Self-Created Tools to convert ONNX files (NCHW) to TensorFlow/TFLite/Keras format (NHWC). The purpose of this tool is to solve the massive Transpose extrapolation problem in onnx-tensorflow (onnx-tf). I don't need a Star, but give me a pull request.
License: MIT License
Others
1.5.36
1.12.0
2.10.1
pip install onnx==1.12.0
noneHi, your library is awesome!
I converted the Yolov7-tiny from PyTorch to TfLite using:
onnx2tf -i yolov7-tiny.onnx -o models-NHWC-final/ -osd -oh5 -cotof
I am trying to use it on an android device. The model works when tested on a PC however the Tensorflow Java API for Android does not support dynamic output models according to their documentation: https://www.tensorflow.org/lite/guide/inference. While the resulting yolo tflite model has a dynamic number of outputs( the number of outputs change with the number of indications/detentions)
On the other hand, if I follow the conversion path PyTorch -> ONNX-> Tensorflow I do get yolov7 with a fixed output size so I suspect it is possible to achieve this with onnx2tf as well while also doing the NcHW to NHWc conversion in the process.
Is there a way to have onnx2tf output a fixed/static output .tflite model for yolov7-tiny?
Thank you
Others
1.4.2
1.12.0
2.10.1
https://drive.google.com/file/d/1vGzO9MZGX-yGz6ATm4yHVJMASZACuy2t/view?usp=share_link
{
"format_version": 1,
"operations": [
{
"op_name": "Add_282",
"param_target": "pre_process_transpose_perm",
"param_name": "perm",
"values": [0, 3, 1, 2]
}
]
}Hi, thank you so much for actively maintaining this useful repo. I am trying to convert my custom pytorch model to onnx and then to keras for further tuning. I exported the model using opset==12. The issue occurred during convertion with an ADD LAYER when I run convert('onnx/py_model.onnx').
In attempting to solve this problem, I provided the parameter replacement file to the converter but got the same issue. It looks like one input for the layer is in NHWC and the other is in NCHW for op_name Add_282 (number 1 on the screenshot) in this case. From the error, I can see that one of the input for add is in correct order(NCHW) but the other one is in NHWC. So I look back to the operation Add_263 (number 2 on the screenshot) that is responsible for generating the wrong input and I found that the keras input for this operation is already in NHWC. I did not trace further up because I think that illustrate my problem: it seems like one of the input for the Add_282 is in correct shape.
So I wonder if there is some quick fix for this, like if I forgot to set some parameter or did incorrectly for parameter replacement, or version issue? I also find #17 but at the very end you said no need to do parameter replacement for transpose operation after 1.1.38 and after opset>11. I apologize in advance if I missed something super obvious, but could you provide some tips on how to fix this? Thanks again!
Traceback (most recent call last):
File "C:\Users\qz796\anaconda3\envs\phoodify\lib\site-packages\onnx2tf\utils\common_functions.py", line 267, in print_wrapper_func
result = func(*args, **kwargs)
File "C:\Users\qz796\anaconda3\envs\phoodify\lib\site-packages\onnx2tf\utils\common_functions.py", line 329, in inverted_operation_enable_disable_wrapper_func
result = func(*args, **kwargs)
File "C:\Users\qz796\anaconda3\envs\phoodify\lib\site-packages\onnx2tf\utils\common_functions.py", line 37, in get_replacement_parameter_wrapper_func
func(*args, **kwargs)
File "C:\Users\qz796\anaconda3\envs\phoodify\lib\site-packages\onnx2tf\ops\Add.py", line 126, in make_node
tf.math.add(
File "C:\Users\qz796\anaconda3\envs\phoodify\lib\site-packages\tensorflow\python\util\traceback_utils.py", line 153, in error_handler
raise e.with_traceback(filtered_tb) from None
File "C:\Users\qz796\anaconda3\envs\phoodify\lib\site-packages\keras\layers\core\tf_op_layer.py", line 119, in handle
return TFOpLambda(op)(*args, **kwargs)
File "C:\Users\qz796\anaconda3\envs\phoodify\lib\site-packages\keras\utils\traceback_utils.py", line 70, in error_handler
raise e.with_traceback(filtered_tb) from None
ValueError: Exception encountered when calling layer "tf.math.add_101" (type TFOpLambda).
Dimensions must be equal, but are 56 and 16 for '{{node tf.math.add_101/Add}} = AddV2[T=DT_FLOAT](Placeholder, Placeholder_1)' with input shapes: [1,56,56,16], [1,16,56,56].
Call arguments received by layer "tf.math.add_101" (type TFOpLambda):
• x=tf.Tensor(shape=(1, 56, 56, 16), dtype=float32)
• y=tf.Tensor(shape=(1, 16, 56, 56), dtype=float32)
• name='Add_248'
Others
1.7.x
1.13.0
2.10.0
N/A
N/ASince the internal processing has already implemented a number of fairly complex workarounds to avoid conversion errors, we have not dared to implement a correction process for accuracy errors at this time in order not to further increase the complexity of the logic. Instead, I have implemented a mechanism whereby the user visually identifies which operation's channel transpositions cause problems, and the user changes the behavior of the tool himself.
However, I cannot overlook the situation where accuracy errors remain despite successful model transformations, and I hope to eradicate them in the future. The work to change the behavior of the tool by the users themselves with the users' visibility of the problem areas is very costly.
Since the internal processing mechanism needs to be significantly revised, we intend to implement many additional test modifications gradually in minor version upgrades to the extent that they do not affect the existing processing.
onnx2tf/onnx2tf/utils/common_functions.py
Lines 2937 to 3062 in 33e65dc
Transpose OP is excluded from the search because the value to be compared does not change.Constant and ConstantOfShape.pre_process_transpose_perm is used to check all combinations that pre-transpose the input tensor to the corresponding OP.parameter_replacement.json, omit processing.MatMul.Softmax, attempt correction in the following order: attribute value change first, then tensor transposition.Others
1.1.45
onnx/onnx-tensorflow#656
https://github.com/PINTO0309/onnx2tf/releases/tag/1.1.28
https://github.com/PINTO0309/onnx2tf/releases/download/1.1.28/detr_demo.onnx
.onnx2tf -i detr_demo.onnx
Others
Generic issue
Clip operation failing as min max are assumed to be inputs .. But. in onnx graph. they may be set as attributes.
Others
1.4.2
1.12.0
2.10
https://drive.google.com/file/d/1QQLf83JwLI_EoQFU5xOGKUvtFPsFSWCN/view
N/A
common_functions.py. In this case, intermediate Sub layer returned wrong output because it didn't re-swapped the parameters. But I didn't fully understand why these lines are needed. What is the bug you mentioned in PR #99?onnx2tf/onnx2tf/utils/common_functions.py
Lines 849 to 855 in f5eb872
Others
1.1.46
https://github.com/DeepranjanG/Image_Background_Removal
https://github.com/DeepranjanG/Image_Background_Removal/raw/main/pretrained/modnet.onnx
{
"format_version": 1,
"operations": [
{
"op_name": "Expand_187",
"param_target": "outputs",
"param_name": "685",
"post_process_transpose_perm": [0,2,3,1]
}
]
}1.Personal development
2.Error occured.
onnx2tf -i modnet.onnx -ois "input:1,3,480,640" -prf replace.json
Dimensions must be equal, but are 15 and 1280 for '{{node tf.math.multiply_17/Mul}} = Mul[T=DT_FLOAT](Placeholder, Placeholder_1)' with input shapes: [1,15,20,1280], [1,1280,15,20]
3.replace.json
5.[Experimental] Automatic transposition of input values (valid only if immediately preceding is "Conv") #36
Others
1.5.18
1.13.0
2.10.0
Not usedMatmul ops are not equal. Error log is attached. error_log.txtparam_replacement.json. But For Matmul ops, only axis transposing is suported, so I am not able to add dimension on the input.
INFO: input_name.1: input.35 shape: [1, 256, 66] dtype: float32
INFO: input_name.2: onnx::MatMul_73 shape: [66, 2] dtype: <class 'numpy.float32'>
We are thinking if we can add one more dimension in input_name.2: onnx::MatMul_73 to [1, 66, 2], it can be resolved.
Others
1.1.17
Noneonnx2tf -i japan_PP-OCRv3_rec_infer.onnx
ERROR: The trace log is below.
Traceback (most recent call last):
File "/home/systemkjapan/.local/lib/python3.8/site-packages/onnx2tf/utils/common_functions.py", line 261, in print_wrapper_func
result = func(*args, **kwargs)
File "/home/systemkjapan/.local/lib/python3.8/site-packages/onnx2tf/utils/common_functions.py", line 323, in inverted_operation_enable_disable_wrapper_func
result = func(*args, **kwargs)
File "/home/systemkjapan/.local/lib/python3.8/site-packages/onnx2tf/ops/Clip.py", line 109, in make_node
if (min_value is not None and min_value.shape is not None) \
AttributeError: 'float' object has no attribute 'shape'
I don't think the code here will work if 'min_value' is 'float'?
Or is it wrong to be 'float'?
Others
1.5.24
1.13.0
2.10.0
https://tfhub.dev/tensorflow/lite-model/mobilebert/1/metadata/1?lite-format=tflite
I cannot run a transformer model with token-level output on accelerated hardware in TF Lite #59232
Tasks
--optimization_for_gpu_delegate optionOthers
1.1.25
gist for reproduction : https://colab.research.google.com/gist/Hyunseok-Kim0/d0aaf6e9ac6fbe461c5f2364db4bc0b2/onnx2tf_20221117.ipynb
N/Achannel_transpose in common_functions.py is used in arithmetic operations like Add, Sub, Mul, etc. What is the main purpose of this function? When second input has more dimension, channel_transpose adds additional squeeze layer and changes the output shape, not vice versa.x = x + y, onnx and tflite has same output shape. Converted tflite has wrong shape for x = y + x.| ONNX | Correct tflite (x = x + y) | Wrong tflite (x = y + x) |
|---|---|---|
 |
 |
 |
Feature Request
1.3.14
1.12.0
2.10.0
https://drive.google.com/file/d/1XeM7XpjOsL37qAdpWa1j6x9KMzYbiHgW/view?usp=sharing
{}
Others
1.1.33
Please use code below to reproduce bug.
import torch
from einops import rearrange
from torch import nn
import numpy as np
from onnx2tf import convert
import shutil
import torchvision
import tensorflow as tf
import cv2
import matplotlib.pyplot as plt
import requests
import onnxruntime
class dummy_network(nn.Module):
def __init__(self, output_size, spatial_scale):
super(dummy_network, self).__init__()
self.output_size = output_size
self.spatial_scale = spatial_scale
def forward(self, x, roi):
x = torchvision.ops.roi_align(x, boxes=roi, output_size=self.output_size, spatial_scale=self.spatial_scale)
return x
def visualize(outputs):
plt.figure(figsize=(8, 8))
col = 4
row = len(outputs) // 4 + 1
for i, o in enumerate(outputs, start=1):
plt.subplot(row, col, i)
plt.axis("off")
plt.imshow(o.astype(int))
plt.show()
def main():
model = dummy_network(output_size=(64, 64), spatial_scale=1.0)
model.eval()
dummy_name = "dummy_roi_align"
onnx_save_path = f"tflite/{dummy_name}.onnx"
temp_tflite = "tflite/model_float32.tflite"
tflite_save_path = f"tflite/{dummy_name}.tflite"
url = "https://static01.nyt.com/images/2021/09/14/science/07CAT-STRIPES/07CAT-STRIPES-mediumSquareAt3X-v2.jpg"
image_nparray = np.asarray(bytearray(requests.get(url).content), dtype=np.uint8)
image = cv2.imdecode(image_nparray, cv2.IMREAD_COLOR)
dummy_input_x = np.expand_dims(cv2.resize(cv2.cvtColor(image, cv2.COLOR_BGR2RGB), (256, 256)), axis=0)
dummy_input_x = rearrange(torch.Tensor(dummy_input_x), "n h w c -> n c h w")
# dummy_input_roi = [[0, 0, 0, 0.5, 0.5], [0, 0.5, 0, 1, 0.5], [0, 0, 0.5, 0.5, 1], [0, 0.5, 0.5, 1, 1]]
dummy_input_roi = [[0, 0, 0, 128, 128], [0, 128, 0, 256, 128], [0, 0, 128, 128, 256], [0, 128, 128, 256, 256]]
# dummy_input_roi = [[0, 0, 1, 2, 3]]
# dummy_input_roi = [[0, 0, 0, 256, 256]]
dummy_input_roi = torch.Tensor(dummy_input_roi)
torch.onnx.export(model,
args=(dummy_input_x, dummy_input_roi),
f=onnx_save_path,
input_names=["x", "roi"],
opset_version=11)
convert(onnx_save_path, output_folder_path="tflite")
shutil.move(temp_tflite, tflite_save_path)
# get torch output
# -----------------------------------------------------------------------------------------------
with torch.no_grad():
torch_output = model(dummy_input_x, dummy_input_roi)
torch_output = rearrange(torch_output, "n c h w -> n h w c")
# visualize(torch_output)
# get onnx output
# -----------------------------------------------------------------------------------------------
onnx_session = onnxruntime.InferenceSession(onnx_save_path, providers=['CPUExecutionProvider'])
onnx_inputs = dict(x=dummy_input_x.numpy(), roi=dummy_input_roi.numpy())
onnx_output = onnx_session.run(None, onnx_inputs)[0]
onnx_output = rearrange(onnx_output, "n c h w -> n h w c")
# compare torch output and onnx output
np.testing.assert_allclose(onnx_output, torch_output.numpy())
# get tflite output
# -----------------------------------------------------------------------------------------------
tflite_onnx2tf = tf.lite.Interpreter(model_path=tflite_save_path)
tflite_onnx2tf.allocate_tensors()
tflite_onnx2tf.set_tensor(tflite_onnx2tf.get_input_details()[0]['index'],
rearrange(dummy_input_x.numpy(), "n c h w -> n h w c"))
tflite_onnx2tf.set_tensor(tflite_onnx2tf.get_input_details()[1]['index'], dummy_input_roi.numpy())
tflite_onnx2tf.invoke()
tflite_output = [tflite_onnx2tf.get_tensor(i['index']) for i in tflite_onnx2tf.get_output_details()][0]
total_output = np.concatenate([torch_output, onnx_output, tflite_output], axis=0)
visualize(total_output)
print("convert done")
return
if __name__ == "__main__":
main()N/APurpose: Personal development
What: When I tested RoiAlign, it showed unexpected output as below.
