jasper pytorch模型直接加载或转onnx或转tf模型均不成功

rknn  v1.3.0
pytorch v1.2
tensorflow  v1.14onnx  v1.4.1
onnx-tf  v1.3.0  

代码及模型
.pt是原始模型
链接：https://pan.baidu.com/s/10xjD5cuDKNkN_c-AWlxksg
提取码：1okw

1. .pt模型是可以正确运行的;
   但rknn加载pytorch模型提示错误:

1. E Catch exception when loading pytorch model: toml_fp32_acoustic.pt!
   
2. E Traceback (most recent call last):
   
3. E File "rknn\api\rknn_base.py", line 567, in rknn.api.rknn_base.RKNNBase.load_pytorch
   
4. E File "rknn\base\RKNNlib\app\importer\import_pytorch.py", line 95, in rknn.base.RKNNlib.app.importer.import_pytorch.ImportPytorch.run
   
5. E File "rknn\base\RKNNlib\converter\convert_pytorch.py", line 517, in rknn.base.RKNNlib.converter.convert_pytorch.convert_pytorch.__init__
   
6. E File "rknn\base\RKNNlib\converter\convert_pytorch.py", line 601, in rknn.base.RKNNlib.converter.convert_pytorch.convert_pytorch.model_simplify
   
7. E File "rknn\base\RKNNlib\converter\convert_pytorch.py", line 104, in rknn.base.RKNNlib.converter.convert_pytorch.torch_inference_engine.shape_pick
   
8. ……   类似重复n次
   
9. E File "rknn\base\RKNNlib\converter\convert_pytorch.py", line 104, in rknn.base.RKNNlib.converter.convert_pytorch.torch_inference_engine.shape_pick
   
10. E File "rknn\base\RKNNlib\converter\convert_pytorch.py", line 139, in rknn.base.RKNNlib.converter.convert_pytorch.torch_inference_engine.__ir_shape_inference
    
11. E File "rknn\base\RKNNlib\converter\convert_pytorch.py", line 251, in rknn.base.RKNNlib.converter.convert_pytorch.torch_inference_engine.convolution_shape
    
12. E File "rknn\base\RKNNlib\converter\convert_pytorch.py", line 242, in rknn.base.RKNNlib.converter.convert_pytorch.torch_inference_engine._pad_pick
    
13. E IndexError: list index out of range
    
14. Traceback (most recent call last):

复制代码
2. .onnx模型也可以正确运行,直接使用onnx-tf运行结果与.pt结果一致  (onnx-tf需要更新为v1.3.0)
  onnx转为.rknn 未量化, 可运行,但结果错误输出为全0
3.  通过onnx转为tf模型然后再转为.rknn, 运行提示模型错误失败:

1. --> Init runtime environment
   
2. rknn run target self
   
3. I npu_transfer_proxy pid: 748, status: sleeping
   
4. D NPUTransfer: efuseid = 3399:4632a3468cf39f3b
   
5. I NPUTransfer: Starting NPU Transfer Client, Transfer version 1.9.8 (cab3961@2019-12-12T09:54:26)
   
6. D NPUTransfer: Transfer spec = local:transfer_proxy
   
7. D NPUTransfer: Transfer interface successfully opened, fd = 4
   
8. E RKNNAPI: rknn_init, msg_load_ack fail, ack = 1(ACK_FAIL), expect 0(ACK_SUCC)!
   
9. E RKNNAPI: ==============================================
   
10. E RKNNAPI: RKNN VERSION:
    
11. E RKNNAPI: API: 1.3.0 (c5654ea build: 2019-12-25 12:40:55)
    
12. E RKNNAPI: DRV: 1.3.0 (c4f8c23 build: 2019-11-25 10:39:29)
    
13. E RKNNAPI: ==============================================
    
14. D NPUTransfer: Transfer client closed, fd = 4
    
15. E Catch exception when init runtime!
    
16. E Traceback (most recent call last):
    
17. E File "rknn/api/rknn_base.py", line 988, in rknn.api.rknn_base.RKNNBase.init_runtime
    
18. E File "rknn/api/rknn_runtime.py", line 320, in rknn.api.rknn_runtime.RKNNRuntime.build_graph
    
19. E Exception: RKNN init failed. error code: RKNN_ERR_MODEL_INVALID
    
20. Init runtime environment failed

复制代码


代码:

1. 
   
2. import platform
   
3. import os
   
4. import torch
   
5. import numpy as np
   
6. 
   
7. from rknn.api import RKNN
   
8. from timeit import default_timer as timer
   
9. 
   
10. import onnx
    
11. from onnx_tf.backend import prepare
    
12. 
    
13. def rm(file):
    
14.     if os.path.exists(file):
    
15.         os.remove(file)
    
16. 
    
17. if __name__ == '__main__':
    
18. 
    
19.     rknn = RKNN(verbose=True)
    
20.     rknn.config(batch_size=1,
    
21.                 epochs=1)  # asymmetric_quantized-u8,  quantized_dtype='dynamic_fixed_point-16' , channel_mean_value='0 0 0 1', reorder_channel='0 1 2',
    
22.     system = platform.system()
    
23. 
    
24.     seq_len = 64
    
25.     isize = 672
    
26.     pt_file = 'toml_fp32_acoustic.pt'
    
27.     m = torch.jit.load(pt_file)
    
28.     m.eval()
    
29.     #x = torch.rand((1, seq_len, isize))
    
30.     x = torch.linspace(-1, 1, steps=(isize * seq_len))
    
31.     x = x.view(1, 1, seq_len, -1)
    
32.     input_size_list = [[1, 1, seq_len, isize]]
    
33.     #x1 = torch.full((1, hsize), fill_value=0.2, dtype=torch.float)
    
34.     ix = x.numpy()
    
35.     t0 = timer()
    
36.     out = m.forward(x)
    
37.     t1 = timer()
    
38.     print("torch out, time:", t1 - t0)
    
39.     print(out.detach().numpy())
    
40. 
    
41.     use_pytorch = False
    
42.     if use_pytorch:
    
43.         rknn.load_pytorch(model=pt_file, input_size_list=input_size_list)
    
44.     else:
    
45.         onnx_file = pt_file + '.onnx'
    
46.         #onnx_file = 'toml_fp32_acoustic.pt_op10.onnx'
    
47.         tf_model = onnx.load(onnx_file)
    
48.         tf_rep = prepare(tf_model)
    
49. 
    
50.         t0 = timer()
    
51.         tf_out = tf_rep.run(ix)
    
52.         t1 = timer()
    
53.         tf_out = np.array(tf_out)
    
54.         print("onnx out:", ix.shape, ', ', tf_out.shape, " time:", t1 - t0)
    
55.         print(tf_out)
    
56.         rknn.load_onnx(model=onnx_file)
    
57. 
    
58.         # onnx转为tf .pb并加载
    
59.         # pb_file = onnx_file + '.pb'
    
60.         # rm(pb_file)
    
61.         # tf_rep.export_graph(pb_file)
    
62.         # input_size_list = [[1, seq_len, isize]]
    
63.         # rknn.load_tensorflow(tf_pb=pb_file,
    
64.         #                      inputs=['input'],
    
65.         #                      outputs=['transpose_326'],
    
66.         #                      input_size_list=input_size_list)
    
67. 
    
68.     is_quant = False
    
69.     if is_quant:
    
70.         rknn_file = pt_file + '_quant.rknn'
    
71.     else:
    
72.         rknn_file = pt_file + '.rknn'
    
73. 
    
74.     print('--> Building model')
    
75.     ret = rknn.build(do_quantization=is_quant, dataset='./dataset.txt')
    
76.     if ret != 0:
    
77.         print('rknn.build failed!')
    
78.         exit(ret)
    
79.     print('build done')
    
80.     rm(rknn_file)
    
81.     rknn.export_rknn(rknn_file)
    
82. 
    
83.     print('--> Init runtime environment')
    
84.     if system == 'Windows':
    
85.         print("rknn run target rk1808")
    
86.         ret = rknn.init_runtime(target='rk1808', target_sub_class='AICS')
    
87.     else:
    
88.         print("rknn run target self")
    
89.         ret = rknn.init_runtime()
    
90.     if ret != 0:
    
91.         print('Init runtime environment failed')
    
92.         exit(ret)
    
93.     print('done')
    
94. 
    
95.     print('input:')
    
96.     print(ix)
    
97. 
    
98.     t0 = timer()
    
99.     torch_out = m.forward(x)
    
100.     t1 = timer()
     
101.     print("torch [out]:", x.shape, ', ', torch_out.shape, " time:", t1 - t0)
     
102.     print(torch_out.detach().numpy())
     
103. 
     
104.     t0 = timer()
     
105.     rknn_out = rknn.inference(inputs=[ix], data_type='float32', data_format='nchw')   # , inputs_pass_through=[1]
     
106.     t1 = timer()
     
107. 
     
108.     if rknn_out is None:
     
109.         print('rknn inference failed')
     
110.         exit(-100)
     
111.     y = np.array(rknn_out[0])
     
112.     print("rknn out:", ix.shape, ', ', y.shape, " time:", t1 - t0)
     
113.     print(y)
     
114. 
     
115.     rknn.release()
     

复制代码

已上报问题

jefferyzhang 发表于 2020-3-17 08:55
已上报问题

麻烦能不能帮问一下这个有什么头绪吗，现在着急评估项目，谢谢啊

拆开算子试了一下
.load_pytorch应该是对conv1d不支持

onnx可以正确运行conv1d
但
conv1d → relu → bn  load_onnx运行结果正确
conv1d → bn → relu：pytorch/onnx输出shape为(1, 64, 672)但rknn输出shape为(1, 672, 64)
主楼中的模型就是类似这样的结构重复多次，这里输出shape会被改变，再往后传输的时候会错乱吧

1. 
   
2. import platform
   
3. import os
   
4. import torch
   
5. import numpy as np
   
6. import copy
   
7. 
   
8. from rknn.api import RKNN
   
9. from timeit import default_timer as timer
   
10. 
    
11. import onnx
    
12. from onnx_tf.backend import prepare
    
13. 
    
14. def rm(file):
    
15.     if os.path.exists(file):
    
16.         os.remove(file)
    
17. 
    
18. 
    
19. class MyConv1d(torch.nn.Module):
    
20.     def __init__(self, seq_len, isize):
    
21.         super(MyConv1d, self).__init__()
    
22. 
    
23.         self._conv1d = torch.nn.Conv1d(seq_len, seq_len, kernel_size=11, stride=1, padding=5, dilation=1, groups=1, bias=False)
    
24.         self._seq_len = seq_len
    
25.         self._isize = isize
    
26.         self._bn = torch.nn.BatchNorm1d(seq_len, eps=1e-3, momentum=0.1)
    
27.         self._relu = torch.nn.ReLU()
    
28. 
    
29.     def forward(self, x):
    
30.         x = x.view(1, self._seq_len, self._isize)
    
31.         x = self._conv1d(x)
    
32.         x = self._bn(x)
    
33.         x = self._relu(x)
    
34.         return x
    
35. 
    
36. 
    
37. if __name__ == '__main__':
    
38. 
    
39.     rknn = RKNN(verbose=True)
    
40.     rknn.config(batch_size=1,
    
41.                 epochs=1)  # asymmetric_quantized-u8,  quantized_dtype='dynamic_fixed_point-16' , channel_mean_value='0 0 0 1', reorder_channel='0 1 2',
    
42.     system = platform.system()
    
43. 
    
44.     seq_len = 64
    
45.     isize = 672
    
46.     pt_file = 'conv1d.pt'
    
47. 
    
48.     x = torch.rand((1, seq_len, isize))
    
49.     #x = torch.linspace(-1, 1, steps=(isize * seq_len))
    
50.     x = x.view(1, 1, seq_len, -1)
    
51.     input_size_list = [[1, seq_len, isize]]
    
52.     #x1 = torch.full((1, hsize), fill_value=0.2, dtype=torch.float)
    
53.     ix = x.numpy()
    
54.     pt_m = MyConv1d(seq_len, isize)
    
55.     pt_m.eval()
    
56.     t0 = timer()
    
57.     out = pt_m.forward(x)
    
58.     t1 = timer()
    
59.     print("torch out, ", out.shape, "time:", t1 - t0)
    
60.     print(out.detach().numpy())
    
61. 
    
62.     jit_m = torch.jit.trace(pt_m, x)
    
63.     rm(pt_file)
    
64.     jit_m.save(pt_file)
    
65. 
    
66.     jit_m = torch.jit.load(pt_file)
    
67.     jit_m.eval()
    
68. 
    
69.     t0 = timer()
    
70.     out = jit_m.forward(x)
    
71.     t1 = timer()
    
72.     print("torch out, ", out.shape, "time:", t1 - t0)
    
73.     print(out.detach().numpy())
    
74. 
    
75.     use_pytorch = False
    
76.     if use_pytorch:
    
77.         rknn.load_pytorch(model=pt_file, input_size_list=input_size_list)
    
78.     else:
    
79.         onnx_file = pt_file + '.onnx'
    
80.         rm(onnx_file)
    
81.         torch.onnx.export(pt_m, x, onnx_file,
    
82.                           export_params=True, verbose=False,
    
83.                           input_names=['input'],
    
84.                           output_names=['output'],
    
85.                           opset_version=10
    
86.                           )
    
87.         tf_model = onnx.load(onnx_file)
    
88.         tf_rep = prepare(tf_model)
    
89. 
    
90.         tf_ix = ix
    
91.         t0 = timer()
    
92.         tf_out = tf_rep.run(tf_ix)
    
93.         t1 = timer()
    
94.         tf_out = np.array(tf_out)
    
95.         print("onnx out:", tf_ix.shape, ', ', tf_out.shape, " time:", t1 - t0)
    
96.         print(tf_out)
    
97.         rknn.load_onnx(model=onnx_file)
    
98. 
    
99.         # # onnx转为tf .pb并加载
    
100.         # pb_file = onnx_file + '.pb'
     
101.         # rm(pb_file)
     
102.         # tf_rep.export_graph(pb_file)
     
103.         # input_size_list = [[1, seq_len, isize]]
     
104.         # rknn.load_tensorflow(tf_pb=pb_file,
     
105.         #                      inputs=['input'],
     
106.         #                      outputs=['transpose_191'],
     
107.         #                      input_size_list=input_size_list)
     
108. 
     
109.     is_quant = False
     
110.     if is_quant:
     
111.         rknn_file = pt_file + '_quant.rknn'
     
112.     else:
     
113.         rknn_file = pt_file + '.rknn'
     
114. 
     
115.     print('--> Building model')
     
116.     ret = rknn.build(do_quantization=is_quant, dataset='./dataset.txt')
     
117.     if ret != 0:
     
118.         print('rknn.build failed!')
     
119.         exit(ret)
     
120.     print('build done')
     
121.     rm(rknn_file)
     
122.     rknn.export_rknn(rknn_file)
     
123. 
     
124.     print('--> Init runtime environment')
     
125.     if system == 'Windows':
     
126.         print("rknn run target rk1808")
     
127.         ret = rknn.init_runtime(target='rk1808', target_sub_class='AICS')
     
128.     else:
     
129.         print("rknn run target self")
     
130.         ret = rknn.init_runtime()
     
131.     if ret != 0:
     
132.         print('Init runtime environment failed')
     
133.         exit(ret)
     
134.     print('done')
     
135. 
     
136.     print('input:')
     
137.     print(ix)
     
138. 
     
139.     pt_m_fp16 = pt_m.cuda()
     
140.     pt_m_fp16.half()
     
141.     pt_m_fp16.eval()
     
142.     t0 = timer()
     
143.     pt_y = pt_m_fp16.forward(x.cuda().half())
     
144.     t1 = timer()
     
145.     print("torch out, cuda:fp16", pt_y.shape, "time:", t1 - t0)
     
146.     pt_y = pt_y.cpu().float().detach().numpy()
     
147.     print(pt_y)
     
148.     # t0 = timer()
     
149.     # torch_out = m.forward(x)
     
150.     # t1 = timer()
     
151.     # print("torch [out]:", x.shape, ', ', torch_out.shape, " time:", t1 - t0)
     
152.     # print(torch_out.detach().numpy())
     
153. 
     
154.     t0 = timer()
     
155.     rknn_out = rknn.inference(inputs=[ix], data_type='float32', data_format='nchw')   # , inputs_pass_through=[1]
     
156.     t1 = timer()
     
157. 
     
158.     if rknn_out is None:
     
159.         print('rknn inference failed')
     
160.         exit(-100)
     
161.     y = np.array(rknn_out[0])
     
162.     print("rknn out:", ix.shape, ', ', y.shape, " time:", t1 - t0)
     
163.     print(y)
     
164. 
     
165.     if y.shape[1] != pt_y.shape[1] or  y.shape[2] != pt_y.shape[2]:
     
166.         print("******* shape err: expect ", pt_y.shape, " but rknn out ", y.shape, " reshape:")
     
167. #        y = y.reshape(pt_y.shape[0], pt_y.shape[1], -1)
     
168. #        print(y)
     
169. 
     
170.     diff = np.abs(y - pt_y)
     
171.     cnt = y.shape[0] * y.shape[1] * y.shape[2]
     
172.     print("diff:")
     
173.     print(diff)
     
174.     d = diff[np.where(diff > 0.05)]
     
175.     print("warning values(diff > 0.05):", d.shape, " %:", 100.0 * np.array(d.shape, dtype=np.float) // cnt)
     
176.     print(d)
     
177.     d = diff[np.where(diff > 0.001)]
     
178.     print("warning values(diff > 0.001):", d.shape, " %:", 100.0 * np.array(d.shape, dtype=np.float) // cnt)
     
179.     print(d)
     
180.     d = diff[np.where(diff > 0.0001)]
     
181.     print("warning values(diff > 0.0001):", d.shape,  " %:", 100.0 * np.array(d.shape, dtype=np.float) // cnt)
     
182.     print(d)
     
183. 
     
184.     rknn.release()
     

复制代码

NPU部门回复：
目前pytorch只支持NCHW这种4维输入的卷积，这个模型的卷积输入是3维，暂不支持。