cnn2.py

import os
import sys
import random
import numpy as np
import pandas as pd
import tensorflow as tf
from sklearn.model_selection import StratifiedKFold
from tensorflow.keras.optimizers import Adam
from sklearn.metrics import accuracy_score
from compare import compare
from utils import data_aug, plot_history
from models import build_cnn2_model

os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"
physical_devices = tf.config.list_physical_devices("GPU")
tf.config.experimental.set_memory_growth(physical_devices[0], True)

train = pd.read_csv('input/train.csv')
test = pd.read_csv('input/test.csv')
sample_submission = pd.read_csv('input/sample_submission.csv')

X = train.drop(columns=['label'], axis=1)
y = train['label']

kf = StratifiedKFold(n_splits=20, shuffle=True, random_state=42)
train_idx, test_idx = list(kf.split(X, y))[0]

X_train, X_test, y_train, y_test = X.loc[train_idx], X.loc[test_idx], y.loc[train_idx], y.loc[test_idx]
X_train = X_train.astype("float32") / 255.0
X_test = X_test.astype("float32") / 255.0
test = test.astype('float32') / 255.0

X_train.reset_index(drop=True, inplace=True)
X_test.reset_index(drop=True, inplace=True)
y_train.reset_index(drop=True, inplace=True)
y_test.reset_index(drop=True, inplace=True)

X_test = X_test.values.reshape(-1, 28, 28)
test = test.values.reshape(-1, 28, 28)

kf = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)

models = []
ans = pd.read_csv('./input/sample_submission.csv')
fans = None
for fold, (train_idx, val_idx) in enumerate(kf.split(X_train, y_train)):
    print(fold)
    train_x, train_y = X_train.loc[train_idx].values, y_train.loc[train_idx].values
    val_x, val_y = X_train.loc[val_idx].values, y_train.loc[val_idx].values
    train_x = train_x.reshape(-1, 28, 28)
    val_x = val_x.reshape(-1, 28, 28)
    print(train_x.shape, train_y.shape)
    train_x, train_y = data_aug(train_x, train_y, 15)
    print(train_x.shape, train_y.shape)

    model = build_cnn2_model()
    callback = tf.keras.callbacks.EarlyStopping(monitor='loss', patience=10)
    model.compile(optimizer=Adam(learning_rate=3e-4),
                  loss='sparse_categorical_crossentropy', metrics=['accuracy'])
    history = model.fit(train_x, train_y, validation_data=(val_x, val_y), epochs=250, batch_size=128, verbose=1,
                        callbacks=[callback])

    model.evaluate(train_x, train_y)
    model.evaluate(val_x, val_y)
    model.evaluate(X_test, y_test)
    pred_val = np.argmax(model.predict(val_x), axis=-1)
    pred_test = np.argmax(model.predict(X_test), axis=-1)
    test_score = accuracy_score(y_test.values, pred_test, )
    models.append((fold, model, accuracy_score(val_y, pred_val, ), test_score))

    pred_test = np.argmax(model.predict(test), axis=-1)
    if fans is None:
        ans['Label'] = pred_test
        fans = test_score
    else:
        if test_score > fans:
            ans['Label'] = pred_test
            fans = test_score

    # plot_history(history)

print(models)
ans.to_csv('./output/sub_cnn.csv', index=False)
compare('sub_cnn')