| pytorch output |
|---|
 |
| onnx output |
 |
| wrong tflite output |
 |
| correct tflite output after modification |
 |
x0, x1, y1, y0 = tf.split(boxes, 4, axis=1), I could get correct result as shown above.onnx2tf/onnx2tf/ops/RoiAlign.py
Lines 116 to 126 in 0b14623
x0, y0, x1, y1 to x0, x1, y1, y0.
Others
1.2.13
{
"format_version": 1,
"operations": [
{
"op_name": "StatefulPartitionedCall/Tile_4",
"param_target": "inputs", # attributes or inputs
"param_name": "const_fold_opt__677",
"values": [1,1,17] # Disable parameter transposition or overwrite parameters
},
{
"op_name": "StatefulPartitionedCall/Cast_3",
"param_target": "attributes", # attributes or inputs
"param_name": "to",
"values": 1 # Disable parameter transposition or overwrite "to" parameters
},
{
"op_name": "Resize__697",
"param_target": "inputs",
"param_name": "Concat__696:0",
"values": [26,26] # Replacement of unk__x (Resize OP, sizes height/width parameter)
},
{
"op_name": "Transpose__927",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3] # Disable parameter transposition or overwrite "perm" parameters
},
{
"op_name": "StatefulPartitionedCall/functional_1/max_unpooling2d_2/Reshape_1",
"param_target": "inputs",
"param_name": "const_fold_opt__911",
"values": [4,131072] # Overwrite "shape" parameters
},
{
"op_name": "Reshape_25",
"param_target": "outputs",
"param_name": "onnx::InstanceNormalization_270",
"post_process_transpose_perm": [0,2,1] # Extrapolate 3D Transpose after Reshape
},
{
"op_name": "Reshape_30",
"param_target": "outputs",
"param_name": "onnx::Mul_275",
"post_process_transpose_perm": [0,2,3,1] # Extrapolate 4D Transpose after Reshape
},
{
"op_name": "flatten_1127",
"param_target": "inputs",
"param_name": "dropout0",
"pre_process_transpose_perm": [0,3,1,2]
}
]
}Failed to run test sample in CLI. Use this:
wget https://github.com/PINTO0309/onnx2tf/releases/download/1.1.27/human_segmentation_pphumanseg_2021oct.onnx
wget https://github.com/PINTO0309/onnx2tf/releases/download/1.1.27/replace.json
onnx2tf -i human_segmentation_pphumanseg_2021oct.onnx -prf replace.json
Encountered:
Model optimizing started ============================================================
WARNING: Failed to optimize the onnx file.
Traceback (most recent call last):
File "C:\Users\zhugc\AppData\Roaming\Python\Python39\site-packages\onnx2tf\onnx2tf.py", line 427, in convert
result = subprocess.check_output(
File "D:\Anaconda3\lib\subprocess.py", line 424, in check_output
return run(*popenargs, stdout=PIPE, timeout=timeout, check=True,
File "D:\Anaconda3\lib\subprocess.py", line 528, in run
raise CalledProcessError(retcode, process.args,
subprocess.CalledProcessError: Command '['onnxsim', 'human_segmentation_pphumanseg_2021oct.onnx', 'human_segmentation_pphumanseg_2021oct.onnx']' returned non-zero exit status 1.
Automatic generation of each OP name started ========================================
Automatic generation of each OP name complete!
Model loaded ========================================================================
Model convertion started ============================================================
INFO: input_op_name: x shape: [1, 3, 192, 192] dtype: float32
ERROR: The trace log is below.
Traceback (most recent call last):
File "C:\Users\zhugc\AppData\Roaming\Python\Python39\site-packages\onnx2tf\utils\common_functions.py", line 262, in print_wrapper_func
result = func(*args, **kwargs)
File "C:\Users\zhugc\AppData\Roaming\Python\Python39\site-packages\onnx2tf\ops\Input.py", line 81, in make_node
if graph_input.shape != tf.TensorShape(None) and len(graph_input.shape) in [3, 4, 5] \
File "C:\Users\zhugc\AppData\Roaming\Python\Python39\site-packages\tensorflow\python\framework\tensor_shape.py", line 1294, in __ne__
raise ValueError("The inequality of unknown TensorShapes is undefined.")
ValueError: The inequality of unknown TensorShapes is undefined.
ERROR: Read this and deal with it. https://github.com/PINTO0309/onnx2tf#parameter-replacement
ERROR: Alternatively, if the input OP has a dynamic dimension, use the -b or -ois option to rewrite it to a static shape and try again.
ERROR: If the input OP of ONNX before conversion is NHWC, use the -kt option.
Others
1.2.2
https://github.com/AISIN-TRC/E2Pose/blob/main/pretrains/download.sh
https://github.com/PINTO0309/PINTO_model_zoo/tree/main/333_E2Pose
{
"format_version": 1,
"operations": [
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_decode_01_norm/FusedBatchNormV3__930",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_decode_02_norm/FusedBatchNormV3__1055",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_decode_03_norm/FusedBatchNormV3__1180",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_decode_04_norm/FusedBatchNormV3__1305",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_decode_02_p_seed/Conv2D__1065",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_decode_03_p_seed/Conv2D__1190",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_decode_04_p_seed/Conv2D__1315",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_decode_05_p_seed/Conv2D__1408",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_output_04_cyx_seed_split/split",
"param_target": "attributes",
"param_name": "axis",
"values": 3
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_output_03_cyx_seed_split/split",
"param_target": "attributes",
"param_name": "axis",
"values": 3
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_output_02_cyx_seed_split/split",
"param_target": "attributes",
"param_name": "axis",
"values": 3
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_output_01_cyx_seed_split/split",
"param_target": "attributes",
"param_name": "axis",
"values": 3
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_output_05_cyx_seed_split/split",
"param_target": "attributes",
"param_name": "axis",
"values": 3
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_output_04_cyx/concat",
"param_target": "attributes",
"param_name": "axis",
"values": 3
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_output_03_cyx/concat",
"param_target": "attributes",
"param_name": "axis",
"values": 3
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_output_02_cyx/concat",
"param_target": "attributes",
"param_name": "axis",
"values": 3
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_output_01_cyx/concat",
"param_target": "attributes",
"param_name": "axis",
"values": 3
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_output_05_cyx/concat",
"param_target": "attributes",
"param_name": "axis",
"values": 3
},
{
"op_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_output_04_cyx_reshape/Reshape",
"param_target": "inputs",
"param_name": "e2pose/E2Pose_Inference/E2Pose_pose_stage00_output_04_cyx/concat:0",
"pre_process_transpose_perm": [0,2,3,1]
}
]
}Others
1.1.15
https://github.com/PINTO0309/onnx2tf/releases/download/1.1.15/vae_encoder.onnx
{
"format_version": 1,
"operations": [
{
"op_name": "Reshape_448",
"param_target": "outputs",
"param_name": "onnx::Add_817",
"post_process_transpose_perm": [0,2,3,1]
}
]
}vae_encoder.onnx. For StableDiffusion [Hugging Face].onnx2tf -i vae_encoder.onnxERROR: The trace log is below.
Traceback (most recent call last):
File "/home/xxxxx/git/onnx2tf/onnx2tf/utils/common_functions.py", line 261, in print_wrapper_func
result = func(*args, **kwargs)
File "/home/xxxxx/git/onnx2tf/onnx2tf/utils/common_functions.py", line 323, in inverted_operation_enable_disable_wrapper_func
result = func(*args, **kwargs)
File "/home/xxxxx/git/onnx2tf/onnx2tf/ops/Add.py", line 79, in make_node
tf.math.add(
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/util/traceback_utils.py", line 153, in error_handler
raise e.with_traceback(filtered_tb) from None
File "/usr/local/lib/python3.8/dist-packages/keras/layers/core/tf_op_layer.py", line 119, in handle
return TFOpLambda(op)(*args, **kwargs)
File "/usr/local/lib/python3.8/dist-packages/keras/utils/traceback_utils.py", line 70, in error_handler
raise e.with_traceback(filtered_tb) from None
ValueError: Exception encountered when calling layer "tf.math.add_56" (type TFOpLambda).
Dimensions must be equal, but are 512 and 64 for '{{node tf.math.add_56/Add}} = AddV2[T=DT_FLOAT](Placeholder, Placeholder_1)' with input shapes: [1,512,64,64], [1,64,64,512].
Call arguments received by layer "tf.math.add_56" (type TFOpLambda):
• x=tf.Tensor(shape=(1, 512, 64, 64), dtype=float32)
• y=tf.Tensor(shape=(1, 64, 64, 512), dtype=float32)
• name='Add_449'
ERROR: Read this and deal with it. https://github.com/PINTO0309/onnx2tf#parameter-replacement
ERROR: Alternatively, if the input OP has a dynamic dimension, use the -b or -ois option to rewrite it to a static shape and try again.
ERROR: If the input OP of ONNX before conversion is NHWC, use the -kt option.
replace.json and run it.
onnx2tf -i vae_encoder.onnx -prf replace.jsonThe final conversion was successful. https://github.com/PINTO0309/onnx2tf/releases/tag/1.1.15
Others
1.2.9
Any onnx file with NonMaxSuppression can be used. Please use code below for simple debugging.
import os
import shutil
import onnxruntime
import tensorflow as tf
import torch
import torchvision
from torch import nn
from onnx2tf import convert
class dummy_network(nn.Module):
def __init__(self):
super(dummy_network, self).__init__()
def forward(self, boxes, scores):
# unique with sort
result = torchvision.ops.nms(boxes, scores, iou_threshold=0.3)
return result
def main():
os.makedirs("tflite", exist_ok=True)
model = dummy_network()
model.eval()
dummy_name = "dummy_nms"
onnx_save_path = f"tflite/{dummy_name}.onnx"
temp_tflite = "tflite/model_float32.tflite"
tflite_save_path = f"tflite/{dummy_name}.tflite"
dummy_input_boxes = torch.abs(torch.randn(32, 4, dtype=torch.float32))
dummy_input_boxes = torch.concat([dummy_input_boxes, dummy_input_boxes], dim=0)
dummy_scores = torch.abs(torch.randn(32, dtype=torch.float32))
dummy_scores = torch.concat([dummy_scores, dummy_scores], dim=0)
torch.onnx.export(model,
args=(dummy_input_boxes, dummy_scores),
f=onnx_save_path,
input_names=["boxes", "scores"],
output_names=["nms_result"],
opset_version=11)
# get onnx output
# -----------------------------------------------------------------------------------------------
onnx_session = onnxruntime.InferenceSession(onnx_save_path, providers=['CPUExecutionProvider'])
onnx_inputs = dict(boxes=dummy_input_boxes.numpy(), scores=dummy_scores.numpy())
onnx_outputs = onnx_session.run(None, onnx_inputs)
convert(onnx_save_path, output_folder_path="tflite")
shutil.move(temp_tflite, tflite_save_path)
# # compare torch output and onnx output
# np.testing.assert_allclose(onnx_output, torch_output.numpy())
#
# get tflite output
# -----------------------------------------------------------------------------------------------
tflite_onnx2tf = tf.lite.Interpreter(model_path=tflite_save_path)
tflite_onnx2tf.allocate_tensors()
tflite_onnx2tf.set_tensor(tflite_onnx2tf.get_input_details()[0]['index'], dummy_input_boxes.numpy())
tflite_onnx2tf.set_tensor(tflite_onnx2tf.get_input_details()[1]['index'], dummy_scores.numpy())
tflite_onnx2tf.invoke()
tflite_output = [tflite_onnx2tf.get_tensor(i['index']) for i in tflite_onnx2tf.get_output_details()][0]
print("convert done")
return
if __name__ == "__main__":
main()N/APurpose: Personal development
What: onnx NonMaxSuppression operator is converted using tf.image.non_max_suppression. However, it looks there is bug in tensorflow operation.
tensorflow/tensorflow#51629
For now, tf.image.non_max_suppression always pad the number of output to max_output_boxes_per_class. The result below shows the difference between onnx and tflite output.
