state].iterrows():\n os.rename(os.path.join(source, row['image name']), os.path.join(source, state, row['label'], row['image name']))\n\nsegment_images(paths['train'], 'train')\nsegment_images(paths['train'], 'eval')","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:40:18.883872Z","iopub.execute_input":"2024-04-15T07:40:18.884277Z","iopub.status.idle":"2024-04-15T07:40:25.364338Z","shell.execute_reply.started":"2024-04-15T07:40:18.884240Z","shell.execute_reply":"2024-04-15T07:40:25.363510Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"
\n
\n 3. Implement VGG16 from Scratch\n ¶\n
\n
","metadata":{}},{"cell_type":"code","source":"VGG16 = [64,64,\"M\",128,128,\"M\",256,256,256,\"M\",512,512,512,\"M\",512,512,512,\"M\"]\n\nclass VGGNet(nn.Module):\n def __init__(self, inchannel=3, numclass=10):\n super(VGGNet,self).__init__()\n \n self.inchannel = inchannel\n self.numclass = numclass\n self.conv_later = self.create_conv_layers(VGG16)\n self.fcs= nn.Sequential(\n nn.Linear(512*7*7,4096),\n nn.ReLU(),\n nn.Dropout(p=0.5),\n nn.Linear(4096,4096),\n nn.ReLU(),\n nn.Dropout(p=0.5),\n nn.Linear(4096,self.numclass),\n )\n \n \n def forward(self, x):\n x=self.conv_later(x)\n x= x.reshape(x.shape[0],-1)\n x=self.fcs(x)\n return x\n \n \n def create_conv_layers(self, architectire):\n layers= []\n inchannel = self.inchannel\n \n for x in architectire:\n if type(x) == int:\n outchannel = x\n layers+=[\n nn.Conv2d(\n in_channels = inchannel,\n out_channels=outchannel,\n kernel_size= (3,3),\n stride = (1,1),\n padding = (1,1),\n ),\n nn.BatchNorm2d(x),# Barch norm orginally indtroduce in 2016 however VGG introduced in 2014\n nn.ReLU(), \n ]\n inchannel = outchannel\n \n else:\n layers+=[nn.MaxPool2d(kernel_size=(2,2),stride=(2,2))]\n \n return nn.Sequential(*layers)","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.778234Z","iopub.status.idle":"2024-04-15T07:34:16.778581Z","shell.execute_reply.started":"2024-04-15T07:34:16.778412Z","shell.execute_reply":"2024-04-15T07:34:16.778429Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"
\n
\n 4. Data Prepration and Training VGG16\n ¶\n
\n
","metadata":{}},{"cell_type":"code","source":"def train(model, loss_func, optimizer, reduce_lr, num_epochs=10):\n start_time = time.time()\n \n with TemporaryDirectory() as temp_dir:\n dir_path = os.path.join(temp_dir, 'best_model_weights.pt')\n best_acc = 0\n best_loss = 0\n \n for epoch in range(num_epochs):\n print('epoch {0} started:\\n'.format(epoch))\n print('-' * 30)\n \n for phase in ['train', 'eval']:\n if phase == 'train':\n model.train()\n \n else:\n model.eval()\n \n running_loss = 0\n running_acc = 0\n \n for inputs, labels in dataLoader[phase]:\n inputs = inputs.to(device)\n labels = labels.to(device)\n optimizer.zero_grad()\n \n with torch.set_grad_enabled(mode= (phase=='train') ):\n # Forward\n outputs = model(inputs)\n loss = loss_func(outputs, labels)\n _, preds = torch.max(outputs, 1)\n \n # Backward\n if phase == 'train':\n loss.backward()\n optimizer.step() \n \n running_loss += loss.item() * inputs.size(0)\n running_acc += torch.sum(preds == labels.data)\n \n if phase == 'train':\n reduce_lr.step()\n \n epoch_loss = running_loss / dataset_size[phase]\n epoch_acc = running_acc / dataset_size[phase]\n print(f'Phase {phase}:\\n Loss : {epoch_loss}\\n Accuracy : {epoch_acc}')\n \n if phase == 'eval' and epoch_loss < best_loss:\n best_acc = epoch_acc\n best_loss = epoch_loss\n torch.save(model.state_dict(), dir_path)\n print()\n \n print(f\"Training is completed in {time.time() - start_time}\\n\")\n# print(f\"The best model:\\n Loss : {best_loss}\\n Accuracy : {best_acc}\")\n# model.load_state_dict(torch.load(dir_path))\n\n return model\n\n\nclass CustomDataset(Dataset):\n def __init__(self, root_dir, transform=None):\n self.root_dir = root_dir\n self.image_files = os.listdir(root_dir)\n self.transform = transform\n\n def __len__(self):\n return len(self.image_files)\n\n def __getitem__(self, idx):\n image_path = os.path.join(self.root_dir, self.image_files[idx])\n image = Image.open(image_path)\n\n if self.transform is not None:\n image = self.transform(image)\n\n return image, []\n\n \n# using gradient checkpointing techniques, we can checkpoint specific parts of your computation graph and reduce the overall memory consumption.