TFLite Conversion Issue: Missing "builtin_options_type" and Correct Conversion Method

[songzhibo1]

I am currently working with a PyTorch model and converting it to TFLite. However, when I input the resulting TFLite model into ModelObfuscator, an error arises. The JSON file generated for the converted TFLite model contains many operators that lack the "builtin_options_type" field. This leads to issues during the conversion process, as I cannot locate the "builtin_options_type" for the operators. I have observed the same issue with the SqueezeNet model. However, the provided SqueezeNet TFLite file includes the "builtin_options_type" for each operator, while the TFLite model I converted does not.

1. Could you kindly clarify where the TFLite file in your code is sourced from?
2. What would be the correct approach for converting a TFLite file from PyTorch?

Traceback (most recent call last):
File "/Code275/obfuscation.py", line 109, in
jsontext = lib_generator(model_json, interpreter, inout_list)
File "/Code275/model_parser.py", line 508, in lib_generator
kwargs = get_attributes_params(op, interpreter)
File "/Code275/model_parser.py", line 69, in get_attributes_params
if op['builtin_options_type'] == 'Conv2DOptions' or op['builtin_options_type'] == 'DepthwiseConv2DOptions':
KeyError: 'builtin_options_type'

"operators": [
{
"inputs": [
0,
53
],
"outputs": [
56
]
},
{
"opcode_index": 1,
"inputs": [
56,
52,
18
],
"outputs": [
57
],
"builtin_options_type": "Conv2DOptions",
"builtin_options": {
"padding": "VALID",
"stride_w": 2,
"stride_h": 2,
"fused_activation_function": "RELU"
}
},

I would appreciate your guidance on this matter.

[zhoumingyi]

Hi songzhibo1, thanks for pointing out this.

1. Our TFLite models are collected from TF-Hub, but now TF-Hub has removed the official download site and recommends using Kaggle Models (https://www.kaggle.com/models?tfhub-redirect=true). The "builtin_options_type" is used to let the TFLite library know what the operator is. Some models (converted models do not use the "standard" format) may use "opcode_index" to do a similar step. However, in this case, our tool has a parsing step to translate some unknown "opcode_index" to "builtin_options_type". It seems our parsing steps fail in the converted model from Pytorch.

2. For the converting process, if the TFLite model can be run on your machine, it should be correct. To solve the issue, could you please send me your converted model ("SqueezeNet") if it is possible? If your model can be loaded by the official TFLite library, I think we just need to solve the mapping issue between the "opcode_index" to "builtin_options_type"

[songzhibo1]

Thank you for your response. I have sent my files and questions to your @monash.edu email. I truly appreciate your help, and I would be grateful for any further guidance you can provide.

[pablojoii]

Hi, any suggestions on how to fix the issue? We have similar problems here. Errors we have are:

Warning: didn't find the builtin_options_type for this op

Warning: no activation function found

Warning: no paddings found and using default padding SAME

Warning: no keep_dims of mean_op found

And the error that happens comes as:

Traceback (most recent call last):
File "obfuscation.py", line 104, in
jsontext = lib_generator(model_json, interpreter, inout_list)
File "/Code275/model_parser.py", line 508, in lib_generator
kwargs = get_attributes_params(op, interpreter)
File "/Code275/model_parser.py", line 327, in get_attributes_params
shape_tensor = interpreter.get_tensor(tensor_details["index"]).squeeze()
File "/root/anaconda3/envs/code275/lib/python3.8/site-packages/tensorflow/lite/python/interpreter.py", line 850, in get_tensor
return self._interpreter.GetTensor(tensor_index)
ValueError: Tensor data is null. Run allocate_tensors() first

Odd thing is that the line is 104 and not 109 as above. We have ran using the docker container, is it updated?