There is another operation, tf.image.non_max_suppression_padded which using extra parameter pad_to_max_output_size. I thought this problem can be solved by setting pad_to_max_output_size to False, but it also didn't work.
Returned indices are not correct, looks like the implementation is totally different between tf.image.non_max_suppression_padded and tf.image.non_max_suppression.
How: This problem can be solved by adding single line of code selected_indices = tf.unique(selected_indices).y after line number 208. However, sometimes last element of indices can be wrong value since this solution is just removing duplicated indices.
onnx2tf/onnx2tf/ops/NonMaxSuppression.py
Lines 202 to 208 in db8c722
Others
1.1.38
https://github.com/triptec/detection-models/blob/master/yolov7/yolov7-tiny.onnx
{}python export.py --weights ../models/yolov7/yolov7-tiny.pt --grid --simplifyonnx2tf -i yolov7-tiny.onnx -ioqd uint8 -oiqtpython detect.py --weights ../models/yolov7/yolov7-tiny.tflite --conf 0.25 --img-size 640 --source data/images/zidane.jpgAnd it results in this:
tensor([2.39083e+05, 1.02942e+05, 3.69536e+05, 3.17344e+05, 7.92245e-01, 0.00000e+00])
tensor([1.40084e+05, 2.27828e+05, 1.72367e+05, 3.18328e+05, 5.62911e-01, 2.70000e+01])
tensor([4.15838e+04, 1.53934e+05, 3.06103e+05, 3.16576e+05, 5.42062e-01, 0.00000e+00])
[tensor(1280.), tensor(720.), tensor(1280.), tensor(720.)] tensor(0.54206) tensor(0.)
[tensor(1280.), tensor(720.), tensor(1280.), tensor(720.)] tensor(0.56291) tensor(27.)
[tensor(1280.), tensor(720.), tensor(1280.), tensor(720.)] tensor(0.79225) tensor(0.)
image 1/1 /home/andreas/Development/priv/cat-detection/yolov5/data/images/zidane.jpg: 640x640 2 persons, 1 tie, 1561.9ms
Speed: 2.7ms pre-process, 1561.9ms inference, 1.6ms NMS per image at shape (1, 3, 640, 640)
Results saved to runs/detect/exp38
As you can see it finds 2 people and 1 tie but the coordinates are not correct.
Here's the correct result with the onnx model:
In the readme it's stated that YOLOv7-tiny with Post-Process (NMS) ONNX to TFLite Float32 has been converted. Do you have the command and a "detection.py" what works with the result? Perhaps if I could get a hold of that I could solve this myself?
Others
1.1.38
none
nonedocker run --rm -it -v `pwd`:/workdir -w /workdir ghcr.io/pinto0309/onnx2tf:1.1.38 I get the following error:Unable to find image 'ghcr.io/pinto0309/onnx2tf:1.1.38' locally
docker: Error response from daemon: Head "https://ghcr.io/v2/pinto0309/onnx2tf/manifests/1.1.38": unauthorized.
Also I wonder what would be the best way to then go from tensorflow to tflite?
Others
Log added below: One of the inputs. to Mul operation is transformed and the one from Unsqueeze is not transformed
INFO: tf_op_type: Unsqueeze input_name.1: tf.math.divide_1/truediv:0 shape: (1, 72) dtype: <dtype: 'float32'>
INFO: tf_op_type: Unsqueeze output_name.1: 466 shape: (1, 72, 1) dtype: <dtype: 'float32'>
starting to make node
<module 'onnx2tf.ops.Unsqueeze' from '~/onnx2tf/onnx2tf/ops/Unsqueeze.py'>
INFO: onnx_op_type: Unsqueeze onnx_op_name: Unsqueeze_49
INFO: input_name.1: 466 shape: [1, 72, 1] dtype: float32
INFO: output_name.1: 467 shape: [1, 72, 1, 1] dtype: float32
[3]
[1, 72, 1, 1]
[1, 72, 1, 1]
tf.reshape_1/Reshape:0
(1, 72, 1, 1)
True
INFO: tf_op_type: Unsqueeze input_name.1: tf.reshape/Reshape:0 shape: (1, 72, 1) dtype: <dtype: 'float32'>
INFO: tf_op_type: Unsqueeze output_name.1: 467 shape: (1, 72, 1, 1) dtype: <dtype: 'float32'>
starting to make node
<module 'onnx2tf.ops.Mul' from '~/onnx2tf/onnx2tf/ops/Mul.py'>
INFO: onnx_op_type: Mul onnx_op_name: Mul_50
INFO: input_name.1: 453 shape: [1, 72, 60, 60] dtype: float32
INFO: input_name.2: 467 shape: [1, 72, 1, 1] dtype: float32
INFO: output_name.1: 468 shape: [1, 72, 60, 60] dtype: float32
[1, 72, 60, 60]
Traceback (most recent call last):
File “/onnx2tf/onnx2tf/utils/common_functions.py", line 49, in print_wrapper_func/onnx2tf/onnx2tf/utils/common_functions.py", line 105, in inverted_operation_enable_disable_wrapper_func
result = func(*args, **kwargs)
File "
result = func(*args, **kwargs)
File “/onnx2tf/onnx2tf/ops/Mul.py", line 61, in make_node/model-conversion/git/env/lib/python3.8/site-packages/tensorflow/python/util/traceback_utils.py", line 153, in error_handler
tf.math.multiply(
File "
raise e.with_traceback(filtered_tb) from None
File "/model-conversion/git/env/lib/python3.8/site-packages/keras/layers/core/tf_op_layer.py", line 119, in handle/model-conversion/git/env/lib/python3.8/site-packages/keras/utils/traceback_utils.py", line 70, in error_handler
return TFOpLambda(op)(*args, **kwargs)
File "
raise e.with_traceback(filtered_tb) from None
ValueError: Exception encountered when calling layer "tf.math.multiply_3" (type TFOpLambda).
Dimensions must be equal, but are 60 and 72 for '{{node tf.math.multiply_3/Mul}} = Mul[T=DT_FLOAT](Placeholder, Placeholder_1)' with input shapes: [1,60,60,72], [1,72,1,1].
Call arguments received by layer "tf.math.multiply_3" (type TFOpLambda):
• x=tf.Tensor(shape=(1, 60, 60, 72), dtype=float32)
• y=tf.Tensor(shape=(1, 72, 1, 1), dtype=float32)
• name='Mul_50'
Others
1.5.16
1.13.0
2.10.0
tflite
https://tfhub.dev/tensorflow/lite-model/mobilebert/1/metadata/1?lite-format=tflite
tflite to onnx
python -m tf2onnx.convert \
--opset 11 \
--tflite lite-model_mobilebert_1_metadata_1.tflite \
--output lite_model_mobilebert_1_metadata_1.onnx
onnxsim lite_model_mobilebert_1_metadata_1.onnx lite_model_mobilebert_1_metadata_1.onnx
onnxsim lite_model_mobilebert_1_metadata_1.onnx lite_model_mobilebert_1_metadata_1.onnx
onnxsim lite_model_mobilebert_1_metadata_1.onnx lite_model_mobilebert_1_metadata_1.onnx
onnxsim lite_model_mobilebert_1_metadata_1.onnx lite_model_mobilebert_1_metadata_1.onnx
onnxsim lite_model_mobilebert_1_metadata_1.onnx lite_model_mobilebert_1_metadata_1.onnxonnx to tflite
onnx2tf -i lite_model_mobilebert_1_metadata_1.onnx -cotof -cotoa 1e-1 -cgdc
N/AThe model structure was too complex and overflowed python's Maximum Recursion Depth.
The following must be added to the code to avoid this.
sys.setrecursionlimit(10000)
Traceback (most recent call last):
File "/home/b920405/.local/bin/onnx2tf", line 33, in <module>
sys.exit(load_entry_point('onnx2tf', 'console_scripts', 'onnx2tf')())
File "/usr/local/lib/python3.8/dist-packages/onnx2tf/onnx2tf.py", line 1698, in main
model = convert(
File "/usr/local/lib/python3.8/dist-packages/onnx2tf/onnx2tf.py", line 718, in convert
model = tf.keras.Model(inputs=inputs, outputs=outputs)
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/trackable/base.py", line 205, in _method_wrapper
result = method(self, *args, **kwargs)
File "/usr/local/lib/python3.8/dist-packages/keras/engine/functional.py", line 165, in __init__
self._init_graph_network(inputs, outputs)
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/trackable/base.py", line 205, in _method_wrapper
result = method(self, *args, **kwargs)
File "/usr/local/lib/python3.8/dist-packages/keras/engine/functional.py", line 264, in _init_graph_network
nodes, nodes_by_depth, layers, _ = _map_graph_network(
File "/usr/local/lib/python3.8/dist-packages/keras/engine/functional.py", line 1046, in _map_graph_network
nodes_in_decreasing_depth, layer_indices = _build_map(outputs)
File "/usr/local/lib/python3.8/dist-packages/keras/engine/functional.py", line 1176, in _build_map
_build_map_helper(
File "/usr/local/lib/python3.8/dist-packages/keras/engine/functional.py", line 1219, in _build_map_helper
_build_map_helper(
File "/usr/local/lib/python3.8/dist-packages/keras/engine/functional.py", line 1219, in _build_map_helper
_build_map_helper(
File "/usr/local/lib/python3.8/dist-packages/keras/engine/functional.py", line 1219, in _build_map_helper
_build_map_helper(
[Previous line repeated 986 more times]
File "/usr/local/lib/python3.8/dist-packages/keras/engine/functional.py", line 1199, in _build_map_helper
node = layer._inbound_nodes[node_index]
RecursionError: maximum recursion depth exceeded
Others
1.2.14
1.12.0
2.10
Please use code below for debugging.
import os
import shutil
import onnxruntime
import tensorflow as tf
import torch
from einops import rearrange
from torch import nn
import numpy as np
from onnx2tf import convert
class dummy_network(nn.Module):
def __init__(self):
super(dummy_network, self).__init__()
# segmentation fault with wrong padding mode
# self.tconv = torch.nn.ConvTranspose2d(3, 64, (3, 3), stride=3, padding=1) # case 1: wrong output
# self.tconv = torch.nn.ConvTranspose2d(3, 64, (4, 4), stride=3, padding=1) # case 1: wrong output
self.tconv = torch.nn.ConvTranspose2d(3, 64, (3, 3), stride=2, padding=1) # case 2: segmentation fault
def forward(self, x):
result = self.tconv(x)
return result
def main():
os.makedirs("tflite", exist_ok=True)
model = dummy_network()
model.eval()
dummy_name = "dummy_tconv"
onnx_save_path = f"tflite/{dummy_name}.onnx"
temp_tflite = "tflite/model_float32.tflite"
tflite_save_path = f"tflite/{dummy_name}.tflite"
# y1, x1, y2, x2
dummy_x = torch.randn(1, 3, 224, 224, dtype=torch.float32)
torch.onnx.export(model,
args=(dummy_x,),
f=onnx_save_path,
input_names=["x"],
output_names=["result"],
opset_version=11)
# get onnx output
# -----------------------------------------------------------------------------------------------
onnx_session = onnxruntime.InferenceSession(onnx_save_path, providers=['CPUExecutionProvider'])
onnx_inputs = dict(x=dummy_x.numpy())
onnx_outputs = onnx_session.run(None, onnx_inputs)
convert(onnx_save_path, output_folder_path="tflite")
shutil.move(temp_tflite, tflite_save_path)
# get tflite output
# -----------------------------------------------------------------------------------------------
tflite_onnx2tf = tf.lite.Interpreter(model_path=tflite_save_path)
tflite_onnx2tf.allocate_tensors()
tflite_onnx2tf.set_tensor(tflite_onnx2tf.get_input_details()[0]['index'],
rearrange(dummy_x.numpy(), "n c h w -> n h w c"))
tflite_onnx2tf.invoke()
tflite_output = [tflite_onnx2tf.get_tensor(i['index']) for i in tflite_onnx2tf.get_output_details()][0]
np.testing.assert_allclose(rearrange(tflite_output, "n h w c -> n c h w"), onnx_outputs[0], atol=1e-3, rtol=0.5)
print("convert done")
return
if __name__ == "__main__":
main()N/APurpose: Personal development
What: Converted tflite contains wrong padding mode which occurs wrong output or segmentation fault when stride is larger than 2 and padding is larger than 0.