\ndef forward_pass(model, inputs):\n return model(inputs)\n\n\ndef predict(model, state:str):\n states = {\n 'test': paths['test'],\n 'eval': os.path.join(paths['train'], 'eval')\n }\n device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n model.to(device)\n optimizer = optim.SGD(model.parameters(), lr=.001, momentum=.9)\n \n def implementation(dataLoader):\n actual_labels = []\n predicted_labels = []\n for inputs, labels in dataLoader:\n inputs = inputs.to(device)\n if labels != []:\n labels = labels.to(device)\n actual_labels.extend(labels)\n \n model.eval()\n with torch.no_grad(): # avoid storing gradients\n # Apply checkpointing to the forward pass\n outputs = checkpoint(forward_pass, model, inputs)\n _, preds = torch.max(outputs, 1)\n predicted_labels.extend(preds)\n \n return predicted_labels, actual_labels\n\n if state.lower() == 'test':\n data_transforms = transforms.Compose([\n transforms.Resize(227),\n transforms.ToTensor(),\n transforms.Normalize(mean=[.485, .456, .406], std=[.229, .224, .225])\n ])\n dataset = CustomDataset(states[state], transform=data_transforms)\n dataLoader_test = torch.utils.data.DataLoader(dataset=dataset, num_workers=4, batch_size=64, shuffle=True)\n predicted_labels, actual_labels = implementation(dataLoader_test)\n return predicted_labels, actual_labels\n \n elif state.lower() == 'eval':\n data_transform = transforms.Compose([\n transforms.Resize(227),\n transforms.ToTensor(),\n transforms.Normalize(mean=[.485, .456, .406], std=[.229, .224, .225])\n ])\n dataset = datasets.ImageFolder(root=os.path.join(states[state]), transform=data_transform)\n dataLoader = torch.utils.data.DataLoader(dataset=dataset, num_workers=4, batch_size=64, shuffle=True)\n predicted_labels, actual_labels = implementation(dataLoader)\n return predicted_labels, actual_labels\n \n else:\n raise ValueError(f\"Expected test or eval, not {state} for the state parameter!\")\n \n \ndef imshow(tensor):\n plt.figure(figsize=(12, 12))\n img = np.transpose(tensor.numpy(), axes=(1, 2, 0)) # tensor : (3, 227, 227) --> image : (227, 227, 3)\n \n # remove the Train Phase normalization from images\n mean = np.array([.485, .456, .406])\n std = np.array([.229, .224, .225])\n img = (mean + img) * std\n \n # normal pixels between 0 and 1\n img = np.clip(a=img, a_min=0, a_max=1)\n plt.imshow(img)\n plt.pause(20)\n\n\ndef get_dataset_size_and_labels(dataset):\n return {x : len(dataset[x]) for x in ['train', 'eval']}, {len(dataset[x].classes) : dataset[x].classes for x in ['train', 'eval']}\n\n\ndef save_data(data, file_name):\n with open(os.path.join('/kaggle', 'working', file_name), 'wb') as f:\n pickle.dump(data, f)\n f.close()\n \ndef load_data(file_name):\n return pickle.load(open(os.path.join('/kaggle', 'working', file_name), 'rb'))","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.780100Z","iopub.status.idle":"2024-04-15T07:34:16.780445Z","shell.execute_reply.started":"2024-04-15T07:34:16.780277Z","shell.execute_reply":"2024-04-15T07:34:16.780294Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"if __name__ == \"__main__\":\n # 1. Implement the AlexNet deep neural network \n device = \"cuda:0\" if torch.cuda.is_available() else \"cpu\"\n model = VGGNet().to(device)\n \n # 2. Data Prepration\n # determine transformations for train and test data \n data_transform = {\n \"train\" : transforms.Compose([\n transforms.RandomResizedCrop(size=227),\n transforms.RandomVerticalFlip(),\n transforms.ToTensor(),\n transforms.Normalize(mean=[.485, .456, .406], std=[.229, .224, .225])\n ]),\n \n \"eval\" : transforms.Compose([\n transforms.Resize(227),\n transforms.ToTensor(),\n