DenseNet与Keras UnknownError:未能得到卷积算法

Question

我想在GPU中运行这段代码

tensorflow-gpu:2.3.1

Cuda版本: 10.2

结果: UnknownError:未能得到卷积算法。这可能是因为cuDNN未能初始化，所以尝试查看上面是否打印了警告日志消息。[节点泛函_1/conv/conv2D(定义为:129) ]

函数调用堆栈: train_function

你能帮帮我吗？

先谢谢你

def build_model():
#include_top: whether to include the fully-connected layer at the top of the network.
#weights: one of None (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded.
    base_model = densenet.DenseNet121(input_shape= (128, 128, 3),
                                     weights=None, 
                                     include_top=True,
                                     pooling='avg', classes=3,)
    #Layers & models also feature a boolean attribute trainable. Its value can be changed. Setting layer.trainable to False moves all the layer's weights from trainable to non-trainable. This is called "freezing" the layer: the state of a frozen layer won't be updated during training (either when
    #training with fit() or when training with any custom loop that relies on trainable_weights to apply gradient updates)
    #pour modifier les poids lors de trainement 
    for layer in base_model.layers:
        layer.trainable = True 
        
# définir les parametres nécessaires 
#La régularisation est une technique qui apporte de légères modifications à l'algorithme d'apprentissage de sorte que le 
#modèle se généralise mieux. Cela améliore également les performances du modèle sur les données invisibles.
    x = base_model.output
    x = Dense(50, kernel_regularizer=regularizers.l1_l2(0.00001), activity_regularizer=regularizers.l2(0.00001))(x)
    x = Activation('relu')(x)
    x = Dense(25, kernel_regularizer=regularizers.l1_l2(0.00001), activity_regularizer=regularizers.l2(0.00001))(x)
    x = Activation('relu')(x)
    predictions = Dense(n_classes, activation='softmax')(x)
    model = Model(inputs=base_model.input, outputs=predictions)
    return model

model = build_model()


# l'utlisation de Adam optimizer + sparese _sparse_categorical_crossentropy
keras.optimizers.Adam(learning_rate=0.001)

#optimizer = Adam(lr=0.01)
#model.compile(loss='sparse_categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['acc'])
#Early stopping is a method that allows you to specify an arbitrary large number of training epochs and stop training once the
#model performance stops improving on a hold out validation dataset
early_stop = EarlyStopping(monitor='val_loss', patience=8, verbose=2, min_delta=1e-3)
#Reduce learning rate when a metric has stopped improving.
reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=4, verbose=1, min_delta=1e-3)
callbacks_list = [early_stop, reduce_lr]
print(" Build model --- %s seconds ---" % (time.time() - start_time))
print('###################### training step #############')
trainy = keras.utils.to_categorical(trainy)
yvalidation = keras.utils.to_categorical(yvalidation)


with tf.device('/device:GPU:0'):
    trainx = tf.constant(trainx)
    trainy = tf.constant(trainy)
    xvalidation = tf.constant(xvalidation)
    yvalidation = tf.constant(yvalidation)
    model_history = model.fit(trainx, trainy,
          validation_data=(xvalidation, yvalidation),
          batch_size=68, 
          epochs=7000,
          verbose=1)

Answer 1

您的CuDNN+Cuda+TensorFlow版本不匹配。为了解决您的问题，请参考我的答案：Tensorflow 2.0 can't use GPU, something wrong in cuDNN? :Failed to get convolution algorithm. This is probably because cuDNN failed to initialize

在您的例子中，很可能TensorFlow 2.3.1不支持Cuda 10.2。您应该降级到Cuda 10.1，然后再试一次。