import tensorflow as tfimport numpy as npfrom sklearn.model_selection import train_test_splitnp.random.seed(4213)data = np.random.randint(low=1,high=29, size=(500, 160, 160, 10)) labels = np.random.randint(low=0,high=5, size=(500, 160, 160)) nclass = len(np.unique(labels))print (nclass)samples, width, height, nbands = data.shapeX_train, X_test, y_train, y_test = train_test_split(data, labels, test_size=0.25, random_state=421)print (X_train.shape)print (y_train.shape)arch = tf.keras.applications.VGG16(input_shape=[width, height, nbands], include_top=False, weights=None)model = tf.keras.Sequential()model.add(arch)model.add(tf.keras.layers.Flatten())model.add(tf.keras.layers.Dense(nclass))model.compile(optimizer = tf.keras.optimizers.Adam(0.0001), loss=tf.keras.losses.SparseCategoricalCrossentropy(), metrics=[tf.keras.metrics.SparseCategoricalAccuracy()]) model.fit(X_train, y_train, epochs=3, batch_size=32, verbose=2)res = model.predict(X_test)print(res.shape)當(dāng)運(yùn)行上面的代碼時(shí)，semantic segmentation我得到異常發(fā)生：InvalidArgumentError Incompatible shapes: [32,160,160] vs. [32] [[node Equal (defined at c...:38) ]] [Op:__inference_train_function_1815]tensorflow.python.framework.errors_impl.InvalidArgumentError
1 回答
烙印99
TA貢獻(xiàn)1829條經(jīng)驗(yàn) 獲得超13個(gè)贊
您的問題來自最后一層的大?。楸苊膺@些錯(cuò)誤，始終希望對(duì)N_IMAGES、WIDTH、HEIGHT和使用 python 常量）：N_CHANNELSN_CLASSES
用于圖像分類
您應(yīng)該為每張圖片分配一個(gè)標(biāo)簽。嘗試切換labels：
import tensorflow as tf
import numpy as np
from sklearn.model_selection import train_test_split
np.random.seed(4213)
N_IMAGES, WIDTH, HEIGHT, N_CHANNELS = (500, 160, 160, 10)
N_CLASSES = 5
data = np.random.randint(low=1,high=29, size=(N_IMAGES, WIDTH, HEIGHT, N_CHANNELS))
labels = np.random.randint(low=0,high=N_CLASSES, size=(N_IMAGES))
#...
用于語義分割
確保您的分類器（網(wǎng)絡(luò)的最后一層）大小相應(yīng)。在這種情況下，每個(gè)像素需要 1 個(gè)類別：
#...
model = tf.keras.Sequential()
model.add(arch)
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(width * height))
model.add(tf.keras.layers.Reshape([width , height]))
#...
這是您可以獲得的最簡單的方法。相反，您可以設(shè)置多個(gè)反卷積層作為分類器，或者您甚至可以翻轉(zhuǎn)架構(gòu)arch并使用它來生成分類結(jié)果。正交地，您可以one_hot對(duì)標(biāo)簽執(zhí)行編碼，從而將它們擴(kuò)展一個(gè)因子N_CLASSES，有效地乘以最后一層中的神經(jīng)元數(shù)量。
添加回答
舉報(bào)
0/150
提交
取消