transforms.Normalize(mean=[.485, .456, .406], std=[.229, .224, .225])\n ])\n }\n # apply transformations on train and test data after being loaded on CPU or GPU\n root = os.path.join('/kaggle', 'working', 'cifar-10')\n transformed_dataset = {\n x: datasets.ImageFolder(root=os.path.join(root, 'train', x), transform=data_transform[x]) for x in ['train', 'eval']\n }\n # create a continuous flow of batches and determine how data should load on CPU or GPU\n # image : (227, 227, 3) --> tensor : (3, 227, 227) --> batch : (500, 3, 227, 227)\n dataLoader = {\n x: torch.utils.data.DataLoader(dataset=transformed_dataset[x], num_workers=4, batch_size=64, shuffle=True) for x in ['train', 'eval']\n }\n \n dataset_size, labels = get_dataset_size_and_labels(transformed_dataset)\n # print(dataset_size)\n # print(labels)\n \n # show the images of the first bacth\n# inputs, lables = next(iter(dataLoader['train']))\n# tensor = torchvision.utils.make_grid(inputs)\n# imshow(tensor)\n \n # 3. Training VGG16\n loss = nn.CrossEntropyLoss()\n optimizer = optim.SGD(model.parameters(), lr=.001, momentum=.9)\n reduce_lr = optim.lr_scheduler.StepLR(optimizer=optimizer, step_size=5, gamma=.1)\n trained_model = train(model, loss, optimizer, reduce_lr, num_epochs=30) # fit the AlexNet NN\n save_data(trained_model, 'trained_model')","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.781327Z","iopub.status.idle":"2024-04-15T07:34:16.781712Z","shell.execute_reply.started":"2024-04-15T07:34:16.781496Z","shell.execute_reply":"2024-04-15T07:34:16.781513Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"
\n
\n 5. Evaluating VGG16 and Predicting Labels\n ¶\n
\n
","metadata":{}},{"cell_type":"code","source":"trained_model = load_data('trained_model')","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.783600Z","iopub.status.idle":"2024-04-15T07:34:16.783990Z","shell.execute_reply.started":"2024-04-15T07:34:16.783821Z","shell.execute_reply":"2024-04-15T07:34:16.783837Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"transformed_dataset['train'].class_to_idx","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.785359Z","iopub.status.idle":"2024-04-15T07:34:16.785745Z","shell.execute_reply.started":"2024-04-15T07:34:16.785535Z","shell.execute_reply":"2024-04-15T07:34:16.785552Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def convert_to_idx(pred_tensors):\n preds = [pred_tensor.cpu().numpy().tolist() for pred_tensor in pred_tensors]\n return preds","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.787147Z","iopub.status.idle":"2024-04-15T07:34:16.787499Z","shell.execute_reply.started":"2024-04-15T07:34:16.787331Z","shell.execute_reply":"2024-04-15T07:34:16.787347Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"predicted_labels, actual_labels = predict(trained_model,'eval')","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.788892Z","iopub.status.idle":"2024-04-15T07:34:16.789237Z","shell.execute_reply.started":"2024-04-15T07:34:16.789071Z","shell.execute_reply":"2024-04-15T07:34:16.789088Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"predicted_labels = convert_to_idx(predicted_labels)\nactual_labels = convert_to_idx(actual_labels)","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.790840Z","iopub.status.idle":"2024-04-15T07:34:16.791189Z","shell.execute_reply.started":"2024-04-15T07:34:16.791021Z","shell.execute_reply":"2024-04-15T07:34:16.791038Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(\n classification_report(\n y_true=actual_labels,\n y_pred=predicted_labels,\n target_names=transformed_dataset['train'].class_to_idx.keys()\n )\n)","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.792086Z","iopub.status.idle":"2024-04-15T07:34:16.792425Z","shell.execute_reply.started":"2024-04-15T07:34:16.792257Z","shell.execute_reply":"2024-04-15T07:34:16.792274Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"conf_mat = confusion_matrix(\n y_true=actual_labels,\n y_pred=predicted_labels\n)\n\nplt.figure(figsize=(8, 