Cannot export onnx

[YaoJiawei329]

An Extraordinary work! Well， I try to export onnx, but error occurs. If opset=11, 12, 13, error message is: RuntimeError: Exporting the operator repeat_interleave to ONNX opset version 13 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub. else if opset=14, 15, 16, 17, error message is: ValueError: Unsupported ONNX opset version: 14

win11 12700H 3070ti-laptop pytorch1.8.2 onnx1.12

[AllenZYJ]

I get the same error: ValueError: Unsupported ONNX opset version: 17

[YaoJiawei329]

I create a new conda env, use pytorch=1.12, and opset=1, solve the problem.

[HighPoint]

Try PyTorch 2.0. The requirements are likely PyTorch 2.0 and opset version 17.

[julinfn]

I got the same problem,17 to 12 is ok,but got new problem:torch_C.value object is not iterable. a problem about pytorch version?

torch:1.10 py:3.9.2

[wavelet2008]

Exporing onnx model to out/dd.onnx... Traceback (most recent call last): File "/home/ubuntu/seg/segment-anything/scripts/export_onnx_model.py", line 180, in run_export( File "/home/ubuntu/seg/segment-anything/scripts/export_onnx_model.py", line 154, in run_export torch.onnx.export( File "/home/ubuntu/anaconda3/envs/face19/lib/python3.9/site-packages/torch/onnx/init.py", line 275, in export return utils.export(model, args, f, export_params, verbose, training, File "/home/ubuntu/anaconda3/envs/face19/lib/python3.9/site-packages/torch/onnx/utils.py", line 88, in export _export(model, args, f, export_params, verbose, training, input_names, output_names, File "/home/ubuntu/anaconda3/envs/face19/lib/python3.9/site-packages/torch/onnx/utils.py", line 689, in _export _model_to_graph(model, args, verbose, input_names, File "/home/ubuntu/anaconda3/envs/face19/lib/python3.9/site-packages/torch/onnx/utils.py", line 463, in _model_to_graph graph = _optimize_graph(graph, operator_export_type, File "/home/ubuntu/anaconda3/envs/face19/lib/python3.9/site-packages/torch/onnx/utils.py", line 200, in _optimize_graph graph = torch._C._jit_pass_onnx(graph, operator_export_type) File "/home/ubuntu/anaconda3/envs/face19/lib/python3.9/site-packages/torch/onnx/init.py", line 313, in _run_symbolic_function return utils._run_symbolic_function(*args, *kwargs) File "/home/ubuntu/anaconda3/envs/face19/lib/python3.9/site-packages/torch/onnx/utils.py", line 994, in _run_symbolic_function return symbolic_fn(g, inputs, **attrs) File "/home/ubuntu/anaconda3/envs/face19/lib/python3.9/site-packages/torch/onnx/symbolic_opset11.py", line 922, in repeat_interleave return torch.onnx.symbolic_opset9.repeat_interleave(g, self, repeats, final_dim) File "/home/ubuntu/anaconda3/envs/face19/lib/python3.9/site-packages/torch/onnx/symbolic_opset9.py", line 2064, in repeat_interleave for idx, r_split in enumerate(r_splits): TypeError: 'torch._C.Value' object is not iterable (Occurred when translating repeat_interleave).

[MolianWH]

It seems that it only supports pytorch version 2.0(cuda11.7). I updated my cuda and pytorch（failed with others versions）and it works. By the way, I modified onnxruntime.inferenceSession parameters at line168

ort_session = onnxruntime.InferenceSession(output,providers=['CUDAExecutionProvider'])

[AllenZYJ]

It is not throwing any errors now,when I updated pytorch to 2.0 and onnx 1.13.1.

[UNeedCryDear]

Hey，guys！ In this PR: https://github.com/facebookresearch/segment-anything/pull/210 After changing torch.repeat_interleave() to torch.expand(),, I successfully exported it under torch1.8.2+opset=12, But I'm not sure how this will affect performance.

[lauraset]

@UNeedCryDear . It deed works.

[InterstellarFang]

Hey，guys！ In this PR: #210 After changing torch.repeat_interleave() to torch.expand(),, I successfully exported it under torch1.8.2+opset=12, But I'm not sure how this will affect performance. @UNeedCryDear Hello！Which python file is this function (torch.repeat_interleave()) in,Can you tell me the location of this file？thanks a lot!

[UNeedCryDear]

@UNeedCryDear Hello！Which python file is this function (torch.repeat_interleave()) in,Can you tell me the location of this file？thanks a lot!

https://github.com/facebookresearch/segment-anything/pull/210/files

[InterstellarFang]

@UNeedCryDear Hello！Which python file is this function (torch.repeat_interleave()) in,Can you tell me the location of this file？thanks a lot!

https://github.com/facebookresearch/segment-anything/pull/210/files

@UNeedCryDearThank you! I made changes based on the code you provided(https://github.com/facebookresearch/segment-anything/pull/210/files), add four lines of code, but still reported an error(RuntimeError: Exporting the operator repeat_interleave to ONNX opset version 12 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub.) under torch1.8.1+opset=12

[UNeedCryDear]

The code with a pink background has been replaced and you need to remove it.

[InterstellarFang]

The code with a pink background has been replaced and you need to remove it.

@UNeedCryDear I have added # to these two sentences with a pink background,but still reported an error(RuntimeError: Exporting the operator repeat_interleave to ONNX opset version 12 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub.)under torch1.8.1+opset=12 Is this a problem with the torch version? I would like to confirm if changing the onnx opt default value is in these two files(notebooks/onnx_model_example.ipynb,scripts/export_onnx_model.py)

[UNeedCryDear]

show me the code you modefied.

