Cityscapes SOTA 模型导出报错'Config' object has no attribute 'model'

[Siiiiiigma]

问题确认 Search before asking

• 我已经查询历史issue(包括open与closed)，没有发现相似的bug。I have searched the open and closed issues and found no similar bug report.

Bug描述 Describe the Bug

python export.py --config configs/mscale_ocr_cityscapes_autolabel_mapillary.yml --save_dir ./output --input_shape 1 3 2048 1024
按照readme要求在制定位置下载了模型参数和预训练参数，使用以上命令导出预训练的模型网络时，出现以下报错
尝试了历史issue中提到的几种方法，例如通过源码安装开发版paddleseg，问题仍然存在
利用飞浆ai studio的notebook也同样存在此问题，和配置环境应该无关

报错内容
d:\deeplearning\paddleseg\paddleseg\cvlibs\manager.py:113: UserWarning: MscaleOCRNet exists already! It is now updated to <class 'models.mscale_ocrnet.MscaleOCRNet'> !!!
warnings.warn("{} exists already! It is now updated to {} !!!".
Traceback (most recent call last):
File "D:\DeepLearning\PaddleSeg\contrib\CityscapesSOTA\export.py", line 140, in
main(args)
File "D:\DeepLearning\PaddleSeg\contrib\CityscapesSOTA\export.py", line 84, in main
net = cfg.model
AttributeError: 'Config' object has no attribute 'model'

复现环境 Environment

paddlepaddle-gpu 2.4.2.post117
paddleseg 2.8.0 d:\deeplearning\paddleseg

Bug描述确认 Bug description confirmation

• 我确认已经提供了Bug复现步骤、代码改动说明、以及环境信息，确认问题是可以复现的。I confirm that the bug replication steps, code change instructions, and environment information have been provided, and the problem can be reproduced.

是否愿意提交PR？ Are you willing to submit a PR?

• 我愿意提交PR！I'd like to help by submitting a PR!

[Asthestarsfalll]

@Siiiiiigma 你好，这应该是一个bug，问题在于CityscapesSOTA使用了paddleseg中的模块，而后续paddleseg更新时没有及时修改。可以尝试使用更早之前的版本，稍后我将会修复这个问题。

[Asthestarsfalll]

@Siiiiiigma 我已经提交了一个PR，你可以尝试克隆我的修改试试

[Siiiiiigma]

@Asthestarsfalll
感谢修复，我尝试导出第一个配置（mscale_ocr_cityscapes_autolabel_mapillary.yml）时，出现如下警告，请问是正常的吗？
(Paddle) D:\DeepLearning\PaddleSeg\contrib\CityscapesSOTA>python export.py --config configs/mscale_ocr_cityscapes_autolabel_mapillary.yml --save_dir ./output --input_shape 1 3 2048 1024
d:\deeplearning\paddleseg\paddleseg\cvlibs\manager.py:113: UserWarning: MscaleOCRNet exists already! It is now updated to <class 'models.mscale_ocrnet.MscaleOCRNet'> !!!
warnings.warn("{} exists already! It is now updated to {} !!!".
2023-07-10 16:43:06 [WARNING] Add the in_channels in train_dataset class to model config. We suggest you manually set in_channels in model config.
2023-07-10 16:43:06 [INFO] Use the following config to build model
model:
backbone:
in_channels: 3
type: HRNet_W48_NV
backbone_indices:

0
n_scales:
0.5
1.0
2.0
num_classes: 19
pretrained: pretrain/pretrained.pdparams
type: MscaleOCRNet
W0710 16:43:06.020490 7732 gpu_resources.cc:61] Please NOTE: device: 0, GPU Compute Capability: 7.5, Driver API Version: 12.1, Runtime API Version: 11.7
W0710 16:43:06.046422 7732 gpu_resources.cc:91] device: 0, cuDNN Version: 8.4.
2023-07-10 16:43:10 [INFO] Loading pretrained model from pretrain/pretrained.pdparams
2023-07-10 16:43:13 [WARNING] [SKIP] Shape of pretrained params ocrnet.head.cls_head.weight doesn't match.(Pretrained: (65, 512, 1, 1), Actual: [19, 512, 1, 1])
2023-07-10 16:43:13 [WARNING] [SKIP] Shape of pretrained params ocrnet.head.cls_head.bias doesn't match.(Pretrained: (65,), Actual: [19])
2023-07-10 16:43:13 [WARNING] [SKIP] Shape of pretrained params ocrnet.head.aux_head.1.weight doesn't match.(Pretrained: (65, 720, 1, 1), Actual: [19, 720, 1, 1])
2023-07-10 16:43:13 [WARNING] [SKIP] Shape of pretrained params ocrnet.head.aux_head.1.bias doesn't match.(Pretrained: (65,), Actual: [19])
2023-07-10 16:43:13 [WARNING] scale_attn.atten_head.0._conv.weight is not in pretrained model
2023-07-10 16:43:13 [WARNING] scale_attn.atten_head.0._batch_norm.weight is not in pretrained model
2023-07-10 16:43:13 [WARNING] scale_attn.atten_head.0._batch_norm.bias is not in pretrained model
2023-07-10 16:43:13 [WARNING] scale_attn.atten_head.0._batch_norm._mean is not in pretrained model
2023-07-10 16:43:13 [WARNING] scale_attn.atten_head.0._batch_norm._variance is not in pretrained model
2023-07-10 16:43:13 [WARNING] scale_attn.atten_head.1._conv.weight is not in pretrained model
2023-07-10 16:43:13 [WARNING] scale_attn.atten_head.1._batch_norm.weight is not in pretrained model
2023-07-10 16:43:13 [WARNING] scale_attn.atten_head.1._batch_norm.bias is not in pretrained model
2023-07-10 16:43:13 [WARNING] scale_attn.atten_head.1._batch_norm._mean is not in pretrained model
2023-07-10 16:43:13 [WARNING] scale_attn.atten_head.1._batch_norm._variance is not in pretrained model
2023-07-10 16:43:13 [WARNING] scale_attn.atten_head.2.weight is not in pretrained model
2023-07-10 16:43:14 [INFO] There are 1572/1587 variables loaded into MscaleOCRNet.
2023-07-10 16:43:48 [INFO] The inference model is saved in ./output

[Asthestarsfalll]

@Siiiiiigma
第一处警告是因为MscaleOCRNet在paddleseg.model中被注册过了，会在CityscapesSOTA重新注册一遍，没有影响。
第二处pretrained params是因为线性层的权重形状不一致，预训练的head通道数和微调不一致也很正常，没有影响。
第三处scale_attn的警告是因为你加载的是预训练权重，所以不存在scale_attn这个模块，deploy应该加载在下游任务训练好的权重。

[Siiiiiigma]

谢谢，明白了，修改为加载之前下载的saved_model/model.pdparams之后就没有警告了

[Siiiiiigma]

@Asthestarsfalll
你好，我想测试该模型在任意街景图上的效果，准备了一张2048*1024的JPG图像，放在image文件夹内，当我在飞桨ai studio的notebook下运行以下命令时：
python deploy/python/infer.py
--config /home/aistudio/PaddleSeg-2.6.0/output/deploy.yaml
--image_path /home/aistudio/PaddleSeg-2.6.0/image
--save_dir /home/aistudio/PaddleSeg-2.6.0/result