How:
onnx2tf/onnx2tf/ops/ConvTranspose.py
Lines 103 to 109 in 686fa3d
SAME by comparing input shape and output shape. However, output shape is not same with input shape when stride is not 1.[int((s * (i - 1) + k - o) / 2) for s, i, k, o in zip(strides, graph_node_input_shape[2:], kernel_shape, graph_node_output_shape[2:])] == pads[1:3], the error is gone in some cases. But this is not enough, there should be more lines of code for comparing given padding value and calculated padding value to check padding mode is SAME or not.
Others
1.1.22
https://drive.google.com/file/d/1HyUDTYSTgS7_Gs29SxUSoxs3irqvIsXM/view?usp=sharing
{}onnx2tf -i nanodet-plus-m_416.onnx -oh5 -k "data"INFO: onnx_op_type: Conv onnx_op_name: /backbone/stage2/stage2.1/branch2/branch2.0/Conv
INFO: input_name.1: /backbone/stage2/stage2.1/Split_output_1 shape: [1, 58, 52, 52] dtype: float32
INFO: input_name.2: onnx::Conv_1392 shape: [58, 58, 1, 1] dtype: <class 'numpy.float32'>
INFO: input_name.3: onnx::Conv_1393 shape: [58] dtype: <class 'numpy.float32'>
INFO: output_name.1: /backbone/stage2/stage2.1/branch2/branch2.0/Conv_output_0 shape: [1, 58, 52, 52] dtype: float32
ERROR: The trace log is below.
Traceback (most recent call last):
File "/mnt/hdd10tb/Users/thaovu/.conda/envs/onnx2tf/lib/python3.8/site-packages/onnx2tf/utils/common_functions.py", line 261, in print_wrapper_func
result = func(*args, **kwargs)
File "/mnt/hdd10tb/Users/thaovu/.conda/envs/onnx2tf/lib/python3.8/site-packages/onnx2tf/utils/common_functions.py", line 323, in inverted_operation_enable_disable_wrapper_func
result = func(*args, **kwargs)
File "/mnt/hdd10tb/Users/thaovu/.conda/envs/onnx2tf/lib/python3.8/site-packages/onnx2tf/ops/Conv.py", line 153, in make_node
tf.nn.convolution(
File "/mnt/hdd10tb/Users/thaovu/.conda/envs/onnx2tf/lib/python3.8/site-packages/tensorflow/python/util/traceback_utils.py", line 153, in error_handler
raise e.with_traceback(filtered_tb) from None
File "/mnt/hdd10tb/Users/thaovu/.conda/envs/onnx2tf/lib/python3.8/site-packages/keras/layers/core/tf_op_layer.py", line 119, in handle
return TFOpLambda(op)(*args, **kwargs)
File "/mnt/hdd10tb/Users/thaovu/.conda/envs/onnx2tf/lib/python3.8/site-packages/keras/utils/traceback_utils.py", line 70, in error_handler
raise e.with_traceback(filtered_tb) from None
ValueError: Exception encountered when calling layer "tf.nn.convolution_4" (type TFOpLambda).
Depth of input (52) is not a multiple of input depth of filter (58) for '{{node tf.nn.convolution_4/convolution}} = Conv2D[T=DT_FLOAT, data_format="NHWC", dilations=[1, 1, 1, 1], explicit_paddings=[], padding="SAME", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true](Placeholder, tf.nn.convolution_4/convolution_internal/filters)' with input shapes: [1,58,52,52], [1,1,58,58].
Question 1: Shape of onnx::Conv_1393 layer must be [58,58,1,1], but in here it is [1,1,58,58]. Any cause from data_format?
Question 2: I set -k "data" but data_format of this layer is NHWC. Correct me if I wrong
3. How:
Firstly, I ran without -k option and get the same issue, I tried to use -k but still the the same
I also check parameter replacement but cant find suitable solution.
4. Why: I wanna convert NanoDet-plus-m model to Keras (h5 format)
5. Resources: Pytorch model from NanoDet model zoo https://github.com/RangiLyu/nanodet
Thank you for your hard working. Your projects are very useful <3
Others
1.5.29
1.13.0
2.10.0
https://s3.ap-northeast-2.wasabisys.com/temp-models/onnx2tf_146/yolox_x.onnx
https://github.com/PINTO0309/onnx2tf/releases/download/1.1.28/faster_rcnn-10.onnx
https://github.com/PINTO0309/onnx2tf/releases/download/1.1.28/yolov7_tiny_head_0.768_post_480x640.onnx
N/Aonnx2tf -i yolox_x.onnx -oh5
ValueError: This Keras op layer was generated from <function non_max_suppression at 0x7fd6a96ff1f0>,
a method that is not publicly exposed in the TensorFlow API.
This may have happened if the method was explicitly decorated to add dispatching support,
and it was used during Functional model construction.
To ensure cross-version compatibility of Keras models that use op layers,
only op layers produced from public TensorFlow API symbols can be serialized.
Others
1.5.7
1.13.0
2.10.0
{
"format_version": 1,
"operations": [
{
"op_name": "Resize_0",
"param_target": "inputs",
"param_name": "Concat_0",
"values": [48,48]
},
{
"op_name": "Resize_1",
"param_target": "inputs",
"param_name": "Concat_1",
"values": [48,48]
},
{
"op_name": "Resize_2",
"param_target": "inputs",
"param_name": "Concat_2",
"values": [48,48]
},
{
"op_name": "Resize_3",
"param_target": "inputs",
"param_name": "Concat_3",
"values": [24,24]
},
{
"op_name": "Resize_4",
"param_target": "inputs",
"param_name": "Concat_4",
"values": [48,48]
},
{
"op_name": "Resize_5",
"param_target": "inputs",
"param_name": "Concat_5",
"values": [48,48]
},
{
"op_name": "Resize_6",
"param_target": "inputs",
"param_name": "Concat_6",
"values": [48,48]
},
{
"op_name": "Resize_7",
"param_target": "inputs",
"param_name": "Concat_7",
"values": [24,24]
},
{
"op_name": "Resize_8",
"param_target": "inputs",
"param_name": "Concat_8",
"values": [24,24]
},
{
"op_name": "Resize_9",
"param_target": "inputs",
"param_name": "Concat_9",
"values": [12,12]
},
{
"op_name": "Resize_10",
"param_target": "inputs",
"param_name": "Concat_10",
"values": [48,48]
},
{
"op_name": "Resize_11",
"param_target": "inputs",
"param_name": "Concat_11",
"values": [48,48]
},
{
"op_name": "Resize_12",
"param_target": "inputs",
"param_name": "Concat_12",
"values": [48,48]
},
{
"op_name": "Resize_13",
"param_target": "inputs",
"param_name": "Concat_14",
"values": [192,192]
},
{
"op_name": "Reshape_0",
"param_target": "outputs",
"param_name": "Reshape_0",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Reshape_1",
"param_target": "outputs",
"param_name": "Reshape_1",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Transpose_0",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
},
{
"op_name": "Transpose_1",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
},
{
"op_name": "Softmax_0",
"param_target": "attributes",
"param_name": "axis",
"values": 3
}
]
}If the output layer to be verified is a Constant value, Keras appears to Abort. Therefore, we believe that Constant and ConstantOfShape should be excluded from the validation target. This should not be determined by the type of OP, but rather by whether it is an np.ndarray, in order to make a generic description.
onnx2tf -i human_segmentation_pphumanseg_2021oct.onnx -prf replace.json -cotof -cotoa 1e-1
ValueError: Output tensors of a Functional model must be the output of a TensorFlow `Layer` (thus holding past layer metadata). Found: [[[[-4.45949372e-05 2.32167356e-03 3.76565661e-03 ...
Others
1.1.24
https://drive.google.com/file/d/15TV0y-Go0VYdyV4FKNhciVDWa3frECfm/view?usp=sharing
{
"format_version": 1,
"operations": [
{
"op_name": "/encoder/stages.0/layers/layers.0/Transpose_1",
"param_target": "outputs",
"param_name": "/encoder/stages.0/layers/layers.0/Transpose_1_output_0",
"post_process_transpose_perm": [0,2,3,1]
}
]
}Hi - thanks for the super useful library.
I'm having trouble diagnosing the problem with an add op - I've narrowed it down to the op, but can't seem to get the right adjustment in the parameter_replacement.json file to fix it, and am not sure exactly how to spot the issue in the attached screenshots.
Many thanks!
