Converts the mask-rcnn keras model https://github.com/matterport/Mask_RCNN/releases/tag/v2.0 to tflite

convert.py

import tensorflow as tf
import numpy as np
import mrcnn.model as modellib # https://github.com/matterport/Mask_RCNN/
from mrcnn.config import Config
import keras.backend as keras

PATH_TO_SAVE_FROZEN_PB ="./"
FROZEN_NAME ="saved_model.pb"


def load_model(Weights):
        global model, graph
        class InferenceConfig(Config):
                NAME = "coco"
                NUM_CLASSES = 1 + 80
                IMAGE_META_SIZE = 1 + 3 + 3 + 4 + 1 + NUM_CLASSES
                DETECTION_MAX_INSTANCES = 100
                DETECTION_MIN_CONFIDENCE = 0.7
                DETECTION_NMS_THRESHOLD = 0.3
                GPU_COUNT = 1
                IMAGES_PER_GPU = 1

        config = InferenceConfig()
        Weights = Weights
        Logs = "./logs"
        model = modellib.MaskRCNN(mode="inference", config=config,
		                  model_dir=Logs)
        model.load_weights(Weights, by_name=True)
        graph = tf.get_default_graph()


# Reference https://github.com/bendangnuksung/mrcnn_serving_ready/blob/master/main.py
def freeze_session(session, keep_var_names=None, output_names=None, clear_devices=True):
    graph = session.graph

    with graph.as_default():
        freeze_var_names = list(set(v.op.name for v in tf.global_variables()).difference(keep_var_names or []))

        output_names = output_names or []
        input_graph_def = graph.as_graph_def()

        if clear_devices:
            for node in input_graph_def.node:
                node.device = ""

        frozen_graph = tf.graph_util.convert_variables_to_constants(
            session, input_graph_def, output_names, freeze_var_names)
        return frozen_graph


def freeze_model(model, name):
    frozen_graph = freeze_session(
        sess,
        output_names=[out.op.name for out in model.outputs][:4])
    directory = PATH_TO_SAVE_FROZEN_PB
    tf.train.write_graph(frozen_graph, directory, name , as_text=False)

def keras_to_tflite(in_weight_file, out_weight_file):
    sess = tf.Session()
    keras.set_session(sess)
    load_model(in_weight_file)
    global model
    freeze_model(model.keras_model, FROZEN_NAME)
    # https://github.com/matterport/Mask_RCNN/issues/2020#issuecomment-596449757
    input_arrays = ["input_image"]
    output_arrays = ["mrcnn_class/Softmax","mrcnn_bbox/Reshape"]
    converter = tf.contrib.lite.TocoConverter.from_frozen_graph(
        PATH_TO_SAVE_FROZEN_PB+"/"+FROZEN_NAME,
        input_arrays, output_arrays,
        input_shapes={"input_image":[1,256,256,3]}
        )
    converter.target_ops = [tf.lite.OpsSet.TFLITE_BUILTINS,tf.lite.OpsSet.SELECT_TF_OPS]
    converter.post_training_quantize = True
    tflite_model = converter.convert()
    open(out_weight_file, "wb").write(tflite_model)
    print("*"*80)
    print("Finished converting keras model to Frozen tflite")
    print('PATH: ', out_weight_file)
    print("*" * 80)

keras_to_tflite("./mask_rcnn_coco.h5","./mask_rcnn_coco.tflite")

Hi @bmabir17
Thanks your code that I can convert the tflite successfully. But when I ran the tflite on iOS, I got bellow error
TensorFlow Lite Error: Input tensor 314 lacks data

I look through the code and notice this code in keras_to_tflite function:
input_arrays = ["input_image"] output_arrays = ["mrcnn_class/Softmax","mrcnn_bbox/Reshape"]

But the real keras model needs more inputs when predict function was executed. I don't know if that has anything to do with it. What do think?

hi @bmabir17
I cant seem to move ahead from this point
ValueError: Invalid tensors 'input_image' were found.
i checked the model using model.keras_model.summary() and it shows that the first layer has the name input_image. What do i do?

@kaamlaS i got past this , check this out matterport/Mask_RCNN#2020

let me know if you get it working
also i cant get the output names to work so i left it out but that is stopping the conversion , how did you pass the output names to the parameter