解決keras使用cov1D函數的輸入問題

解決了以下錯誤：

1.ValueError: Input 0 is incompatible with layer conv1d_1: expected ndim=3, found ndim=4

2.ValueError: Error when checking target: expected dense_3 to have 3 dimensions, but got array with …

1.ValueError: Input 0 is incompatible with layer conv1d_1: expected ndim=3, found ndim=4

錯誤代碼：

model.add(Conv1D(8, kernel_size=3, strides=1, padding='same', input_shape=(x_train.shape))

或者

model.add(Conv1D(8, kernel_size=3, strides=1, padding='same', input_shape=(x_train.shape[1:])))

這是因為模型輸入的維數有誤，在使用基於tensorflow的keras中，cov1d的input_shape是二維的，應該：

1、reshape x_train的形狀

x_train=x_train.reshape((x_train.shape[0],x_train.shape[1],1))
 x_test = x_test.reshape((x_test.shape[0], x_test.shape[1],1))

2、改變input_shape

model = Sequential()
 model.add(Conv1D(8, kernel_size=3, strides=1, padding='same', input_shape=(x_train.shape[1],1)))

大神原文：

The input shape is wrong, it should be input_shape = (1, 3253) for Theano or (3253, 1) for TensorFlow. The input shape doesn't include the number of samples.

Then you need to reshape your data to include the channels axis:

x_train = x_train.reshape((500000, 1, 3253))

Or move the channels dimension to the end if you use TensorFlow. After these changes it should work.

2.ValueError: Error when checking target: expected dense_3 to have 3 dimensions, but got array with …

出現此問題是因為ylabel的維數與x_train x_test不符，既然將x_train x_test都reshape了，那麼也需要對y進行reshape。

解決辦法：

同時對照x_train改變ylabel的形狀

t_train=t_train.reshape((t_train.shape[0],1))
 t_test = t_test.reshape((t_test.shape[0],1))

