TensorFlow README

July 29, 2020 · View on GitHub

TensorFlow parser (Tensorflow -> IR part) is an experimental product, since the granularity of TensorFlow checkpoint graph is much finer than other platform. We have to use graph matching-like method to retrieve operators.

We tested the slim pre-trained models. Any contribution is welcome.

Models	Caffe	CoreML	CNTK	Keras	MXNet	PyTorch	ONNX
Vgg19	√	√	√	√	√	√	√
Inception_v1	√	√	o	√	√	√	√
Inception_v3	√	√	√	o	√	√	√
ResNet V1	√	√	o	√	√	√	√
ResNet V2	√	√	√	√	√	√	√
MobileNet V1	x	√	o	√	√	√	√
MobileNet V2	x	√	o	√	√	√	√
NasNet-A	x	x	x	x	√	√	√

√ - Correctness tested

o - Some difference after conversion

x - Can not be converted

Usage

Checkpoint File Conversion

We will give an example to convert TensorFlow resnet slim model with checkpoint files to caffe.

# Download TensorFlow pre-trained model first
$ mmdownload -f tensorflow -n resnet_v2_152

Downloading file [./resnet_v2_152_2017_04_14.tar.gz] from [http://download.tensorflow.org/models/resnet_v2_152_2017_04_14.tar.gz]
100% [......................................................................] 675629399 / 675629399
Model saved in file: ./imagenet_resnet_v2_152.ckpt

# Convert the TensorFlow model to Caffe
$ mmconvert -sf tensorflow -in imagenet_resnet_v2_152.ckpt.meta -iw imagenet_resnet_v2_152.ckpt --dstNodeName MMdnn_Output -df caffe -om tf_resnet
.
.
Caffe model files are saved as [tf_resnet.prototxt] and [tf_resnet.caffemodel], generated by [203e03ef187a42f59942737dace8773d.py] and [203e03ef187a42f59942737dace8773d.npy].

-in is used to specify the ".ckpt.meta" file.

-iw is used to specify the ".ckpt" file.

--dstNodeName is used to specify the output node of your model, which can be found in your code or tensorboard graph. We provide a tool vis_meta to help visualize your meta graph.

Frozen Graph Conversion

We will give an example to convert TensorFlow mobilenet slim model with frozen_pb to caffe

# Download TensorFlow pre-trained model first
$ mmdownload -f tensorflow -n mobilenet_v1_1.0_frozen

Downloading file [mobilenet_v1_1.0_224_frozen.tgz] from [https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_1.0_224_frozen.tgz]
progress: 18712.0 KB downloaded, 100%k

# Convert the TensorFlow model to Caffe
$ mmconvert -sf tensorflow -iw mobilenet_v1_1.0_224/frozen_graph.pb --inNodeName input --inputShape 224,224,3 --dstNodeName MobilenetV1/Predictions/Softmax -df caffe -om tf_mobilenet
.
.
Caffe model files are saved as [tf_mobilenet.prototxt] and [tf_mobilenet.caffemodel], generated by [e96550a4c55141afa8cd94372b858613.py] and [e96550a4c55141afa8cd94372b858613.npy].

For frozen graph parser, --inNodeName and --inputShape are required, and don't need to set -in.

How to prepare your model and conversion parameters

You can refer Slim Model Extractor to extract your own tensorflow model, which is a sample tool to extract both architecture and weights from slim pre-trained models.

$ mmdownload -f tensorflow

Supported models: ['inception_v3_frozen', 'resnet_v2_200', 'inception_v1', 'mobilenet_v1_1.0', 'mobilenet_v2_1.0_224', 'resnet_v2_152', 'vgg16', 'mobilenet_v1_1.0_frozen', 'resnet_v1_50', 'resnet_v2_50', 'inception_v3', 'inception_resnet_v2', 'resnet_v1_152', 'inception_v1_frozen', 'vgg19', 'nasnet-a_large']

[Note!] The extractor create a Squeeze operator named MMdnn_Output as the output node of the models.

