[SAR] Exported model still throws “shape size 4 ≠ perm size 3” / “Conv expects 4-D” — leftover transpose2 after fine-tuning (Paddle 2.6.2)

[AliShafiee2003]

Hi. I'm having trouble with inference. It seems like different layers have different dimensions as input.

1. Environment

item	value
PaddlePaddle	2.6.2-gpu-cuda12.0-cudnn8.9-trt8.6 Docker image (paddlepaddle/paddle:2.6.2-gpu-cuda12.0-cudnn8.9-trt8.6)
cuDNN on host	8.8 (so runtime warns about 8.9 vs 8.8 mismatch)
GPU	RTX 4090 (compute 8.9)
PaddleOCR	current develop (June 2025) cloned 2 weeks ago
Model	SAR / ResNet31 fine-tuned on Persian dataset, starting from official English SAR weights
Checkpoint	output/rec/fa_r31_sar_finetune/best_accuracy.pdparams
Extra	custom extended dictionary combined_dict.txt

---

2. Training YAML (excerpt)

Architecture:
  model_type: rec
  algorithm: SAR
  Backbone: { name: ResNet31 }
  Head:     { name: SARHead, max_text_length: 30 }

Train.transforms[-1].SARRecResizeImg.image_shape: [3, 64, 64, 256]
PostProcess:
  name: SARLabelDecode
  character_dict_path: /workspace/DataSet/rec/combined_dict.txt

So H=64 and max W=256 during training.

---

3. Export steps tried

A. first export (default IR-optim on)

python3 tools/export_model.py \
  -c configs/rec/rec_fa_train.yml \
  -o Global.checkpoints=output/rec/fa_r31_sar_finetune/best_accuracy \
     Global.save_inference_dir=inference/fa_r31_sar

Inference with a trivial script (NCHW input) gave:

InvalidArgument: shape should have the same dim with perm, but received shape size is:4, perm size is:3.
  [operator < transpose2 > error]

If I feed CHW instead, I get:

InvalidArgument: The input of Op(Conv) should be a 4-D or 5-D Tensor. But received: input's dimension is 3, input's shape is [3, 64, 256].

So whichever shape I send, one of the two ops fails.

B. predictor with IR-optim off

Added cfg.switch_ir_optim(False) in the inference script → same errors.

C. re-export with Global.ir_optim=False

python3 tools/export_model.py \
  -c configs/rec/rec_fa_train.yml \
  -o Global.checkpoints=output/rec/fa_r31_sar_finetune/best_accuracy \
     Global.save_inference_dir=inference/fa_r31_sar_noopt \
     Global.ir_optim=False

Log shows “Ir optimization is turned off, no ir pass will be executed.”
But the new model still throws exactly the same two errors depending on 4-D vs 3-D input.

---

4. Minimal inference script

import cv2, numpy as np, paddle.inference as pi
cfg = pi.Config("inference/fa_r31_sar_noopt/inference.pdmodel",
                "inference/fa_r31_sar_noopt/inference.pdiparams")
cfg.switch_ir_optim(False)
pred = pi.create_predictor(cfg)

img = cv2.imread("line_000001.png")
...  # resize to 64×≤256, norm
x = img.transpose(2,0,1)         # CHW  OR  x[None] -> NCHW
ratio = np.array([x.shape[-1] / 256], "float32")

pred.get_input_handle("x").copy_from_cpu(x)          # or x[None]
pred.get_input_handle("data_0").copy_from_cpu(ratio) # valid_ratio
pred.run()

• If x is NCHW → transpose2 perm mismatch (shape 4 vs 3).
• If x is CHW → first conv says “input must be 4-D”.

---

5. Suspected cause

A stray transpose2 op with perm=[2,1,0] is still present before the first conv2d.
So the graph expects:

x (CHW) → transpose2 (CHW) → unsqueeze2 (→ NCHW) → conv2d (NCHW)

but my export contains both transpose2 and unsqueeze2.
This makes it impossible to satisfy both ops at once.

Several issues (#8970, #9631, #10934) mention the same leftover transpose after SAR export.

---

6. What I need

• Is this a known bug in SAR export on Paddle 2.6.x?
• Is there a recommended flag (other than ir_optim=False) to avoid inserting that 3-D transpose?
• Any safe way to drop or bypass the offending transpose2 without manually editing the graph?
• What recommendations do you have if I have to make a change manually?

Thanks for any guidance!

[cuicheng01]

Hello, the performance of this model is not very good, and it is indeed not one of the high-priority models maintained by PaddleOCR at the moment. We highly recommend that you use the PP-OCRv5 configuration for model fine-tuning.

[AliShafiee2003]

Thank you very much.
I intend to expand the Persian dictionary of one of the models.
Ultimately, I plan to fine-tune a bilingual Persian-English model.
I have also prepared a relatively large dataset.
Could you please advise me on which architecture would be most suitable?