c2 to onnx and hybrid frontend initial commit

Author: inkawhich

beginner_source/caffe2_onnx_primer_tutorial.py

@@ -0,0 +1,201 @@
+# -*- coding: utf-8 -*-
+"""
+Caffe2 ONNX Primer
+==================
+**Author**: `Nathan Inkawhich <https://github.com/inkawhich>`_
+
+This tutorial is a brief look at how to use Caffe2 and
+`ONNX <http://onnx.ai/about>`_ together. More specifically, we will
+show how to export a model from Caffe2 to ONNX and how to import a model
+from ONNX into Caffe2. Hopefully, the motivation is clear but this
+tutorial shows how to use the very fast and efficient Caffe2 framework
+with the flexibility enabling ONNX framework. One important fact to keep
+in mind is that ONNX is designed to enable deployment and *inference* in
+frameworks other than where the model was trained. Currently, there is
+no streamlined way to finetune ONNX models. The workflow for this
+document is as follows:
+
+-  Run prediction with a Caffe2 model and collect initial prediction
+-  Export the Caffe2 model to ONNX format
+-  Import the saved ONNX model back into Caffe2
+-  Run prediction on imported model and verify results
+
+Let's get started with some imports.
+
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+import numpy as np
+import operator
+from caffe2.proto import caffe2_pb2
+from caffe2.python import core, workspace, models
+import onnx
+import caffe2.python.onnx.frontend # Required for Caffe2->ONNX export
+import caffe2.python.onnx.backend # Required for ONNX->Caffe2 import
+
+
+######################################################################
+# Inputs
+# ------
+#
+# Now we will specify the inputs. The *MODELS\_DIR* is where the
+# downloaded Caffe2 models are saved, the *MODEL\_NAME* is the name of the
+# model we want to use, and *SIZE* is the size of image the model expects.
+# For more information about downloading a pretrained Caffe2 model, see
+# the `Loading Pretrained Models
+# Tutorial <https://github.com/caffe2/tutorials/blob/master/Loading_Pretrained_Models.ipynb>`__.
+#
+
+# User Inputs
+MODELS_DIR = "../models"
+MODEL_NAME = "squeezenet" # e.g. [squeezenet, bvlc_alexnet, bvlc_googlenet, bvlc_reference_caffenet]
+SIZE = 224
+
+# Construct path strings from inputs
+INIT_NET = "{}/{}/init_net.pb".format(MODELS_DIR, MODEL_NAME)
+PREDICT_NET = "{}/{}/predict_net.pb".format(MODELS_DIR, MODEL_NAME)
+ONNX_MODEL = "{}/{}/my_model.onnx".format(MODELS_DIR, MODEL_NAME) # we will create this
+
+
+######################################################################
+# Load Caffe2 Model
+# -----------------
+#
+# Before we perform the export we will first load the pretrained init and
+# predict nets, then create a *Predictor*. Next, we will create a random
+# input to get a baseline result for comparision later. Take note of the
+# predicted label and confidence.
+#
+
+# Generate random NCHW input to run model
+#   This is a placeholder for any real image that is processed and
+#   put in NCHW order.
+image = np.random.rand(1,3,SIZE,SIZE).astype(np.float32)
+print("Input Shape: ",image.shape)
+
+# Prepare the nets
+predict_net = caffe2_pb2.NetDef()
+with open(PREDICT_NET, 'rb') as f:
+    predict_net.ParseFromString(f.read())
+init_net = caffe2_pb2.NetDef()
+with open(INIT_NET, 'rb') as f:
+    init_net.ParseFromString(f.read())
+
+# Initialize the predictor from the nets
+p = workspace.Predictor(init_net, predict_net)
+
+#### Run the sample data
+
+# Run the net and return prediction
+results = p.run({'data': image})
+results = np.asarray(results)
+print("Results Shape: ", results.shape)
+
+# Quick way to get the top-1 prediction result
+curr_pred, curr_conf = max(enumerate(np.squeeze(results)), key=operator.itemgetter(1))
+print("Top-1 Prediction: {} @ {}".format(curr_pred, curr_conf))
+
+
+
+######################################################################
+# Caffe2 :math:`\rightarrow` ONNX Export
+# --------------------------------------
+#
+# Finally, we have reached the interesting stuff. It is not hard to
+# imagine why one may want to export a Caffe2 model to ONNX. Maybe you
+# have a cool idea for an iPhone app and want to use a model trained in
+# Caffe2 with CoreML as part of the app. Or, maybe you have a system built
+# in Tensorflow but want to test out a model from the Caffe2 Model Zoo.
