.. note:: :class: sphx-glr-download-link-note Click :ref:`here ` to download the full example code .. rst-class:: sphx-glr-example-title .. _sphx_glr_topic_vta_tutorials_frontend_deploy_classification.py: Deploy Pretrained Vision Model from MxNet on VTA ================================================ **Author**: `Thierry Moreau `_ This tutorial provides an end-to-end demo, on how to run ImageNet classification inference onto the VTA accelerator design to perform ImageNet classification tasks. It showcases Relay as a front end compiler that can perform quantization (VTA only supports int8/32 inference) as well as graph packing (in order to enable tensorization in the core) to massage the compute graph for the hardware target. Install dependencies -------------------- To use the autotvm package in tvm, we need to install some extra dependencies. (change "3" to "2" if you use python2): .. code-block:: bash pip3 install --user mxnet requests "Pillow<7" Now return to the python code. Import packages. .. code-block:: default from __future__ import absolute_import, print_function import argparse, json, os, requests, sys, time from io import BytesIO from os.path import join, isfile from PIL import Image from mxnet.gluon.model_zoo import vision import numpy as np from matplotlib import pyplot as plt import tvm from tvm import te from tvm import rpc, autotvm, relay from tvm.contrib import graph_executor, utils, download from tvm.contrib.debugger import debug_executor from tvm.relay import transform import vta from vta.testing import simulator from vta.top import graph_pack # Make sure that TVM was compiled with RPC=1 assert tvm.runtime.enabled("rpc") Define the platform and model targets ------------------------------------- Execute on CPU vs. VTA, and define the model. .. code-block:: default # Load VTA parameters from the 3rdparty/vta-hw/config/vta_config.json file env = vta.get_env() # Set ``device=arm_cpu`` to run inference on the CPU # or ``device=vta`` to run inference on the FPGA. device = "vta" target = env.target if device == "vta" else env.target_vta_cpu # Dictionary lookup for when to start/end bit packing pack_dict = { "resnet18_v1": ["nn.max_pool2d", "nn.global_avg_pool2d"], "resnet34_v1": ["nn.max_pool2d", "nn.global_avg_pool2d"], "resnet18_v2": ["nn.max_pool2d", "nn.global_avg_pool2d"], "resnet34_v2": ["nn.max_pool2d", "nn.global_avg_pool2d"], "resnet50_v2": ["nn.max_pool2d", "nn.global_avg_pool2d"], "resnet101_v2": ["nn.max_pool2d", "nn.global_avg_pool2d"], } # Name of Gluon model to compile # The ``start_pack`` and ``stop_pack`` labels indicate where # to start and end the graph packing relay pass: in other words # where to start and finish offloading to VTA. model = "resnet18_v1" assert model in pack_dict Obtain an execution remote -------------------------- When target is 'pynq', reconfigure FPGA and runtime. Otherwise, if target is 'sim', execute locally. .. code-block:: default if env.TARGET not in ["sim", "tsim", "intelfocl"]: # Get remote from tracker node if environment variable is set. # To set up the tracker, you'll need to follow the "Auto-tuning # a convolutional network for VTA" tutorial. tracker_host = os.environ.get("TVM_TRACKER_HOST", None) tracker_port = os.environ.get("TVM_TRACKER_PORT", None) # Otherwise if you have a device you want to program directly from # the host, make sure you've set the variables below to the IP of # your board. device_host = os.environ.get("VTA_RPC_HOST", "192.168.2.99") device_port = os.environ.get("VTA_RPC_PORT", "9091") if not tracker_host or not tracker_port: remote = rpc.connect(device_host, int(device_port)) else: remote = autotvm.measure.request_remote( env.TARGET, tracker_host, int(tracker_port), timeout=10000 ) # Reconfigure the JIT runtime and FPGA. # You can program the FPGA with your own custom bitstream # by passing the path to the bitstream file instead of None. reconfig_start = time.time() vta.reconfig_runtime(remote) vta.program_fpga(remote, bitstream=None) reconfig_time = time.time() - reconfig_start print("Reconfigured FPGA and RPC runtime in {0:.2f}s!".format(reconfig_time)) # In simulation mode, host the RPC server locally. else: remote = rpc.LocalSession() if env.TARGET in ["intelfocl"]: # program intelfocl aocx vta.program_fpga(remote, bitstream="vta.bitstream") # Get execution context from remote ctx = remote.ext_dev(0) if device == "vta" else remote.cpu(0) Build the inference graph executor ---------------------------------- Grab vision model from Gluon model zoo and compile with Relay. The compilation steps are: 1. Front end translation from MxNet into Relay module. 2. Apply 8-bit quantization: here we skip the first conv layer, and dense layer which will both be executed in fp32 on the CPU. 3. Perform graph packing to alter the data layout for tensorization. 4. Perform constant folding to reduce number of operators (e.g. eliminate batch norm multiply). 5. Perform relay build to object file. 6. Load the object file onto remote (FPGA device). 7. Generate graph executor, `m`. .. code-block:: default # Load pre-configured AutoTVM schedules with autotvm.tophub.context(target): # Populate the shape and data type dictionary for ImageNet classifier input dtype_dict = {"data": "float32"} shape_dict = {"data": (env.BATCH, 3, 224, 224)} # Get off the shelf gluon model, and convert to relay gluon_model = vision.get_model(model, pretrained=True) # Measure build start time build_start = time.time() # Start front end compilation mod, params = relay.frontend.from_mxnet(gluon_model, shape_dict) # Update shape and type dictionary shape_dict.update({k: v.shape for k, v in params.items()}) dtype_dict.update({k: str(v.dtype) for k, v in params.items()}) if target.device_name == "vta": # Perform quantization in Relay # Note: We set opt_level to 3 in order to fold batch norm with tvm.transform.PassContext(opt_level=3): with relay.quantize.qconfig(global_scale=8.0, skip_conv_layers=[0]): mod = relay.quantize.quantize(mod, params=params) # Perform graph packing and constant folding for VTA target assert env.BLOCK_IN == env.BLOCK_OUT # do device annotation if target is intelfocl or sim relay_prog = graph_pack( mod["main"], env.BATCH, env.BLOCK_OUT, env.WGT_WIDTH, start_name=pack_dict[model][0], stop_name=pack_dict[model][1], device_annot=(env.TARGET == "intelfocl"), ) else: relay_prog = mod["main"] # Compile Relay program with AlterOpLayout disabled if target.device_name != "vta": with tvm.transform.PassContext(opt_level=3, disabled_pass={"AlterOpLayout"}): graph, lib, params = relay.build( relay_prog, target=target, params=params, target_host=env.target_host ) else: if env.TARGET == "intelfocl": # multiple targets to run both on cpu and vta target = {"cpu": env.target_vta_cpu, "ext_dev": target} with vta.build_config(opt_level=3, disabled_pass={"AlterOpLayout"}): graph, lib, params = relay.build( relay_prog, target=target, params=params, target_host=env.target_host ) # Measure Relay build time build_time = time.time() - build_start print(model + " inference graph built in {0:.2f}s!".format(build_time)) # Send the inference library over to the remote RPC server temp = utils.tempdir() lib.export_library(temp.relpath("graphlib.tar")) remote.upload(temp.relpath("graphlib.tar")) lib = remote.load_module("graphlib.tar") if env.TARGET == "intelfocl": ctxes = [remote.ext_dev(0), remote.cpu(0)] m = graph_executor.create(graph, lib, ctxes) else: # Graph runtime m = graph_executor.create(graph, lib, ctx) .. rst-class:: sphx-glr-script-out Out: .. code-block:: none /workspace/python/tvm/relay/build_module.py:333: DeprecationWarning: Please use input parameter mod (tvm.IRModule) instead of deprecated parameter mod (tvm.relay.function.Function) DeprecationWarning, /workspace/vta/tutorials/frontend/deploy_classification.py:211: DeprecationWarning: legacy graph executor behavior of producing json / lib / params will be removed in the next release. Please see documents