0

我正在将运行此笔记本时获得的冻结推理图转换为 TFLite 模型。

我修改了从 TensorFlow文档中找到的代码，并在 Google Colab 中运行它：

import tensorflow as tf

path = "/content/drive/MyDrive/real_frozen_inference_graph.pb"

input = ["image_tensor"]
output = ["detection_boxes", "detection_scores", "detection_multiclass_scores", "detection_classes", "num_detections", "raw_detection_boxes", "raw_detection_scores"]

# Convert the model
converter = tf.compat.v1.lite.TFLiteConverter.from_frozen_graph(path, input_arrays=input, output_arrays=output)
print("starting conversion")

tflite_model = converter.convert()

print("done converting")

# Save the model.
with open('/content/drive/MyDrive/model.tflite', 'wb') as f:
  f.write(tflite_model)

运行单元后，运行时几乎立即断开连接，这特别奇怪，因为我为 colab pro 付费......

我更改了上面的代码以在我的本地开发环境上运行，但仍然没有运气。我得到一长串原始二进制代码和类似这样的奇怪的东西打印到我的标准输出，但脚本终止时没有写入文件：

  %220 = "tfl.add"(%219, %cst_12) {fused_activation_function = "NONE"} : (tensor<?x100xf32>, tensor<f32>) -> tensor<?x100xf32>
  %221 = "tf.TensorArraySizeV3"(%handle_170, %210#5) {_class = ["loc:@Postprocessor/BatchMultiClassNonMaxSuppression/map/TensorArray_12"], device = ""} : (tensor<2x!tf.resource<tensor<*xf32>>>, tensor<f32>) -> tensor<i32>
  %222 = "tfl.range"(%cst_10, %221, %cst_13) : (tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<?xi32>
  %223 = "tf.TensorArrayGatherV3"(%handle_170, %222, %210#5) {_class = ["loc:@Postprocessor/BatchMultiClassNonMaxSuppression/map/TensorArray_12"], device = "", element_shape = #tf.shape<100x2>} : (tensor<2x!tf.resource<tensor<*xf32>>>, tensor<?xi32>, tensor<f32>) -> tensor<?x100x2xf32>
  %224 = "tf.TensorArraySizeV3"(%handle_172, %210#6) {_class = ["loc:@Postprocessor/BatchMultiClassNonMaxSuppression/map/TensorArray_13"], device = ""} : (tensor<2x!tf.resource<tensor<*xi32>>>, tensor<f32>) -> tensor<i32>
  %225 = "tfl.range"(%cst_10, %224, %cst_13) : (tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<?xi32>
  %226 = "tf.TensorArrayGatherV3"(%handle_172, %225, %210#6) {_class = ["loc:@Postprocessor/BatchMultiClassNonMaxSuppression/map/TensorArray_13"], device = "", element_shape = #tf.shape<>} : (tensor<2x!tf.resource<tensor<*xi32>>>, tensor<?xi32>, tensor<f32>) -> tensor<?xi32>
  %227 = "tfl.cast"(%226) : (tensor<?xi32>) -> tensor<?xf32>
  "std.return"(%213, %216, %223, %220, %227, %181, %186) : (tensor<?x100x4xf32>, tensor<?x100xf32>, tensor<?x100x2xf32>, tensor<?x100xf32>, tensor<?xf32>, tensor<?x?x4xf32>, tensor<?x?x2xf32>) -> ()
}) {sym_name = "main", tf.entry_function = {control_outputs = "", inputs = "image_tensor", outputs = "detection_boxes,detection_scores,detection_multiclass_scores,detection_classes,num_detections,raw_detection_boxes,raw_detection_scores"}, type = (tensor<?x?x?x3x!tf.quint8>) -> (tensor<?x100x4xf32>, tensor<?x100xf32>, tensor<?x100x2xf32>, tensor<?x100xf32>, tensor<?xf32>, tensor<?x?x4xf32>, tensor<?x?x2xf32>)} : () -> ()

有人对这里出了什么问题有任何想法吗？是否将 TFLite 文件打印到我的标准输出？也许我忽略了一些明显的东西？我是 TensorFlow 的新手，因此不胜感激。

4

1 回答 1

1

这是指示转换失败位置的转换器的输出。这是本机应用程序，python 包装器只是翻译这个输出。这可能是由各种原因引起的：未实现的操作、未实现的操作输入/输出等。我将如何开始：

  1. 尽可能选择最新的 TF 版本（稳定，尝试所有步骤，如果没有帮助 - 每晚）

  2. 添加选定的 TF 操作- 这将扩大您可以使用的操作数量。申请并尝试转换：

    converter.target_spec.supported_ops = [ tf.lite.OpsSet.TFLITE_BUILTINS, # enable TensorFlow Lite ops. tf.lite.OpsSet.SELECT_TF_OPS # enable TensorFlow ops. ]

  3. 允许自定义操作并为他们调查 tflite 模型：converter.allow_custom_ops = True. 虽然您仍然无法推断模型，但您会发现应该更改哪些操作。

于 2020-12-11T21:11:00.850 回答