def test_sym_nnvm(batch_size, iter_num):
logger = logging.getLogger("log.test.nnvm")
logger.info("=== Log Test NNVM ===")
sym_file, param_file, ext_file = load_fname("mrt.all.quantize", True)
sym, params = mx.sym.load(sym_file), nd.load(param_file)
inputs_ext, _ = sim.load_ext(ext_file)
val_data = dataset.load_voc(1, 512)
val_data_iter = iter(val_data)
data, _ = next(val_data_iter)
if False:
data = sim.load_real_data(data, 'data', inputs_ext)
inputs_ext['data']['data'] = data
spass.sym_dump_ops(sym,
params,
inputs_ext,
datadir="/data/wlt",
ctx=mx.gpu(1),
cleanDir=True,
ops=[
"broadcast_div0",
])
else:
_mrt.std_dump(sym, params, inputs_ext, data, "ssd_ryt", max_num=100)
def test_sym_nnvm(batch_size=10):
dump_sym, dump_params, dump_ext = load_fname(version, "sym.quantize", True)
sym, params = mx.sym.load(dump_sym), nd.load(dump_params)
(inputs_ext, ) = sim.load_ext(dump_ext)
data_iter = utils.load_dataset(batch_size)
data = data_iter.next().data[0]
_mrt.std_dump(sym, params, inputs_ext, data, "resnet" + version)
def load(model_name, datadir="./data"):
# pylint: disable=unbalanced-tuple-unpacking
sym_file, params_file, ext_file = \
utils.extend_fname(path.join(datadir, model_name), True)
mrt = MRT(Model.load(sym_file, params_file))
mrt.old_names, mrt.th_dict, mrt.precs, mrt.scales = \
sim.load_ext(ext_file)
return mrt
def load_fname(prefix, suffix=None):
suffix = "." + suffix if suffix is not None else ""
load_prefix = prefix + suffix
names = list(utils.extend_fname(load_prefix, True))
names, ext_file = names[:-1], names[-1]
(inputs_ext, ) = sim.load_ext(ext_file)
dump_prefix = prefix + ".nnvm.compile"
names.extend(utils.extend_fname(dump_prefix, False))
return names, inputs_ext
def test_sym_nnvm(batch_size, iter_num):
logger = logging.getLogger("log.test.nnvm")
logger.info("=== Log Test NNVM ===")
sym_file, param_file, ext_file = load_fname("_darknet53_voc", "mrt.all.quantize", True)
sym, params = mx.sym.load(sym_file), nd.load(param_file)
inputs_ext, _ = sim.load_ext(ext_file)
nnvm_sym, nnvm_params = spass.mxnet_to_nnvm(sym, params, inputs_ext)
spass.cvm_build(nnvm_sym, nnvm_params, inputs_ext, *load_fname("_darknet53_voc", "nnvm"))
def test_sym_nnvm(batch_size, iter_num):
logger = logging.getLogger("log.test.nnvm")
logger.info("=== Log Test NNVM ===")
sym_file, param_file, ext_file = load_fname("_darknet53_voc",
"all.quantize", True)
dump_sym, dump_params = load_fname("_darknet53_voc", "all.nnvm.compile")
sym, params = mx.sym.load(sym_file), nd.load(param_file)
inputs_ext, _ = sim.load_ext(ext_file)
spass.mxnet_to_nnvm(sym, params, inputs_ext, dump_sym, dump_params)
def test_nnvm_pass(iter_num=10):
logger = logging.getLogger("log.test.nnvm")
logger.info("=== Log Test NNVM ===")
dump_sym, dump_params, dump_ext = load_fname(version, "sym.quantize", True)
sym, params = mx.sym.load(dump_sym), nd.load(dump_params)
(inputs_ext,) = sim.load_ext(dump_ext)
data_iter = iter(val_loader)
data, _ = next(data_iter)
_mrt.std_dump(sym, params, inputs_ext, data, "cvm_mnist")
def test_sym_nnvm():
