CasiaFan/mxnet2tensorflow

convert mxnet model (>=1.3) to tensorflow V1 model (1.5)

Convert mxnet model to tensorflow V1 model

Convert mxnet model (Version 1.5) to tensorflow V1 model (Version 1.15) or tflite model for deployment

MMdnn pipeline (mxnet2tf)

1. Install MMdnn pip install mmdnn

2. MMdnn one-step conversion
   Use insightface mxnet r50 model as an example (download from model zoo)
   The supported framework could be found here: https://github.com/microsoft/MMdnn#support-frameworks

mmconvert -sf mxnet -in model-symbol.json -iw model-0000.params -df tensorflow -om model.pb --inputShape 3,112,112

Then the converted model is saved in this format for serving: saved_model.pb + variables

3. MMdnn step-by-step conversion
   1) convert model from MXnet to IR

mmtoir -f mxnet -n model-symbol.json -w model-0000.params -d ir_model --inputShape 3,112,112

It generates network structure ir_model.json, ir_model.pb and weights ir_model.npy

2) convert model from IR to tensorflow code snippet

mmtocode -f tensorflow --IRModelPath ir_model.pb --IRWeightPath ir_model.npy --dstModelPath ir_model.py

Parse network to save it in ir_model.py

3) convert model form IR to tensorflow

mmtomodel -f tensorflow -in ir_model.py -iw ir_model.npy -o ir_model --dump_tag SERVING

Tensorflow model is saved in ir_model directory with saved_model.pb and variables. Choose dump_tag as SERVING or TRAINING

Convert tensorflow model to tflite (tf2tflite)

Run pb2lite.py to convert pb to lite

python pb2lite_fire.py --model=ir_model    # pb saved_model path
                       --input_size=112    # input node size
                       --input_node="data" # input node name 
                       --output_nodes="fc1/add_1" # output node name 
                       --save_name=ir_model      # saved model name 
                       --tune_image_dir=face_tune_image # tune image directory path
                       --preprocess_norm=False    # use norm during preprocessing or not
                       --color_fmt=RGB     # input image color format
                       --quant=False       # quantize weight or not  

Check inference result

Run test_inference.py to check model output so that their results should be consistent.

python infer_fire.py --model ir_mdoel.tflite   # model path
                     --model_arch=tflite       # modle archtecture for inference 
                     --input_size=112          # input model path 
                     --input_node=data         # input node name 
                     --output_nodes=fc1/add_1  # output node name 
                     --test_image_path=face_tune_image/00c11cdc71eaafb14fef53f032bf3ae7.jpg # test image path

Add custom layer

For example, if we want to add softmax layer after the last layer, we could modify the ir_model.py generated in MMdnn step-by-step conversion and re-freeze it. We add the following lines:

a). Add with tf.Session() as sess: @Line 24 of r50.py to initialize the session
b). Add the following lines @Line 261 of r50.py to save the extra node into graph

...
# add softmax
fc1 = tf.nn.softmax(fc1, axis=1, name="softmax")
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
saver.save(sess, 'model_add_softmax')  # ckpt file save path
tf.io.write_graph(sess.graph_def, "/tmp/my_model", "model_add_softmax.pbtxt", as_text=True)  # new graph model name

Now the new graph is saved in model_add_softmax.pbtxt. Then freeze it with freez_graph tool:

import tensorflow as tf
from tensorflow.python.tools import freeze_graph

output_nodes = ["softmax"]
output_nodes = ",".join(output_nodes)
graph_filepath = "model_add_softmax.pbtxt"
ckpt_filepath = "model_add_softmax"
pb_filepath = "model_add_softmax.pb"

freeze_graph.freeze_graph(input_graph=graph_filepath, 
                          input_saver='', 
                          input_binary=False, 
                          input_checkpoint=ckpt_filepath, 
                          output_node_names=output_nodes, 
                          restore_op_name='save/restore_all', 
                          filename_tensor_name='save/Const:0', 
                          output_graph=pb_filepath, 
                          clear_devices=True, 
                          initializer_nodes='')

Now the softmax layer appended model is frozen into model_add_softmax.pb.

BUT NOTE that if new layers with trainable parameters are added like conv2D, its weights is initialzied based on the initializer. It doesn't make sense. So what we usually add after training is layers withou parameters like softmax, reshape, etc.