如何获取TOCO tf_convert的冻结Tensorflow模型的input_shape

Question

I'm trying to convert a frozen model i got from davidsandberg/facenet to .tflite on Ubuntu 18.04.1 LTS (VirtualBox) using the TF Lite Converter (this is the specific model i am using) . When i try to run the command:

/home/nils/.local/bin/tflite_convert 
--output_file=/home/nils/Documents/frozen.tflite 
--graph_def_file=/home/nils/Documents/20180402-114759/20180402-114759.pb 
--input_arrays=input --output_array=embeddings

i get the following error:

2018-11-29 16:36:21.774098: I 
tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports 
instructions that this TensorFlow binary was not compiled to use: AVX2
Traceback (most recent call last):
File "/home/nils/.local/bin/tflite_convert", line 11, in <module>
 sys.exit(main())
File 
"/home/nils/.local/lib/python3.6/site-packages/tensorflow/contrib   /lite/python/tflite_convert.py", 
line 412, in main
 app.run(main=run_main, argv=sys.argv[:1])
   File 
"/home/nils/.local/lib/python3.6/site-packages/tensorflow/python/platform/app.py", 
line 125, in run
 _sys.exit(main(argv))
File 
"/home/nils/.local/lib/python3.6/site-packages/tensorflow/contrib/lite/python/tflite_convert.py", 
line 408, in run_main
 _convert_model(tflite_flags)
File 
"/home/nils/.local/lib/python3.6/site-packages/tensorflow/contrib/lite/python/tflite_convert.py", 
line 162, in _convert_model
 output_data = converter.convert()
File 
"/home/nils/.local/lib/python3.6/site-packages/tensorflow/contrib/lite/python/lite.py", 
line 404, in convert
 "'{0}'.".format(_tensor_name(tensor)))
ValueError: Provide an input shape for input array 'input'.

Since i havn't trained the model myself, i don't know what exact shape the input had. Probably one can extract it from the classifier.py and the facenet.py found in David Sandberg's GitHubRep., at facenet/src but i don't understand the code enough to do this myself. I even tried to analyse the graph via tensorboard. I couldn't figure it out anyway, but maybe u can: Tensorboard-Screenshot As you might have already noticed, I'm quite new to Ubuntu, Tensorflow and everything related, so I am happy to take any kind of advice on this issue. Thank you in advance!

This is the relevant part of classifier.py, where the model gets loaded and set up:

 # Load the model
        print('Loading feature extraction model')
        facenet.load_model(args.model)

        # Get input and output tensors
        images_placeholder = tf.get_default_graph().get_tensor_by_name("input:0")
        embeddings = tf.get_default_graph().get_tensor_by_name("embeddings:0")
        phase_train_placeholder = tf.get_default_graph().get_tensor_by_name("phase_train:0")
        embedding_size = embeddings.get_shape()[1]

        # Run forward pass to calculate embeddings
        print('Calculating features for images')
        nrof_images = len(paths)
        nrof_batches_per_epoch = int(math.ceil(1.0*nrof_images / args.batch_size))
        emb_array = np.zeros((nrof_images, embedding_size))
        for i in range(nrof_batches_per_epoch):
            start_index = i*args.batch_size
            end_index = min((i+1)*args.batch_size, nrof_images)
            paths_batch = paths[start_index:end_index]
            images = facenet.load_data(paths_batch, False, False, args.image_size)
            feed_dict = { images_placeholder:images, phase_train_placeholder:False }
            emb_array[start_index:end_index,:] = sess.run(embeddings, feed_dict=feed_dict)

        classifier_filename_exp = os.path.expanduser(args.classifier_filename)

Answer 1

Thank you for your help, i did like Alan Chiao said and followed load_data() to the facenet.py, where i eventually found the shape [1,160, 160, 3]. Also, Tensorflow's command line reference for the tf lite converter showed me what i had to look out for:

--input_shapes. Type: colon-separated list of comma-separated lists of integers. Each comma-separated list of integers gives the shape of one of the input arrays specified in TensorFlow convention.

Example: --input_shapes=1,60,80,3 for a typical vision model means a batch size of 1, an input image height of 60, an input image widthof 80, and an input image depth of 3 (representing RGB channels).

Answer 2

I went through the code for tflite converter. I found out that you need to give the input shape as a dictionary in the format {"input_tensor_name": [input shape]} .

Here's an example that resolves the issue:

`graph_def_file = "20180402-114759/20180402-114759.pb"
input_arrays = ["input"]
output_arrays = ["embeddings"]

converter = tf.lite.TFLiteConverter.from_frozen_graph(
  graph_def_file, input_arrays, output_arrays,input_shapes={"input":[1,160,160,3]})

tflite_model = converter.convert()
open("model.tflite", "wb").write(tflite_model)
`

Answer 3

If you start Tensorboard again, go back to the graph that you've seen, there should be a search icon (I think at the upperleft corner) where you can type in "input" and find the input tensor. It'll give you the desired shape. I'm guessing it'll be something in the form of '[1,image_size, image_size, 3]'.

Alternatively, you can examine the code

feed_dict = { images_placeholder:images, phase_train_placeholder:False }

Note that we're feeding in the 'images' object into the images_placeholder, which maps to the "input:0" tensor. You essentially then want the shape of the images object.

images comes from the call to facenet.load_data(). If you go into facenet.py and examine the load_data function, you can observe that the shape is something like what I suggested above. If you print the image_size value, it should match what you see in Tensorboard.