logger = logging.getLogger("log.test.nnvm")
logger.info("=== Log Test NNVM ===")
dump_sym, dump_params, dump_ext = load_fname(version, "sym.quantize", True)
sym, params = mx.sym.load(dump_sym), nd.load(dump_params)
(inputs_ext,) = sim.load_ext(dump_ext)
data_iter = ds.load_imagenet_rec(1)
data = data_iter.next().data[0]
_mrt.std_dump(sym, params, inputs_ext, data, "mobilenet"+version)
def test_sym_nnvm(batch_size=10, iter_num=10):
logger = logging.getLogger("log.test.nnvm")
logger.info("=== Log Test NNVM ===")
dump_sym, dump_params, dump_ext = load_fname(version, "sym.quantize", True)
sym, params = mx.sym.load(dump_sym), nd.load(dump_params)
(inputs_ext, ) = sim.load_ext(dump_ext)
data_iter = ds.load_imagenet_rec(batch_size, 224)
data = data_iter.next().data[0]
_mrt.std_dump(sym, params, inputs_ext, data, "shufflenet", max_num=100)
def test_sym_pass(iter_num=10):
inputs_ext = { 'data': {
'shape': (batch_size, 1, 28, 28),
} }
inputs = [mx.sym.var(n) for n in inputs_ext]
data_iter = iter(val_loader)
def data_iter_func():
return next(data_iter)
data, _ = data_iter_func()
net1 = utils.load_model(*load_fname(version), inputs, ctx=ctx)
def graph_func(data):
return net1.forward(data.as_in_context(ctx))
sym_file, param_file = load_fname(version)
sym, params = mx.sym.load(sym_file), nd.load(param_file)
sym, params = spass.sym_quant_prepare(sym, params, inputs_ext)
if True:
mrt = _mrt.MRT(sym, params, inputs_ext)
mrt.set_data('data', data)
mrt.calibrate(ctx=ctx)
mrt.set_output_prec(8)
qsym, qparams, inputs_ext = mrt.quantize()
else:
inputs_ext['data']['data'] = data
th_dict = calib.sym_calibrate(sym, params, inputs_ext, ctx=ctx)
qsym, qparams, precs, _ = calib.sym_simulate(sym, params, inputs_ext, th_dict)
qsym, qparams = calib.sym_realize(qsym, qparams, inputs_ext, precs, "cvm")
dump_sym, dump_params, dump_ext = load_fname(version, "sym.quantize", True)
sim.save_ext(dump_ext, inputs_ext)
nd.save(dump_params, qparams)
open(dump_sym, "w").write(qsym.tojson())
dump_sym, dump_params, dump_ext = load_fname(version, "sym.quantize", True)
(inputs_ext,) = sim.load_ext(dump_ext)
inputs = [mx.sym.var(n) for n in inputs_ext]
net2 = utils.load_model(dump_sym, dump_params, inputs, ctx=ctx)
def cvm_quantize(data):
data = sim.load_real_data(data, 'data', inputs_ext)
return net2.forward(data.as_in_context(ctx))
utils.multi_eval_accuracy(graph_func, data_iter_func,
cvm_quantize,
iter_num=iter_num)
def test_sym_nnvm(batch_size=10, iter_num=10):
logger = logging.getLogger("log.test.nnvm")
logger.info("=== Log Test NNVM ===")
version = "v3"
dump_sym, dump_params, dump_ext = load_fname(version, "mrt", True)
'''
byr---------
./data/tf_inceptionv3.mrt.json
./data/tf_inceptionv3.mrt.params
./data/tf_inceptionv3.mrt.ext
byr--------
'''
sym, params = mx.sym.load(dump_sym), nd.load(dump_params)
(inputs_ext, ) = sim.load_ext(dump_ext)
data_iter = ds.load_imagenet_rec(batch_size, 299)
data = data_iter.next().data[0]
_mrt.std_dump(sym, params, inputs_ext, data, "inception_v3")
def test_sym_nnvm(batch_size=10, iter_num=10):
logger = logging.getLogger("log.test.nnvm")
