使用Keras 2的ResNet50出现NaN损失值

Question

自从升级到Keras 2后，我在尝试微调ResNet50时发现出现了nan损失。如果我使用单个卷积层（如下所示的注释）而不是ResNet，损失和准确率看起来都没问题。我是否错过了Keras 2中的一些变化？

from keras.applications.resnet50 import ResNet50
from keras.layers import Flatten, Dense, Input, Conv2D, Activation, Flatten
from keras.layers.pooling import MaxPooling2D
from keras.models import Model
from keras.optimizers import SGD
import numpy as np

inp = Input(batch_shape=(32, 224, 224, 3), name='input_image')

### resnet
modelres = ResNet50(weights="imagenet", include_top=False, input_tensor=inp)
x = modelres.output
x = Flatten()(x)

### single convolutional layer
#x = Conv2D(32, (3,3))(inp)
#x = Activation('relu')(x)
#x = MaxPooling2D(pool_size=(3,3))(x)
#x = Flatten()(x)
#x = Dense(units=32)(x)
predictions = Dense(units=2, kernel_initializer="he_normal", activation="softmax")(x) 

model = Model(inputs=inp, outputs=predictions)
model.compile(SGD(lr=.001, momentum=0.9), "categorical_crossentropy", metrics=["accuracy"])

# generate images of all ones with the same label
def gen():
    while True:
        x_data = np.ones((32,224,224,3)).astype('float32')
        y_data = np.zeros((32,2)).astype('float32')
        y_data[:,1]=1.0
        yield x_data, y_data

model.fit_generator(gen(), 10, validation_data=gen(), validation_steps=1)

model.summary()的开头和结尾如下所示：

____________________________________________________________________________________________________
Layer (type)                     Output Shape          Param #     Connected to
====================================================================================================
input_image (InputLayer)         (32, 224, 224, 3)     0
____________________________________________________________________________________________________
zero_padding2d_1 (ZeroPadding2D) (32, 230, 230, 3)     0
____________________________________________________________________________________________________
conv1 (Conv2D)                   (32, 112, 112, 64)    9472

...

avg_pool (AveragePooling2D)      (32, 1, 1, 2048)      0
____________________________________________________________________________________________________
flatten_1 (Flatten)              (32, 2048)            0
____________________________________________________________________________________________________
dense_1 (Dense)                  (32, 2)               4098
====================================================================================================

训练输出为:

Epoch 1/1
10/10 [==============================] - 30s - loss: nan - acc: 0.0000e+00 - val_loss: nan - val_acc: 0.0000e+00

- Chris K

---

这些 nan 是从一开始就出现还是在某些批次/时期之后出现的？ - Marcin Możejko

从一开始 - Chris K

你的输入大小是多少？ - Marcin Możejko

尝试将np.zeros也作为x输入。 - Marcin Możejko

你能打印出 model.summary() 吗？是的，它看起来非常奇怪。 - Marcin Możejko

Answer 1

当我将后端切换到tensorflow而不是theano时，一切正常。看起来keras 2中的theano实现出了问题。

- Chris K