出现了如下报错：
2023-07-10 18:47:58 [INFO] Use GPU
--- Running analysis [ir_graph_build_pass]
I0710 18:48:00.875998 2513 executor.cc:187] Old Executor is Running.
--- Running analysis [ir_analysis_pass]
--- Running IR pass [map_op_to_another_pass]
--- Running IR pass [identity_scale_op_clean_pass]
--- Running IR pass [is_test_pass]
--- Running IR pass [simplify_with_basic_ops_pass]
--- Running IR pass [delete_quant_dequant_linear_op_pass]
--- Running IR pass [delete_weight_dequant_linear_op_pass]
--- Running IR pass [constant_folding_pass]
--- Running IR pass [silu_fuse_pass]
--- Running IR pass [conv_bn_fuse_pass]
--- Running IR pass [conv_eltwiseadd_bn_fuse_pass]
--- Running IR pass [embedding_eltwise_layernorm_fuse_pass]
--- Running IR pass [multihead_matmul_fuse_pass_v2]
--- Running IR pass [vit_attention_fuse_pass]
--- Running IR pass [fused_multi_transformer_encoder_pass]
--- Running IR pass [fused_multi_transformer_decoder_pass]
--- Running IR pass [fused_multi_transformer_encoder_fuse_qkv_pass]
--- Running IR pass [fused_multi_transformer_decoder_fuse_qkv_pass]
--- Running IR pass [multi_devices_fused_multi_transformer_encoder_pass]
--- Running IR pass [multi_devices_fused_multi_transformer_encoder_fuse_qkv_pass]
--- Running IR pass [multi_devices_fused_multi_transformer_decoder_fuse_qkv_pass]
--- Running IR pass [fuse_multi_transformer_layer_pass]
--- Running IR pass [gpu_cpu_squeeze2_matmul_fuse_pass]
--- Running IR pass [gpu_cpu_reshape2_matmul_fuse_pass]
--- Running IR pass [gpu_cpu_flatten2_matmul_fuse_pass]
--- Running IR pass [gpu_cpu_map_matmul_v2_to_mul_pass]
--- Running IR pass [gpu_cpu_map_matmul_v2_to_matmul_pass]
--- Running IR pass [matmul_scale_fuse_pass]
--- Running IR pass [multihead_matmul_fuse_pass_v3]
--- Running IR pass [gpu_cpu_map_matmul_to_mul_pass]
--- Running IR pass [fc_fuse_pass]
--- Running IR pass [fc_elementwise_layernorm_fuse_pass]
--- Running IR pass [conv_elementwise_add_act_fuse_pass]
--- Running IR pass [conv_elementwise_add2_act_fuse_pass]
--- Running IR pass [conv_elementwise_add_fuse_pass]
I0710 18:48:47.483732 2513 fuse_pass_base.cc:59] --- detected 12 subgraphs
--- Running IR pass [transpose_flatten_concat_fuse_pass]
--- Running IR pass [conv2d_fusion_layout_transfer_pass]
--- Running IR pass [transfer_layout_elim_pass]
--- Running IR pass [auto_mixed_precision_pass]
--- Running IR pass [inplace_op_var_pass]
I0710 18:48:47.669679 2513 fuse_pass_base.cc:59] --- detected 3 subgraphs
--- Running analysis [save_optimized_model_pass]
W0710 18:48:47.685402 2513 save_optimized_model_pass.cc:28] save_optim_cache_model is turned off, skip save_optimized_model_pass
--- Running analysis [ir_params_sync_among_devices_pass]
I0710 18:48:47.685453 2513 ir_params_sync_among_devices_pass.cc:51] Sync params from CPU to GPU
--- Running analysis [adjust_cudnn_workspace_size_pass]
--- Running analysis [inference_op_replace_pass]
--- Running analysis [memory_optimize_pass]
I0710 18:48:50.664584 2513 memory_optimize_pass.cc:222] Cluster name : shape_28.tmp_0_slice_0 size: 8
I0710 18:48:50.664654 2513 memory_optimize_pass.cc:222] Cluster name : shape_0.tmp_0_slice_0 size: 8
