标题: pytorch to rknn transform error [打印本页] 作者: rakidis 时间: 2020-4-30 18:30 标题: pytorch to rknn transform error Hello,
I am trying to convert a pytorch model to rknn. Even though I didnt see any documentation on it, I found out my model needs to be in torch jit format. So I converted my mode to torch jit and saved it using "trace" from torch.jit
Here is my model file in the correct format for rknn toolkit as far as I am aware. https://mylifebox.com/shr/f45eed ... 93f&language=en
When I try to convert it using rknn toolkit visualition gui with 256x256 pictures in dataset.txt file, then 3x256x256 in the input node spot, I have received the following error.
"
-> Config model
-> Loading pytorch model
/home/kadir/anaconda3/envs/rknn/lib/python3.6/site-packages/onnx_tf/common/__init__.py:87: UserWarning: FrontendHandler.get_outputs_names is deprecated. It will be removed in future release .. Use node .outputs instead.
warnings.warn (message)
/home/kadir/Desktop/rknn-toolkit-v1.3.0/test_resimleri/net_trace.pt ********************
WARNING: Token 'COMMENT' defined, but not used
WARNING: There is 1 unused token
Syntax error in input! LexToken (NAMED_IDENTIFIER, 'str', 8,406)
D import clients finished
2020-04-30 11: 53: 06.641320: I tensorflow / core / platform / cpu_feature_guard.cc: 141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2020-04-30 11: 53: 06.733713: I tensorflow / stream_executor / cuda / cuda_gpu_executor.cc: 964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2020-04-30 11: 53: 06.734282: I tensorflow / core / common_runtime / gpu / gpu_device.cc: 1411] Found device 0 with properties:
name: GeForce GTX 1050 Ti major: 6 minor: 1 memoryClockRate (GHz): 1.62
pciBusID: 0000: 01: 00.0
totalMemory: 3.95GiB freeMemory: 3.27GiB
2020-04-30 11: 53: 06.734298: I tensorflow / core / common_runtime / gpu / gpu_device.cc: 1490] Adding visible gpu devices: 0
2020-04-30 11: 53: 07.405678: I tensorflow / core / common_runtime / gpu / gpu_device.cc: 971] Device interconnect StreamExecutor with strength 1 edge matrix:
2020-04-30 11: 53: 07.405702: I tensorflow / core / common_runtime / gpu / gpu_device.cc: 977] 0
2020-04-30 11: 53: 07.405710: I tensorflow / core / common_runtime / gpu / gpu_device.cc: 990] 0: N
2020-04-30 11: 53: 07.406076: I tensorflow / core / common_runtime / gpu / gpu_device.cc: 1103] Created TensorFlow device (/ job: localhost / replica: 0 / task: 0 / device: GPU: 0 with 2981 MB memory)-> physical GPU (device: 0, name: GeForce GTX 1050 Ti, pci bus id: 0000: 01: 00.0, compute capability: 6.1)
I Build net_trace complete.
D Optimizing network with force_1d_tensor, swapper, merge_layer, auto_fill_bn, resize_nearest_transformer, auto_fill_multiply, merge_avgpool_conv1x1, auto_fill_zero_bias, proposal_opt_import
D Optimizing network with conv2d_big_kernel_size_transform
-> Load pytorch model succeed!
-> Model process
W Genreate input meta fail, please check model.
W External input meta file "/tmp/tmp78uchgmi/net_trace_inputmeta.yml" is not exists.
Process Process-1: 2:
Traceback (most recent call last):
File "/home/kadir/anaconda3/envs/rknn/lib/python3.6/multiprocessing/process.py", line 258, in _bootstrap
self.run ()
File "/home/kadir/anaconda3/envs/rknn/lib/python3.6/multiprocessing/process.py", line 93, in run
self._target (* self._args, ** self._kwargs)
File "/home/kadir/anaconda3/envs/rknn/lib/python3.6/site-packages/rknn/visualization/server/rknn_func.py", line 260, in start_quantization_convert
ret = rknn.hybrid_quantization_step1 (dataset = dataset)
File "/home/kadir/anaconda3/envs/rknn/lib/python3.6/site-packages/rknn/api/rknn.py", line 266, in hybrid_quantization_step1
ret = self.rknn_base.hybrid_quantization_step1 (dataset = dataset)
File "rknn / api / rknn_base.py", line 773, in rknn.api.rknn_base.RKNNBase.hybrid_quantization_step1
File "rknn / api / rknn_base.py", line 2249, in rknn.api.rknn_base.RKNNBase._generate_inputmeta
IndexError: list index out of range
"
Could you shed some light on why this error could be occuring? What are my possible solutions to avoid this type of error.
-By the way, the model I used doesn't have a default input size. It will work with different input sizes and its output size will be ([1,69,2],[1,69]), but I used 3x256x256 because rknn toolkit asks for input node size.
It's been a long question, I tried to explain it the best I could.
Regards, Kadir
作者: tigertang 时间: 2020-5-1 16:35
It is a notorious that the RK team hasn't fixed. I have been bothered by this "list index out of range" for lots of times. There should be some problems with your model (maybe some ops are not supported or some op combinations are not supported). However, under this situation, basically you can't do anything since no further model compatibility modification instruction is given by the output log. So the conclusion is: now let's wait for further lib update.作者: jefferyzhang 时间: 2020-5-1 21:24
Support for pytorch is just beginning from 1.3.0, there are many issues with it, you may try the lastest(1.3.2) rknn-toolkit again with your problem, which has resolved many known issues in 1.3.0.
You can find the latest version here: https://github.com/rockchip-linux/rknn-toolkit 作者: rakidis 时间: 2020-5-5 16:06
I installed rknn 1.3.2 but the results are same. Same error persists
Am I missing something here? Could you examine the .pt file I uploaded to see if you observe the same behaviour?作者: jefferyzhang 时间: 2020-5-5 16:52