[InterstellarFang]

show me code you modefied.

    # Expand per-image data in batch direction to be per-mask
    # src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
    src_shape = (tokens.shape[0],*image_embeddings.shape[1:])
    src = image_embeddings.expand(src_shape)
    src = src + dense_prompt_embeddings
    # pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)
    pos_src_shape = (tokens.shape[0],*image_pe.shape[1:])
    pos_src = image_pe.expand(pos_src_shape)
    b, c, h, w = src.shape

@UNeedCryDear

[UNeedCryDear]

Search for repeat_interleave in the project, only here was calling. Your code is correct unless it is different from the function you are calling. So, have you saved your modifications?

[InterstellarFang]

Search for repeat_interleave in the project, only here was calling. Your code is correct unless it is different from the function you are calling. So, have you saved your modifications?

@UNeedCryDear yes,I saved my modifications.Should I try other values besides 12(onnx opset)?

[InterstellarFang]

Search for repeat_interleave in the project, only here was calling. Your code is correct unless it is different from the function you are calling. So, have you saved your modifications?

@UNeedCryDear hello!I don't know how to fix the error,can you give me some advice?thanks a lot!

[UNeedCryDear]

hello!I don't know how to fix the error,can you give me some advice?thanks a lot!

According to the error, it is a problem that was not successfully modified.You can search it like me throughout the project and find out where the modifications were not made correctly At the same time, if you are using Jupyter Notebook and colab, you may encounter issues with modified files being inconsistent with the actual running files. The correct approach is to git clone the code and make local modifications instead of pips.

Finally, I will provide you .py that I can export which based on the modification of PR210, If you still cannot export, I suggest you try changing to PyTorch2.0. Good luck!

python export.py --checkpoint path/to/checkpoint --type vit_b --opset 12

import torch
import warnings
from segment_anything import sam_model_registry, SamPredictor
from segment_anything.utils.onnx import SamOnnxModel
import argparse
import onnx

def export_onnx( sam_checkpoint="sam_vit_b_01ec64.pth", model_type = "vit_b", opset=12, onnx_model_path="sam_onnx_example_maskdeocde.onnx"): sam = sam_model_registrymodel_type onnx_model = SamOnnxModel(sam, return_single_mask=True) dynamic_axes = { "point_coords": {1: "num_points"}, "point_labels": {1: "num_points"}, }

embed_dim = sam.prompt_encoder.embed_dim
embed_size = sam.prompt_encoder.image_embedding_size
mask_input_size = [4 * x for x in embed_size]
img_size=sam.image_encoder.img_size

img=torch.randn(1, 3, img_size,img_size, dtype=torch.float)
dynamic_shape = {'images': {0: 'batch', 2: 'height', 3: 'width'}}
dummy_inputs = {
    "image_embeddings": torch.randn(1, embed_dim, *embed_size, dtype=torch.float),
    "point_coords": torch.randint(low=0, high=1024, size=(1, 5, 2), dtype=torch.float),
    "point_labels": torch.randint(low=0, high=4, size=(1, 5), dtype=torch.float),
    "mask_input": torch.randn(1, 1, *mask_input_size, dtype=torch.float),
    "has_mask_input": torch.tensor([1], dtype=torch.float),
    "orig_im_size": torch.tensor([1500, 2250], dtype=torch.float),
}
output_names = ["masks", "iou_predictions", "low_res_masks"]

with warnings.catch_warnings():
    warnings.filterwarnings("ignore", category=torch.jit.TracerWarning)
    warnings.filterwarnings("ignore", category=UserWarning)
    with open(onnx_model_path, "wb") as f:
        torch.onnx.export(
            onnx_model,
            tuple(dummy_inputs.values()),
            f,
            export_params=True,
            verbose=False,
            opset_version=opset,
            do_constant_folding=True,
            input_names=list(dummy_inputs.keys()),
            output_names=output_names,
            dynamic_axes=dynamic_axes,
        )
model_onnx = onnx.load(onnx_model_path)  # load onnx model
onnx.checker.check_model(model_onnx)  # check onnx model
print("Done!")

def parse_opt(): parser = argparse.ArgumentParser() parser.add_argument("--checkpoint", type=str, default ="./model_weights/sam_vit_b_01ec64.pth", help="The path to the SAM model checkpoint.") parser.add_argument("--type", type=str, default="vit_b", help="In ['default', 'vit_h', 'vit_l', 'vit_b']. Which type of SAM model to export.") parser.add_argument("--opset",type=int,default=12,help="The ONNX opset version to use") parser.add_argument("--output", type=str,default ="sam_onnx_example_maskdeocde.onnx", help="The ONNX opset version to use") opt = parser.parse_args() return opt if name == 'main': opt = parse_opt() export_onnx(opt.checkpoint, opt.type,opt.opset,opt.output)

[captainIT]

Hey，guys！ In this PR: #210 After changing torch.repeat_interleave() to torch.expand(),, I successfully exported it under torch1.8.2+opset=12, But I'm not sure how this will affect performance.

It is true that onnx can be exported successfully, but the web demo cannot be used normally.

[UNeedCryDear]

It is true that onnx can be exported successfully, but the web demo cannot be used normally.

I'm sorry I can't help you,I am not familiar with the web at all. If you need to use the web side, it is best to use the original code.