Mainly extract code like:

with slim.arg_scope(...):
    data_input = tf.placeholder(name='input', dtype=tf.float32, shape=[...])
    logits = your_own_network_builder(data_input)
    if logits.op.type == 'Squeeze':
        labels = tf.identity(logits, name='MMdnn_Output')
    else:
        labels = tf.squeeze(logits, name='MMdnn_Output')

Meta File Graph Visualization

When you prepared your checkpoint, you can find the output node name from the graph by Tensorboard.

$ mmvismeta imagenet_resnet_v2_152.ckpt.meta ./logs/

TensorBoard 0.4.0rc3 at http://kit-station:6006 (Press CTRL+C to quit)

tensorboard

Step-by-step conversion for debugging

We will give an example to convert TensorFlow mobilenet slim model with frozen_pb to caffe.

# Convert frozen graph to IR
$ mmtoir -f tensorflow -w mobilenet_v1_1.0_224/frozen_graph.pb --inNodeName input --inputShape 224,224,3 --dstNodeName MobilenetV1/Predictions/Softmax -o mobilenet_v1

IR network structure is saved as [mobilenet_v1.json].
IR network structure is saved as [mobilenet_v1.pb].
IR weights are saved as [mobilenet_v1.npy].

# Convert IR to Caffe network building code
$ mmtocode -f caffe -n mobilenet_v1.pb -w mobilenet_v1.npy -o tf_mobilenet.py -ow tf_mobilenet.npy

Parse file [mobilenet_v1.pb] with binary format successfully.
Target network code snippet is saved as [tf_mobilenet.py].
Target weights are saved as [tf_mobilenet.npy].

# Use Caffe network building code to generate an original Caffe model
$ mmtomodel -f caffe -in tf_mobilenet.py -iw tf_mobilenet.npy -o tf_mobilenet
.
.
.
Caffe model files are saved as [tf_mobilenet.prototxt] and [tf_mobilenet.caffemodel], generated by [tf_mobilenet.py] and [tf_mobilenet.npy].

The you can use tf_mobilenet.prototxt and tf_mobilenet.caffemodel in Caffe directly.

Convert models from IR to Tensorflow code snippet

We will give you an example to convert models from other frameworks to tensorflow.

$ mmtocode -f tensorflow --IRModelPath resnet152.pb --IRWeightPath resnet152.npy --dstModelPath tf_resnet152.py

Parse file [resnet152.pb] with binary format successfully.
Target network code snippet is saved as [tf_resnet152.py].

Convert models from IR to Tensorflow model

After generating the Tensorflow code snippet, you can convert the Tensorflow code snippet and IR weights file to Tensorflow original model for further usage.There are two types of dump tags: SERVING and TRAINING.

$ mmtomodel -f tensorflow -in tf_resnet152.py -iw resnet152.npy -o tf_resnet152 --dump_tag SERVING
.
.
.
Tensorflow file is saved as [tf_resnet152/saved_model.pb], generated by [tf_resnet152.py] and [resnet152.npy].

FAQ

Reuse the converted Tensorflow model

If you want to retrain the converted model, you can change the variable is_train from "False" to "True" in the converted code file and then use '--dump_tag TRAINING' to dump it.

export_dir = "./tf_resnet152"
with tf.Session(graph=tf.Graph()) as sess:
    tf.saved_model.loader.load(sess, [tf.saved_model.tag_constants.TRAINING], export_dir)

    x = sess.graph.get_tensor_by_name('input:0')
    y = sess.graph.get_tensor_by_name('xxxxxx:0')
    ......
    _y = sess.run(y, feed_dict={x: _x})

Develop version

Ubuntu 16.04 with

Tensorflow 1.8

@ 2018/07/05

Limitation

Currently no RNN related operations support.
Frozen graph parser is not well tested.