logger.info("=== Log Test NNVM ===")
target = "llvm"
tvm_ctx = tvm.context(target, 1)
mx_ctx = mx.gpu(2)
inputs_ext = {
'data': {
'shape': (batch_size, 3, 224, 224),
}
}
inputs = [mx.sym.var(name) for name in inputs_ext]
inputs_shape = {k: v['shape'] for k, v in inputs_ext.items()}
data_iter = load_dataset(batch_size)
def data_iter_func():
data = data_iter.next()
return data.data[0], data.label[0]
data_iter_func()
version = ""
dump_sym, dump_params, dump_ext = load_fname(version, "sym.quantize", True)
(inputs_ext, ) = sim.load_ext(dump_ext)
sym, params = mx.sym.load(dump_sym), nd.load(dump_params)
# sim.load_ins_ext(params, inputs_ext)
# nnvm_sym, _ = nnvm.frontend.from_mxnet(sym)
# with open('debug_nnvm_sym_after_load_from_mxnet.json', 'w') as fout:
# fout.write(nnvm_sym.debug_str())
dump_sym, dump_params = load_fname(version, "nnvm.compile", False)
spass.mxnet_to_nnvm(sym,
params,
inputs_ext,
dump_sym,
dump_params,
target='llvm')
def run_mx(modelname,
batch_size=160,
quantized=False,
mx_check_point=None,
float_dtype="float32",
evaluate=False):
input_size = model_input_size[modelname]
suffix = ".mrt.quantize" if quantized else ""
symbol_file = path.join(mx_dir, modelname + suffix + ".json")
params_file = path.join(mx_dir, modelname + suffix + ".params")
symbol = mx.sym.load(symbol_file)
params = mx.nd.load(params_file)
params = tpass.convert_params_dtype(params, dest_dtype="float32")
if mx_check_point is None:
mx_check_point = sutils.topo_sort(symbol)[-1].attr('name')
mx_data, mx_label = load_data_3(modelname,
batch_size=batch_size,
input_size=input_size,
layout='NCHW',
quantized=quantized)
if quantized:
ext_file = path.join(mx_dir, modelname + suffix + ".ext")
_, _, _, scales = sim.load_ext(ext_file)
else:
op_names = tpass.collect_op_names(symbol, params)
print(op_names)
scales = None
mx_outs = get_mxnet_outs(symbol, params, mx_data, mx_check_point)
if scales and mx_check_point in scales:
scale_factor = scales[mx_check_point]
print('multiply scale factor: ', scale_factor)
mx_outs /= scale_factor
print('mxnet check with shape: ', mx_outs.shape)
if evaluate:
get_metric(mx_outs, mx_label)
return mx_outs
def test_sym_pass(batch_size=10, iter_num=10, quantize=True):
logger = logging.getLogger("log.test.sym.pass")
calib_ctx = mx.gpu(1)
ctx = [mx.gpu(int(i)) for i in "1,2,3,4".split(',') if i.strip()]
inputs_ext = {
'data': {
'shape': (batch_size, 3, 224, 224),
}
}
inputs = [mx.sym.var(name) for name in inputs_ext]
logger.info("load dataset, symbol and parameters")
# load dataset and iter function
data_iter = ds.load_imagenet_rec(batch_size)
def data_iter_func():
data = data_iter.next()
return data.data[0], data.label[0]
data, _ = data_iter_func()
# load original model for accuracy
net1 = utils.load_model(*load_fname(version), inputs, ctx=ctx)
acc_top1 = mx.metric.Accuracy()
acc_top5 = mx.metric.TopKAccuracy(5)
acc_top1.reset()
acc_top5.reset()
def shufflenet(data, label):
data = gluon.utils.split_and_load(data,
ctx_list=ctx,
batch_axis=0,
even_split=False)
res = [net1.forward(d) for d in data]
