我正在嘗試獲取通過 Vgg16 網(wǎng)絡(luò)傳遞的圖像（訓(xùn)練和驗(yàn)證）的輸出include_top = false，然后添加最后幾層，如下面的代碼所示。我想x存儲(chǔ)完整的模型，以便我可以從中創(chuàng)建一個(gè) tflite 文件（包括 vgg 和我添加的圖層）from tensorflow.keras.models import Modelimport osx= vgg16.outputprint(x.shape)x = GlobalAveragePooling2D()(x)x = Flatten()(x)x = Dense(100)(x)x = tf.keras.layers.LeakyReLU(alpha=0.2)(x)x = (Dropout(0.5)) (x)x = (Dense(50)) (x) x = tf.keras.layers.LeakyReLU(alpha=0.3)(x)x = Dropout(0.3)(x)x = Dense(num_classes, activation='softmax')(x)# this is the model we will trainmodel = Model(inputs=vgg16.input, outputs=x)# first: train only the top layers (which were randomly initialized)# i.e. freeze all convolutional InceptionV3 layersfor layer in vgg16.layers: layer.trainable = Falsemodel.compile(loss='categorical_crossentropy', optimizer=optimizers.RMSprop(lr=1e-4), metrics=['acc'])# train the model on the new data for a few epochshistory = model.fit(train_data, train_labels, epochs=15, batch_size=batch_size, validation_data=(validation_data, validation_labels))model.save(top_model_weights_path)(eval_loss, eval_accuracy) = model.evaluate( validation_data, validation_labels, batch_size=batch_size, verbose=1)的輸出x.shape是 (?, ?, ?, 512)train_data.shape (1660, 2, 2, 512)train_labels.shape (1660, 4)validation_data.shape (137, 4)validation_labels.shape (137, 2, 2, 512)錯(cuò)誤：ValueError：檢查輸入時(shí)出錯(cuò)：預(yù)期 input_3 具有形狀 (None, None, 3) 但得到形狀為 (2, 2, 512) 的數(shù)組此錯(cuò)誤發(fā)生在以下行：52 驗(yàn)證數(shù)據(jù)=（驗(yàn)證數(shù)據(jù)，驗(yàn)證標(biāo)簽））如下所示的先前代碼片段工作得非常好，并提供準(zhǔn)確的輸出。train_data存儲(chǔ)一個(gè) numpy 數(shù)組vgg16.predict_generator()
1 回答
子衿沉夜
TA貢獻(xiàn)1828條經(jīng)驗(yàn) 獲得超3個(gè)贊
出色地...
您定義了target_size=(img_width, img_height)，如果(img_width, img_height)不是(224, 224)，那么您還需要target_size在 VGG 模型中進(jìn)行定義：
vgg16 = applications.VGG16(
include_top=False,
weights='imagenet',
target_size=(img_width, img_height, 3))
你為什么用class_mode=None在datagen.flow_from_directory？None是默認(rèn)值。如果你想讓它明確write class_mode='categorical'，但 usingclass_mode=None根本沒有意義。
predict_generator返回預(yù)測(cè)?，F(xiàn)在predict_generator已棄用，但您可以使用predict它與生成器一起工作。但predict應(yīng)在訓(xùn)練后使用。正確使用生成器的方法是：
datagen = ImageDataGenerator(rescale=1. / 255)
generator = datagen.flow_from_directory(
train_data_dir,
target_size=(img_width, img_height),
batch_size=batch_size,
shuffle=False)
// ...
history = model.fit(
generator,
epochs=15,
steps_per_epoch=len(generator),
batch_size=batch_size,
validation_data=validation_generator,
validation_steps=len(validation_generator))
稍后，如果您想進(jìn)行預(yù)測(cè)，請(qǐng)使用：model.predict(test_generator)
在這種情況下，您不需要使用Flattenafter 。將輸出減少到一維數(shù)組。GlobalAveragePooling2DGlobalAveragePooling2D
添加回答
舉報(bào)
0/150
提交
取消