关于TensorFlow Inception PB模型转TFLite及ckpt文件获取的咨询

Can TensorFlow Inception PB models be converted to TFLite? #

Absolutely! The TFLite conversion tooling supports most standard TensorFlow models, including all variants of the Inception family (like InceptionV3, InceptionV4, Inception-ResNet-V2). While the official guide uses MobileNet as an example, the conversion workflow is universal for frozen PB graphs.

To convert your Inception PB model to TFLite, you can use either the command-line tflite_convert tool or the TensorFlow Lite Python API. Here's a basic command-line example (make sure to replace the input/output array names with those specific to your Inception model):

tflite_convert \
  --graph_def_file=your_inception_frozen.pb \
  --output_file=inception_model.tflite \
  --input_arrays=input_tensor_name \
  --output_arrays=output_tensor_name

You can find the correct input/output tensor names by inspecting your PB model with TensorBoard or the saved_model_cli tool.

Where to find Inception checkpoint (ckpt) files? #

You're not missing anything— the slim/nets directory only contains model architecture code, not pre-trained weights. Pre-trained ckpt files for Inception models (and other Slim models) are hosted separately as downloadable packages.

That said, if you already have a frozen PB model (one that includes both the graph structure and trained weights), you don't need a ckpt file to convert to TFLite. The frozen PB alone is sufficient for conversion.

If you do need the ckpt files (for fine-tuning, re-exporting the PB, etc.), you can download them from the official TensorFlow Slim pre-trained model repository. Each Inception variant has its own dedicated download link for the ckpt bundle, which includes files like .ckpt.data-00000-of-00001, .ckpt.index, and .ckpt.meta.

内容的提问来源于stack exchange，提问作者Elye