怎么pb转tflite

[panguxiaoshen]

调用函数：tf.contrib.lite.TFLiteConverter.from_frozen_graph 报错： 2019-07-10 03:18:33.807555: F tensorflow/lite/toco/graph_transformations/resolve_strided_slice_attributes.cc:95] Check failed: start_indices_size <= num_input_axes (2 vs. 1)StridedSlice op requires no more than 1 start indices

作者尝试过转换吗？

[dvlee1024]

pb_file = "../my_yolov3_coco.pb"
input_arrays = ["input/input_data"]
output_arrays = ["pred_sbbox/concat_2", "pred_mbbox/concat_2", "pred_lbbox/concat_2"]

converter = tf.lite.TFLiteConverter.from_frozen_graph(pb_file, input_arrays, output_arrays, {"input/input_data" :[1, 416, 416, 3]})
tflite_model = converter.convert()
open("myModel.tflite", "wb").write(tflite_model)

遇到同样的问题

[lukaisheng1203]

@dvlee1024 您好我也想转一下部署在安卓上，之前没有做过，请问你转好了吗

[panguxiaoshen]

找到问题所在了，在模型里面作者扩展tensor维数通过tf.newaxis实现，转tflite会一直报上述的错误。通过tf.expand_dims来实现tensor扩维就可以转成功了。（这部分在def decode（）函数里面） @dvlee1024 @airu924123

[panguxiaoshen]

补充：虽然可以生成tflite,但在调用是会报._interpreter.allocate_tensors()的相关错误：StridedSlice op only supports 1D-4D input arrays.Node number 43 (STRIDED_SLICE) failed to prepare. 所以相关运算中需要把5维改成4维，应该就可以了，尝试中。。。

[yuanzhedong]

@panguxiaoshen 大佬您好，请问有什么update吗？感谢！

[panguxiaoshen]

@panguxiaoshen 大佬您好，请问有什么update吗？感谢！

按照我上面说的改就行，我已经成功转换。tiny_tolov3可以在板子上跑50ms

[panguxiaoshen]

@panguxiaoshen 大佬您好，请问有什么update吗？感谢！

我已经上传到 https://github.com/panguxiaoshen/tiny_yolov3

[yuanzhedong]

@panguxiaoshen 万分感谢！我回来试一下

[pauqilaz]

Hi, I encounter the same error message than in the original issue. Have you found a solution or a workaround ?

[panguxiaoshen]

Hi, I encounter the same error message than in the original issue. Have you found a solution or a workaround ?

There are two places to change in the code. When you are expanding the dimensions of tensor，you should use "tf.expand_dims" rather than "tf.newaxis". Another thing is , so far ,tflite op only supports 1D-4D input arrays, but in Yunyang1994's code, there exists 5D operations. If you want to solve the problem, you can refer to the link mentioned above.

[peace195]

Solution is here: https://github.com/peace195/tensorflow-lite-yolo-v3 The .pb file must be in right format of SavedModel. Please try it. I would appreciate if you give me a star for this project 👍

[kerolos]

dear @peace195 :+1: how can I get the right format (.pb) from training the Yolo v3 model here, not from the original https://pjreddie.com/darknet/yolo/. thanks in advance.

[peace195]

@kerolos Please use this code: https://github.com/peace195/tensorflow-lite-yolo-v3/blob/master/convert_weights_pb.py

[774219268]

转成功了我在为这个问题烦恼，可以加下微信吗？15348365038

[tongda]

@panguxiaoshen 大佬您好，请问有什么update吗？感谢！

我已经上传到 https://github.com/panguxiaoshen/tiny_yolov3

参考了你的代码，绕过5D -> 4D的方法很棒。

[netist123]

@panguxiaoshen 大佬您好，请问有什么update吗？感谢！

我已经上传到 https://github.com/panguxiaoshen/tiny_yolov3

参考了你的代码，绕过5D -> 4D的方法很棒。

请问，哪个文件里有5D操作？ train.py吗？

[774219268]

不是什么大佬，互相学习，可以加我微信15348365038互相探讨，都是慢慢测试出来的

------------------ 原始邮件 ------------------ 发件人: "netist123"<[email protected]>; 发送时间: 2020年3月6日(星期五) 晚上9:46 收件人: "YunYang1994/tensorflow-yolov3"<[email protected]>; 抄送: "烧火棍"<[email protected]>; "Comment"<[email protected]>; 主题: Re: [YunYang1994/tensorflow-yolov3] 怎么pb转tflite (#232)

@panguxiaoshen 大佬您好，请问有什么update吗？感谢！

我已经上传到 https://github.com/panguxiaoshen/tiny_yolov3

参考了你的代码，绕过5D -> 4D的方法很棒。

请问，哪个文件里有5D操作？ train.py吗？

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

[tongda]

@panguxiaoshen 大佬您好，请问有什么update吗？感谢！

我已经上传到 https://github.com/panguxiaoshen/tiny_yolov3

参考了你的代码，绕过5D -> 4D的方法很棒。

请问，哪个文件里有5D操作？ train.py吗？

yolov3.py里面，从下面这行开始：

conv_output = tf.reshape(conv_output, (batch_size, output_size, output_size, 2, 5 + NUM_CLASS))

就都是5D操作了。