Open thanhlct opened 2 years ago
Describe the bug
The outputs of the converted model with onnx_tf.backend.prepare is different from the original onnx model
To Reproduce `import numpy as np ort_inputs = {'input1': np.random.rand(1, 3, 48, 160).astype(np.float32), 'input2': np.random.randint(2, size=(1,240))==1}
print('test onnx') import onnxruntime as ort osession = ort.InferenceSession("nocr.onnx") ort_outputs = osession.run(None, ort_inputs) ort_outputs = ort_outputs[0]
print('test onnx_tf') import onnx from onnx_tf.backend import prepare onnx_model = onnx.load("./nocr.onnx") tf_rep = prepare(onnx_model) toutputs = tf_rep.run(ort_inputs) toutputs = toutputs.output1 print('===>> results onnx vs onnx_tf backend, max gap', np.max(np.abs(ort_outputs - toutputs)), ', max onnx output:', np.max(ort_outputs), ', min onnx ouput:', np.min(ort_outputs))`
ONNX model file The onnx model attached
Python, ONNX, ONNX-TF, Tensorflow version
Additional context
Tested with some lower versions of tensorflow, but the models output gap is still occurs
Please give us some clues, we can't figure out how to find the root of the problem
Describe the bug
The outputs of the converted model with onnx_tf.backend.prepare is different from the original onnx model
To Reproduce `import numpy as np ort_inputs = {'input1': np.random.rand(1, 3, 48, 160).astype(np.float32), 'input2': np.random.randint(2, size=(1,240))==1}
Test onnx
print('test onnx') import onnxruntime as ort osession = ort.InferenceSession("nocr.onnx") ort_outputs = osession.run(None, ort_inputs) ort_outputs = ort_outputs[0]
tf run
print('test onnx_tf') import onnx from onnx_tf.backend import prepare onnx_model = onnx.load("./nocr.onnx") tf_rep = prepare(onnx_model) toutputs = tf_rep.run(ort_inputs) toutputs = toutputs.output1 print('===>> results onnx vs onnx_tf backend, max gap', np.max(np.abs(ort_outputs - toutputs)), ', max onnx output:', np.max(ort_outputs), ', min onnx ouput:', np.min(ort_outputs))`
ONNX model file The onnx model attached
Python, ONNX, ONNX-TF, Tensorflow version
Additional context
Tested with some lower versions of tensorflow, but the models output gap is still occurs