附：

修改完的代碼：
 

  import warnings  warnings.filterwarnings("ignore")  import os  os.environ["CUDA_VISIBLE_DEVICES"] = "0"    import pandas as pd  import numpy as np  import matplotlib  # matplotlib.use('Agg')  import matplotlib.pyplot as plt    from sklearn.model_selection import train_test_split  from sklearn import preprocessing    from keras.models import Sequential  from keras.layers import Dense, Dropout, BatchNormalization, Activation, Flatten, Conv1D  from keras.callbacks import LearningRateScheduler, EarlyStopping, ModelCheckpoint, ReduceLROnPlateau  from keras import optimizers  from keras.regularizers import l2  from keras.models import load_model  df_train = pd.read_csv('./input/train_V2.csv')  df_test = pd.read_csv('./input/test_V2.csv')  df_train.drop(df_train.index[[2744604]],inplace=True)#去掉nan值  df_train["distance"] = df_train["rideDistance"]+df_train["walkDistance"]+df_train["swimDistance"]  # df_train["healthpack"] = df_train["boosts"] + df_train["heals"]  df_train["skill"] = df_train["headshotKills"]+df_train["roadKills"]  df_test["distance"] = df_test["rideDistance"]+df_test["walkDistance"]+df_test["swimDistance"]  # df_test["healthpack"] = df_test["boosts"] + df_test["heals"]  df_test["skill"] = df_test["headshotKills"]+df_test["roadKills"]    df_train_size = df_train.groupby(['matchId','groupId']).size().reset_index(name='group_size')  df_test_size = df_test.groupby(['matchId','groupId']).size().reset_index(name='group_size')    df_train_mean = df_train.groupby(['matchId','groupId']).mean().reset_index()  df_test_mean = df_test.groupby(['matchId','groupId']).mean().reset_index()    df_train = pd.merge(df_train, df_train_mean, suffixes=["", "_mean"], how='left', on=['matchId', 'groupId'])  df_test = pd.merge(df_test, df_test_mean, suffixes=["", "_mean"], how='left', on=['matchId', 'groupId'])  del df_train_mean  del df_test_mean    df_train = pd.merge(df_train, df_train_size, how='left', on=['matchId', 'groupId'])  df_test = pd.merge(df_test, df_test_size, how='left', on=['matchId', 'groupId'])  del df_train_size  del df_test_size    target = 'winPlacePerc'  train_columns = list(df_test.columns)  """ remove some columns """  train_columns.remove("Id")  train_columns.remove("matchId")  train_columns.remove("groupId")  train_columns_new = []  for name in train_columns:   if '_' in name:    train_columns_new.append(name)  train_columns = train_columns_new  # print(train_columns)    X = df_train[train_columns]  Y = df_test[train_columns]  T = df_train[target]    del df_train  x_train, x_test, t_train, t_test = train_test_split(X, T, test_size = 0.2, random_state = 1234)    # scaler = preprocessing.MinMaxScaler(feature_range=(-1, 1)).fit(x_train)  scaler = preprocessing.QuantileTransformer().fit(x_train)    x_train = scaler.transform(x_train)  x_test = scaler.transform(x_test)  Y = scaler.transform(Y)  x_train=x_train.reshape((x_train.shape[0],x_train.shape[1],1))  x_test = x_test.reshape((x_test.shape[0], x_test.shape[1],1))  t_train=t_train.reshape((t_train.shape[0],1))  t_test = t_test.reshape((t_test.shape[0],1))    model = Sequential()  model.add(Conv1D(8, kernel_size=3, strides=1, padding='same', input_shape=(x_train.shape[1],1)))  model.add(BatchNormalization())  model.add(Conv1D(8, kernel_size=3, strides=1, padding='same'))  model.add(Conv1D(16, kernel_size=3, strides=1, padding='valid'))  model.add(BatchNormalization())  model.add(Conv1D(16, kernel_size=3, strides=1, padding='same'))  model.add(Conv1D(32, kernel_size=3, strides=1, padding='valid'))  model.add(BatchNormalization())  model.add(Conv1D(32, kernel_size=3, strides=1, padding='same'))  model.add(Conv1D(32, kernel_size=3, strides=1, padding='same'))  model.add(Conv1D(64, kernel_size=3, strides=1, padding='same'))  model.add(Activation('tanh'))  model.add(Flatten())  model.add(Dropout(0.5))  # model.add(Dropout(0.25))  model.add(Dense(512,kernel_initializer='he_normal', activation='relu', W_regularizer=l2(0.01)))  model.add(Dense(128,kernel_initializer='he_normal', activation='relu', W_regularizer=l2(0.01)))  model.add(Dense(1, kernel_initializer='normal', activation='sigmoid'))    optimizers.Adam(lr=0.01, epsilon=1e-8, decay=1e-4)    model.compile(optimizer=optimizer, loss='mse', metrics=['mae'])  model.summary()    ng = EarlyStopping(monitor='val_mean_absolute_error', mode='min', patience=4, verbose=1)  # model_checkpoint = ModelCheckpoint(filepath='best_model.h5', monitor='val_mean_absolute_error', mode = 'min', save_best_only=True, verbose=1)  # reduce_lr = ReduceLROnPlateau(monitor='val_mean_absolute_error', mode = 'min',factor=0.5, patience=3, min_lr=0.0001, verbose=1)  history = model.fit(x_train, t_train,       validation_data=(x_test, t_test),       epochs=30,       batch_size=32768,       callbacks=[early_stopping],       verbose=1)predict(Y)  pred = pred.ravel()

補充知識：Keras Conv1d 參數及輸入輸出詳解

Conv1d(in_channels，out_channels，kernel_size，stride=1，padding=0，dilation=1，groups=1，bias=True)

filters:卷積核的數目(即輸出的維度)

kernel_size: 整數或由單個整數構成的list/tuple，卷積核的空域或時域窗長度

strides: 整數或由單個整數構成的list/tuple，為卷積的步長。任何不為1的strides均為任何不為1的dilation_rata均不兼容

padding: 補0策略，為”valid”，”same”或”casual”，”casual”將產生因果(膨脹的)卷積，即output[t]不依賴於input[t+1:]。當對不能違反事件順序的時序信號建模時有用。“valid”代表只進行有效的卷積，即對邊界數據不處理。“same”代表保留邊界處的卷積結果，通常會導致輸出shape與輸入shape相同。

activation:激活函數，為預定義的激活函數名,或逐元素的Theano函數。如果不指定該函數，將不會使用任何激活函數(即使用線性激活函數:a(x)=x)
 

  model.add(Conv1D(filters=nn_params["input_filters"],        kernel_size=nn_params["filter_length"],        strides=1,        padding='valid',        activation=nn_params["activation"],        kernel_regularizer=l2(nn_params["reg"])))

例：輸入維度為（None，1000，4）

第一維度：None

第二維度：

output_length = int((input_length - nn_params["filter_length"] + 1))

在此情況下為：

output_length = （1000 + 2*padding - filters +1）/ strides = （1000 + 2*0 -32 +1）/1 = 969

第三維度：filters