{"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_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"# This Python 3 environment comes with many helpful analytics libraries installed\n# It is defined by the kaggle/python Docker image: https://github.com/kaggle/docker-python\n# For example, here's several helpful packages to load\n\nimport numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\n\n# Input data files are available in the read-only \"../input/\" directory\n# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory\n\nimport os\nfor dirname, _, filenames in os.walk('/kaggle/input'):\n    for filename in filenames:\n        print(os.path.join(dirname, filename))\n\n# You can write up to 20GB to the current directory (/kaggle/working/) that gets preserved as output when you create a version using \"Save & Run All\" \n# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of the current session","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from pathlib import Path\nfrom PIL import Image\nimport random\nimport tensorflow as tf\nfrom tensorflow import keras\nfrom keras.models import Sequential\nfrom keras.layers import Conv2D, Dense, Flatten, Dropout, normalization, Input\nfrom keras.losses import CategoricalCrossentropy, SparseCategoricalCrossentropy\nfrom keras.optimizers import Adam, SGD\nfrom keras.metrics import SparseCategoricalAccuracy\nfrom keras.regularizers import l2, l1, l1_l2\nimport matplotlib.pyplot as plt","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# https://github.com/dariush-bahrami/gravity.optimizer\nclass Gravity(tf.keras.optimizers.Optimizer):\n    def __init__(self,\n                 learning_rate=0.1,\n                 alpha=0.01,\n                 beta=0.9,\n                 name=\"Gravity\",\n                 **kwargs):\n        super(Gravity, self).__init__(name, **kwargs)\n        self._set_hyper('learning_rate', kwargs.get('lr', learning_rate))\n        self._set_hyper('decay', self._initial_decay)\n        self._set_hyper('alpha', alpha)\n        self._set_hyper('beta', beta)\n        self.epsilon = 1e-7\n\n    def _create_slots(self, var_list):\n        alpha = self._get_hyper(\"alpha\")\n        stddev = alpha / self.learning_rate\n        initializer = tf.keras.initializers.RandomNormal(mean=0.0,\n                                                         stddev=stddev,\n                                                         seed=None)\n        for var in var_list:\n            self.add_slot(var, \"velocity\", initializer=initializer)\n\n    @tf.function\n    def _resource_apply_dense(self, grad, var):\n        # Get Data\n        var_dtype = var.dtype.base_dtype\n        lr_t = self._decayed_lr(var_dtype)\n        beta = self._get_hyper(\"beta\", var_dtype)\n        t = tf.cast(self.iterations, float)\n        beta_hat = (beta * t + 1) / (t + 2)\n        velocity = self.get_slot(var, \"velocity\")\n\n        # Calculations\n        max_step_grad = 1 / tf.math.reduce_max(tf.math.abs(grad))\n        gradient_term = grad / (1 + (grad / max_step_grad)**2)\n\n        # update variables\n        updated_velocity = velocity.assign(beta_hat * velocity +\n                                           (1 - beta_hat) * gradient_term)\n        updated_var = var.assign(var - lr_t * updated_velocity)\n\n        # updates = [updated_var, updated_velocity]\n        # return tf.group(*updates)\n    def _resource_apply_sparse(self, grad, var):\n        raise NotImplementedError\n\n    def get_config(self):\n        config = super(Gravity, self).get_config()\n        config.update({\n            'learning_rate':\n            self._serialize_hyperparameter('learning_rate'),\n            'decay':\n            self._serialize_hyperparameter('decay'),\n            'alpha':\n            self._serialize_hyperparameter('alpha'),\n            'beta':\n            self._serialize_hyperparameter('beta'),\n            'epsilon':\n            self.epsilon,\n        })\n        return config","metadata":{"jupyter":{"source_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Find Data\nroot = Path('/kaggle')\ninput_dir = Path.joinpath(root, Path('input'))\ncsv_paths = {i.name: i for i in input_dir.rglob('*.csv')}\n\n# Create Dataframes\ndf = pd.read_csv(csv_paths['train.csv'])\n\n# Create Arrays\nx = df.drop(columns=['label']).to_numpy(dtype=np.uint8)\ndata_num = x.shape[0]\ny = df['label'].to_numpy(dtype=np.uint8)\nx = x.reshape(data_num, 28, 28)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test_df = pd.read_csv(csv_paths['test.csv'])\nx_test = test_df.to_numpy(dtype=np.uint8)\ntest_data_num = x_test.shape[0]\nx_test = x_test.reshape(test_data_num, 28, 28)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Show random test image\nindex = random.randint(0, test_data_num-1)\nshow_size = (256, 256)\nim = Image.fromarray(x_test[index]).resize(show_size)\nim","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"x_test_normalize = (x_test / 255).reshape(-1, 28, 28, 1)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Show random image\nindex = random.randint(0, data_num-1)\nshow_size = (256, 256)\nim = Image.fromarray(x[index]).resize(show_size)\nprint(y[index])\nim","metadata":{"jupyter":{"outputs_hidden":true},"collapsed":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Normalize & Add channel dimension\nx = (x/255).reshape(-1, 28, 28, 1)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Design Model\n\nlayers = [Input(shape=x.shape[1:]),\n          Conv2D(filters=8, kernel_size=3, padding='valid', activation='relu'),\n          Conv2D(filters=16, kernel_size=3, padding='valid', activation='relu'),\n          Conv2D(filters=32, kernel_size=3, padding='valid', activation='relu'),\n          Conv2D(filters=64, kernel_size=3, padding='valid', activation='relu'),\n          Conv2D(filters=128, kernel_size=3, padding='valid', activation='relu'),\n          Conv2D(filters=256, kernel_size=3, padding='valid', activation='relu'),\n          Flatten(),\n#           Dense(64),\n#           Dense(64),\n#           Dropout(0.25),\n#           Dense(64),\n#           Dropout(0.25),\n          Dense(64, activation='relu',\n                kernel_regularizer= l1_l2(l1=1e-5, l2=1e-4),\n                bias_regularizer= l1_l2(1e-4),\n                activity_regularizer= l1_l2(1e-5)),\n          Dense(64, activation='relu',\n                kernel_regularizer= l1_l2(l1=1e-5, l2=1e-4),\n                bias_regularizer= l1_l2(1e-4),\n                activity_regularizer= l1_l2(1e-5)),\n          Dense(64, activation='relu',\n                kernel_regularizer= l1_l2(l1=1e-5, l2=1e-4),\n                bias_regularizer= l1_l2(1e-4),\n                activity_regularizer= l1_l2(1e-5)),\n          Dense(10)]\n\nmodel = Sequential(layers)\nmodel.summary()","metadata":{"collapsed":true,"jupyter":{"outputs_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Compile Model\n\nsgd_hypers = {'learning_rate': 0.01, 'momentum': 0.0, 'nesterov':False}\nsgd_optimizer = SGD(**sgd_hypers)\ngravity_hypers = {'learning_rate':0.1, 'alpha':0.01, 'beta':0.9,}\ngravity_optimizer = Gravity(**gravity_hypers)\nloss_fn = SparseCategoricalCrossentropy(from_logits=True)\nmodel.compile(optimizer=gravity_optimizer, loss=loss_fn, metrics=[SparseCategoricalAccuracy()])","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Train Model\n\nfit_hypers = {'batch_size': 64, 'epochs': 100, 'validation_split': 0.25, }\nhistory = model.fit(x, y, **fit_hypers).history","metadata":{"collapsed":true,"jupyter":{"outputs_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"accuracy = history['sparse_categorical_accuracy']\nloss = history['loss']\ndev_accuracy = history['val_sparse_categorical_accuracy']\ndev_loss = history['val_loss']\n# history.keys()","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"fig, axis = plt.subplots(nrows=2, ncols=1, sharex=True)\nfig.set_size_inches(8, 7)\nfig.suptitle('MNIST Network Training Data')\naxis[0].plot(accuracy, label='Accuracy', color='blue')\naxis[0].plot(dev_accuracy, label = 'Dev Set Accuracy', color='red')\naxis[0].set_ylabel('Accuracy')\naxis[0].legend(frameon=False)\n\naxis[1].plot(loss, label='Loss', color='blue')\naxis[1].plot(dev_loss, label = 'Dev Set Loss', color='red')\naxis[1].set_xlabel('Epoch')\naxis[1].set_ylabel('Loss')\naxis[1].legend(frameon=False)\nfig.savefig('training_data.png', dpi=300, transparent=True)\nplt.show()","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"model.save('mnist_model')\n# !zip -r mnist_model.zip mnist_model","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"df = pd.DataFrame(history).rename(columns={'sparse_categorical_accuracy': 'accuracy', 'val_sparse_categorical_accuracy': 'val_accuracy'})\n# df.to_csv('train_history.csv', index=False)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"y_pred = np.argmax(tf.nn.softmax(model.predict(x_test_normalize)), axis=-1)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Show random test image\nindex = random.randint(0, test_data_num-1)\nshow_size = (256, 256)\nim = Image.fromarray(x_test[index]).resize(show_size)\nprint(y_pred[index])\nim","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"image_id = np.arange(1, 28001, 1)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"submission = pd.DataFrame({'ImageId': image_id, 'Label': y_pred})","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"submission.to_csv('mnist_submission.csv', index=False)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}