RuntimeError: could not execute a primitive error one of the solutions

[vaguest]

Using the flask library to implement the web service, you can replace the fastapi library, which solves the problem.

[GreatV]

The error "RuntimeError: could not execute a primitive" appears while training the ParseQ algorithm in PaddleOCR.

From the provided GitHub issue (Issue #13951), the problem seems to stem from an index-related issue in the rec_parseq_head.py implementation. Specifically, at line 498, an invalid indexing operation is being used, likely on a scalar where an array is expected.

✅ Possible Solutions:

1. Check for Scalar Indexing Issue

   • The bug might be occurring because targets[0] or targets[1] is a scalar value instead of a tensor or list. Try printing their shapes before the line where the error occurs:

     print("targets[0] shape:", targets[0].shape if hasattr(targets[0], 'shape') else type(targets[0]))
     print("targets[1] shape:", targets[1].shape if hasattr(targets[1], 'shape') else type(targets[1]))

   • If they are scalars instead of tensors, modify the code to ensure correct indexing.

2. Modify Data Conversion

   • If the issue is related to tensor conversion, explicitly convert targets to a tensor using paddle.to_tensor before indexing.

3. Using Flask Instead of FastAPI

   • The user suggested replacing FastAPI with Flask when implementing the web service. If the error occurs in the API handling logic, testing with Flask can help determine whether the issue is FastAPI-specific.

4. Ensure Data Is Properly Loaded

   • According to the issue description, the dataset is stored as an LMDBDataSet. The correctness of the LMDB files should be verified.
   • Run:

     import os
     print(os.listdir("/mnt/workspace/workgroup/sukunming/code/parseq/data/train/synth"))

     Ensure data.mdb and lock.mdb exist.

5. Try Running with a Smaller Batch Size

   • batch_size_per_card: 192 might be too high for your environment. Reduce it to 64 or 32 and re-test:

     Train:
       dataset:
         batch_size_per_card: 64

---

🔗 Relevant Issue:

• The full conversation and debugging details can be found in the GitHub Issue #13951.

If none of these solutions resolve the issue, consider reporting it directly on PaddleOCR's GitHub with a reproducible example. 🚀

---

Response generated by 🤖 feifei-bot | chatgpt-4o-latest

⚠️ This is an AI-generated response. Please verify any critical information.