I0710 18:48:50.664659 2513 memory_optimize_pass.cc:222] Cluster name : concat_1.tmp_0 size: -2147483648
I0710 18:48:50.664661 2513 memory_optimize_pass.cc:222] Cluster name : transpose_0.tmp_0 size: 1073741824
I0710 18:48:50.664664 2513 memory_optimize_pass.cc:222] Cluster name : relu_78.tmp_0 size: 50331648
I0710 18:48:50.664673 2513 memory_optimize_pass.cc:222] Cluster name : batch_norm_305.tmp_2 size: 1509949440
I0710 18:48:50.664676 2513 memory_optimize_pass.cc:222] Cluster name : batch_norm_196.tmp_2 size: 50331648
I0710 18:48:50.664680 2513 memory_optimize_pass.cc:222] Cluster name : relu_227.tmp_0 size: 12582912
I0710 18:48:50.664685 2513 memory_optimize_pass.cc:222] Cluster name : batch_norm_200.tmp_2 size: 25165824
I0710 18:48:50.664688 2513 memory_optimize_pass.cc:222] Cluster name : x size: 25165824
I0710 18:48:50.664702 2513 memory_optimize_pass.cc:222] Cluster name : relu_171.tmp_0 size: 25165824
I0710 18:48:50.664711 2513 memory_optimize_pass.cc:222] Cluster name : batch_norm_930.tmp_1 size: 768
I0710 18:48:50.664716 2513 memory_optimize_pass.cc:222] Cluster name : concat_0.tmp_0 size: 1509949440
I0710 18:48:50.664718 2513 memory_optimize_pass.cc:222] Cluster name : tmp_310 size: 3145728
I0710 18:48:50.664721 2513 memory_optimize_pass.cc:222] Cluster name : bilinear_interp_v2_35.tmp_0 size: 76
--- Running analysis [ir_graph_to_program_pass]
I0710 18:48:51.751169 2513 analysis_predictor.cc:1660] ======= optimize end =======
I0710 18:48:51.776242 2513 naive_executor.cc:164] --- skip [feed], feed -> x
I0710 18:48:51.808507 2513 naive_executor.cc:164] --- skip [argmax_0.tmp_0], fetch -> fetch
W0710 18:48:51.966293 2513 gpu_resources.cc:119] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.2, Runtime API Version: 11.6
W0710 18:48:51.974512 2513 gpu_resources.cc:149] device: 0, cuDNN Version: 8.4.
Traceback (most recent call last):
File "/home/aistudio/PaddleSeg-2.6.0/deploy/python/infer.py", line 430, in
main(args)
File "/home/aistudio/PaddleSeg-2.6.0/deploy/python/infer.py", line 418, in main
predictor.run(imgs_list)
File "/home/aistudio/PaddleSeg-2.6.0/deploy/python/infer.py", line 375, in run
self.predictor.run()
ValueError: (InvalidArgument) The 2-th dimension of input[0] and input[1] is expected to be equal.But received input[0]'s shape = [1, 512, 1024, 512], input[1]'s shape = [1, 512, 512, 1024].
[Hint: Expected inputs_dims[0][j] == inputs_dims[i][j], but received inputs_dims[0][j]:1024 != inputs_dims[i][j]:512.] (at ../paddle/phi/kernels/funcs/concat_funcs.h:83)
[operator < concat > error]

请问是我输入数据的形状问题吗，还是模型的问题？

[Asthestarsfalll]

@Siiiiiigma 应该是输入数据的形状问题

[Siiiiiigma]

https://github.com/PaddlePaddle/PaddleSeg/blob/release/2.8/docs/deployment/inference/python_inference_cn.md
我使用该链接提供的cityscapes_demo.png仍然报同样的问题，感觉不像是形状的问题，是我漏了什么预处理步骤吗

[Asthestarsfalll]

https://github.com/PaddlePaddle/PaddleSeg/blob/release/2.8/docs/deployment/inference/python_inference_cn.md 我使用该链接提供的cityscapes_demo.png仍然报同样的问题，感觉不像是形状的问题，是我漏了什么预处理步骤吗

看报错是模型内部concat时tensor形状不一样，使用develop分支试试呢？

[Siiiiiigma]

@Asthestarsfalll
我在本地使用了源码安装的开发者版本（2.8.0），以及在ai studio使用notebook提供的2.6.0版本，且均使用cityscapes_demo.png测试，该问题仍然存在，报错位置相同，请检查一下模型内部是否存在bug