8))\nsns.heatmap(\n conf_mat,\n cmap='Oranges',\n annot=True,\n cbar=False,\n fmt='.3g'\n);\n\n\nxlocs, _ = plt.xticks()\nplt.xticks(\n ticks=xlocs,\n labels=transformed_dataset['train'].class_to_idx.keys(),\n rotation=30\n);\n\nylocs, _ = plt.yticks()\nplt.yticks(\n ticks=ylocs,\n labels=transformed_dataset['train'].class_to_idx.keys(),\n rotation=360\n);","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.794329Z","iopub.status.idle":"2024-04-15T07:34:16.794745Z","shell.execute_reply.started":"2024-04-15T07:34:16.794529Z","shell.execute_reply":"2024-04-15T07:34:16.794547Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"preds, _ = predict(trained_model,'test')","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.796552Z","iopub.status.idle":"2024-04-15T07:34:16.796934Z","shell.execute_reply.started":"2024-04-15T07:34:16.796763Z","shell.execute_reply":"2024-04-15T07:34:16.796780Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"preds[:10]","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.798333Z","iopub.status.idle":"2024-04-15T07:34:16.798822Z","shell.execute_reply.started":"2024-04-15T07:34:16.798557Z","shell.execute_reply":"2024-04-15T07:34:16.798578Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"preds = convert_to_idx(preds)\npreds[:10]","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.800132Z","iopub.status.idle":"2024-04-15T07:34:16.800482Z","shell.execute_reply.started":"2024-04-15T07:34:16.800314Z","shell.execute_reply":"2024-04-15T07:34:16.800331Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def idx_to_class(preds):\n li = []\n map_preds = {0 : 'airplane',\n 1 : 'automobile',\n 2 : 'bird',\n 3 : 'cat',\n 4 : 'deer',\n 5 : 'dog',\n 6 : 'frog',\n 7 : 'horse',\n 8 : 'ship',\n 9 : 'truck'\n }\n for pred in preds:\n if pred in map_preds.keys():\n li.append(map_preds[pred])\n return li","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.802001Z","iopub.status.idle":"2024-04-15T07:34:16.802366Z","shell.execute_reply.started":"2024-04-15T07:34:16.802188Z","shell.execute_reply":"2024-04-15T07:34:16.802206Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"preds = idx_to_class(preds)\npreds[:10]","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.803753Z","iopub.status.idle":"2024-04-15T07:34:16.804093Z","shell.execute_reply.started":"2024-04-15T07:34:16.803928Z","shell.execute_reply":"2024-04-15T07:34:16.803944Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"length = len(preds)\nlength","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.805305Z","iopub.status.idle":"2024-04-15T07:34:16.805669Z","shell.execute_reply.started":"2024-04-15T07:34:16.805479Z","shell.execute_reply":"2024-04-15T07:34:16.805495Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"
\n
\n 6. Preprocessing and Training ResNet50\n ¶\n
\n
","metadata":{}},{"cell_type":"code","source":"batch_tfms = [*aug_transforms(), Normalize.from_stats(*imagenet_stats)]\ndls=ImageDataLoaders.from_folder(\n path=paths['train'],\n train='train',\n valid='eval',\n item_tfms=Resize(224),\n batch_tfms=batch_tfms,bs=64\n)\nlearn = vision_learner(dls, resnet50, metrics=accuracy, wd=0.01)","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:40:58.364619Z","iopub.execute_input":"2024-04-15T07:40:58.365025Z","iopub.status.idle":"2024-04-15T07:41:04.248898Z","shell.execute_reply.started":"2024-04-15T07:40:58.364993Z","shell.execute_reply":"2024-04-15T07:41:04.247900Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"learn.lr_find()","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:41:07.026903Z","iopub.execute_input":"2024-04-15T07:41:07.027862Z","iopub.status.idle":"2024-04-15T07:41:32.823857Z","shell.execute_reply.started":"2024-04-15T07:41:07.027823Z","shell.execute_reply":"2024-04-15T07:41:32.822678Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"learn.fit_one_cycle(3, 