Question

我正在尝试使用Grad-CAM方法可视化（部分重新训练的）ResNet-152的主要响应焦点。我在下面的代码行部分中花了很多心思来理解教科书，对它们进行了略微简化以方便阅读：

...

# last convolution layer
last_conv_layer = model.get_layer('res5c_branch2c')

# visualize response foci
for index in range(dataset_validation['images'].shape[0]):

    # get input 
    img  = np.expand_dims(dataset_validation['images'][index], axis=0)

    # already tried (*): tf.contrib.keras.backend.set_learning_phase(False)

    # query network response
    estimate = model.predict(img)

    # get top-3 result indicees
    t3_indicees = (np.squeeze(estimate.argsort()))[-3:][::-1]

    # loop over best 3 responses    
    for resi in range(3):

        m_out = model.output[:, t3_indicees[resi]]

        # already tried (**): k.set_learning_phase(False)

        grads = k.gradients(m_out, last_conv_layer.output)[0]
        pooled_grads = k.mean(grads, axis=(0, 1, 2))

        iterate = k.function([model.input], [pooled_grads, last_conv_layer.output[0]])

        pooled_grads_value, conv_layer_output_value = iterate([img])

...

代码在最后一行崩溃：

tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'bn_conv1/keras_learning_phase' with dtype bool
     [[Node: bn_conv1/keras_learning_phase = Placeholder[dtype=DT_BOOL, shape=<unknown>, _device="/job:localhost/replica:0/task:0/device:GPU:0"]()]]
     [[Node: strided_slice/_4465 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device_incarnation=1, tensor_name="edge_14658_strided_slice", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:CPU:0"]()]]

我已经进行了各种尝试来解决或至少更好地理解该问题，包括：

添加命令（*，**）来设置缺少的占位符值（在类似的帖子中建议）...这没有帮助
实施教科书中的玩具示例，以确保（基于VGG16的）示例正确无误……
逐步比较我的代码和玩具示例之间的输入数据格式，以发现任何明显的差异……我找不到

什么是进一步进行调试的好方法？我真的很想了解这个问题，但实际上我现在的想法已经用光了。

非常感谢您提供任何有益的建议！

使用ResNet-152的Grad-CAM导致InvalidArgumentError错误

0 个答案: