Zheng-Bicheng commented on PR #16651:
URL: https://github.com/apache/tvm/pull/16651#issuecomment-1968519462
@jiangjiajun 我添加了对PaddleSlim量化模型的支持，但是TVM在这方面好像是存在问题的。
## Paddle测试
我使用了PaddleSlim训练的MobileNetV1_QAT进行测试，测试代码为:
```python
import paddle
import tvm
from tvm import relay
from tvm.contrib import graph_executor
import numpy as np
log_file = "tune.json"
if __name__ == "__main__":
input_shape = [1, 3, 224, 224]
input_name = "inputs"
paddle.enable_static()
prefix = "MobileNetV1_QAT/inference"
params_file_path = prefix + ".pdiparams"
exe = paddle.static.Executor(paddle.CPUPlace())
prog, feed_target_names, fetch_targets =
paddle.static.load_inference_model(prefix, exe)
# build
mod, params = relay.frontend.from_paddle(prog, shape_dict={input_name:
input_shape})
with tvm.transform.PassContext(opt_level=5):
lib = relay.build(mod, target="llvm", params=params)
# create input data
input_data = np.random.randn(1, 3, 224, 224).astype(np.float32)
# tvm inference
ctx = tvm.cpu()
tvm_model = graph_executor.GraphModule(lib['default'](ctx))
tvm_model.set_input(input_name, input_data)
tvm_model.run()
tvm_output = tvm_model.get_output(0).asnumpy()
# paddle inference
paddle_output, = exe.run(prog, feed={feed_target_names[0]: input_data},
fetch_list=fetch_targets)
print(np.argmax(tvm_output[0]), np.argmax(paddle_output[0]))
np.testing.assert_allclose(tvm_output[0], paddle_output[0], rtol=1e-5,
atol=1e-5)
```
发现无法通过测试，误差如下:
```text
AssertionError:
Not equal to tolerance rtol=1e-05, atol=1e-05
Mismatched elements: 5 / 1000 (0.5%)
Max absolute difference: 0.02359476
Max relative difference: 1.
x: array([0. , 0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. , 0. ,...
y: array([0. , 0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. , 0. ,...
```
## ONNX测试
为了验证是否是TVM和Paddle框架的推理机制不同导致的这个问题，我补充测试了ONNX框架，输入模型是Paddle2ONNX导出的同一个模型，测试代码如下：
```python
import tvm
from tvm import relay
from tvm.contrib import graph_executor
import numpy as np
import onnx
import onnxruntime as rt
onnx_model_path = "MobileNetV1_QAT/inference.onnx"
log_file = "tune.json"
if __name__ == "__main__":
input_shape = [1, 3, 224, 224]
input_name = "inputs"
# build
onnx_model = onnx.load_model(onnx_model_path)
mod, params = relay.frontend.from_onnx(onnx_model, shape={input_name:
input_shape})
with tvm.transform.PassContext(opt_level=5):
lib = relay.build(mod, target="llvm", params=params)
# create input data
input_data = np.random.randn(1, 3, 224, 224).astype(np.float32)
# tvm inference
ctx = tvm.cpu()
tvm_model = graph_executor.GraphModule(lib['default'](ctx))
tvm_model.set_input(input_name, input_data)
tvm_model.run()
tvm_output = tvm_model.get_output(0).asnumpy()
sess = rt.InferenceSession(onnx_model_path,None)
input_name = sess.get_inputs()[0].name
out_name = sess.get_outputs()[0].name
onnx_output = sess.run([out_name], {input_name:input_data})[0]
print(np.max(tvm_output[0] - onnx_output[0]))
print(np.argmax(tvm_output[0] - onnx_output[0]))
np.testing.assert_allclose(tvm_output[0], onnx_output[0], rtol=1e-5,
atol=1e-5)
```
发现仍然通不过测试，且误差非常大
```text
Mismatched elements: 270 / 1000 (27%)
Max absolute difference: 0.01025282
Max relative difference: 0.52028346
x: array([6.539907e-06, 6.727985e-05, 5.355392e-05, 4.267169e-06,
1.363640e-04, 1.172559e-04, 1.836461e-04, 6.608244e-06,
7.928120e-06, 3.821332e-06, 1.304903e-05, 4.503604e-05,...
y: array([6.145719e-06, 7.556988e-05, 5.032599e-05, 4.009968e-06,
1.281447e-04, 1.101883e-04, 1.725770e-04, 6.209937e-06,
7.450255e-06, 4.292187e-06, 1.465689e-05, 4.232152e-05,...
```
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