res = nd.concatenate(res)
acc_top1.update(label, res)
_, top1 = acc_top1.get()
acc_top5.update(label, res)
_, top5 = acc_top5.get()
return "top1={:6.2%} top5={:6.2%}".format(top1, top5)
if quantize:
# load original model
sym_fname, param_fname = load_fname(version)
sym, params = mx.sym.load(sym_fname), nd.load(param_fname)
sym, params = spass.sym_quant_prepare(sym, params, inputs_ext)
# quantize process
mrt = _mrt.MRT(sym, params, inputs_ext) # initialize
mrt.set_data('data', data) # set input data
mrt.calibrate(ctx=calib_ctx) # calibration
mrt.set_output_prec(8) # set output prec, do nothing by default
qsym, qparams, inputs_ext = mrt.quantize() # quantization
# dump quantized model
dump_sym, dump_params, dump_ext = load_fname(version, "sym.quantize",
True)
sim.save_ext(dump_ext, inputs_ext)
nd.save(dump_params, qparams)
open(dump_sym, "w").write(qsym.tojson())
if False:
# convert to cvm executor model
inputs_ext['data']['shape'] = (1, 3, 224, 224)
nnvm_sym, nnvm_params = spass.mxnet_to_nnvm(qsym, qparams, inputs_ext)
spass.cvm_build(nnvm_sym, nnvm_params, inputs_ext,
*load_fname(version, "nnvm"))
# load quantized model for accuracy
dump_sym, dump_params, dump_ext = load_fname(version, "sym.quantize", True)
(inputs_ext, ) = sim.load_ext(dump_ext)
inputs = [mx.sym.var(n) for n in inputs_ext]
net3 = utils.load_model(dump_sym, dump_params, inputs, ctx=ctx)
# net3 = mx.gluon.nn.SymbolBlock(qsym, inputs)
# utils.load_parameters(net3, qparams, ctx=ctx)
qacc_top1 = mx.metric.Accuracy()
qacc_top5 = mx.metric.TopKAccuracy(5)
qacc_top1.reset()
qacc_top5.reset()
def cvm_quantize(data, label):
data = sim.load_real_data(data, 'data', inputs_ext)
data = gluon.utils.split_and_load(data,
ctx_list=ctx,
batch_axis=0,
even_split=False)
res = [net3.forward(d) for d in data]
res = nd.concatenate(res)
qacc_top1.update(label, res)
_, top1 = qacc_top1.get()
qacc_top5.update(label, res)
_, top5 = qacc_top5.get()
return "top1={:6.2%} top5={:6.2%}".format(top1, top5)
# compare accuracy between models
utils.multi_validate(shufflenet,
data_iter_func,
cvm_quantize,
iter_num=iter_num,
logger=logger)
# zoo.save_model('yolo3_darknet53_voc')
# name = "yolo3_resnet18_v1_voc"
# net = zoo.load_resnet18_v1_yolo()
# sym = net(mx.sym.var('data'))
# if isinstance(sym, tuple):
# sym = mx.sym.Group([*sym])
# open("./data/%s.json"%name, "w").write(sym.tojson())
# exit()
if False:
val_data = dataset.load_voc(1, 416)
sym_file, param_file, ext_file = load_fname("_darknet53_voc",
"all.quantize", True)
sym, params = mx.sym.load(sym_file), nd.load(param_file)
inputs_ext, _ = sim.load_ext(ext_file)
if False:
for data, _ in val_data:
data = sim.load_real_data(data, 'data', inputs_ext)
inputs_ext['data']['data'] = data
spass.sym_dump_ops(sym,
params,
inputs_ext,
datadir="/data/wlt",
ctx=mx.gpu(2))
else:
val_data_iter = iter(val_data)
data, _ = next(val_data_iter)
data = sim.load_real_data(data, 'data', inputs_ext)
inputs_ext['data']['data'] = data
spass.sym_dump_layer_outputs(sym,
def test_mrt_quant(batch_size=1, iter_num=10, from_scratch=0):
