\n The goal of this notebook is to build a model to accurately identify digits from a dataset of tens of thousands \n of handwritten images. Using both the Kaggle and Keras-provided MNIST datasets, this project will explore \n neural network approaches to determine effective techniques in digit recognition.\n
\n\n\n More details about the MNIST dataset, including algorithms that have been tried on it and their levels of \n success, can be found at this link. \n The dataset is made available under a Creative Commons Attribution-Share Alike 3.0 license.\n
\n","metadata":{}},{"cell_type":"code","source":"import pandas as pd\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom keras.datasets import mnist\nfrom tensorflow.keras.utils import to_categorical\nimport matplotlib.cm as cm\nfrom keras.models import Sequential,load_model\nfrom keras.layers import Dense,Dropout,Flatten,Conv2D,MaxPooling2D\nfrom keras.callbacks import ReduceLROnPlateau,ModelCheckpoint\n\nimport warnings \nwarnings.filterwarnings('ignore')\n%matplotlib inline","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","execution":{"iopub.status.busy":"2024-10-30T11:57:36.668813Z","iopub.execute_input":"2024-10-30T11:57:36.669960Z","iopub.status.idle":"2024-10-30T11:57:50.816625Z","shell.execute_reply.started":"2024-10-30T11:57:36.669884Z","shell.execute_reply":"2024-10-30T11:57:50.815420Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"df=pd.read_csv('/kaggle/input/digit-recognizer/train.csv')\ndt=pd.read_csv('/kaggle/input/digit-recognizer/test.csv')\ndf.shape,dt.shape","metadata":{"execution":{"iopub.status.busy":"2024-10-30T11:57:50.818462Z","iopub.execute_input":"2024-10-30T11:57:50.819124Z","iopub.status.idle":"2024-10-30T11:57:56.028307Z","shell.execute_reply.started":"2024-10-30T11:57:50.819084Z","shell.execute_reply":"2024-10-30T11:57:56.027279Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":" ##\n Before feeding our data into a neural network, we need to format it to match the input requirements. \n Since neural networks expect a consistent format and data type, the following transformations are applied:\n
\n\nfloat32 \n to ensure compatibility with the neural network’s processing and optimize memory usage during training.\n (28, 28, 1) to match the structure of \n the MNIST images—28x28 pixels and a single grayscale channel—so the neural network can interpret each image correctly.\n \n For validation data, we similarly convert to float32, \n reshape it to fit the 28x28 image format, and extract the labels as integer values. The final shape of \n X_val and \n y_val is printed to confirm the data preparation.\n
\n In this section, we are loading the MNIST dataset and preparing it for our neural network model. \n The goal is to create a unified training dataset that combines both training and testing data, allowing us to leverage \n the full range of available handwritten digit images for better model training.\n
\n\n\n Finally, we check the shapes of the merged dataset and labels to confirm everything is prepared correctly for training.\n
\n","metadata":{}},{"cell_type":"code","source":"(X_train,y_train),(X_test,y_test)=mnist.load_data()\nX_train=np.vstack((X_train,X_test))\ny_train=np.concatenate([y_train,y_test])\nX_train=X_train.reshape(-1,28,28,1)\nX_train.shape,y_train.shape","metadata":{"execution":{"iopub.status.busy":"2024-10-30T11:57:56.141468Z","iopub.execute_input":"2024-10-30T11:57:56.141938Z","iopub.status.idle":"2024-10-30T11:57:57.552300Z","shell.execute_reply.started":"2024-10-30T11:57:56.141889Z","shell.execute_reply":"2024-10-30T11:57:57.551003Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":" ##\n In this section, we create a function to visualize handwritten digits from the dataset. This is crucial for understanding \n how well our model is performing in recognizing different numbers. By displaying the images alongside their predicted \n values, we can evaluate the model's accuracy visually.\n
\n\nimg_to_num function displays an \n image of the digit, annotating it with its pixel values for better insight into the data being processed.\n \n After defining the function, we visualize one of the images from the validation set, allowing us to confirm that \n the input data is being processed correctly.\n
\n","metadata":{}},{"cell_type":"code","source":"def img_to_num(img, ax):\n ax.imshow(img, cmap='gray')\n w, h = img.shape\n thresh = img.max() / 2.5\n for x in range(w):\n for y in range(h):\n ax.annotate(\n str(round(img[x, y])),\n xy=(y, x),\n horizontalalignment='center',\n verticalalignment='center',\n color='white' if img[x, y] < thresh else 'black'\n )\n\nfig = plt.figure(figsize=(12, 12))\nax = fig.add_subplot(111)\nimg_to_num(X_val[3].reshape(28, 28), ax) \nplt.show()\n ","metadata":{"execution":{"iopub.status.busy":"2024-10-30T12:00:19.275717Z","iopub.execute_input":"2024-10-30T12:00:19.276306Z","iopub.status.idle":"2024-10-30T12:00:22.389976Z","shell.execute_reply.started":"2024-10-30T12:00:19.276260Z","shell.execute_reply":"2024-10-30T12:00:22.388089Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":" ##\n In this section, we normalize our datasets by scaling the pixel values to a range of [0, 1]. \n This step is crucial for training neural networks as it helps in faster convergence and improved \n performance by ensuring that all input features contribute equally during training.\n
