{"cells":[{"metadata":{},"cell_type":"markdown","source":"![](http://)Thank you @ulrich07 for this lovely notebook, don't forget upvode it -->\nhttps://www.kaggle.com/ulrich07/osic-multiple-quantile-regression-starter","execution_count":null},{"metadata":{"_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","trusted":true},"cell_type":"code","source":"import numpy as np\nimport pandas as pd\nimport pydicom\nimport os\nimport random\nimport matplotlib.pyplot as plt\nfrom tqdm import tqdm\nfrom PIL import Image\nfrom sklearn.metrics import mean_absolute_error\nfrom sklearn.metrics import mean_squared_error\nfrom sklearn.model_selection import KFold\nfrom math import sqrt\nfrom statistics import mean, pstdev","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"import tensorflow as tf\nimport tensorflow.keras.backend as K\nimport tensorflow.keras.layers as L\nimport tensorflow.keras.models as M","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def seed_everything(seed=2020):\n    random.seed(seed)\n    os.environ['PYTHONHASHSEED'] = str(seed)\n    np.random.seed(seed)\n    tf.random.set_seed(seed)\n    \nseed_everything(42)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"ROOT = \"../input/osic-pulmonary-fibrosis-progression\"\nBATCH_SIZE=128","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"tr = pd.read_csv(f\"{ROOT}/train.csv\")\ntr.drop_duplicates(keep=False, inplace=True, subset=['Patient','Weeks'])\nchunk = pd.read_csv(f\"{ROOT}/test.csv\")\n\nprint(\"add infos\")\nsub = pd.read_csv(f\"{ROOT}/sample_submission.csv\")\nsub['Patient'] = sub['Patient_Week'].apply(lambda x:x.split('_')[0])\nsub['Weeks'] = sub['Patient_Week'].apply(lambda x: int(x.split('_')[-1]))\nsub =  sub[['Patient','Weeks','Confidence','Patient_Week']]\nsub = sub.merge(chunk.drop('Weeks', axis=1), on=\"Patient\")","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"tr['WHERE'] = 'train'\nchunk['WHERE'] = 'val'\nsub['WHERE'] = 'test'\ndata = tr.append([chunk, sub])","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"print(tr.shape, chunk.shape, sub.shape, data.shape)\nprint(tr.Patient.nunique(), chunk.Patient.nunique(), sub.Patient.nunique(), \n      data.Patient.nunique())\n#","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"data['min_week'] = data['Weeks']\ndata.loc[data.WHERE=='test','min_week'] = np.nan\ndata['min_week'] = data.groupby('Patient')['min_week'].transform('min')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"base = data.loc[data.Weeks == data.min_week]\nbase = base[['Patient','FVC']].copy()\nbase.columns = ['Patient','min_FVC']\nbase['nb'] = 1\nbase['nb'] = base.groupby('Patient')['nb'].transform('cumsum')\nbase = base[base.nb==1]\nbase.drop('nb', axis=1, inplace=True)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"data = data.merge(base, on='Patient', how='left')\ndata['base_week'] = data['Weeks'] - data['min_week']\ndel base","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"COLS = ['Sex','SmokingStatus'] #,'Age'\nFE = []\nfor col in COLS:\n    for mod in data[col].unique():\n        FE.append(mod)\n        data[mod] = (data[col] == mod).astype(int)\n#=================","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"#\ndata['age'] = (data['Age'] - data['Age'].min() ) / ( data['Age'].max() - data['Age'].min() )\ndata['BASE'] = (data['min_FVC'] - data['min_FVC'].min() ) / ( data['min_FVC'].max() - data['min_FVC'].min() )\ndata['week'] = (data['base_week'] - data['base_week'].min() ) / ( data['base_week'].max() - data['base_week'].min() )\ndata['percent'] = (data['Percent'] - data['Percent'].min() ) / ( data['Percent'].max() - data['Percent'].min() )\nFE += ['age','percent','week','BASE']","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"tr = data.loc[data.WHERE=='train']\nchunk = data.loc[data.WHERE=='val']\nsub = data.loc[data.WHERE=='test']\ndel data","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"tr.shape, chunk.shape, sub.shape","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"### BASELINE NN ","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"C1, C2 = tf.constant(70, dtype='float32'), tf.constant(1000, dtype=\"float32\")\n#=============================#\ndef score(y_true, y_pred):\n    #K.print_tensor(y_true)\n    #K.print_tensor(y_pred)\n    tf.dtypes.cast(y_true, tf.float32)\n    tf.dtypes.cast(y_pred, tf.float32)\n    sigma = y_pred[:, 5] - y_pred[:, 1]\n    fvc_pred = y_pred[:, 3]\n    \n    #sigma_clip = sigma + C1\n    sigma_clip = tf.maximum(sigma, C1)\n    delta = tf.abs(y_true[:, 0] - fvc_pred)\n    delta = tf.minimum(delta, C2)\n    sq2 = tf.sqrt( tf.dtypes.cast(2, dtype=tf.float32) )\n    metric = (delta / sigma_clip)*sq2 + tf.math.log(sigma_clip* sq2)\n    #K.print_tensor(K.mean(metric))\n    return K.mean(metric)\n#============================#\ndef qloss(y_true, y_pred):\n    #print(K.int_shape(y_true))\n    #print(K.int_shape(y_pred))\n    # Pinball loss for multiple quantiles\n    #K.print_tensor(y_true[:1])\n    #K.print_tensor(y_pred[:1])\n    qs = [0.09, 0.159, 0.34, 0.50, 0.66, 0.841, 0.91]\n    q = tf.constant(np.array([qs]), dtype=tf.float32)\n    e = y_true - y_pred\n    #K.print_tensor(e[:1])\n    #K.print_tensor((q*e)[:1])\n    #K.print_tensor(((q-1)*e)[:1])\n    v = tf.maximum(q*e, (q-1)*e)\n    #K.print_tensor(v[:1])\n    #K.print_tensor(K.mean(v))\n    return K.mean(v)\n#=============================#\ndef mloss(_lambda):\n    def loss(y_true, y_pred):\n        return _lambda * qloss(y_true, y_pred) + (1 - _lambda)*score(y_true, y_pred)\n        #return _lambda * qloss(y_true, y_pred) + (1 - _lambda)*tf.metrics.mean_absolute_error(y_true[:,0], y_pred[:,3])\n    return loss\n#=================\ndef make_model(nh):\n    z = L.Input((nh,), name=\"Patient\")\n    x = L.Dense(100, activation=\"relu\", name=\"d1\")(z)\n    x = L.Dense(100, activation=\"relu\", name=\"d2\")(x)\n    #x = L.Dense(100, activation=\"relu\", name=\"d3\")(x)\n    p1 = L.Dense(7, activation=\"linear\", name=\"p1\")(x)\n    p2 = L.Dense(7, activation=\"relu\", name=\"p2\")(x)\n    preds = L.Lambda(lambda x: x[0] + tf.cumsum(x[1], axis=1), name=\"preds\")([p1, p2])\n    #preds = L.Lambda(lambda x: x[0], name=\"preds\")([p1, p2])\n    #preds = L.Lambda(lambda x: tf.cumsum(x[1], axis=1), name=\"preds\")([p1, p2]) \n                     \n    \n    model = M.Model(z, preds, name=\"CNN\")\n    #model.compile(loss=qloss, optimizer=\"adam\", metrics=[score])\n    model.compile(loss=mloss(0.1), optimizer=tf.keras.optimizers.Adam(lr=0.1, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.01, amsgrad=False), metrics=[score])\n    return model\n\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"y = tr['FVC'].values\nz = tr[FE].values\nze = sub[FE].values\nnh = z.shape[1]\npe = np.zeros((ze.shape[0], 7))\npred = np.zeros((z.shape[0], 7))","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"net = make_model(nh)\nprint(net.summary())\nprint(net.count_params())","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"NFOLD = 5\nkf = KFold(n_splits=NFOLD)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"%%time\ncnt = 0\nEPOCHS = 400\ntr_scores = []\nval_scores = []\nfor tr_idx, val_idx in kf.split(z):\n    cnt += 1\n    print(f\"FOLD {cnt}\")\n    net = make_model(nh)\n    net.fit(z[tr_idx], y[tr_idx], batch_size=BATCH_SIZE, epochs=EPOCHS, \n            validation_data=(z[val_idx], y[val_idx]), verbose=0) #\n    print(\"train\", net.evaluate(z[tr_idx], y[tr_idx], verbose=0, batch_size=BATCH_SIZE))\n    print(\"val\", net.evaluate(z[val_idx], y[val_idx], verbose=0, batch_size=BATCH_SIZE))\n    print(\"predict val...