logger = logging.getLogger("log.test.mrt.quantize")
flag = [False]*from_scratch + [True]*(4-from_scratch)
ctx = mx.gpu(4)
qctx = mx.gpu(3)
input_size = 416
input_shape = (batch_size, 3, input_size, input_size)
# define data iter function, get:
# get_iter_func
val_data = dataset.load_voc(batch_size, input_size)
val_data_iter = iter(val_data)
def data_iter_func():
data, label = next(val_data_iter)
return data, label
# split model, get:
# base, base_params, top, top_params, top_inputs_ext
base, base_params, top, top_params, top_inputs_ext = \
None, None, None, None, None
if flag[0]:
sym_file, param_file = load_fname("_darknet53_voc")
sym, params = mx.sym.load(sym_file), nd.load(param_file)
mrt = MRT(sym, params, input_shape)
keys = [
'yolov30_yolooutputv30_expand_dims0',
'yolov30_yolooutputv31_expand_dims0',
'yolov30_yolooutputv32_expand_dims0',
'yolov30_yolooutputv30_tile0',
'yolov30_yolooutputv31_tile0',
'yolov30_yolooutputv32_tile0',
'yolov30_yolooutputv30_broadcast_add1',
'yolov30_yolooutputv31_broadcast_add1',
'yolov30_yolooutputv32_broadcast_add1',
]
base, base_params, top, top_params, top_inputs_ext \
= split_model(mrt.csym, mrt.cprm, {'data': input_shape}, keys)
dump_sym, dump_params = load_fname("_darknet53_voc", "mrt.base")
open(dump_sym, "w").write(base.tojson())
nd.save(dump_params, base_params)
dump_sym, dump_params, dump_ext = \
load_fname("_darknet53_voc", "mrt.top", True)
open(dump_sym, "w").write(top.tojson())
nd.save(dump_params, top_params)
sim.save_ext(dump_ext, top_inputs_ext)
else:
dump_sym, dump_params = load_fname("_darknet53_voc", "mrt.base")
base, base_params = mx.sym.load(dump_sym), nd.load(dump_params)
dump_sym, dump_params, dump_ext = \
load_fname("_darknet53_voc", "mrt.top", True)
top, top_params = mx.sym.load(dump_sym), nd.load(dump_params)
(top_inputs_ext,) = sim.load_ext(dump_ext)
base_graph = mx.gluon.nn.SymbolBlock(base, [mx.sym.var('data')])
nbase_params = convert_params_dtype(base_params, src_dtypes="float64",
dest_dtype="float32")
utils.load_parameters(base_graph, nbase_params, ctx=ctx)
top_graph = mx.gluon.nn.SymbolBlock(top,
[mx.sym.var(n) for n in top_inputs_ext])
ntop_params = convert_params_dtype(top_params, src_dtypes="float64",
dest_dtype="float32")
utils.load_parameters(top_graph, ntop_params, ctx=ctx)
# calibrate split model, get:
# th_dict
th_dict = None
if flag[1]:
mrt = MRT(base, base_params, input_shape)
for i in range(1):
data, _ = data_iter_func()
mrt.set_data(data)
mrt.calibrate(ctx=ctx)
_, _, dump_ext = load_fname("_darknet53_voc", "mrt.dict", True)
th_dict = mrt.th_dict
sim.save_ext(dump_ext, th_dict)
else:
_, _, dump_ext = load_fname("_darknet53_voc", "mrt.dict", True)
(th_dict,) = sim.load_ext(dump_ext)
# quantize split model, get:
# qbase, qbase_params, qbase_inputs_ext, oscales, maps
qbase, qbase_params, qbase_inputs_ext, oscales, maps = \
None, None, None, None, None
if flag[2]:
mrt = MRT(base, base_params, input_shape)
mrt.set_th_dict(th_dict)
mrt.set_threshold('data', 2.64)
mrt.set_threshold('yolov30_yolooutputv30_expand_dims0', 1)
