{"cells":[{"metadata":{},"cell_type":"markdown","source":"

🏃🏽‍♂️ Implementing the RepVGG 🏃🏽‍♀️

"},{"metadata":{},"cell_type":"markdown","source":"

Okay, so let's get this right. Recently a paper popped up in paperswithcode's website.
It talked about Making VGG-style ConvNets Great Again.

\nNo, I am not kidding that is literally the title of that paper.
\nI kind of smirked when I read the title. Dude, it's high time we start replacing ConvNets with Pretrained Transformer Models because well they do an equally or sometimes even better of a job when it comes to Image Classification.

\nYet here it seems like that Transformers might be met with some sort of competition? Maybe they don't?
No one knows let's jump into this.\n

\nThe Purpose of this Notebook is to help everyone use the RepVGG Model.

"},{"metadata":{},"cell_type":"markdown","source":"

Also, if you already upvoted the Notebook without even going through it: We are automatically Best Friends. 😋

"},{"metadata":{},"cell_type":"markdown","source":"# Importing Libraries\n

Importing all the necessary libraries.

"},{"metadata":{"trusted":true,"collapsed":true},"cell_type":"code","source":"# Cloning the Github Repo for the Model.\n!git clone https://github.com/DingXiaoH/RepVGG.git","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Some Native Python Tools.\nimport glob\nfrom itertools import chain\nimport os\nimport random\n\n# These ones are for dealing with our Data.\nimport matplotlib.pyplot as plt\nimport pandas as pd\nfrom PIL import Image\nimport numpy as np\n\n# PyTorch is like that one boss who actually does everything he can to help you.\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.optim as optim\nfrom torch.optim.lr_scheduler import StepLR\nfrom torch.utils.data import DataLoader, Dataset\nfrom torchvision import datasets, transforms\n\n# Keeping Time in check.\nfrom tqdm.notebook import tqdm\n\n# Tools for Preprocessing and stuff.\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.preprocessing import OrdinalEncoder\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Heading to the directory with model files.\ncd ./RepVGG","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# For Grabbing the RepVGG Model.\nfrom repvgg import repvgg_model_convert, create_RepVGG_B2","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Heading back to where the other data is stored.\ncd ../","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Setting up the Hyperparameters\n\nbatch_size = 64\nepochs = 15\nlr = 3e-4\ngamma= 0.7\nseed = 42","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"

Seeding Everything

\n

Seeding makes things reproduceable so let's go!

"},{"metadata":{"trusted":true},"cell_type":"code","source":"# Definign a function to seed everything.\ndef seed_it(seed):\n random.seed(seed)\n os.environ[\"PYTHONSEED\"] = str(seed)\n np.random.seed(seed)\n torch.cuda.manual_seed(seed)\n torch.cuda.manual_seed_all(seed)\n torch.backends.cudnn.deterministic = True\n \n# Running the Function.\nseed_it(seed)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Moving stuff to the GPU.\ndevice = \"cuda\"","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"

Preprocessing

\n

Here, we will make a list of all images' paths and put them in a tuple along with thier labels. Also, let's gather all those tuples and put them into a list.

"},{"metadata":{"trusted":true},"cell_type":"code","source":"# Defining the paths to files.\ntrainimages = \"../input/cassava-leaf-disease-classification/train_images\"\n\ntraindf = pd.read_csv(\"../input/cassava-leaf-disease-classification/train.csv\")\n\ntrain_list = []\n# Converting the Image IDs to their paths.\nfor i in traindf.index:\n \n a = traindf[\"image_id\"].loc[i]\n \n b = trainimages + \"/\" + a\n \n train_list.append((b, traindf['label'].loc[i]))\n\n# Taking a look.\ntrain_list[:3]","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Visualizing the Data\n

Time to take a look at our Data!

\n"},{"metadata":{"trusted":true},"cell_type":"code","source":"random_idx = np.random.randint(1, len(train_list), size=16)\nfig, axes = plt.subplots(6,6, figsize=(13,13))\n\nfor idx, ax in enumerate(axes.ravel()):\n img = Image.open(train_list[idx][0])\n ax.set_title(train_list[idx][1])\n ax.imshow(img)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Dataloader\n

Time to set up a DataLoader with the resources we have.

"},{"metadata":{"trusted":true},"cell_type":"code","source":"# Splitting the data into training and validation datasets.\ntrain, valid = train_test_split(train_list,\n test_size=0.2,\n random_state=42)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"

Transforms

\n

Producing some transformations for our training and validation datasets.

