0

I'm trying to visualize important regions for a classification task with CNN.

I'm using VGG16 + my own top layers (A global average pooling layer and a Dense layer)

model_vgg16_conv = VGG16(weights='imagenet', include_top=False, input_shape=(100, 100, 3))

model = models.Sequential()

model.add(model_vgg16_conv)

model.add(Lambda(global_average_pooling, output_shape=global_average_pooling_shape))
model.add(Dense(4, activation = 'softmax', init='uniform'))

After compiling and fitting the model I'm trying to use Grad-CAM for a new image:

image = cv2.imread("data/example_images/test.jpg")
# Resize to 100x100
image = resize(image,(100,100),anti_aliasing=True, mode='constant')
# Because it's a grey scale image extend the dimensions
image = np.repeat(image.reshape(1,100, 100, 1), 3, axis=3)

class_weights = model.get_layer("dense_1").get_weights()[0]
final_conv_layer = model.get_layer("vgg16").get_layer("block5_conv3")
input1 = model.get_layer("vgg16").layers[0].input
output1 = model.get_layer("dense_1").output
get_output = K.function([input1], [final_conv_layer.output, output1])

After that I'm executing

[conv_outputs, predictions] = get_output([image])

Leading to the following error:

InvalidArgumentError: You must feed a value for placeholder tensor 'vgg16_input' with dtype float and shape [?,100,100,3] [[{{node vgg16_input}}]] [[dense_1/Softmax/_233]]

Additional information

def global_average_pooling(x):
    return K.mean(x, axis = (2, 3))

def global_average_pooling_shape(input_shape):
    return input_shape[0:2]

Model summary:

Layer (type)                 Output Shape              Param #   
=================================================================
vgg16 (Model)                (None, 3, 3, 512)         14714688  
_________________________________________________________________
lambda_1 (Lambda)            (None, 3)                 0         
_________________________________________________________________
dense_1 (Dense)              (None, 4)                 16        
=================================================================
Total params: 14,714,704
Trainable params: 16
Non-trainable params: 14,714,688

VGG-Model Summary:

Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, 100, 100, 3)       0   
...

I'm new to Grad-CAM and I'm not sure if I'm just overseeing something or if I misunderstood the entire concept.

4

1 回答 1

2

使用 Sequential，使用 add() 方法添加层。在这种情况下，由于直接添加了模型对象，因此模型现在有两个输入 - 一个通过 Sequential，另一个通过 model_vgg16_conv。

>>> layer = model.layers[0]
>>> layer.get_input_at(0)
<tf.Tensor 'input_1:0' shape=(?, ?, ?, 3) dtype=float32>
>>> layer.get_input_at(1)
<tf.Tensor 'vgg16_input:0' shape=(?, ?, ?, 3) dtype=float32>

由于使用 K.function，只提供了一个输入，因此出现了关于缺少“vgg16_input”输入的错误。这会奏效，

get_output = K.function([input1] + [model.input], [final_conv_layer.output, output1])

[conv_outputs, predictions] = get_output([image, image])

但是函数式 API 可以像这样在这种情况下使用：

model_vgg16_conv = VGG16(weights='imagenet', include_top=False, input_shape=(100, 100, 3))
gavg = Lambda(global_average_pooling, output_shape=global_average_pooling_shape)(model_vgg16_conv.output)
output = Dense(4, activation = 'softmax', init='uniform')(gavg)
model_f = Model(model_vgg16_conv.input, output)

final_conv_layer = model_f.get_layer("block5_conv3")
get_output = K.function([model_f.input], [final_conv_layer.output, model_f.output])
[conv_outputs, predictions] = get_output([image])
于 2019-04-02T06:33:35.197 回答