3e-3)","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:41:49.603321Z","iopub.execute_input":"2024-04-15T07:41:49.603790Z","iopub.status.idle":"2024-04-15T07:50:47.694355Z","shell.execute_reply.started":"2024-04-15T07:41:49.603751Z","shell.execute_reply":"2024-04-15T07:50:47.693193Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"learn.unfreeze()\nlearn.lr_find()","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:50:50.140381Z","iopub.execute_input":"2024-04-15T07:50:50.140810Z","iopub.status.idle":"2024-04-15T07:51:14.717966Z","shell.execute_reply.started":"2024-04-15T07:50:50.140771Z","shell.execute_reply":"2024-04-15T07:51:14.716820Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"learn.fit_one_cycle(5, lr_max=slice(1e-6,3e-4))","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:53:52.384318Z","iopub.execute_input":"2024-04-15T07:53:52.385302Z","iopub.status.idle":"2024-04-15T08:12:36.614237Z","shell.execute_reply.started":"2024-04-15T07:53:52.385259Z","shell.execute_reply":"2024-04-15T08:12:36.613239Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"learn.unfreeze()\nlearn.lr_find()","metadata":{"execution":{"iopub.status.busy":"2024-04-15T08:12:49.455506Z","iopub.execute_input":"2024-04-15T08:12:49.456865Z","iopub.status.idle":"2024-04-15T08:13:11.986720Z","shell.execute_reply.started":"2024-04-15T08:12:49.456811Z","shell.execute_reply":"2024-04-15T08:13:11.985691Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"
\n
\n 7. Predicting Labels using ResNet50\n ¶\n
\n
","metadata":{}},{"cell_type":"code","source":"test_img_paths = [os.path.join(paths['test'], file) for file in os.listdir(paths['test'])]\ntest_dl = learn.dls.test_dl(test_img_paths)\nclass_score, y =learn.get_preds(dl=test_dl)\nclass_score = np.argmax(class_score, axis=1)\nclassScore = class_score.tolist()","metadata":{"execution":{"iopub.status.busy":"2024-04-15T08:14:18.269051Z","iopub.execute_input":"2024-04-15T08:14:18.269439Z","iopub.status.idle":"2024-04-15T08:22:59.496118Z","shell.execute_reply.started":"2024-04-15T08:14:18.269405Z","shell.execute_reply":"2024-04-15T08:22:59.494951Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"learn.dls.vocab","metadata":{"execution":{"iopub.status.busy":"2024-04-15T08:26:40.715880Z","iopub.execute_input":"2024-04-15T08:26:40.716733Z","iopub.status.idle":"2024-04-15T08:26:40.723727Z","shell.execute_reply.started":"2024-04-15T08:26:40.716697Z","shell.execute_reply":"2024-04-15T08:26:40.722623Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"predicted_labels = idx_to_class(classScore)\nprint(classScore[:10])\nprint(predicted_labels[:10])","metadata":{"execution":{"iopub.status.busy":"2024-04-15T08:26:59.875519Z","iopub.execute_input":"2024-04-15T08:26:59.876233Z","iopub.status.idle":"2024-04-15T08:26:59.927502Z","shell.execute_reply.started":"2024-04-15T08:26:59.876201Z","shell.execute_reply":"2024-04-15T08:26:59.926134Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"ImageId = [ (''.join(filter(str.isdigit, name))) for name in os.listdir(paths['test'])]\nsubmission_df = pd.DataFrame({\n 'id' : ImageId,\n 'label' : predicted_labels\n})\nsubmission_df.head()\nsubmission_df.to_csv('/kaggle/working/submission_df.csv', index=False)","metadata":{"execution":{"iopub.status.busy":"2024-04-15T07:34:16.821916Z","iopub.status.idle":"2024-04-15T07:34:16.822268Z","shell.execute_reply.started":"2024-04-15T07:34:16.822093Z","shell.execute_reply":"2024-04-15T07:34:16.822110Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"
\n
\n 8. Conclusion\n ¶\n
\n
","metadata":{}},{"cell_type":"markdown","source":"
\n \nUsing ResNet50, we can acheive a reliable and accurate image classifier with respect to VGG16 because it is a modern naural network and avoid vanishing gradient using skip connections commonly encountered in deep neural networks.
\n\n
If you find this notebook valuable, your support by giving an upvote would be greatly appreciated. THANK YOU :)\n \n
\n","metadata":{}},{"cell_type":"code","source":"","metadata":{},"outputs":[],"execution_count":null}]}