def quantize_relay_module(mod, params, qconfig=None):
""" Quantize the relay module with qconfig options.
Parameters:
------
mod : tvm.relay.module
The original module.
qconfig : tvm.relay.quantize.quantize.QConfig
The quantization configuration
Returns:
------
qfunc : vm.relay.expr.Function
The graph after quantization
"""
# default qconfig
if not qconfig:
qconfig = qtz.qconfig()
with qconfig:
logging.debug('current quantize config')
logging.debug(qtz.current_qconfig())
mod['main'] = qtz.quantize(mod['main'], params=params)
logging.debug('after quantize')
logging.debug(mod['main'].astext(show_meta_data=False))
return mod
def get_model(model_name,
batch_size,
qconfig,
target=None,
original=False,
simulated=False):
gluon_model = gluon.model_zoo.vision.get_model(model_name, pretrained=True)
img_size = 299 if model_name == 'inceptionv3' else 224
data_shape = (batch_size, 3, img_size, img_size)
mod, params = relay.frontend.from_mxnet(gluon_model, {"data": data_shape})
net = mod['main']
with tvm.transform.PassContext(opt_level=3):
qfunc = relay.quantize.prerequisite_optimize(net, params=params)
logging.debug('original')
logging.debug(qfunc.astext(show_meta_data=False))
if original:
return qfunc
with qconfig:
logging.debug('current quantize config')
logging.debug(qtz.current_qconfig())
qfunc = qtz.quantize(qfunc)
logging.debug('after quantize')
logging.debug(qfunc.astext(show_meta_data=False))
return qfunc
def get_onnx_model(model_name, batch_size, qconfig, original=False, dataset=None):
assert model_name == "vit32", "Only support vit32 model!"
base = "https://github.com/TheGreatCold/tvm-vit/raw/d2aa1e60eef42e2fdedbd1e13aa85ac5faf0a7fc"
logfile = "gtx1660_vit_B32_224.log"
onnx_path = "vit_B32_224.onnx"
download_file(base, logfile)
download_file(base, onnx_path)
onnx_graph = onnx.load(open(onnx_path, "rb"))
data_shape = (batch_size, 3, 224, 224)
mod, params = relay.frontend.from_onnx(onnx_graph, {"data": data_shape})
with tvm.transform.PassContext(opt_level=3):
qfunc = relay.quantize.prerequisite_optimize(mod, params=params)
logging.debug("original")
logging.debug(qfunc.astext(show_meta_data=False))
if original:
return qfunc, params, logfile
with qconfig:
logging.debug("current quantize config")
logging.debug(qtz.current_qconfig())
if dataset is not None:
with tvm.target.cuda():
with tvm.autotvm.apply_history_best(logfile):
qfunc = qtz.quantize(qfunc, params, dataset=dataset)
else:
qfunc = qtz.quantize(qfunc, params)
logging.debug("after quantize")
logging.debug(qfunc.astext(show_meta_data=False))
return qfunc, params, logfile
def get_model(model_name, batch_size, qconfig, target=None, original=False, simulated=False):
gluon_model = gluon.model_zoo.vision.get_model(model_name, pretrained=True)
img_size = 299 if model_name == 'inceptionv3' else 224
input_shape = (batch_size, 3, img_size, img_size)
mod, params = relay.frontend.from_mxnet(gluon_model, {"data": input_shape})
qfunc = mod['main']
start_time = time.time()
with relay.build_config(opt_level=3):
qfunc = relay.quantize.prerequisite_optimize(qfunc, params=params)
logging.debug('original')
logging.debug(qfunc.astext(show_meta_data=False))
if original:
return qfunc
with qconfig:
logging.debug('current quantize config')
logging.debug(qtz.current_qconfig())
qfunc = qtz.quantize(qfunc,params=params)
logging.debug('after quantize')
logging.debug(qfunc.astext(show_meta_data=False))
# os._exit(-1)
return qfunc, params, input_shape
def build_model(args, gluon_model):
"""Build with relay."""
