使用TensorFlow Serving部署Keras构建的TensorFlow模型时的拼接层错误

Hey there, let's work through this AbortionError you're facing when calling TensorFlow Serving's predict interface. The core issue here is a dimension mismatch in a concatenation operation, so let's break down what's happening and how to fix it.

What the Error Means #

The error message points directly to the lys_conc/concat node:

ConcatOp : Expected concatenating dimensions in the range [-1, 1), but got 1

This tells us that the ConcatV2 operation is trying to concatenate tensors along axis 1, but TensorFlow expects the concatenation axis to fall within the range [-1, 1) (valid integer values here are -1 or 0). Either the axis specified in your model's concat layer is incorrect, or the input tensors you're sending don't have the right shape to support concatenation along axis 1.

Step-by-Step Fixes #

Here are the key areas to investigate:

• Verify Input Tensor Shapes vs. Model Expectations
  Double-check that every tensor you're sending in your predict request matches the shape the model was trained with. For example, if your model expects an input of shape [batch_size, 28, 28] but you're sending a flattened tensor of shape [batch_size, 784], this can throw off downstream operations like concatenation. Use this command to list your model's expected input shapes and names:

  saved_model_cli show --dir <your_model_dir> --tag_set serve --signature_def serving_default
  

• Inspect the Concatenation Layer in Your Model
  Locate the lys_conc/concat node in your model graph. Check what axis it's configured to concatenate along. If the axis is set to 1, confirm that all four input tensors listed in the error (_arg_lys_in_0_4, _arg_lyb_in_0_0, flatten_1/Reshape, batch_normalization_2/batchnorm_1/add_1) have a valid second dimension (axis 1) and that their sizes along this axis are identical. If one of the tensors is 1-dimensional (e.g., shape [batch_size] instead of [batch_size, N]), concatenating along axis 1 will fail.

• Validate Input Names and Count
  Ensure that the input tensor names in your predict request exactly match the names the model expects (case-sensitive!). The error lists four inputs being passed to the concat operation—make sure you're sending all required inputs, and that none are missing or misnamed.

• Test Local Inference First
  Run a prediction directly using TensorFlow (without TensorFlow Serving) to see if the error reproduces. If it works locally but fails with TF Serving, the issue might be related to model export or version compatibility. For example, if you trained with TensorFlow 2.x, ensure you exported the model using the correct SavedModel format, and that your TF Serving version supports that format.

• Check Tensor Compatibility for Concatenation
  For concatenation to work, all tensors must have the same rank (number of dimensions) except along the concatenation axis. If one tensor has fewer dimensions than the others, you'll need to reshape it to match the rank of the others before concatenation (e.g., adding a dummy dimension with tf.expand_dims()).

Example Scenario #

Suppose one of your input tensors is a 1D tensor of shape [batch_size], while the others are 2D tensors of shape [batch_size, 10]. Trying to concatenate along axis 1 will fail because the 1D tensor doesn't have a second dimension. Fix this by reshaping the 1D tensor to [batch_size, 1] before concatenation, or adjust the concat axis to 0 if that makes sense for your model logic.

内容的提问来源于stack exchange，提问作者willycs40