\")\n    pred[val_idx] = net.predict(z[val_idx], batch_size=BATCH_SIZE, verbose=0)\n    print(\"predict test...\")\n    pe += net.predict(ze, batch_size=BATCH_SIZE, verbose=0) / NFOLD\n    tr_scores.append(net.evaluate(z[tr_idx], y[tr_idx],verbose=0, batch_size=BATCH_SIZE)[1])\n    val_scores.append(net.evaluate(z[val_idx], y[val_idx],verbose=0, batch_size=BATCH_SIZE)[1])\n#==============\nprint(\"\")\nprint(\"Mean training score: \", mean(tr_scores))\nprint(\"Mean validation score: \", mean(val_scores))\nprint(\"\")\nprint(\"STD training score: \", pstdev(tr_scores))\nprint(\"STD validation score: \", pstdev(val_scores))\nprint(\"\")\nprint('Overfitting metric - mean(val) - mean(tr):', mean(val_scores) - mean(tr_scores))\nprint('Overfitting metric - STD(val) - STD(tr):', pstdev(val_scores) - pstdev(tr_scores))\nprint(\"\")","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"pe[:10]","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"sigma_opt_mae = mean_absolute_error(y, pred[:, 3])\nsigma_opt_mse = sqrt(mean_squared_error(y, pred[:, 3]))\nunc = pred[:,5] - pred[:, 1]\nunc_real = abs(y - pred[:,3])\nsigma_mean = np.mean(unc)\nprint(\"MAE, SIGmean: \")\nprint(sigma_opt_mae, sigma_mean)\nprint(\"sqrt(MSE), SIGmean: \")\nprint(sigma_opt_mse, sigma_mean)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"idxs = np.random.randint(0, y.shape[0], 100)\nplt.plot(y[idxs], label=\"ground truth\")\n#plt.figure(figsize=(100,60))\nplt.plot(pred[idxs, 0], label=\"q09\")\nplt.plot(pred[idxs, 1], label=\"q18\")\nplt.plot(pred[idxs, 2], label=\"q34\")\nplt.plot(pred[idxs, 3], label=\"q50\")\nplt.plot(pred[idxs, 4], label=\"q66\")\nplt.plot(pred[idxs, 5], label=\"q82\")\nplt.plot(pred[idxs, 6], label=\"q91\")\nplt.legend(loc=\"best\")","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"print(unc.min(), unc.mean(), unc.max(), (unc>=0).mean())\nprint(unc_real.min(), unc_real.mean(), unc_real.max(), (unc_real>=0).mean())","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"plt.hist(unc, bins=100, label='guess')\n#plt.hist(unc_real, bins=100, label='real')\nplt.title(\"uncertainty in prediction\")\nplt.show()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"### PREDICTION","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"sub.head()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"sub['FVC1'] = 0.996*pe[:, 3]\nsub['Confidence1'] = pe[:, 5] - pe[:, 1]","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"subm = sub[['Patient_Week','FVC','Confidence','FVC1','Confidence1']].copy()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"subm.loc[~subm.FVC1.isnull()].head(10)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"subm.loc[~subm.FVC1.isnull(),'FVC'] = subm.loc[~subm.FVC1.isnull(),'FVC1']\nif sigma_mean<70:\n    subm['Confidence'] = sigma_opt\nelse:\n    subm.loc[~subm.FVC1.isnull(),'Confidence'] = subm.loc[~subm.FVC1.isnull(),'Confidence1']","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"subm.head()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"subm.describe().T","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"otest = pd.read_csv('../input/osic-pulmonary-fibrosis-progression/test.csv')\nfor i in range(len(otest)):\n    subm.loc[subm['Patient_Week']==otest.Patient[i]+'_'+str(otest.Weeks[i]), 'FVC'] = otest.FVC[i]\n    subm.loc[subm['Patient_Week']==otest.Patient[i]+'_'+str(otest.Weeks[i]), 'Confidence'] = 0.1","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"subm[[\"Patient_Week\",\"FVC\",\"Confidence\"]].to_csv(\"submission.csv\", index=False)","execution_count":null,"outputs":[]}],"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}