{"cells":[{"metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"cell_type":"code","source":"import numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\nfrom statistics import mean\nimport gc\n\n\nfrom sklearn.preprocessing import StandardScaler,OneHotEncoder\n\nimport os\nfrom tqdm import tqdm\nimport random\n\nimport warnings\nwarnings.filterwarnings('ignore')\n","execution_count":25,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"from keras import backend as K\nK.tensorflow_backend._get_available_gpus()","execution_count":3,"outputs":[{"output_type":"stream","text":"Using TensorFlow backend.\n","name":"stderr"},{"output_type":"execute_result","execution_count":3,"data":{"text/plain":"['/job:localhost/replica:0/task:0/device:GPU:0']"},"metadata":{}}]},{"metadata":{"trusted":true},"cell_type":"code","source":"from keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Conv1D, MaxPooling1D, GlobalAveragePooling1D, Input, CuDNNLSTM, Flatten\nfrom keras.optimizers import Adam\nfrom keras.losses import mean_squared_error\nfrom keras.callbacks import History","execution_count":4,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"train_data = pd.read_csv('../input/train.csv',dtype = {'acoustic_data':np.float32,'time_to_failure':np.float32})\n\n\n","execution_count":10,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"rows = 150000\nsegments = int(np.floor(train_data.shape[0] / rows))\n\nX_train = np.zeros((segments,150000))\ny_train = pd.DataFrame(index = range(segments),dtype = np.float32,columns = ['time_to_failure'])\n\nfor segment in tqdm(range(segments)):\n x = train_data.iloc[segment*rows:segment*rows+rows]\n y = x['time_to_failure'].values[-1]\n x = x['acoustic_data'].values\n y_train.loc[segment,'time_to_failure'] = y\n X_train[segment] = x\ndel train_data","execution_count":15,"outputs":[{"output_type":"stream","text":"100%|██████████| 4194/4194 [00:04<00:00, 864.75it/s]\n","name":"stderr"}]},{"metadata":{"trusted":true},"cell_type":"code","source":"gc.collect()","execution_count":29,"outputs":[{"output_type":"execute_result","execution_count":29,"data":{"text/plain":"252"},"metadata":{}}]},{"metadata":{"trusted":true},"cell_type":"code","source":"X_train.shape\ny_train['time_to_failure'] = round(y_train['time_to_failure'])","execution_count":19,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"y_train['time_to_failure'] = y_train['time_to_failure'].astype(np.int32)","execution_count":22,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"ohe = OneHotEncoder()\ny_train = ohe.fit_transform(np.array(y_train['time_to_failure']).reshape(-1,1))","execution_count":28,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"y_train.shape","execution_count":32,"outputs":[{"output_type":"execute_result","execution_count":32,"data":{"text/plain":"(4194, 17)"},"metadata":{}}]},{"metadata":{"trusted":true},"cell_type":"code","source":"model = Sequential()\nmodel.add(Conv1D(filters=40, kernel_size=20, strides=2, activation='relu', input_shape=(150000,1)))\nmodel.add(MaxPooling1D(3))\nmodel.add(Conv1D(filters=40, kernel_size=20, strides=1, activation='relu'))\nmodel.add(MaxPooling1D(3))\nmodel.add(Conv1D(filters=40, kernel_size=20, strides=1, activation='relu'))\nmodel.add(MaxPooling1D(3))\nmodel.add(CuDNNLSTM(8,return_sequences=True))\nmodel.add(CuDNNLSTM(8,return_sequences=True))\n#model.add(Flatten())\nmodel.add(Conv1D(filters=40, kernel_size=10, strides=1, activation='relu'))\nmodel.add(GlobalAveragePooling1D())\nmodel.add(Dropout(rate=0.1))\nmodel.add(Dense(17,activation = 'softmax'))\nprint(model.summary())","execution_count":35,"outputs":[{"output_type":"stream","text":"_________________________________________________________________\nLayer (type) Output Shape Param # \n=================================================================\nconv1d_7 (Conv1D) (None, 74991, 40) 840 \n_________________________________________________________________\nmax_pooling1d_4 (MaxPooling1 (None, 24997, 40) 0 \n_________________________________________________________________\nconv1d_8 (Conv1D) (None, 24978, 40) 32040 \n_________________________________________________________________\nmax_pooling1d_5 (MaxPooling1 (None, 8326, 40) 0 \n_________________________________________________________________\nconv1d_9 (Conv1D) (None, 8307, 40) 32040 \n_________________________________________________________________\nmax_pooling1d_6 (MaxPooling1 (None, 