tflite2tensorflow

【WIP】 Generate saved_model, tfjs, tf-trt, EdgeTPU, CoreML, quantized tflite and .pb from .tflite.

GitHub

1. Supported Layers

No.	TFLite Layer	TF Layer	Remarks
1	CONV_2D	tf.nn.conv2d
2	DEPTHWISE_CONV_2D	tf.nn.depthwise_conv2d
3	MAX_POOL_2D	tf.nn.max_pool
4	PAD	tf.pad
5	MIRROR_PAD	tf.raw_ops.MirrorPad
6	RELU	tf.nn.relu
7	PRELU	tf.keras.layers.PReLU
8	RELU6	tf.nn.relu6
9	RESHAPE	tf.reshape
10	ADD	tf.add
11	SUB	tf.math.subtract
12	CONCATENATION	tf.concat
13	LOGISTIC	tf.math.sigmoid
14	TRANSPOSE_CONV	tf.nn.conv2d_transpose
15	MUL	tf.multiply
16	HARD_SWISH	xtf.nn.relu6(x+3)0.16666667 Or xtf.nn.relu6(x+3)0.16666666
17	AVERAGE_POOL_2D	tf.keras.layers.AveragePooling2D
18	FULLY_CONNECTED	tf.keras.layers.Dense
19	RESIZE_BILINEAR	tf.image.resize Or tf.image.resize_bilinear
20	RESIZE_NEAREST_NEIGHBOR	tf.image.resize Or tf.image.resize_nearest_neighbor
21	MEAN	tf.math.reduce_mean
22	SQUARED_DIFFERENCE	tf.math.squared_difference
23	RSQRT	tf.math.rsqrt
24	DEQUANTIZE	(const)
25	FLOOR	tf.math.floor
26	TANH	tf.math.tanh
27	DIV	tf.math.divide
28	FLOOR_DIV	tf.math.floordiv
29	SUM	tf.math.reduce_sum
30	POW	tf.math.pow
31	SPLIT	tf.split
32	SOFTMAX	tf.nn.softmax
33	STRIDED_SLICE	tf.strided_slice
34	TRANSPOSE	ttf.transpose
35	SPACE_TO_DEPTH	tf.nn.space_to_depth
36	DEPTH_TO_SPACE	tf.nn.depth_to_space
37	REDUCE_MAX	tf.math.reduce_max
38	Convolution2DTransposeBias	tf.nn.conv2d_transpose, tf.math.add	CUSTOM, MediaPipe
39	LEAKY_RELU	tf.keras.layers.LeakyReLU
40	MAXIMUM	tf.math.maximum
41	MINIMUM	tf.math.minimum
42	MaxPoolingWithArgmax2D	tf.raw_ops.MaxPoolWithArgmax	CUSTOM, MediaPipe
43	MaxUnpooling2D	tf.cast, tf.shape, tf.math.floordiv, tf.math.floormod, tf.ones_like, tf.shape, tf.concat, tf.reshape, tf.transpose, tf.scatter_nd	CUSTOM, MediaPipe
44	GATHER	tf.gather
45	CAST	tf.cast
46	SLICE	tf.slice
47	PACK	tf.stack
48	UNPACK	tf.unstack

2. Environment

Python3.6+
TensorFlow v2.4.0+ or tf-nightly
TensorFlow Lite v2.4.1 with MediaPipe Custom OP, FlexDelegate and XNNPACK enabled
- Add a custom OP to the TFLite runtime to build the whl installer (for Python), MaxPoolingWithArgmax2D, MaxUnpooling2D, Convolution2DTransposeBias
flatc v1.12.0

3. Setup

To install using the Python Package Index (PyPI), use the following command.

$ pip3 install tflite2tensorflow --upgrade

Or, To install with the latest source code of the main branch, use the following command.

$ pip3 install git+https://github.com/PINTO0309/tflite2tensorflow --upgrade

Installs a customized TensorFlow Lite runtime with support for MediaPipe Custom OP, FlexDelegate, and XNNPACK. If tflite_runtime does not install properly, please follow the instructions in the next article to build a custom build in the environment you are using. Add a custom OP to the TFLite runtime to build the whl installer (for Python), MaxPoolingWithArgmax2D, MaxUnpooling2D, Convolution2DTransposeBias

$ sudo pip3 uninstall tensorboard-plugin-wit tb-nightly tensorboard \
                      tf-estimator-nightly tensorflow-gpu \
                      tensorflow tf-nightly tensorflow_estimator tflite_runtime -y

### Customized version of TensorFlow Lite installation
$ sudo gdown --id 1RWZmfFgtxm3muunv6BSf4yU29SKKFXIh
$ sudo chmod +x tflite_runtime-2.4.1-py3-none-any.whl
$ sudo pip3 install tflite_runtime-2.4.1-py3-none-any.whl

### Install the full TensorFlow package
$ sudo pip3 install tf-nightly
 or
$ sudo pip3 install tensorflow==2.4.1

### Download flatc
$ flatbuffers/1.12.0/download.sh

### Download schema.fbs
$ wget https://github.com/PINTO0309/tflite2tensorflow/raw/main/schema/schema.fbs

If the downloaded flatc does not work properly, please build it in your environment.