of tvm.contrib.graph_executor.GraphModule for the new recommended usage. relay_prog, target=target, params=params, target_host=env.target_host resnet18_v1 inference graph built in 18.69s! Perform image classification inference -------------------------------------- We run classification on an image sample from ImageNet We just need to download the categories files, `synset.txt` and an input test image. .. code-block:: default # Download ImageNet categories categ_url = "https://github.com/uwsampl/web-data/raw/main/vta/models/" categ_fn = "synset.txt" download.download(join(categ_url, categ_fn), categ_fn) synset = eval(open(categ_fn).read()) # Download test image image_url = "https://homes.cs.washington.edu/~moreau/media/vta/cat.jpg" image_fn = "cat.png" download.download(image_url, image_fn) # Prepare test image for inference image = Image.open(image_fn).resize((224, 224)) plt.imshow(image) plt.show() image = np.array(image) - np.array([123.0, 117.0, 104.0]) image /= np.array([58.395, 57.12, 57.375]) image = image.transpose((2, 0, 1)) image = image[np.newaxis, :] image = np.repeat(image, env.BATCH, axis=0) # Set the network parameters and inputs m.set_input(**params) m.set_input("data", image) # Perform inference and gather execution statistics # More on: :py:method:`tvm.runtime.Module.time_evaluator` num = 4 # number of times we run module for a single measurement rep = 3 # number of measurements (we derive std dev from this) timer = m.module.time_evaluator("run", ctx, number=num, repeat=rep) if env.TARGET in ["sim", "tsim"]: simulator.clear_stats() timer() sim_stats = simulator.stats() print("\nExecution statistics:") for k, v in sim_stats.items(): # Since we execute the workload many times, we need to normalize stats # Note that there is always one warm up run # Therefore we divide the overall stats by (num * rep + 1) print("\t{:<16}: {:>16}".format(k, v // (num * rep + 1))) else: tcost = timer() std = np.std(tcost.results) * 1000 mean = tcost.mean * 1000 print("\nPerformed inference in %.2fms (std = %.2f) for %d samples" % (mean, std, env.BATCH)) print("Average per sample inference time: %.2fms" % (mean / env.BATCH)) # Get classification results tvm_output = m.get_output(0, tvm.nd.empty((env.BATCH, 1000), "float32", remote.cpu(0))) for b in range(env.BATCH): top_categories = np.argsort(tvm_output.numpy()[b]) # Report top-5 classification results print("\n{} prediction for sample {}".format(model, b)) print("\t#1:", synset[top_categories[-1]]) print("\t#2:", synset[top_categories[-2]]) print("\t#3:", synset[top_categories[-3]]) print("\t#4:", synset[top_categories[-4]]) print("\t#5:", synset[top_categories[-5]]) # This just checks that one of the 5 top categories # is one variety of cat; this is by no means an accurate # assessment of how quantization affects classification # accuracy but is meant to catch changes to the # quantization pass that would accuracy in the CI. cat_detected = False for k in top_categories[-5:]: if "cat" in synset[k]: cat_detected = True assert cat_detected .. image:: /topic/vta/tutorials/frontend/images/sphx_glr_deploy_classification_001.png :class: sphx-glr-single-img .. rst-class:: sphx-glr-script-out Out: .. code-block:: none Execution statistics: inp_load_nbytes : 5549568 wgt_load_nbytes : 12763136 acc_load_nbytes : 6051840 uop_load_nbytes : 22952 out_store_nbytes: 2433536 gemm_counter : 6623232 alu_counter : 757344 resnet18_v1 prediction for sample 0 #1: tiger cat #2: Egyptian cat #3: tabby, tabby cat #4: lynx, catamount #5: weasel .. _sphx_glr_download_topic_vta_tutorials_frontend_deploy_classification.py: .. only :: html .. container:: sphx-glr-footer :class: sphx-glr-footer-example .. container:: sphx-glr-download :download:`Download Python source code: deploy_classification.py ` .. container:: sphx-glr-download :download:`Download Jupyter notebook: deploy_classification.ipynb ` .. only:: html .. rst-class:: sphx-glr-signature `Gallery generated by Sphinx-Gallery `_