import tvm
from tvm import relay
from tvm.relay import quantize as qtz
img_size = 299 if args.model == 'inceptionv3' else 224
data_shape = (args.batch_size, 3, img_size, img_size)
net, params = relay.frontend.from_mxnet(gluon_model, {"data": data_shape})
target = args.target
if args.original:
# run original model
with relay.build_config(opt_level=3):
graph, lib, params = relay.build(net, target, params=params)
ctx = tvm.nd.context(target, 0)
return graph, lib, params, ctx
# constant folding and scale folding.
print('original')
print(net.astext(show_meta_data=False))
with relay.build_config(opt_level=3):
qgraph = relay.optimize(net, target, params)
# qgraph = relay.optimize(qgraph)
print('after optimize')
print(qgraph.astext(show_meta_data=False))
with qtz.qconfig(skip_k_conv=0,
nbit_input=args.nbit_input,
nbit_weight=args.nbit_input,
global_scale=args.global_scale,
dtype_input=args.dtype_input,
dtype_weight=args.dtype_input,
dtype_activation=args.dtype_output,
store_lowbit_output=False,
debug_enabled_ops=None):
print(qtz.current_qconfig())
qgraph = qtz.annotate(qgraph)
print('after annotate')
print(qgraph.astext(show_meta_data=False))
qgraph = qtz.calibrate(qgraph)
print('after calibrate\n')
print(qgraph.astext(show_meta_data=False))
if not args.simulated:
qgraph = qtz.realize(qgraph)
qgraph = relay.ir_pass.infer_type(qgraph)
print('after realize\n')
print(qgraph.astext(show_meta_data=False))
with relay.build_config(opt_level=3):
graph, lib, params = relay.build(qgraph, target)
ctx = tvm.nd.context(target, 0)
return graph, lib, params, ctx
def get_model(model_name, batch_size, qconfig, original=False):
gluon_model = gluon.model_zoo.vision.get_model(model_name, pretrained=True)
img_size = 299 if model_name == "inceptionv3" else 224
data_shape = (batch_size, 3, img_size, img_size)
mod, params = relay.frontend.from_mxnet(gluon_model, {"data": data_shape})
logging.debug("original")
logging.debug(mod.astext(show_meta_data=False))
if original:
return mod, params
with qconfig:
logging.debug("current quantize config")
logging.debug(qtz.current_qconfig())
qfunc = qtz.quantize(mod, params)
logging.debug("after quantize")
logging.debug(qfunc.astext(show_meta_data=False))
return qfunc, params
def quantize_model(args):
"""Build with relay."""
import tvm
from tvm import relay
from tvm.relay import quantize as qtz
img_size = 224
data_shape = (args.batch_size, 3, img_size, img_size)
mx_sym, mx_args, mx_auxs = mx.model.load_checkpoint(args.model, 0)
net, params = relay.frontend.from_mxnet(mx_sym, {"data": data_shape},
arg_params=mx_args,
aux_params=mx_auxs)
target = args.target
if args.original:
# run original model
with relay.build_config(opt_level=3):
graph, lib, params = relay.build(net, target, params=params)
ctx = tvm.nd.context(target, 0)
return graph, lib, params, ctx
# constant folding and scale folding.
# print('original')
# print(net.astext(show_meta_data=False))
with relay.build_config(opt_level=3):
qgraph = relay.optimize(net, target, params)
# print('after optimize')
# print(qgraph.astext(show_meta_data=False))
with qtz.qconfig(skip_k_conv=0,
nbit_input=args.nbit_input,
nbit_weight=args.nbit_input,
global_scale=args.global_scale,
dtype_input=args.dtype_input,
dtype_weight=args.dtype_input,
dtype_activation=args.dtype_output,
store_lowbit_output=False,
debug_enabled_ops=None):
print(qtz.current_qconfig())
qgraph = qtz.annotate(qgraph)
# print('after annotate')
# print(qgraph.astext(show_meta_data=False))
qgraph = qtz.calibrate(qgraph)
# print('after calibrate\n')
# print(qgraph.astext(show_meta_data=False))
if not args.simulated:
qgraph = qtz.realize(qgraph)
qgraph = relay.ir_pass.infer_type(qgraph)
# print('after realize\n')
# print(qgraph.astext(show_meta_data=False))
with relay.build_config(opt_level=3):
graph, lib, params = relay.build(qgraph, target)
### save/load the graph, lib and params into separate files
# save
lib.export_library(os.path.join(thisdir, "deploy_lib.so"))
with open(os.path.join(thisdir, "deploy_graph.json"), "w") as fo:
fo.write(graph)
with open(os.path.join(thisdir, "deploy_param.params"), "wb") as fo:
fo.write(relay.save_param_dict(params))
# load
graph = open(os.path.join(thisdir, "deploy_graph.json")).read()
lib = tvm.module.load(os.path.join(thisdir, "deploy_lib.so"))
params = bytearray(
open(os.path.join(thisdir, "deploy_param.params"), "rb").read())
ctx = tvm.nd.context(target, 0)
return graph, lib, params, ctx