Thanks,

Pablo

[pablojoii]

Another info: seems all goes well on the processing, but for some reason the reshaping I have an issue being thrown. So I've edit the model_parser to "ignore" the problem which happens only with the Reshape options.

    for tensor_details in interpreter.get_tensor_details():
        if tensor_details['index'] == op['inputs'][1]:
            interpreter.allocate_tensors()
            try:
                shape_tensor = interpreter.get_tensor(op['inputs'][1]).squeeze()
                shape_item_num = shape_tensor.size
                kwargs['shape='] = 'shape' + '{' + str(shape_tensor.tolist()).strip('[').strip(']') + '}'
                kwargs['shape_size='] = 'shape_size=' + str(shape_item_num)
            except:
                print("Error with variable")
                print(op['inputs'][1])

This bypassed the problem and at least I have a JSON here now, but I've noticed this:

Some shapes are just an empty array. Is this a correct thing?

[pablojoii]

Hi, is there anyway to put the required files somewhere that we can access? I had 404's on dependencies and compilation didn't work probably because of that. Or, are there any plans to update the library?

Thanks.

[zhoumingyi]

Thank you for your response. I have sent my files and questions to your @monash.edu email. I truly appreciate your help, and I would be grateful for any further guidance you can provide.

Hi, I am just back to work and I don't find your email that has your files. You can send it to me again if you still have the problem.

[songzhibo1]

Hello, thank you so much for your kind assistance. Below are the three TFLite files I converted: My actual model is the one with WMNet, which is a new model component I plan to integrate into the original squeezenet1_1 model. This new component, WMNet, represents an improvement over the original model. The conversion of this model was done in two different ways: squeezenet1_1.tflite: This is the original pre-trained model, converted using ai_edge_torch with weights specified as models.squeezenet1_1(weights=models.SqueezeNet1_1_Weights.IMAGENET1K_V1). It serves as a baseline to ensure that any potential issues are not due to modifications in the model itself. squeezenet1_1_wmnet.tflite: This file represents my actual model, which incorporates WMNet as an improvement over the original squeezenet1_1. It was converted using the ai_edge_torch library. squeezenet1_1_wmnet3.tflite: This file was converted through the pipeline: py -> onnx -> tf -> tflite. Unfortunately, neither of these files have been successfully converted so far. Additionally, when attempting to parse the squeezenet1_1_wmnet.tflite file into JSON, multiple operations are appearing. I truly appreciate your help and would be grateful for any further guidance you can provide.

…

------------------ 原始邮件 ------------------ 发件人: "zhoumingyi/ModelObfuscator" ***@***.***>; 发送时间: 2025年2月5日(星期三) 上午9:50 ***@***.***>; ***@***.******@***.***>; 主题: Re: [zhoumingyi/ModelObfuscator] TFLite Conversion Issue: Missing "builtin_options_type" and Correct Conversion Method (Issue #6) Thank you for your response. I have sent my files and questions to your @monash.edu email. I truly appreciate your help, and I would be grateful for any further guidance you can provide. Hi, I am just back to work and I don't find your email that has your files. You can send it to me again if you still have the problem. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***> 从QQ邮箱发来的超大附件 squeezenet1_1.tflite (4.73M, 无限期)进入下载页面：http://mail.qq.com/cgi-bin/ftnExs_download?t=exs_ftn_download&k=2c616533c10cd19a33dd99b04732574a1d1501560703060148575403571f52070151485055515d48515403025706515d0053555761266516141400561b570b0011503a024f4603090c1500335c&code=eae3a2ee squeezenet1_1_with_WMNet.tflite (47.39M, 无限期)进入下载页面：http://mail.qq.com/cgi-bin/ftnExs_download?t=exs_ftn_download&k=3a63366490ad3ec834dfcae71663051a1a17575454540f024f550155534e050c53011b0605550f18525450540653010d5b57555630403746131653014a065950165269556f145e410a3c61297e0643154a511f4a44055b5c16063659&code=bc6d0c75 squeezenet1_1_with_WMNet3.tflite (47.41M, 无限期)进入下载页面：http://mail.qq.com/cgi-bin/ftnExs_download?t=exs_ftn_download&k=60313235461239c4638dceb6146102164d4503050055555818030406044c020d57571f510a535214515457015402515f540153563241304a4444575048045e5c41006d046d16594d5d6e65787c04440a1b4554595b15553908&code=51252a09