what is the version of fairseq ?

[King-king424]

[King-king424]

CoGenAV/models/cogenav.py", line 130, in init
self.encoder = TransformerEncoder(cfg)

fairseq/fairseq/models/wav2vec/wav2vec2.py", line 1020, in init
if pos_conv_depth > 1:
TypeError: '>' not supported between instances of 'NoneType' and 'int'

---

I encountered the above error

[HarukiK2]

I encountered the same issue, and although I’m not one of the developers, I was able to get it working by adding the following line to config/base.yaml or config/large.yaml:

pos_conv_depth: 1

However, I’ve noticed that the performance is still lower than the values reported in the paper, which is somewhat concerning. Just wanted to share this in case it helps with further debugging or configuration checks.

Also, for reference, the version of fairseq I’m using is 0.12.2. Just in case this affects the behavior or performance.

[HarukiK2]

Additionally, I ran inference on the VSR task and obtained a score of 24.2 with the Large model and 29.8 with the Base model. Both are noticeably worse than the values reported in the paper (22.0 for Large and 24.8 for Base), which I find concerning.

[HarukiK2]

Hi @OdingdongO,

Thank you for releasing this work.

As mentioned in my earlier comment, I was able to run inference by manually setting pos_conv_depth: 1 in the config file. However, I’ve observed that the VSR performance is significantly worse than what was reported in the paper:

• Base model: I got 29.8 (vs. 24.8 in the paper)
• Large model: I got 24.2 (vs. 22.0 in the paper)

Given that these numbers are notably higher (i.e., worse), I wanted to ask if you could clarify a few points:

• Are there any additional config settings, scripts, or preprocessing steps that are required to reproduce the reported results?
• Could this gap be due to differences in the model checkpoints or fairseq version?
• Is it possible the released checkpoints differ in any way from those used for the results in the paper?

I’d really appreciate any guidance you could provide to help us understand the performance gap.

Best,
HarukiK2