$ git clone -b v1.12.0 https://github.com/google/flatbuffers.git
$ cd flatbuffers && mkdir build && cd build
$ cmake -G "Unix Makefiles" -DCMAKE_BUILD_TYPE=Release ..
$ make -j$(nproc)

vvtvsu0y1791ow2ybdk61s9fv7e4 saxqukktcjncsk2hp7m8p2cns4q4

4. Usage / Execution sample

4-1. Step 1 : Generating saved_model and FreezeGraph (.pb)

$ tflite2tensorflow \
  --model_path magenta_arbitrary-image-stylization-v1-256_fp16_prediction_1.tflite \
  --flatc_path ./flatc \
  --schema_path schema.fbs \
  --output_pb True

$ tflite2tensorflow \
  --model_path magenta_arbitrary-image-stylization-v1-256_fp16_prediction_1.tflite \
  --flatc_path ./flatc \
  --schema_path schema.fbs \
  --output_pb True \
  --optimizing_hardswish_for_edgetpu True

4-2. Step 2 : Generation of quantized tflite, TFJS, TF-TRT, EdgeTPU, and CoreML

$ tflite2tensorflow \
  --model_path magenta_arbitrary-image-stylization-v1-256_fp16_prediction_1.tflite \
  --flatc_path ./flatc \
  --schema_path schema.fbs \
  --output_no_quant_float32_tflite True \
  --output_weight_quant_tflite True \
  --output_float16_quant_tflite True \
  --output_integer_quant_tflite True \
  --string_formulas_for_normalization 'data / 255.0' \
  --output_tfjs True \
  --output_coreml True \
  --output_tftrt True

$ tflite2tensorflow \
  --model_path magenta_arbitrary-image-stylization-v1-256_fp16_prediction_1.tflite \
  --flatc_path ./flatc \
  --schema_path schema.fbs \
  --output_no_quant_float32_tflite True \
  --output_weight_quant_tflite True \
  --output_float16_quant_tflite True \
  --output_integer_quant_tflite True \
  --output_edgetpu True \
  --string_formulas_for_normalization 'data / 255.0' \
  --output_tfjs True \
  --output_coreml True \
  --output_tftrt True

5. Sample image

This is the result of converting MediaPipe's Meet Segmentation model (segm_full_v679.tflite / Float16 / Google Meet) to saved_model and then reconverting it to Float32 tflite. Replace the GPU-optimized Convolution2DTransposeBias layer with the standard TransposeConv and BiasAdd layers in a fully automatic manner. The weights and biases of the Float16 Dequantize layer are automatically back-quantized to Float32 precision. The generated saved_model in Float32 precision can be easily converted to Float16, INT8, EdgeTPU, TFJS, TF-TRT, CoreML, ONNX, and OpenVINO.

Before	After

Metadata

Release files for tflite2tensorflow 0.2.3

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for tflite2tensorflow 0.2.3
File	Size	Uploaded
tflite2tensorflow-0.2.3.tar.gz	18.1 kB	Jan 24, 2021	Details

Built distribution (wheel)

Table of built distributions (wheels) for tflite2tensorflow 0.2.3
File	Interpreter	ABI	Platform	Reset
tflite2tensorflow-0.2.3-py3-none-any.whl 16.9 kB Jan 24, 2021	Python 3	none	any	Details

Total release size: 35.0 kB

Release files / tflite2tensorflow-0.2.3.tar.gz

Download URL	tflite2tensorflow-0.2.3.tar.gz
Size	18.1 kB
Tags	Source
SHA-256 checksum How to use checksums	`38aadaa6604f1d2339aa09df19e5915af606ba896299d798a466657a9a039fe0`
BLAKE2b-256 checksum How to use checksums	`903b49c10f57b4de619dc96205e9e11404765fd7381e6bdc65a4cdf388678932`
Upload date	Jan 24, 2021
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/3.2.0 pkginfo/1.6.0 requests/2.22.0 setuptools/47.1.1 requests-toolbelt/0.9.1 tqdm/4.40.2 CPython/3.6.9`

Release files / tflite2tensorflow-0.2.3-py3-none-any.whl

Download URL	tflite2tensorflow-0.2.3-py3-none-any.whl
Size	16.9 kB
Tags	Python 3
SHA-256 checksum How to use checksums	`df29062bf7200dcded7305dad91e9c4a21a12970d765cf39efbf5d0dc8efba17`
BLAKE2b-256 checksum How to use checksums	`7cd20505016f43788e5d12d12fe57ec8473dc8a5fe5d8fd75f2f8a24cd94c56d`
Upload date	Jan 24, 2021
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/3.2.0 pkginfo/1.6.0 requests/2.22.0 setuptools/47.1.1 requests-toolbelt/0.9.1 tqdm/4.40.2 CPython/3.6.9`