\n\nX_train, \n dt, and X_val by 255.\n This transformation changes the pixel values from a range of [0, 255] to [0, 1], enhancing numerical stability.\n \n Overall, this normalization step is essential for optimizing the performance of our neural network model.\n
\n","metadata":{}},{"cell_type":"code","source":"X_train=X_train/255\ndt=dt/255\nX_val=X_val/255","metadata":{"execution":{"iopub.status.busy":"2024-10-29T12:15:56.353440Z","iopub.execute_input":"2024-10-29T12:15:56.353917Z","iopub.status.idle":"2024-10-29T12:15:56.576837Z","shell.execute_reply.started":"2024-10-29T12:15:56.353846Z","shell.execute_reply":"2024-10-29T12:15:56.575829Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":" ##\n In this section, we transform the integer labels into one-hot encoded vectors. Each digit (0-9) is represented as a vector of 10 elements, where only the index corresponding to the digit is set to 1, and all other indices are set to 0.\n
\n\n\n For example, the digit 3 is transformed as follows:\n
\n\n\n3 becomes \n the vector [0, 0, 0, 1, 0, 0, 0, 0, 0, 0].\n \n This encoding method allows our model to output probabilities for each digit during training and is essential \n for multi-class classification tasks.\n
\n","metadata":{}},{"cell_type":"code","source":"print(\"Before catagorizing \",y_train[0])\ny_train=to_categorical(y_train,10)\ny_val=to_categorical(y_val,10)\nprint(\"Before catagorizing \" ,y_train[0])","metadata":{"execution":{"iopub.status.busy":"2024-10-30T12:01:54.637426Z","iopub.execute_input":"2024-10-30T12:01:54.638027Z","iopub.status.idle":"2024-10-30T12:01:54.653191Z","shell.execute_reply.started":"2024-10-30T12:01:54.637982Z","shell.execute_reply":"2024-10-30T12:01:54.651748Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":" ##\n The architecture of our convolutional neural network (CNN) is designed to efficiently learn and classify handwritten digits from the MNIST dataset. Below are the key types of layers used in the model:\n
\n\n\n Each layer type plays a distinct and essential role in transforming the input data through feature extraction and classification, ultimately enabling accurate digit recognition.\n
\n","metadata":{}},{"cell_type":"code","source":"model=Sequential()\nmodel.add(Conv2D(filters=64,kernel_size=3,padding='same',activation='relu',input_shape=(28,28,1)))\nmodel.add(Conv2D(filters=128,kernel_size=3,padding='same',activation='relu'))\nmodel.add(Conv2D(filters=128,kernel_size=3,padding='same',activation='relu'))\nmodel.add(MaxPooling2D(pool_size=2))\nmodel.add(Conv2D(filters=128,kernel_size=3,padding='same',activation='relu'))\nmodel.add(Conv2D(filters=192,kernel_size=3,padding='same',activation='relu'))\nmodel.add(MaxPooling2D(pool_size=2))\nmodel.add(Conv2D(filters=192,kernel_size=5,padding='same',activation='relu'))\nmodel.add(MaxPooling2D(pool_size=2,padding='same'))\nmodel.add(Flatten())\nmodel.add(Dense(256,activation='relu'))\nmodel.add(Dense(10,activation='softmax'))\nmodel.summary()\n","metadata":{"execution":{"iopub.status.busy":"2024-10-29T12:15:56.591847Z","iopub.execute_input":"2024-10-29T12:15:56.592169Z","iopub.status.idle":"2024-10-29T12:15:57.652637Z","shell.execute_reply.started":"2024-10-29T12:15:56.592137Z","shell.execute_reply":"2024-10-29T12:15:57.651738Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":" ##\n In this section, we compile and train our convolutional neural network model. The following steps are performed:\n
\n\nmodel.compile() to set up the model for training. The loss function categorical_crossentropy is suitable for multi-class classification problems, while the adam optimizer efficiently updates the model weights during training based on the loss.\n ReduceLROnPlateau callback monitors the validation loss, reducing the learning rate by a factor of 0.3 if there is no improvement for a specified number of epochs. This helps the model converge better by allowing it to take smaller steps when approaching optimal weights.\n model.fit(), specifying the training data and labels, batch size, number of epochs, and validation data. The training process allows the model to learn from the data, adjusting weights based on the computed loss and improving accuracy over time.\n model.save('my_model.h5'). This enables us to reuse the trained model in future applications without needing to retrain it.\n \n These steps are crucial for ensuring our model is effectively trained and can be utilized later for predictions.\n
\n","metadata":{}},{"cell_type":"code","source":"model.compile(loss='categorical_crossentropy',optimizer='adam',metrics=['accuracy'])","metadata":{"execution":{"iopub.status.busy":"2024-10-29T12:15:57.653880Z","iopub.execute_input":"2024-10-29T12:15:57.654258Z","iopub.status.idle":"2024-10-29T12:15:57.667822Z","shell.execute_reply.started":"2024-10-29T12:15:57.654221Z","shell.execute_reply":"2024-10-29T12:15:57.666932Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"lr=ReduceLROnPlateau(monitir='loss',factor=.3,verbose=1,patience=2,min_lr=10e-8)\nhistory=model.fit(X_train,y_train,batch_size=128,epochs=21,validation_data=(X_val,y_val),callbacks=[lr],verbose=1,shuffle=True)\nmodel.save('my_model.h5') ","metadata":{"execution":{"iopub.status.busy":"2024-10-29T12:15:57.669056Z","iopub.execute_input":"2024-10-29T12:15:57.670152Z","iopub.status.idle":"2024-10-29T12:16:38.640454Z","shell.execute_reply.started":"2024-10-29T12:15:57.670114Z","shell.execute_reply":"2024-10-29T12:16:38.639553Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":" ##