TensorFlow.js 是一个开源硬件加速 JavaScript 库，用于训练和部署机器学习模型.zip

# Getting started **TensorFlow.js converter** is an open source library to load a pretrained TensorFlow [SavedModel](https://www.tensorflow.org/guide/saved_model) or [TensorFlow Hub module](https://www.tensorflow.org/hub/) into the browser and run inference through [TensorFlow.js](https://js.tensorflow.org). __Note__: _Session bundle format have been deprecated. A 2-step process to import your model: 1. A python pip package to convert a TensorFlow SavedModel or TensorFlow Hub module to a web friendly format. If you already have a converted model, or are using an already hosted model (e.g. MobileNet), skip this step. 2. [JavaScript API](./src/executor/graph_model.ts), for loading and running inference. ## Step 1: Converting a [TensorFlow SavedModel](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/saved_model/README.md), [TensorFlow Hub module](https://www.tensorflow.org/hub/), [Keras HDF5](https://keras.io/getting_started/faq/#what-are-my-options-for-saving-models), [tf.keras SavedModel](https://www.tensorflow.org/api_docs/python/tf/keras/saving/save_model), or [Flax/JAX model](http://github.com/google/flax) to a web-friendly format __0. Please make sure that you run in a Docker container or a virtual environment.__ The script pulls its own subset of TensorFlow, which might conflict with the existing TensorFlow/Keras installation. __Note__: *Check that [`tf-nightly-cpu-2.0-preview`](https://pypi.org/project/tf-nightly-cpu-2.0-preview/#files) is available for your platform.* Most of the times, this means that you have to use Python 3.6.8 in your local environment. To force Python 3.6.8 in your local project, you can install [`pyenv`](https://github.com/pyenv/pyenv) and proceed as follows in the target directory: ```bash pyenv install 3.6.8 pyenv local 3.6.8 ``` Now, you can [create and activate](https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/) a `venv` virtual environment in your current folder: ```bash virtualenv --no-site-packages venv . venv/bin/activate ``` __1. Install the TensorFlow.js pip package:__ Install the library with interactive CLI: ```bash pip install tensorflowjs[wizard] ``` __2. Run the converter script provided by the pip package:__ There are three way to trigger the model conversion, explain below: - The conversion wizard: `tensorflowjs_wizard` ([go to section](https://github.com/tensorflow/tfjs/blob/master/tfjs-converter/README.md#conversion-wizard-tensorflowjs_wizard)) - Regular conversion script: `tensorflowjs_converter` ([go to section](https://github.com/tensorflow/tfjs/tree/master/tfjs-converter#regular-conversion-script-tensorflowjs_converter)) - Calling a converter function in Python (Flax/JAX) ([go to section](https://github.com/tensorflow/tfjs/tree/master/tfjs-converter#calling-a-converter-function-in-python-flaxjax)) ## Conversion wizard: `tensorflowjs_wizard` To start the conversion wizard: ```bash tensorflowjs_wizard ``` This tool will walk you through the conversion process and provide you with details explanations for each choice you need to make. Behind the scene it calls the converter script (`tensorflowjs_converter`) in pip package. This is the recommended way to convert a single model. There is also a dry run mode for the wizard, which will not perform the actual conversion but only generate the command for `tensorflowjs_converter` command. This generated command can be used in your own shell script. Here is an screen capture of the wizard in action.  ```bash tensorflowjs_wizard --dryrun ``` To convert a batch of models or integrate the conversion process into your own script, you should use the tensorflowjs_converter script. ## Regular conversion script: `tensorflowjs_converter` The converter expects a __TensorFlow SavedModel__, __TensorFlow Hub module__, __TensorFlow.js JSON__ format, __Keras HDF5 model__, or __tf.keras SavedModel__ for input. * __TensorFlow SavedModel__ example: ```bash tensorflowjs_converter \ --input_format=tf_saved_model \ --output_format=tfjs_graph_model \ --signature_name=serving_default \ --saved_model_tags=serve \ /mobilenet/saved_model \ /mobilenet/web_model ``` * __TensorFlow Frozen Model__ example: __Note:__ Frozen model is a deprecated format and support is added for backward compatibility purpose. ```bash $ tensorflowjs_converter \ --input_format=tf_frozen_model \ --output_node_names='MobilenetV1/Predictions/Reshape_1' \ /mobilenet/frozen_model.pb \ /mobilenet/web_model ``` * __Tensorflow Hub module__ example: ```bash tensorflowjs_converter \ --input_format=tf_hub \ 'https://tfhub.dev/google/imagenet/mobilenet_v1_100_224/classification/1' \ /mobilenet/web_model ``` * __Keras HDF5 model__ example: ```bash tensorflowjs_converter \ --input_format=keras \ /tmp/my_keras_model.h5 \ /tmp/my_tfjs_model ``` * __tf.keras SavedModel__ example: ```bash tensorflowjs_converter \ --input_format=keras_saved_model \ /tmp/my_tf_keras_saved_model/1542211770 \ /tmp/my_tfjs_model ``` Note that the input path used above is a subfolder that has a Unix epoch time (1542211770) and is generated automatically by tensorflow when it saved a tf.keras model in the SavedModel format. ### Conversion Flags |Positional Arguments | Description | |---|---| |`input_path` | Full path of the saved model directory or TensorFlow Hub module handle or path.| |`output_path` | Path for all output artifacts.| | Options | Description |---|---| |`--input_format` | The format of input model, use `tf_saved_model` for SavedModel, `tf_hub` for TensorFlow Hub module, `tfjs_layers_model` for TensorFlow.js JSON format, and `keras` for Keras HDF5. | |`--output_format`| The desired output format. Must be `tfjs_layers_model`, `tfjs_graph_model` or `keras`. Not all pairs of input-output formats are supported. Please file a [github issue](https://github.com/tensorflow/tfjs/issues) if your desired input-output pair is not supported.| |<nobr>`--saved_model_tags`</nobr> | Only applicable to SavedModel conversion. Tags of the MetaGraphDef to load, in comma separated format. If there are no tags defined in the saved model, set it to empty string `saved_model_tags=''`. Defaults to `serve`.| |`--signature_name` | Only applicable to TensorFlow SavedModel and Hub module conversion, signature to load. Defaults to `serving_default` for SavedModel and `default` for Hub module. See https://www.tensorflow.org/hub/common_signatures/.| |`--strip_debug_ops` | Strips out TensorFlow debug operations `Print`, `Assert`, `CheckNumerics`. Defaults to `True`.| |`--quantization_bytes` | (Deprecated) How many bytes to optionally quantize/compress the weights to. Valid values are 1 and 2. which will quantize int32 and float32 to 1 or 2 bytes respectively. The default (unquantized) size is 4 bytes.| |`--quantize_float16` | Comma separated list of node names to apply float16 quantization. You can also use wildcard symbol (\*) to apply quantization to multiple nodes (e.g., conv/\*/weights). When the flag is provided without any nodes the default behavior will match all nodes. | |`--quantize_uint8` | Comma separated list of node names to apply 1-byte affine quantization. You can also use wildcard symbol (\*) to apply quantization to multiple nodes (e.g., conv/\*/weights). When the flag is provided without any nodes the default behavior will match all nodes. | |`--quantize_uint16` | Comma separated list of node names to apply 2-byte affine quantization. You can also use wildcard symbol (\*) to apply quantization to multiple nodes (e.g., conv/\*/weights). When the flag is provided without any nodes the default behavior will match all nodes. | |`--weight_shard_size_bytes` | Shard size (in bytes) of the weight files. Only supported when `output_format` is `tfjs_layers_model` or `tfjs_graph_model`. Default size is 4 MB (4194304 bytes).| |<nobr>`--output