0

我有一个具有以下架构的预训练模型。

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, 200)               0         
_________________________________________________________________
embedding (Embedding)        (None, 200, 300)          402877800 
_________________________________________________________________
spatial_dropout1d (SpatialDr (None, 200, 300)          0         
_________________________________________________________________
bidirectional (Bidirectional (None, 128)               186880    
_________________________________________________________________
dropout (Dropout)            (None, 128)               0         
_________________________________________________________________
batch_normalization_v1 (Batc (None, 128)               512       
_________________________________________________________________
dense (Dense)                (None, 6)                 774       
_________________________________________________________________
reshape (Reshape)            (None, 6)                 0         
=================================================================
Total params: 403,065,966
Trainable params: 187,910
Non-trainable params: 402,878,056
_________________________________________________________________

该Reshape层确保logits与标签相同

我有一个输入数据，它是Tensors 形状为 (200,) 的张量流对象和标签，它们也是一个形状为 (6,) 的“张量”张量流对象。

我的目标evaluate是使用tensorflow keras 提供的方法评估单个样本。为了更容易，我将Tensor对象转换为 numpy. 由于模型接受 的输入形状[None, 200]，因此我必须在将输入数据输入模型之前对其进行整形。这样，模型将具有形状为 (1,6) 的 logit，然后该Reshape层将重新整形为 (6,)。

# Clone model is the keras model
# sample_data.x is the input
# sample_data.y is the label

clone_model.evaluate([sample_data.x.numpy().reshape(1,200)], [sample_data.y.numpy()])

但最后，我收到以下错误

InvalidArgumentError: logits and labels must have the same first dimension, got logits shape [1,6] and labels shape [6]
     [[{{node loss_4/dense_loss/SparseSoftmaxCrossEntropyWithLogits/SparseSoftmaxCrossEntropyWithLogits}}]] [Op:StatefulPartitionedCall]

为了克服这个问题，我移除了Reshape图层并将标签重塑为 (1,6) ( sample_data.y.numpy().reshape(1,6))。但这并没有帮助，我最终遇到了同样的错误。

我想知道是否有人可以指出我在这里遗漏了什么？提前致谢。

4

0 回答 0