"},{"metadata":{"trusted":true},"cell_type":"code","source":"train_trans = transforms.Compose([\n transforms.RandomResizedCrop(224),\n transforms.RandomHorizontalFlip(p=0.3),\n transforms.ToTensor(),\n transforms.Normalize(mean=[0.485, 0.456, 0.406],\n std=[0.229, 0.224, 0.225])\n ])\n\nvalid_trans = transforms.Compose([\n transforms.ToTensor(),\n transforms.Normalize(mean=[0.485, 0.456, 0.406],\n std=[0.229, 0.224, 0.225])\n ])","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"

Dataset Class

\n

Defining the PyTorch Dataset Class.

"},{"metadata":{"trusted":true},"cell_type":"code","source":"class CassavaDataset(Dataset):\n \n # Grabbing the transforms presets.\n def __init__(self, file_list, transform=None):\n self.file_list = file_list\n self.transform = transform\n \n def __len__(self):\n self.filelength = len(self.file_list)\n return self.filelength\n \n def __getitem__(self, idx):\n \n img_path = self.file_list[idx][0]\n img = Image.open(img_path).convert(\"RGB\")\n img_transformed = self.transform(img)\n \n # The second item in the tuple is the label.\n label = self.file_list[idx][1]\n \n return img_transformed, label","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"

Finally collecting everthing into a two DataLoaders. One for the training data and another for the valid data.

\n"},{"metadata":{"trusted":true},"cell_type":"code","source":"train_data = CassavaDataset(train, transform = train_trans)\nvalid_data = CassavaDataset(valid, transform = valid_trans)\n\ntrain_loader = DataLoader(dataset = train_data, batch_size = batch_size, shuffle = True)\nvalid_loader = DataLoader(dataset = valid_data, batch_size = batch_size, shuffle = True)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Model\n

Grabbing the Model.

"},{"metadata":{"trusted":true,"collapsed":true},"cell_type":"code","source":"model = create_RepVGG_B2(deploy=False).to(device)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Defining some presets.\n\ncriterion = nn.CrossEntropyLoss()\n\noptimizer = optim.Adam(model.parameters(), lr = lr)\n\nscheduler = StepLR(optimizer, step_size=1, gamma=gamma)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Training the Model\n

Let's use everything that we have done above to finally train the model for 20 Epochs!

"},{"metadata":{"trusted":true},"cell_type":"code","source":"for epoch in range(epochs):\n epoch_loss = 0.0\n epoch_accuracy = 0.0\n \n for data, label in tqdm(train_loader):\n data = data.to(device)\n label = label.to(device)\n \n output = model(data)\n \n label = torch.nn.functional.one_hot(label, num_classes = 5)\n label = label.squeeze_()\n label = torch.argmax(label, axis=1)\n \n label = label.type_as(output)\n \n loss = criterion(output, label.long())\n \n optimizer.zero_grad()\n loss.backward()\n optimizer.step()\n \n acc = (output.argmax(dim=1) == label).float().mean()\n epoch_accuracy += acc / len(train_loader)\n epoch_loss += loss / len(train_loader)\n\n \n with torch.no_grad():\n epoch_val_accuracy = 0\n epoch_val_loss = 0 \n \n for data, label in valid_loader:\n data = data.to(device)\n \n label = label.to(device)\n #label = label.squeeze_()\n label = label.type_as(output)\n \n val_output = model(data)\n val_loss = criterion(val_output, label.long())\n \n accc = (val_output.argmax(dim = 1) == label).float().mean()\n epoch_val_accuracy += acc / len(valid_loader)\n epoch_val_loss += val_loss / len(valid_loader)\n\n print(\n f\"Epoch: {epoch+1} - loss: {epoch_loss:.4f} - acc : {epoch_accuracy:.4f} - val_loss : {epoch_val_loss:.4f} - val_acc: {epoch_val_accuracy: .4f}\\n\"\n )","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"\n

Briefing about this Model

\n\n\n

A simple but powerful architecture of convolutional neural network, which has a VGG-like inference-time body composed of nothing but a stack of 3x3 convolution and ReLU, while the training-time model has a multi-branch topology. Such decoupling of the training-time and inference-time architecture is realized by a structural re-parameterization technique so that the model is named RepVGG.\n

\nFor any further reading: Check out the researcher's Repo!

"}],"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat":4,"nbformat_minor":4}