2769, 40) 0 \n_________________________________________________________________\ncu_dnnlstm_3 (CuDNNLSTM) (None, 2769, 8) 1600 \n_________________________________________________________________\ncu_dnnlstm_4 (CuDNNLSTM) (None, 2769, 8) 576 \n_________________________________________________________________\nconv1d_10 (Conv1D) (None, 2760, 40) 3240 \n_________________________________________________________________\nglobal_average_pooling1d_2 ( (None, 40) 0 \n_________________________________________________________________\ndropout_2 (Dropout) (None, 40) 0 \n_________________________________________________________________\ndense_2 (Dense) (None, 17) 697 \n=================================================================\nTotal params: 71,033\nTrainable params: 71,033\nNon-trainable params: 0\n_________________________________________________________________\nNone\n","name":"stdout"}]},{"metadata":{"trusted":true},"cell_type":"code","source":"model.compile(loss='categorical_crossentropy',optimizer='adam')","execution_count":36,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"model.fit(X_train.reshape(-1,150000,1),y_train,epochs = 100, validation_split = 0.1,batch_size = 16)","execution_count":null,"outputs":[{"output_type":"stream","text":"Train on 3774 samples, validate on 420 samples\nEpoch 1/100\n3774/3774 [==============================] - 50s 13ms/step - loss: 2.6606 - val_loss: 2.5530\nEpoch 2/100\n1312/3774 [=========>....................] - ETA: 30s - loss: 2.6503","name":"stdout"}]},{"metadata":{"trusted":true},"cell_type":"markdown","source":"def prepareAd(x):\n x = np.sign(x)*np.log(1 + np.sqrt(np.abs(x)))/4.4\n return x"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"def getTrainBatch(dfl,nrows,batch_size):\n x = np.empty([batch_size,15000,1])\n y = np.empty([batch_size,15000])\n for i,rn in enumerate(np.random.randint(nrows-15000, size=batch_size)):\n df = dfl.loc[rn:rn+14999,:]\n x[i,:,0] = df.acoustic_data.values\n y[i,:] = df.time_to_failure.values\n return x,y"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"random.seed(42)\nhistory = History()\nstep = 150000000\nstop = 600000000\n#sc = StandardScaler()\nfor i in range(0, stop, step):\n print('Reading file chunk :',i)\n train_df = pd.read_csv(\"../input/train.csv\",\n skiprows = i,\n nrows = step,\n dtype={'acoustic_data': np.int16, 'time_to_failure': np.float32}\n )\n train_df.columns = ['acoustic_data','time_to_failure']\n #train_df.acoustic_data =prepareAd(train_df.acoustic_data.values)\n #train_df.time_to_failure = train_df.time_to_failure/16\n \n loss = []\n val_loss = []\n mae = []\n for j in range(20):\n #print('Generating training batch :',j)\n x_train,y_train = getTrainBatch(train_df,step,batch_size=1024)\n history = model.fit(x_train,\n y_train,\n batch_size=16,\n epochs=10,\n validation_split=0.1,\n verbose=0)\n loss = loss + history.history['loss']\n val_loss = val_loss + history.history['val_loss']\n #mae = mae + history.history['mean_absolute_error']\n if (j%5==0):\n print('loss :',mean(loss[-10:]),' val_loss :',mean(val_loss[-10:])) #, ' val_mae :',mean(mae[-10:])*16)\n del x_train, y_train\n gc.collect()\n del train_df\n gc.collect()"},{"metadata":{"trusted":true},"cell_type":"code","source":"def predictSubmission(seg_id):\n test_df = pd.read_csv('../input/test/' + seg_id + '.csv')\n #y = model.predict(prepareAd(test_df.acoustic_data.values).reshape(1,150000,1))*16\n #x = sc.fit_transform(test_df.acoustic_data.values.reshape(-1,1))\n x = test_df.acoustic_data.values\n y = model.predict(x.reshape(1,150000,1))\n return np.argmax(y)","execution_count":38,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"submission = pd.read_csv('../input/sample_submission.csv')\nsubmission['time_to_failure']=submission['seg_id'].apply(predictSubmission)\nsubmission.to_csv('submission_8.csv',index=False)","execution_count":39,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"submission.head()","execution_count":40,"outputs":[{"output_type":"execute_result","execution_count":40,"data":{"text/plain":" seg_id time_to_failure\n0 seg_00030f 1\n1 seg_0012b5 1\n2 seg_00184e 1\n3 seg_003339 1\n4 seg_0042cc 1","text/html":"\n\n
\n \n \n | \n seg_id | \n time_to_failure | \n
\n \n \n \n | 0 | \n seg_00030f | \n 1 | \n
\n \n | 1 | \n seg_0012b5 | \n 1 | \n
\n \n | 2 | \n seg_00184e | \n 1 | \n
\n \n | 3 | \n seg_003339 | \n 1 | \n
\n \n | 4 | \n seg_0042cc | \n 1 | \n
\n \n
\n