[GitHub] [incubator-tvm] leandron commented on a change in pull request #6112: TVMC - a command line driver for TVM

[GitBox <gi...@apache.org>]

leandron commented on a change in pull request #6112:
URL: https://github.com/apache/incubator-tvm/pull/6112#discussion_r461597250



##########
File path: tvmc/README.md
##########
@@ -0,0 +1,122 @@
+<!--- Licensed to the Apache Software Foundation (ASF) under one -->
+<!--- or more contributor license agreements.  See the NOTICE file -->
+<!--- distributed with this work for additional information -->
+<!--- regarding copyright ownership.  The ASF licenses this file -->
+<!--- to you under the Apache License, Version 2.0 (the -->
+<!--- "License"); you may not use this file except in compliance -->
+<!--- with the License.  You may obtain a copy of the License at -->
+
+<!---   http://www.apache.org/licenses/LICENSE-2.0 -->
+
+<!--- Unless required by applicable law or agreed to in writing, -->
+<!--- software distributed under the License is distributed on an -->
+<!--- "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY -->
+<!--- KIND, either express or implied.  See the License for the -->
+<!--- specific language governing permissions and limitations -->
+<!--- under the License. -->
+
+# TVMC
+
+```tvmc``` is a tool that provides useful command line invocations to compile,
+run and tune models using TVM graph runtime.
+
+In order to compile and tune, ```tvmc``` takes a model file and parameters as inputs,
+and outputs a TAR file that contains the TVM modules that represent the
+input model, graph and weights, for the required target. Target can be native or
+cross-compiled.
+
+When running a given model, ```tvmc``` expects a compiled model and input tensor values, so
+that it can produce the outputs, when running on the required target, local or remote.
+
+This document presents an overview and a short tutorial about ```tvmc```.
+
+## Installation
+
+```tvmc``` is a Python tool and - provided TVM and dependencies are available - it can be
+installed in various ways.
+
+The recommended way to install ```tvmc``` is via it's ```setuptools``` configuration file,
+located at ```tvm/tvmc/setup.py```. To do that, go to the the TVM directory and run the
+installation command, as described below:
+
+    cd tvm/tvmc
+    python setup.py install
+
+The command above should install everything needed to get started with ```tvmc```, including
+all the the supported frontends.
+
+Once ```tvmc``` is installed, the main entry-point is the ```tvmc``` command line. A set of
+sub-commands are available, to run the specific tasks offered by ```tvmc```: ```tune```,
+```compile``` and ```run```.
+
+The simplest way to get more information about a specific sub-command is ```tvmc <subcommand>
+-- help```.
+
+    tvmc compile --help
+
+##  Usage
+
+Now, let's compile a network and generate a few predictions using ```tvmc```.
+
+As described above, in order to compile a model using ```tvmc```, the first thing we need is
+a model file. For the sake of this example, let's use a MobileNet V1 model, in TFLite format.
+More information about the model is available on
+[this page](https://www.tensorflow.org/lite/guide/hosted_models).
+
+To download and un-compress the ```.tgz``` file (34Mb), so that we can access the TFLite model,
+run the command lines below:
+
+    wget https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_1.0_224_quant.tgz
+    tar xvzf mobilenet_v1_1.0_224_quant.tgz
+
+With these commands, we should be able to provide the MobileNet V1 file (```mobilenet_v1_1.0_224_quant.tflite```)
+to ```tvmc```, and obtain our TVM compiled model as an output. To do that, run the
+following command line:
+
+    tvmc compile -v mobilenet_v1_1.0_224_quant.tflite -o compiled_model.tar
+
+As an output, you will notice a ```compiled_model.tar```, in the same directory.
+
+Now it is time to feed the model with some input, that will generate a prediction using TVM.
+As models are very diverse in terms of input formats and the source of those inputs (images, streams,
+sensors, sound, to name a few), ```tvmc``` supports ```.npz``` (serialized NumPy arrays) as the
+main format for ```tvmc run```. To learn more about the ```.npz``` format, please read the
+[documentation](https://numpy.org/doc/stable/reference/generated/numpy.savez.html) on NumPy website.
+
+MobileNet V1 expects a ```(224, 224, 3)``` input tensor. The Python code snippet below, can be used
+as an example on how to convert a PNG file into a ```.npz``` file in the expected shape.
+The example below uses [PIL](https://pillow.readthedocs.io/en/stable/) and
+[NumPy](https://numpy.org) functions to import the image and generate the expected file.
+
+    from tvm.contrib.download import download_testdata
+    from PIL import Image
+    import numpy as np
+
+    cat_url = 'https://github.com/dmlc/mxnet.js/blob/master/data/cat.png?raw=true'
+    image_path = download_testdata(cat_url, 'imagenet_cat.png', module='data')
+    resized_image = Image.open(image_path).resize((224, 224))
+    image_data = np.asarray(resized_image).astype("float32")
+    image_data = np.expand_dims(image_data, axis=0)
+
+    np.savez("imagenet_cat", input=image_data)

Review comment:
       We considered that choice of "inside or outside the existing tvm/python", as well as the decision to publish the command line in the same Python package as TVM or in a different one.
   
   The main reason for the current choice, is to have a different (more comprehensive) set of Python dependencies, that would allow the user to have everything they need to run tuning, compilations and predictions, without the need of installing and thinking about ad-hoc dependencies, such as tensorflow, tflite, keras, onnx, pytorch, etc.
   
   As you can see on the proposed `setup.py`, we add the dependencies for all the supported frontends, which I understand are intentionally left out the current TVM Python package.
   
   What do you think? I suggest, if we are to move the driver to be inside the TVM package, would you agree with having the `tvm.driver` with a set of _optional dependencies_[1] (it is a feature from `setuptools`), so that we can provide a simple way for the end user to install these dependencies? Otherwise, I would suggest we keep them as separate packages.
   
   [1] https://setuptools.readthedocs.io/en/latest/setuptools.html#declaring-extras-optional-features-with-their-own-dependencies




----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org