Others
I am trying to convert above model.onnx to tflite format by executing onnx2tf -i EyeNet.onnx, but it throw error at the concat op:
log:
Model loaded ========================================================================
Model convertion started ============================================================
INFO: input_op_name: input shape: [1, 1, 192, 192] dtype: float32
INFO: onnx_op_type: Conv onnx_op_name: Conv_0
INFO: input_name.1: input shape: [1, 1, 192, 192] dtype: float32
INFO: input_name.2: onnx::Conv_456 shape: [8, 1, 3, 3] dtype: <class 'numpy.float32'>
INFO: input_name.3: onnx::Conv_457 shape: [8] dtype: <class 'numpy.float32'>
INFO: output_name.1: input.4 shape: None dtype: None
INFO: tf_op_type: convolution_v2
INFO: input.1.input: name: tf.compat.v1.pad/Pad:0 shape: (1, 194, 194, 1) dtype: <dtype: 'float32'>
INFO: input.2.weights: shape: (3, 3, 1, 8) dtype: float32
INFO: input.3.bias: shape: (8,) dtype: float32
INFO: output.1.output: name: tf.math.add/Add:0 shape: (1, 96, 96, 8) dtype: <dtype: 'float32'>
INFO: onnx_op_type: LeakyRelu onnx_op_name: LeakyRelu_1
INFO: input_name.1: input.4 shape: None dtype: None
INFO: output_name.1: onnx::Conv_282 shape: None dtype: None
INFO: tf_op_type: leaky_relu
INFO: input.1.features: name: tf.math.add/Add:0 shape: (1, 96, 96, 8) dtype: <dtype: 'float32'>
INFO: input.2.alpha: val: 0.10000000149011612
INFO: output.1.output: name: tf.nn.leaky_relu/LeakyRelu:0 shape: (1, 96, 96, 8) dtype: <dtype: 'float32'>
INFO: onnx_op_type: Conv onnx_op_name: Conv_2
INFO: input_name.1: onnx::Conv_282 shape: None dtype: None
INFO: input_name.2: onnx::Conv_459 shape: [8, 1, 3, 3] dtype: <class 'numpy.float32'>
INFO: input_name.3: onnx::Conv_460 shape: [8] dtype: <class 'numpy.float32'>
INFO: output_name.1: input.12 shape: None dtype: None
INFO: tf_op_type: depthwise_conv2d_v2
INFO: input.1.input: name: tf.compat.v1.pad_1/Pad:0 shape: (1, 98, 98, 8) dtype: <dtype: 'float32'>
INFO: input.2.weights: shape: (3, 3, 8, 1) dtype: <dtype: 'float32'>
INFO: input.3.bias: shape: (8,) dtype: float32
INFO: output.1.output: name: tf.math.add_1/Add:0 shape: (1, 48, 48, 8) dtype: <dtype: 'float32'>
INFO: onnx_op_type: LeakyRelu onnx_op_name: LeakyRelu_3
INFO: input_name.1: input.12 shape: None dtype: None
INFO: output_name.1: onnx::Conv_285 shape: None dtype: None
INFO: tf_op_type: leaky_relu
INFO: input.1.features: name: tf.math.add_1/Add:0 shape: (1, 48, 48, 8) dtype: <dtype: 'float32'>
INFO: input.2.alpha: val: 0.10000000149011612
INFO: output.1.output: name: tf.nn.leaky_relu_1/LeakyRelu:0 shape: (1, 48, 48, 8) dtype: <dtype: 'float32'>
INFO: onnx_op_type: Conv onnx_op_name: Conv_4
INFO: input_name.1: onnx::Conv_285 shape: None dtype: None
INFO: input_name.2: onnx::Conv_462 shape: [16, 8, 1, 1] dtype: <class 'numpy.float32'>
INFO: input_name.3: onnx::Conv_463 shape: [16] dtype: <class 'numpy.float32'>
INFO: output_name.1: input.20 shape: None dtype: None
INFO: tf_op_type: convolution_v2
INFO: input.1.input: name: tf.nn.leaky_relu_1/LeakyRelu:0 shape: (1, 48, 48, 8) dtype: <dtype: 'float32'>
INFO: input.2.weights: shape: (1, 1, 8, 16) dtype: float32
INFO: input.3.bias: shape: (16,) dtype: float32
INFO: output.1.output: name: tf.math.add_2/Add:0 shape: (1, 48, 48, 16) dtype: <dtype: 'float32'>
INFO: onnx_op_type: LeakyRelu onnx_op_name: LeakyRelu_5
INFO: input_name.1: input.20 shape: None dtype: None
INFO: output_name.1: onnx::Conv_288 shape: None dtype: None
INFO: tf_op_type: leaky_relu
INFO: input.1.features: name: tf.math.add_2/Add:0 shape: (1, 48, 48, 16) dtype: <dtype: 'float32'>
INFO: input.2.alpha: val: 0.10000000149011612
INFO: output.1.output: name: tf.nn.leaky_relu_2/LeakyRelu:0 shape: (1, 48, 48, 16) dtype: <dtype: 'float32'>
INFO: onnx_op_type: Conv onnx_op_name: Conv_6
INFO: input_name.1: onnx::Conv_288 shape: None dtype: None
INFO: input_name.2: onnx::Conv_465 shape: [16, 1, 3, 3] dtype: <class 'numpy.float32'>
INFO: input_name.3: onnx::Conv_466 shape: [16] dtype: <class 'numpy.float32'>
INFO: output_name.1: input.28 shape: None dtype: None
INFO: tf_op_type: depthwise_conv2d_v2
INFO: input.1.input: name: tf.compat.v1.pad_2/Pad:0 shape: (1, 50, 50, 16) dtype: <dtype: 'float32'>
INFO: input.2.weights: shape: (3, 3, 16, 1) dtype: <dtype: 'float32'>
INFO: input.3.bias: shape: (16,) dtype: float32
INFO: output.1.output: name: tf.math.add_3/Add:0 shape: (1, 48, 48, 16) dtype: <dtype: 'float32'>
INFO: onnx_op_type: LeakyRelu onnx_op_name: LeakyRelu_7
INFO: input_name.1: input.28 shape: None dtype: None
INFO: output_name.1: onnx::Conv_291 shape: None dtype: None
INFO: tf_op_type: leaky_relu
INFO: input.1.features: name: tf.math.add_3/Add:0 shape: (1, 48, 48, 16) dtype: <dtype: 'float32'>
INFO: input.2.alpha: val: 0.10000000149011612
INFO: output.1.output: name: tf.nn.leaky_relu_3/LeakyRelu:0 shape: (1, 48, 48, 16) dtype: <dtype: 'float32'>
INFO: onnx_op_type: Conv onnx_op_name: Conv_8
INFO: input_name.1: onnx::Conv_291 shape: None dtype: None
INFO: input_name.2: onnx::Conv_468 shape: [16, 16, 1, 1] dtype: <class 'numpy.float32'>
INFO: input_name.3: onnx::Conv_469 shape: [16] dtype: <class 'numpy.float32'>
INFO: output_name.1: input.36 shape: None dtype: None
INFO: tf_op_type: convolution_v2
INFO: input.1.input: name: tf.nn.leaky_relu_3/LeakyRelu:0 shape: (1, 48, 48, 16) dtype: <dtype: 'float32'>
INFO: input.2.weights: shape: (1, 1, 16, 16) dtype: float32
INFO: input.3.bias: shape: (16,) dtype: float32
INFO: output.1.output: name: tf.math.add_4/Add:0 shape: (1, 48, 48, 16) dtype: <dtype: 'float32'>
INFO: onnx_op_type: LeakyRelu onnx_op_name: LeakyRelu_9
INFO: input_name.1: input.36 shape: None dtype: None
INFO: output_name.1: onnx::Conv_294 shape: None dtype: None
INFO: tf_op_type: leaky_relu
INFO: input.1.features: name: tf.math.add_4/Add:0 shape: (1, 48, 48, 16) dtype: <dtype: 'float32'>
INFO: input.2.alpha: val: 0.10000000149011612
INFO: output.1.output: name: tf.nn.leaky_relu_4/LeakyRelu:0 shape: (1, 48, 48, 16) dtype: <dtype: 'float32'>
INFO: onnx_op_type: Conv onnx_op_name: Conv_10
INFO: input_name.1: onnx::Conv_294 shape: None dtype: None
INFO: input_name.2: onnx::Conv_471 shape: [8, 16, 3, 3] dtype: <class 'numpy.float32'>
INFO: input_name.3: onnx::Conv_472 shape: [8] dtype: <class 'numpy.float32'>
INFO: output_name.1: onnx::Concat_470 shape: None dtype: None
INFO: tf_op_type: convolution_v2
INFO: input.1.input: name: tf.compat.v1.pad_3/Pad:0 shape: (1, 50, 50, 16) dtype: <dtype: 'float32'>
INFO: input.2.weights: shape: (3, 3, 16, 8) dtype: float32
INFO: input.3.bias: shape: (8,) dtype: float32
INFO: output.1.output: name: tf.math.add_5/Add:0 shape: (1, 48, 48, 8) dtype: <dtype: 'float32'>
INFO: onnx_op_type: Conv onnx_op_name: Conv_11
INFO: input_name.1: onnx::Conv_294 shape: None dtype: None
INFO: input_name.2: onnx::Conv_474 shape: [4, 16, 3, 3] dtype: <class 'numpy.float32'>
INFO: input_name.3: onnx::Conv_475 shape: [4] dtype: <class 'numpy.float32'>
INFO: output_name.1: input.48 shape: None dtype: None
INFO: tf_op_type: convolution_v2
INFO: input.1.input: name: tf.compat.v1.pad_4/Pad:0 shape: (1, 50, 50, 16) dtype: <dtype: 'float32'>
INFO: input.2.weights: shape: (3, 3, 16, 4) dtype: float32
INFO: input.3.bias: shape: (4,) dtype: float32
INFO: output.1.output: name: tf.math.add_6/Add:0 shape: (1, 48, 48, 4) dtype: <dtype: 'float32'>
INFO: onnx_op_type: LeakyRelu onnx_op_name: LeakyRelu_12
INFO: input_name.1: input.48 shape: None dtype: None
INFO: output_name.1: onnx::Conv_299 shape: None dtype: None
INFO: tf_op_type: leaky_relu
INFO: input.1.features: name: tf.math.add_6/Add:0 shape: (1, 48, 48, 4) dtype: <dtype: 'float32'>
INFO: input.2.alpha: val: 0.10000000149011612
INFO: output.1.output: name: tf.nn.leaky_relu_5/LeakyRelu:0 shape: (1, 48, 48, 4) dtype: <dtype: 'float32'>
INFO: onnx_op_type: Conv onnx_op_name: Conv_13
INFO: input_name.1: onnx::Conv_299 shape: None dtype: None
INFO: input_name.2: onnx::Conv_477 shape: [4, 4, 3, 3] dtype: <class 'numpy.float32'>
INFO: input_name.3: onnx::Conv_478 shape: [4] dtype: <class 'numpy.float32'>
INFO: output_name.1: onnx::Concat_476 shape: None dtype: None
INFO: tf_op_type: convolution_v2
INFO: input.1.input: name: tf.compat.v1.pad_5/Pad:0 shape: (1, 50, 50, 4) dtype: <dtype: 'float32'>
INFO: input.2.weights: shape: (3, 3, 4, 4) dtype: float32
INFO: input.3.bias: shape: (4,) dtype: float32
INFO: output.1.output: name: tf.math.add_7/Add:0 shape: (1, 48, 48, 4) dtype: <dtype: 'float32'>
INFO: onnx_op_type: Conv onnx_op_name: Conv_14
INFO: input_name.1: onnx::Conv_299 shape: None dtype: None
INFO: input_name.2: onnx::Conv_480 shape: [4, 4, 3, 3] dtype: <class 'numpy.float32'>
INFO: input_name.3: onnx::Conv_481 shape: [4] dtype: <class 'numpy.float32'>
INFO: output_name.1: input.60 shape: None dtype: None
INFO: tf_op_type: convolution_v2
INFO: input.1.input: name: tf.compat.v1.pad_6/Pad:0 shape: (1, 50, 50, 4) dtype: <dtype: 'float32'>
INFO: input.2.weights: shape: (3, 3, 4, 4) dtype: float32
INFO: input.3.bias: shape: (4,) dtype: float32
INFO: output.1.output: name: tf.math.add_8/Add:0 shape: (1, 48, 48, 4) dtype: <dtype: 'float32'>
INFO: onnx_op_type: LeakyRelu onnx_op_name: LeakyRelu_15
INFO: input_name.1: input.60 shape: None dtype: None
INFO: output_name.1: onnx::Conv_304 shape: None dtype: None
INFO: tf_op_type: leaky_relu
INFO: input.1.features: name: tf.math.add_8/Add:0 shape: (1, 48, 48, 4) dtype: <dtype: 'float32'>
INFO: input.2.alpha: val: 0.10000000149011612
INFO: output.1.output: name: tf.nn.leaky_relu_6/LeakyRelu:0 shape: (1, 48, 48, 4) dtype: <dtype: 'float32'>
INFO: onnx_op_type: Conv onnx_op_name: Conv_16
INFO: input_name.1: onnx::Conv_304 shape: None dtype: None
INFO: input_name.2: onnx::Conv_483 shape: [4, 4, 3, 3] dtype: <class 'numpy.float32'>
INFO: input_name.3: onnx::Conv_484 shape: [4] dtype: <class 'numpy.float32'>
INFO: output_name.1: onnx::Concat_482 shape: None dtype: None
INFO: tf_op_type: convolution_v2
INFO: input.1.input: name: tf.compat.v1.pad_7/Pad:0 shape: (1, 50, 50, 4) dtype: <dtype: 'float32'>
INFO: input.2.weights: shape: (3, 3, 4, 4) dtype: float32
INFO: input.3.bias: shape: (4,) dtype: float32
INFO: output.1.output: name: tf.math.add_9/Add:0 shape: (1, 48, 48, 4) dtype: <dtype: 'float32'>
INFO: onnx_op_type: Concat onnx_op_name: Concat_17
INFO: input_name.1: onnx::Concat_470 shape: None dtype: None
INFO: input_name.2: onnx::Concat_476 shape: None dtype: None
INFO: input_name.3: onnx::Concat_482 shape: None dtype: None
INFO: output_name.1: input.68 shape: None dtype: None
ERROR: The trace log is below.
Traceback (most recent call last):
File "c:\users\user\appdata\local\programs\python\python38\lib\site-packages\onnx2tf\utils\common_functions.py", line 176, in print_wrapper_func
result = func(*args, **kwargs)
File "c:\users\user\appdata\local\programs\python\python38\lib\site-packages\onnx2tf\utils\common_functions.py", line 225, in inverted_operation_enable_disable_wrapper_func
result = func(*args, **kwargs)
File "c:\users\user\appdata\local\programs\python\python38\lib\site-packages\onnx2tf\ops\Concat.py", line 61, in make_node
tensor_rank=len(shape),
TypeError: object of type 'NoneType' has no len()
Others
Default resnet18-v1-7.onnx has the same issue
I used example resnet onnx model... The converted saved model doesnt have required Signatures. I was trying. to use tensorflowjs_converter to convert it into tfjs model and it was failing in the missing signatures.
heres is the. output. of saved_model_cli for resnet:
MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:
signature_def['__saved_model_init_op']:
The given SavedModel SignatureDef contains the following input(s):
The given SavedModel SignatureDef contains the following output(s):
outputs['__saved_model_init_op'] tensor_info:
dtype: DT_INVALID
shape: unknown_rank
name: NoOp
Method name is:
Others
1.3.14
1.12.0
2.10.0
https://drive.google.com/file/d/1XeM7XpjOsL37qAdpWa1j6x9KMzYbiHgW/view?usp=sharing
{}import torch
import tensorflow as tf
import numpy as np
from onnx2tf import convert
class PixelShuffle(torch.nn.Module):
def __init__(self, upscale_factor=2):
super().__init__()
self._upscale_factor = upscale_factor
def forward(self, inp):
out = torch.nn.functional.pixel_shuffle(inp, upscale_factor=self._upscale_factor)
return out
upscale_factor = 2
b, c, h, w = 2, 3, 8, 8
max_val = b*c*h*w*upscale_factor*upscale_factor
ps = PixelShuffle(upscale_factor=upscale_factor)
inp = (torch.arange(max_val)/max_val).view(b, c*upscale_factor*upscale_factor, h, w)
out_torch = ps(inp)
_ = torch.onnx.export(ps,
torch.randn_like(inp),
"ps.onnx",
opset_version=17,
export_params=True,
verbose=False,
input_names=["input"],
output_names=["output"])
convert(input_onnx_file_path="ps.onnx",
output_folder_path="ps_tf",
non_verbose=True)
interpreter = tf.lite.Interpreter(model_path="ps_tf/model_float32.tflite")
interpreter.allocate_tensors()
# Get input and output tensors
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
input_data = (inp).permute(0,2,3,1).cpu().numpy()
interpreter.set_tensor(input_details[0]['index'], input_data)
interpreter.invoke()
out_tf = interpreter.get_tensor(output_details[0]['index'])
print(np.allclose(out_torch.permute(0,2,3,1).numpy(), out_tf))This gives False output.
3. I fixed the bug in the DepthToSpace op. Creating Pull Request now.
Others
N/A
N/A
2.10.0
N/A
N/AI tried to install onnx2tf in colab as instructed
!sudo add-apt-repository -y ppa:deadsnakes/ppa
...
!python3.9 -m pip install tensorflow==2.10.0 \
...
&& python3.9 -m pip install -U onnx2tf \
&& python3.9 -m pip install -U protobuf==3.20.3
But I got the following error
tensorflow 2.10.0 requires protobuf<3.20,>=3.9.2, but you have protobuf 3.20.3 which is incompatible.
tensorboard 2.10.1 requires protobuf<3.20,>=3.9.2, but you have protobuf 3.20.3 which is incompatible.
Is that expected?
Thanks!
Others
1.5.28
1.13.0
2.10.0
https://s3.ap-northeast-2.wasabisys.com/temp-models/onnx2tf_146/yolox_x.onnx
https://github.com/PINTO0309/onnx2tf/releases/download/1.1.28/faster_rcnn-10.onnx
N/Aonnx2tf -i yolox_x.onnx -oh5
'numpy.dtype[int64]' object has no attribute 'item'
File "/home/xxxxx/git/onnx2tf/onnx2tf/onnx2tf.py", line 744, in convert
model.save(f'{output_folder_path}/{output_file_name}_float32.h5')
File "/home/xxxxx/git/onnx2tf/onnx2tf/onnx2tf.py", line 1735, in main
model = convert(
File "/home/xxxxx/git/onnx2tf/onnx2tf/onnx2tf.py", line 1786, in <module>
main()
AttributeError: 'numpy.dtype[int64]' object has no attribute 'item'
Others
1.1.23
gist for reproduction : https://colab.research.google.com/gist/Hyunseok-Kim0/41920f44e994caa96271d604e5831775/onnx2tf_batchnorm_test.ipynb
N/AFirst of all, thank you for sharing another wonderful tool. onnx2tf looks very stable and useful.
Purpose: Personal development
What : I tried to convert centernet (https://github.com/xingyizhou/CenterNet) to tflite. It was very impressive that initial conversion from onnx to tflite was completed without an error using onnx2tf. However, I found that generated tflite using onnx2tf shows different output comparing with the one generated from openvino2tensorflow. Most of network components were identical except the TransposeConv layer, which was followed by Batchnorm layer.
part of tflite from openvino2tensorflow (generates correct output)
part of tflite from onnx2tf (generates wrong output)
As you see in the figures above, the TransposeConv layer from openvino2tensorflow has no bias while onnx2tf version has some bias. Also, the Add layers which is converted from Batchnorm layer has different values which are shown in the right side of images.
I think this problem is related with Batchnorm layer, but couldn't find solution.
target onnx
tflite from openvino2tensorflow
tflite from onnx2tf
In tflite from onnx2tf, bias of Batchnrom layer is moved to Add layer, and another unknown bias value is added to TransposeConv layer. In tflite from openvino2tensorflow, TransposeConv layer has no bias and unknown value is added in Add layer.
Also, I found there is typo in ConvTranspose.py in line 68. tf_layers_dict[input_weights.name] instead of tf_layers_dict[input_bias.name], but it does not look like cause of this problem..
onnx2tf/onnx2tf/ops/ConvTranspose.py
Lines 62 to 70 in b275740
I will keep try to find out the solution, please share your idea about this problem.
Others
1.2.4
https://github.com/snakers4/silero-vad/blob/v3.1/files/silero_vad.onnx
{}INFO: onnx_op_type: Pad
INFO: input shape: [1, 1, 1, 512]
...
INFO: output shape: [1, 1, 1, 768]
INFO: tf_op_type: Pad
INFO: input shape: [1, 1, 1, 512]
...
INFO: output shape: [1, 1, 257, 512]
ValueError: Exception encountered when calling layer "tf.squeeze"
Can not squeeze dim[3], expected a dimension of 1, got 512
So why the Pad's behavior is not the same during conversion and what should I do?
Others
1.0.27
https://drive.google.com/file/d/1Wkb-xi8NBICJttIyctUL3m7ETGHv2BZ8/view
{
"format_version": 1,
"operations": [
{
"op_name": "Flatten_10",
"param_target": "inputs",
"param_name": "input.16",
"pre_process_transpose_perm": [3,1,2,0]
},
{
"op_name": "Flatten_10",
"param_target": "outputs",
"param_name": "onnx::Gemm_31",
"post_process_transpose_perm": [1, 0]
}
]
}onnx2tf -i baseline_simplified.onnx -b 1The onnx model has a Flatten layer. Input shape is ['batch_size', 32, 1, 3], output shape is ['batch_size', 96]. The OP get converted to tf.reshape, the input shape is (1, 1, 3, 32), output shape is (3, 32).
I attempted to solve it with the Parameter Replacement JSON to tranpose both the input and output. The transpose for the output works, but the transpose for the input gives me this error.
I need the tflite model so that I can quantise it and deploy using TFLiteMicro.
I have tried to use openvino library to convert onnx to openvino, then openvino to tf. I can convert but the output of tf model is very different from the onnx output.
Others
1.5.23
1.12.0
2.10
https://github.com/PINTO0309/onnx2tf/releases/download/1.1.28/pidnet_S_cityscapes_192x320.onnx
N/Aonnx_tf_tensor_validation use tensor shape to find out matching pairs between onnx and tflite outputs. Due to this logic, some outputs are failed to matched and generates wrong absolute error.
The output of global average pooling has shape of (1 x 512 x 1 x 1), and another average pooling output also have same shape and same value. When onnx_tf_tensor_validation is running, the output of global average pooling is failed to be matched with correct onnx output, generates Unmatched result as below.
Similar bug can be happen in other onnx files if some outputs have same shape and same value.
Others
0.28
Added screenshots of the models to show bias values
My model. is getting converted properly but I see results mismatch, so. trying to. figure out whats causing differences.
One of the possible issues is due to how. the bias added from Conv2D operator while conversion.
Adding screenshots of the bias in Conv2D. operator. from original. onnx model and corresponding conversion to tf. using the tool.
You can see the bias at index [1] is added in the last. Looks weird. Please correct me If I am wrong.
Others
0.10.0
None
NoneHi @PINTO0309,
Thanks for great repo. I want to ask can you convert a model with grid sample from pytorch to tflite?
Hope get your reply soon. Much thanks
Others
1.4.2
1.12.0
2.10.1
9e: https://drive.google.com/file/d/1vGzO9MZGX-yGz6ATm4yHVJMASZACuy2t/view?usp=share_link
9t: https://drive.google.com/file/d/1Be-a7Pmo6auyAXHChAhC1qtzkbr5EytJ/view?usp=share_link
12e: https://drive.google.com/file/d/1NR-0Fm5q_ludb5MTWgDPzRzqDFKOsCkw/view?usp=share_link
12t: https://drive.google.com/file/d/1ReuXt4gpbq6O7mpUGudaBKjb6wI43ka6/view?usp=share_link
9e: https://drive.google.com/file/d/1av1dL5sWfjlghf2-jtmws-CR-Tcv_WIP/view?usp=share_link
9t: https://drive.google.com/file/d/18qB_VOW4Xp9PvCN4G-eRRnWhh9o1fF3F/view?usp=share_link
12e: https://drive.google.com/file/d/1T0nTkb5Szf9XjaNVVBTzzu5C9dgOURLd/view?usp=share_link
12t: https://drive.google.com/file/d/1JhH8K64PDJong4qkZbzyyJDKRExE7wlJ/view?usp=share_linkHi Master PINTO,
Following up on issue #103, the output values of the converted model are different from what is expected. I tried to export the pytorch model in both eval and training and both opset==9 and opset==12. For the models and parameter replacement files, the number stands for opset version, e stands for eval, and t stands for training. This issue happens in all scenarios. Please see this notebook https://drive.google.com/file/d/1mErgTLFiYGTgUnMRDcNp3HkOkSWa96aL/view?usp=share_link to replicate and for issue demonstration.
Also, I see that since you match every onnx operation with a basic tf operation, is there no way to restore those weights as trainable parameters? Thank you so much for keep digging in on this. If you need the original pytorch model or anything else I could assist on, let me know!
Others
1.1.4
{
"format_version": 1,
"operations": [
{
"op_name": "tower0/network/layer1/conv/Relu6;tower0/network/layer1/conv/BatchNorm/FusedBatchNormV3;tower0/network/layer2/sepconv/BatchNorm/FusedBatchNormV3;tower0/network/layer2/sepconv/depthwise;tower0/network/layer1/conv/Conv2D_prequant__219",
"param_target": "inputs",
"param_name": "new_shape__513",
"values": [1,96,64,1]
}
]
}onnx2tf -i lite-model_yamnet_classification_tflite_1.onnx -prf replace.jsonTypeError: EndVector() missing 1 required positional argument: 'vectorNumElems'From:
pip show flatbuffers
Name: flatbuffers
Version: 1.12
To:
pip install -U flatbuffers
pip show flatbuffers
Name: flatbuffers
Version: 22.10.26
pip install -U flatbuffers
Others
1.0.48
Model:
1.face_recognition_sface_2021dec.onnx
2.https://github.com/opencv/opencv_zoo/blob/master/models/face_recognition_sface/face_recognition_sface_2021dec-act_int8-wt_int8-quantized.onnx
NAHi @PINTO0309
Thanks for your job and it working when ONNX covert to tflite without quantizion.
for quantized ONNX model or tflite model, the tool seems not working for belwo error
ERROR: QLinearMul OP is not yet implemented.
so i would like to check with you
whether the tool support NCHW between NHWC convertion for tflite model?
Thanks
Others
1.1.44
onnx/onnx-tensorflow#693
https://github.com/sayakpaul/Adventures-in-TensorFlow-Lite/releases/download/v0.2.0/tunit_onnx_files.tar.gz
{
"format_version": 1,
"operations": [
{
"op_name": "Reshape_136",
"param_target": "outputs",
"param_name": "194",
"post_process_transpose_perm": [0,2,1]
},
{
"op_name": "Reshape_143",
"param_target": "outputs",
"param_name": "205",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Reshape_170",
"param_target": "outputs",
"param_name": "234",
"post_process_transpose_perm": [0,2,1]
},
{
"op_name": "Reshape_177",
"param_target": "outputs",
"param_name": "245",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Reshape_206",
"param_target": "outputs",
"param_name": "276",
"post_process_transpose_perm": [0,2,1]
},
{
"op_name": "Reshape_213",
"param_target": "outputs",
"param_name": "287",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Reshape_240",
"param_target": "outputs",
"param_name": "316",
"post_process_transpose_perm": [0,2,1]
},
{
"op_name": "Reshape_247",
"param_target": "outputs",
"param_name": "327",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Reshape_303",
"param_target": "outputs",
"param_name": "385",
"post_process_transpose_perm": [0,2,1]
},
{
"op_name": "Reshape_310",
"param_target": "outputs",
"param_name": "396",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Reshape_364",
"param_target": "outputs",
"param_name": "452",
"post_process_transpose_perm": [0,2,1]
},
{
"op_name": "Reshape_371",
"param_target": "outputs",
"param_name": "463",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Reshape_425",
"param_target": "outputs",
"param_name": "519",
"post_process_transpose_perm": [0,2,1]
},
{
"op_name": "Reshape_432",
"param_target": "outputs",
"param_name": "530",
"post_process_transpose_perm": [0,2,3,1]
}
]
}1.Personal development
https://github.com/clovaai/tunit
2.TypeError: 'Variable' object is not iterable
3.replace.json
5.Error log
ERROR: The trace log is below.
Traceback (most recent call last):
File "/home/xxxxx/git/onnx2tf/onnx2tf/utils/common_functions.py", line 262, in print_wrapper_func
result = func(*args, **kwargs)
File "/home/xxxxx/git/onnx2tf/onnx2tf/utils/common_functions.py", line 324, in inverted_operation_enable_disable_wrapper_func
result = func(*args, **kwargs)
File "/home/xxxxx/git/onnx2tf/onnx2tf/ops/BatchNormalization.py", line 41, in make_node
graph_node.outputs = graph_node.outputs[0]
File "/home/xxxxx/.local/lib/python3.8/site-packages/onnx_graphsurgeon/ir/node.py", line 89, in __setattr__
getattr(self, name).extend(value)
File "/home/xxxxx/.local/lib/python3.8/site-packages/onnx_graphsurgeon/util/misc.py", line 111, in extend
super().extend(iterable)
TypeError: 'Variable' object is not iterable
Others
1.7.x
1.13.1
2.10.0
N/ANonMaxSuppressionV4. Perform this workaround only when the --replace_nonmaxsuppression_to_pseudo_nonmaxsuppression option is enabled. However, this model customization should originally be implemented on the PyTorch side of the PyTorch implementation that exports ONNX. Therefore, the full-scratch NMS processing group generated by this tool is a very specific custom that can only be used for very limited models.Others
1.5.12
1.13.0
2.10.0
https://github.com/PINTO0309/onnx2tf/releases/download/1.1.28/fastestdet.onnx
https://github.com/PINTO0309/onnx2tf/releases/download/1.1.28/ppmattingv2_stdc1_human_480x640.onnx
N/AAveragePool is quite large. In addition, there is almost no error in other OPs.
onnx2tf -i fastestdet.onnx -cotof -cotoa 1e-4
Others
1.5.33
1.13.1
2.10.0
https://github.com/PINTO0309/onnx2tf/releases/download/1.1.28/alike_t_opset11_192x320.onnx
N/AMaxPool with padding is very large. However, I believe a separate investigation is needed to determine if this is a problem with the logic used to verify accuracy or with MaxPool's padding process.
Others
1.5.x
1.13.1
2.10.0
N/A
N/AAtan / Atan2
Others
1.1.38
{
"format_version": 1,
"operations": [
{
"op_name": "Gather_18",
"param_target": "outputs",
"param_name": "onnx::Concat_445",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Gather_33",
"param_target": "outputs",
"param_name": "onnx::Concat_463",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Gather_48",
"param_target": "outputs",
"param_name": "onnx::Concat_481",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Gather_72",
"param_target": "outputs",
"param_name": "onnx::Concat_513",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Gather_87",
"param_target": "outputs",
"param_name": "onnx::Concat_531",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Gather_102",
"param_target": "outputs",
"param_name": "onnx::Concat_549",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Gather_117",
"param_target": "outputs",
"param_name": "onnx::Concat_567",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Gather_132",
"param_target": "outputs",
"param_name": "onnx::Concat_585",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Gather_147",
"param_target": "outputs",
"param_name": "onnx::Concat_603",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Gather_162",
"param_target": "outputs",
"param_name": "onnx::Concat_621",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Gather_186",
"param_target": "outputs",
"param_name": "onnx::Concat_653",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Gather_201",
"param_target": "outputs",
"param_name": "onnx::Concat_671",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Gather_216",
"param_target": "outputs",
"param_name": "onnx::Concat_689",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Transpose_261",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
},
{
"op_name": "Transpose_263",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
}
]
}onnx2tf -i FastestDet.onnx
ERROR: The trace log is below.
Traceback (most recent call last):
File "/home/b920405/git/onnx2tf/onnx2tf/utils/common_functions.py", line 262, in print_wrapper_func
result = func(*args, **kwargs)
File "/home/b920405/git/onnx2tf/onnx2tf/utils/common_functions.py", line 324, in inverted_operation_enable_disable_wrapper_func
result = func(*args, **kwargs)
File "/home/b920405/git/onnx2tf/onnx2tf/ops/Conv.py", line 153, in make_node
tf.nn.convolution(
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/util/traceback_utils.py", line 153, in error_handler
raise e.with_traceback(filtered_tb) from None
File "/usr/local/lib/python3.8/dist-packages/keras/layers/core/tf_op_layer.py", line 119, in handle
return TFOpLambda(op)(*args, **kwargs)
File "/usr/local/lib/python3.8/dist-packages/keras/utils/traceback_utils.py", line 70, in error_handler
raise e.with_traceback(filtered_tb) from None
ValueError: Exception encountered when calling layer "tf.nn.convolution_4" (type TFOpLambda).
Depth of input (44) is not a multiple of input depth of filter (24) for '{{node tf.nn.convolution_4/convolution}} = Conv2D[T=DT_FLOAT, data_format="NHWC", dilations=[1, 1, 1, 1], explicit_paddings=[], padding="SAME", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true](Placeholder, tf.nn.convolution_4/convolution_internal/filters)' with input shapes: [1,24,44,44], [1,1,24,24].
Others
1.1.11
https://github.com/opencv/opencv_zoo/tree/master/models/human_segmentation_pphumanseg
{
"format_version": 1,
"operations": [
{
"op_name": "Resize_0",
"param_target": "inputs",
"param_name": "Concat_0",
"values": [48,48]
},
{
"op_name": "Resize_1",
"param_target": "inputs",
"param_name": "Concat_1",
"values": [48,48]
},
{
"op_name": "Resize_2",
"param_target": "inputs",
"param_name": "Concat_2",
"values": [48,48]
},
{
"op_name": "Resize_3",
"param_target": "inputs",
"param_name": "Concat_3",
"values": [24,24]
},
{
"op_name": "Resize_4",
"param_target": "inputs",
"param_name": "Concat_4",
"values": [48,48]
},
{
"op_name": "Resize_5",
"param_target": "inputs",
"param_name": "Concat_5",
"values": [48,48]
},
{
"op_name": "Resize_6",
"param_target": "inputs",
"param_name": "Concat_6",
"values": [48,48]
},
{
"op_name": "Resize_7",
"param_target": "inputs",
"param_name": "Concat_7",
"values": [24,24]
},
{
"op_name": "Resize_8",
"param_target": "inputs",
"param_name": "Concat_8",
"values": [24,24]
},
{
"op_name": "Resize_9",
"param_target": "inputs",
"param_name": "Concat_9",
"values": [12,12]
},
{
"op_name": "Resize_10",
"param_target": "inputs",
"param_name": "Concat_10",
"values": [48,48]
},
{
"op_name": "Resize_11",
"param_target": "inputs",
"param_name": "Concat_11",
"values": [48,48]
},
{
"op_name": "Resize_12",
"param_target": "inputs",
"param_name": "Concat_12",
"values": [48,48]
},
{
"op_name": "Resize_13",
"param_target": "inputs",
"param_name": "Concat_14",
"values": [192,192]
},
{
"op_name": "Reshape_0",
"param_target": "outputs",
"param_name": "Reshape_0",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Reshape_1",
"param_target": "outputs",
"param_name": "Reshape_1",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Transpose_0",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
},
{
"op_name": "Transpose_1",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
},
{
"op_name": "Softmax_0",
"param_target": "attributes",
"param_name": "axis",
"values": 3
}
]
}onnx2tf -i human_segmentation_pphumanseg_2021oct.onnxTraceback (most recent call last):
File "/usr/lib/python3.8/runpy.py", line 194, in _run_module_as_main
return _run_code(code, main_globals, None,
File "/usr/lib/python3.8/runpy.py", line 87, in _run_code
exec(code, run_globals)
File "/home/xxxxx/.vscode/extensions/ms-python.python-2022.16.1/pythonFiles/lib/python/debugpy/adapter/../../debugpy/launcher/../../debugpy/__main__.py", line 39, in <module>
cli.main()
File "/home/xxxxx/.vscode/extensions/ms-python.python-2022.16.1/pythonFiles/lib/python/debugpy/adapter/../../debugpy/launcher/../../debugpy/../debugpy/server/cli.py", line 430, in main
run()
File "/home/xxxxx/.vscode/extensions/ms-python.python-2022.16.1/pythonFiles/lib/python/debugpy/adapter/../../debugpy/launcher/../../debugpy/../debugpy/server/cli.py", line 284, in run_file
runpy.run_path(target, run_name="__main__")
File "/home/xxxxx/.vscode/extensions/ms-python.python-2022.16.1/pythonFiles/lib/python/debugpy/_vendored/pydevd/_pydevd_bundle/pydevd_runpy.py", line 321, in run_path
return _run_module_code(code, init_globals, run_name,
File "/home/xxxxx/.vscode/extensions/ms-python.python-2022.16.1/pythonFiles/lib/python/debugpy/_vendored/pydevd/_pydevd_bundle/pydevd_runpy.py", line 135, in _run_module_code
_run_code(code, mod_globals, init_globals,
File "/home/xxxxx/.vscode/extensions/ms-python.python-2022.16.1/pythonFiles/lib/python/debugpy/_vendored/pydevd/_pydevd_bundle/pydevd_runpy.py", line 124, in _run_code
exec(code, run_globals)
File "/home/xxxxx/git/onnx2tf/onnx2tf/onnx2tf.py", line 1142, in <module>
main()
File "/home/xxxxx/git/onnx2tf/onnx2tf/onnx2tf.py", line 1108, in main
model = convert(
File "/home/xxxxx/git/onnx2tf/onnx2tf/onnx2tf.py", line 558, in convert
concrete_func = run_model.get_concrete_function(
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/eager/def_function.py", line 1239, in get_concrete_function
concrete = self._get_concrete_function_garbage_collected(*args, **kwargs)
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/eager/def_function.py", line 1219, in _get_concrete_function_garbage_collected
self._initialize(args, kwargs, add_initializers_to=initializers)
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/eager/def_function.py", line 785, in _initialize
self._stateful_fn._get_concrete_function_internal_garbage_collected( # pylint: disable=protected-access
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/eager/function.py", line 2523, in _get_concrete_function_internal_garbage_collected
graph_function, _ = self._maybe_define_function(args, kwargs)
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/eager/function.py", line 2760, in _maybe_define_function
graph_function = self._create_graph_function(args, kwargs)
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/eager/function.py", line 2670, in _create_graph_function
func_graph_module.func_graph_from_py_func(
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/framework/func_graph.py", line 1247, in func_graph_from_py_func
func_outputs = python_func(*func_args, **func_kwargs)
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/eager/def_function.py", line 677, in wrapped_fn
out = weak_wrapped_fn().__wrapped__(*args, **kwds)
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/framework/func_graph.py", line 1233, in autograph_handler
raise e.ag_error_metadata.to_exception(e)
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/framework/func_graph.py", line 1222, in autograph_handler
return autograph.converted_call(
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/autograph/impl/api.py", line 439, in converted_call
result = converted_f(*effective_args, **kwargs)
File "/tmp/__autograph_generated_fileldpw01_n.py", line 6, in <lambda>
tf__lam = (lambda *inputs: ag__.with_function_scope((lambda lscope: ag__.converted_call(model, (inputs,), None, lscope)), 'lscope', ag__.ConversionOptions(recursive=True, user_requested=True, optional_features=(), internal_convert_user_code=True)))
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/autograph/core/function_wrappers.py", line 113, in with_function_scope
return thunk(scope)
File "/tmp/__autograph_generated_fileldpw01_n.py", line 6, in <lambda>
tf__lam = (lambda *inputs: ag__.with_function_scope((lambda lscope: ag__.converted_call(model, (inputs,), None, lscope)), 'lscope', ag__.ConversionOptions(recursive=True, user_requested=True, optional_features=(), internal_convert_user_code=True)))
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/autograph/impl/api.py", line 377, in converted_call
return _call_unconverted(f, args, kwargs, options)
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/autograph/impl/api.py", line 459, in _call_unconverted
return f(*args)
File "/usr/local/lib/python3.8/dist-packages/keras/utils/traceback_utils.py", line 70, in error_handler
raise e.with_traceback(filtered_tb) from None
File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/framework/ops.py", line 1969, in _create_c_op
raise ValueError(e.message)
ValueError: in user code:
File "/home/xxxxx/git/onnx2tf/onnx2tf/onnx2tf.py", line 557, in None *
lambda *inputs : model(inputs)
File "/usr/local/lib/python3.8/dist-packages/keras/utils/traceback_utils.py", line 70, in error_handler **
raise e.with_traceback(filtered_tb) from None
ValueError: Exception encountered when calling layer "tf.math.add_6" " f"(type TFOpLambda).
