[Python] NotImplementedError when running forward pass on pre-trained VA-VAE encoder for...

I'm working on a semantic segmentation task where I'm trying to use the encoder from a pre-trained VA-VAE model as the backbone for a UperNet architecture. My goal is to get the VA-VAE encoder to output a list of 4 multi-scale features that UperNet can use for its decoder part.

The issue I'm facing is a persistent NotImplementedError: Module [Module] is missing the required "forward" function. The error occurs during the forward pass of the VA-VAE encoder, specifically when I try to call a sub-module within a loop.

I have already implemented a custom VA_VAE class and MultiscaleEncoder to adapt the pre-trained model for this task. I believe the error points to a fundamental issue with the loaded model's internal structure—it seems that the self.down_layers list contains a module that is not correctly instantiated and therefore lacks a forward method.

❌ Error during encoder loading or testing: Module [Module] is missing the required "forward" function


❌ VA-VAE encoder 加载失败。请检查错误信息并根据提示解决问题。

❌ Encoder failed to initialize due to previous errors. Traceback (most recent call last): File "/tmp/ipython-input-666249121.py", line 276, in <cell line: 0> test_features = vae_encoder(test_img) ^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl return self._call_impl(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1784, in _call_impl return forward_call(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/tmp/ipython-input-666249121.py", line 248, in forward features = self.base_encoder(x) ^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl return self._call_impl(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1784, in _call_impl return forward_call(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/tmp/ipython-input-666249121.py", line 74, in forward current_feature = down_block(current_feature) ^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl return self._call_impl(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1784, in _call_impl return forward_call(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 399, in _forward_unimplemented raise NotImplementedError( NotImplementedError: Module [Module] is missing the required "forward" function

I've checked the code for MultiscaleEncoder and the forward method seems logically correct for iterating over a ModuleList. However, the persistent NotImplementedError suggests that the base_encoder.down list is corrupted or incorrectly populated.

Has anyone encountered a similar issue where a pre-trained model's sub-modules are missing core methods like forward after loading? Could this be related to a mismatch between the checkpoint file and the model's class definition (AutoencoderKL in autoencoder.py)?

I'm looking for guidance on how to debug this deeper, or if there's a known fix or a safer way to extract intermediate features from this specific type of VA-VAE encoder.

Continue reading...