+# ONNX enables this interoperability by allowing models to be imported and
+# exported into different frameworks (for inference!).
+#
+# The code below shows how to **export** a model trained in Caffe2 to ONNX
+# format. Once in ONNX format, the model can be imported into any other
+# compatible framework to be used for *inference*. From the Caffe2 side,
+# we only need the previously loaded *init\_net* and *predict\_net*
+# *caffe2\_pb2.NetDef* objects.
+#
+# There are only a few steps to export once the nets are loaded. First, we
+# must declare (via Python dictionary) the type and shape of inputs and
+# outputs of the model. This information is not explicitly specified in
+# the Caffe2 model architecture but is required by ONNX. Next, we must
+# make sure the model has a name, otherwise the internal model checks in
+# the ONNX converter will fail. Then, all thats left to do is create the
+# ONNX model, check it, and save it.
+#
+
+# We need to provide type and shape of the model inputs
+data_type = onnx.TensorProto.FLOAT
+data_shape = (1, 3, 224, 224)
+value_info = {
+    'data': (data_type, data_shape)
+}
+
+# Make sure the net has a name. Otherwise, the checker will fail.
+if predict_net.name == "":
+    predict_net.name = "ModelNameHere"
+
+# Create the ONNX model
+onnx_model = caffe2.python.onnx.frontend.caffe2_net_to_onnx_model(
+    predict_net,
+    init_net,
+    value_info,
+)
+
+# Check the ONNX model. Exception will be thrown if there is a problem here.
+onnx.checker.check_model(onnx_model)
+
+# Save the ONNX model
+onnx.save(onnx_model, ONNX_MODEL)
+
+
+######################################################################
+# ONNX :math:`\rightarrow` Caffe2 Import
+# --------------------------------------
+#
+# Now suppose someone has trained Alexnet2.0 which gets 99.9% top-1 test
+# accuracy on ImageNet ... *gasp* ... in Tensorflow. As a Caffe2 user, all
+# we have to do is convince them to convert the model to ONNX format, then
+# we can import it and use it. Since we are running out of time in this
+# 5-minute primer, here we will only show how to import the model we just
+# exported back into Caffe2. The import happens in a single load command
+# (``onnx.load``), then we can start feeding the model data in just one
+# more command (``run_model``). Also, note that the predictions from this
+# imported model and the original model are the exact same, indicating
+# nothing was lost in the export/import process.
+#
+
+# Load the ONNX model
+model = onnx.load(ONNX_MODEL)
+
+# Run the ONNX model with Caffe2
+outputs = caffe2.python.onnx.backend.run_model(model, [image])
+print("Output Shape: ", np.array(outputs).shape)
+
+# Get model prediction
+curr_pred, curr_conf = max(enumerate(np.squeeze(results)), key=operator.itemgetter(1))
+print("Top-1 Prediction: {} @ {}".format(curr_pred, curr_conf))
+
+
+
+######################################################################
+# Hopefully it is clear that the caffe2-onnx interface for both importing
+# and exporting is relatively simple. For more information about ONNX and
+# to see more tutorials on using ONNX with different frameworks see the
+# `ONNX Tutorials <https://github.com/onnx/tutorials>`__. Also, although
+# importing and exporting with Caffe2 is supported, and exporting a model
+# from PyTorch to ONNX is supported, *importing* an ONNX model into
+# PyTorch is *NOT*, but is coming soon!
+#
+# Here are some more cool ONNX resources for the curious reader:
+#
+# -  `ONNX Python API
+#    Overview <https://github.com/onnx/onnx/blob/master/docs/PythonAPIOverview.md>`__
+# -  `ONNX Model Zoo <https://github.com/onnx/models>`__
+# -  `ONNX
+#    Operators <https://github.com/onnx/onnx/blob/master/docs/Operators.md>`__
+# -  `ONNX Tutorials <https://github.com/onnx/tutorials>`__
+#

beginner_source/hybrid_frontend/README.txt

@@ -0,0 +1,10 @@
+ Hybrid Frontend Tutorials
+ -------------------------
+ 
+1. learning_hybrid_frontend_through_example_tutorial.py
+	Learning Hybrid Frontend Through Example
+	http://pytorch.org/tutorials/beginner/hybrid_frontend/learning_hybrid_frontend_through_example_tutorial.html
+
+2. introduction_to_hybrid_frontend_tutorial.py
+	Introduction to Hybrid Frontend
+	http://pytorch.org/tutorials/beginner/hybrid_frontend/introduction_to_hybrid_frontend_tutorial.html