Dimensions must be equal, but are 48 and 24 for '{{node model/tf.math.add_6/Add}} = AddV2[T=DT_FLOAT](model/tf.nn.relu_13/Relu, model/tf.image.resize/resize/ResizeBilinear)' with input shapes: [1,48,48,16], [1,24,48,16].
Call arguments received by layer "tf.math.add_6" " f"(type TFOpLambda):
• x=tf.Tensor(shape=(1, 48, 48, 16), dtype=float32)
• y=tf.Tensor(shape=(1, 24, 48, 16), dtype=float32)
• name='Add_5'
replace.json and run it.
onnx2tf -i human_segmentation_pphumanseg_2021oct.onnx -prf replace.jsonThe final conversion was successful. https://github.com/PINTO0309/onnx2tf/releases/tag/1.1.12
Others
1.3.6
1.12.0
2.10.0
INFO: onnx_op_type: MaxPool onnx_op_name: MaxPool_346
INFO: input_name.1: 645 shape: [1, 64, 22, 22] dtype: float32
INFO: output_name.1: 646 shape: [1, 64, 10, 10] dtype: float32
INFO: tf_op_type: max_pool_v2
INFO: input.1.input: name: tf.nn.relu_16/Relu:0 shape: (1, 64, 22, 22) dtype: <dtype: 'float32'>
INFO: input.2.filters:
INFO: input.3.kernel_shape: val: [3, 3]
INFO: input.4.strides: val: [2, 2]
INFO: input.5.dilations: val: [1, 1]
INFO: input.6.padding: val: VALID
INFO: input.7.ceil_mode: val: False
INFO: output.1.output: name: tf.nn.max_pool2d_5/MaxPool2d:0 shape: (1, 31, 10, 22) dtype: <dtype: 'float32'>
{
"format_version": 1,
"operations": [
{
"op_name": "Reshape_54",
"param_target": "outputs",
"param_name": "173",
"post_process_transpose_perm": [
0,
2,
3,
1
]
},
{
"op_name": "Reshape_73",
"param_target": "outputs",
"param_name": "205",
"post_process_transpose_perm": [
0,
2,
3,
1
]
},
{
"op_name": "Reshape_92",
"param_target": "outputs",
"param_name": "237",
"post_process_transpose_perm": [
0,
2,
3,
1
]
},
{
"op_name": "Relu_306",
"param_target": "outputs",
"param_name": "576",
"post_process_transpose_perm": [
0,
2,
3,
1
]
},
{
"op_name": "Relu_150",
"param_target": "outputs",
"param_name": "320",
"post_process_transpose_perm": [
0,
2,
3,
1
]
}
]
}In onnx model, the output of MaxPool2D is [1, 64, 10, 10], but onnx2tf output (1, 31, 10, 22). I don't know what went wrong. Maybe should I transpose the input of MaxPool2D?
Others
1.5.8
1.13.0
2.10.0
{
"format_version": 1,
"operations": [
{
"op_name": "Resize_0",
"param_target": "inputs",
"param_name": "Concat_0",
"values": [48,48]
},
{
"op_name": "Resize_1",
"param_target": "inputs",
"param_name": "Concat_1",
"values": [48,48]
},
{
"op_name": "Resize_2",
"param_target": "inputs",
"param_name": "Concat_2",
"values": [48,48]
},
{
"op_name": "Resize_3",
"param_target": "inputs",
"param_name": "Concat_3",
"values": [24,24]
},
{
"op_name": "Resize_4",
"param_target": "inputs",
"param_name": "Concat_4",
"values": [48,48]
},
{
"op_name": "Resize_5",
"param_target": "inputs",
"param_name": "Concat_5",
"values": [48,48]
},
{
"op_name": "Resize_6",
"param_target": "inputs",
"param_name": "Concat_6",
"values": [48,48]
},
{
"op_name": "Resize_7",
"param_target": "inputs",
"param_name": "Concat_7",
"values": [24,24]
},
{
"op_name": "Resize_8",
"param_target": "inputs",
"param_name": "Concat_8",
"values": [24,24]
},
{
"op_name": "Resize_9",
"param_target": "inputs",
"param_name": "Concat_9",
"values": [12,12]
},
{
"op_name": "Resize_10",
"param_target": "inputs",
"param_name": "Concat_10",
"values": [48,48]
},
{
"op_name": "Resize_11",
"param_target": "inputs",
"param_name": "Concat_11",
"values": [48,48]
},
{
"op_name": "Resize_12",
"param_target": "inputs",
"param_name": "Concat_12",
"values": [48,48]
},
{
"op_name": "Resize_13",
"param_target": "inputs",
"param_name": "Concat_14",
"values": [192,192]
},
{
"op_name": "Reshape_0",
"param_target": "outputs",
"param_name": "Reshape_0",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Reshape_1",
"param_target": "outputs",
"param_name": "Reshape_1",
"post_process_transpose_perm": [0,2,3,1]
},
{
"op_name": "Transpose_0",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
},
{
"op_name": "Transpose_1",
"param_target": "attributes",
"param_name": "perm",
"values": [0,1,2,3]
},
{
"op_name": "Softmax_0",
"param_target": "attributes",
"param_name": "axis",
"values": 3
}
]
}Validation of 1D tensor always results in Unmatch. However, no inference error is introduced. Therefore, there is no problem with the behavior of the model, but only with the accuracy check decision.
ONNX and TF output value validation started =========================================
INFO: validation_conditions: np.allclose(onnx_outputs, tf_outputs, rtol=0.0, atol=1.0, equal_nan=True)
INFO: onnx_output_name: Shape_20 shape: (4,) dtype: int64 validate_result: Unmatched
INFO: onnx_output_name: conv2d_73.tmp_0 shape: (1, 64, 96, 96) dtype: float32 validate_result: Matches
INFO: onnx_output_name: shape_10.tmp_0 shape: (4,) dtype: int32 validate_result: Unmatched
INFO: onnx_output_name: batch_norm_0.tmp_2 shape: (1, 64, 96, 96) dtype: float32 validate_result: Matches
INFO: onnx_output_name: shape_10.tmp_0_slice_0 shape: (2,) dtype: int32 validate_result: Unmatched
INFO: onnx_output_name: relu_0.tmp_0 shape: (1, 64, 96, 96) dtype: float32 validate_result: Matches
INFO: onnx_output_name: Cast_21 shape: (2,) dtype: int64 validate_result: Unmatched
INFO: onnx_output_name: conv2d_74.tmp_0 shape: (1, 64, 48, 48) dtype: float32 validate_result: Matches
INFO: onnx_output_name: batch_norm_1.tmp_2 shape: (1, 64, 48, 48) dtype: float32 validate_result: Matches
INFO: onnx_output_name: relu_1.tmp_0 shape: (1, 64, 48, 48) dtype: float32 validate_result: Matches
INFO: onnx_output_name: conv2d_75.tmp_0 shape: (1, 32, 48, 48) dtype: float32 validate_result: Matches
INFO: onnx_output_name: conv2d_78.tmp_0 shape: (1, 128, 48, 48) dtype: float32 validate_result: Matches
Others
1.5.0
1.13.0
2.10.0
YOLOX-X
https://drive.google.com/file/d/1QQLf83JwLI_EoQFU5xOGKUvtFPsFSWCN/view
N/ARef issue: [YoloX] output differs between onnx and tflite #107
Problems that can occur in most use cases, such as when post-processing results in zero cases. This is not an onnx2tf problem, but rather an onnxruntime specification problem.
ONNX and TF output value validation started =========================================
INFO: validation_conditions: np.allclose(onnx_outputs, tf_outputs, rtol=1e-05, atol=1e-05, equal_nan=True)
2023-01-10 14:18:07.353720422 [E:onnxruntime:, sequential_executor.cc:369 Execute] Non-zero status code returned while running ReduceMax node. Name:'ReduceMax_686' Status Message: /home/b920405/work/onnxruntime/onnxruntime/core/providers/cpu/reduction/reduction_ops.cc:763 void onnxruntime::ValidateKeepDims(const onnxruntime::TensorShape&, int64_t) keepdims was false. Can't reduce on dim with value of 0 if 'keepdims' is false. Invalid output shape would be produced. input_shape:{0,4}
[E] No function: __iter__ registered for opset: 11
Traceback (most recent call last):
File "/usr/lib/python3.8/runpy.py", line 194, in _run_module_as_main
return _run_code(code, main_globals, None,
File "/usr/lib/python3.8/runpy.py", line 87, in _run_code
exec(code, run_globals)
File "/home/b920405/.vscode/extensions/ms-python.python-2022.20.1/pythonFiles/lib/python/debugpy/adapter/../../debugpy/launcher/../../debugpy/__main__.py", line 39, in <module>
cli.main()
File "/home/b920405/.vscode/extensions/ms-python.python-2022.20.1/pythonFiles/lib/python/debugpy/adapter/../../debugpy/launcher/../../debugpy/../debugpy/server/cli.py", line 430, in main
run()
File "/home/b920405/.vscode/extensions/ms-python.python-2022.20.1/pythonFiles/lib/python/debugpy/adapter/../../debugpy/launcher/../../debugpy/../debugpy/server/cli.py", line 284, in run_file
runpy.run_path(target, run_name="__main__")
File "/home/b920405/.vscode/extensions/ms-python.python-2022.20.1/pythonFiles/lib/python/debugpy/_vendored/pydevd/_pydevd_bundle/pydevd_runpy.py", line 321, in run_path
return _run_module_code(code, init_globals, run_name,
File "/home/b920405/.vscode/extensions/ms-python.python-2022.20.1/pythonFiles/lib/python/debugpy/_vendored/pydevd/_pydevd_bundle/pydevd_runpy.py", line 135, in _run_module_code
_run_code(code, mod_globals, init_globals,
File "/home/b920405/.vscode/extensions/ms-python.python-2022.20.1/pythonFiles/lib/python/debugpy/_vendored/pydevd/_pydevd_bundle/pydevd_runpy.py", line 124, in _run_code
exec(code, run_globals)
File "/home/b920405/git/onnx2tf/onnx2tf/onnx2tf.py", line 1521, in <module>
main()
File "/home/b920405/git/onnx2tf/onnx2tf/onnx2tf.py", line 1475, in main
model = convert(
File "/home/b920405/git/onnx2tf/onnx2tf/onnx2tf.py", line 974, in convert
dummy_onnx_outputs: List[np.ndarray] = dummy_onnx_inference(
File "/home/b920405/git/onnx2tf/onnx2tf/utils/common_functions.py", line 2832, in dummy_onnx_inference
outputs = onnx_session.run(None, dummy_datas)
File "/usr/local/lib/python3.8/dist-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 200, in run
return self._sess.run(output_names, input_feed, run_options)
onnxruntime.capi.onnxruntime_pybind11_state.RuntimeException: [ONNXRuntimeError] : 6 : RUNTIME_EXCEPTION : Non-zero status code returned while running ReduceMax node. Name:'ReduceMax_686' Status Message: /home/b920405/work/onnxruntime/onnxruntime/core/providers/cpu/reduction/reduction_ops.cc:763 void onnxruntime::ValidateKeepDims(const onnxruntime::TensorShape&, int64_t) keepdims was false. Can't reduce on dim with value of 0 if 'keepdims' is false. Invalid output shape would be produced. input_shape:{0,4}
keepdims=False, modify as follows to slice the corresponding dimension after performing ReduceXX with keepdims=True and keeping the dimension. https://github.com/PINTO0309/PINTO_model_zoo/blob/efe3e8da69e1ef14014430b1c88c790976ec10ef/342_ALIKE/PINTO_special/make_mnn_matcher.py#L23-L24np.ndarray be input as dummy data when --check_onnx_tf_outputs_elementwise_close is executed. sample: https://github.com/PINTO0309/onnx2tf/releases/download/1.0.49/calibration_image_sample_data_20x128x128x3_float32.npy
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
Django
The Web framework for perfectionists with deadlines.
Laravel
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
Microsoft
Open source projects and samples from Microsoft.
Google
Google ❤️ Open Source for everyone.
Alibaba
Alibaba Open Source for everyone
D3
Data-Driven Documents codes.
Tencent
China tencent open source team.