mrt.set_threshold('yolov30_yolooutputv31_expand_dims0', 1)
mrt.set_threshold('yolov30_yolooutputv32_expand_dims0', 1)
mrt.set_threshold('yolov30_yolooutputv30_tile0', 416)
mrt.set_threshold('yolov30_yolooutputv31_tile0', 416)
mrt.set_threshold('yolov30_yolooutputv32_tile0', 416)
mrt.set_output_prec(30)
qbase, qbase_params, qbase_inputs_ext = mrt.quantize()
oscales = mrt.get_output_scales()
maps = mrt.get_maps()
dump_sym, dump_params, dump_ext = load_fname("_darknet53_voc", "mrt.quantize", True)
open(dump_sym, "w").write(qbase.tojson())
nd.save(dump_params, qbase_params)
sim.save_ext(dump_ext, qbase_inputs_ext, oscales, maps)
else:
qb_sym, qb_params, qb_ext = load_fname("_darknet53_voc", "mrt.quantize", True)
qbase, qbase_params = mx.sym.load(qb_sym), nd.load(qb_params)
qbase_inputs_ext, oscales, maps = sim.load_ext(qb_ext)
# merge quantized split model, get:
# qsym, qparams, oscales2
qsym, qparams = None, None
if flag[3]:
def box_nms(node, params, graph):
name, op_name = node.attr('name'), node.attr('op_name')
childs, attr = sutils.sym_iter(node.get_children()), node.list_attr()
if op_name == '_contrib_box_nms':
valid_thresh = sutils.get_attr(attr, 'valid_thresh', 0)
attr['valid_thresh'] = int(valid_thresh * oscales[3])
node = sutils.get_mxnet_op(op_name)(*childs, **attr, name=name)
return node
qsym, qparams = merge_model(qbase, qbase_params,
top, top_params, maps, box_nms)
oscales2 = [oscales[1], oscales[0], oscales[2]]
sym_file, param_file, ext_file = \
load_fname("_darknet53_voc", "mrt.all.quantize", True)
open(sym_file, "w").write(qsym.tojson())
nd.save(param_file, qparams)
sim.save_ext(ext_file, qbase_inputs_ext, oscales2)
else:
dump_sym, dump_params, dump_ext = \
load_fname("_darknet53_voc", "mrt.all.quantize", True)
qsym, qparams = mx.sym.load(dump_sym), nd.load(dump_params)
_, oscales2 = sim.load_ext(dump_ext)
if False:
compile_to_cvm(qsym, qparams, "yolo_tfm",
datadir="/data/ryt", input_shape=(1, 3, 416, 416))
exit()
metric = dataset.load_voc_metric()
metric.reset()
def yolov3(data, label):
def net(data):
tmp = base_graph(data.as_in_context(ctx))
outs = top_graph(*tmp)
return outs
acc = validate_data(net, data, label, metric)
return "{:6.2%}".format(acc)
net2 = mx.gluon.nn.SymbolBlock(qsym,
[mx.sym.var(n) for n in qbase_inputs_ext])
utils.load_parameters(net2, qparams, ctx=qctx)
net2_metric = dataset.load_voc_metric()
net2_metric.reset()
def mrt_quantize(data, label):
def net(data):
data = sim.load_real_data(data, 'data', qbase_inputs_ext)
outs = net2(data.astype("float64").as_in_context(qctx))
outs = [o.as_in_context(ctx) / oscales2[i] \
for i, o in enumerate(outs)]
return outs
acc = validate_data(net, data, label, net2_metric)
return "{:6.2%}".format(acc)
utils.multi_validate(yolov3, data_iter_func,
mrt_quantize,
iter_num=iter_num, logger=logger)
def test_mrt_quant(batch_size=1, iter_num=10):
logger = logging.getLogger("log.test.mrt.quantize")
base_ctx = mx.gpu(1)
ctx = mx.gpu(2)
qctx = mx.gpu(3)
input_size = 416
h, w = input_size, input_size
inputs_ext = {
'data': {
'shape': (batch_size, 3, h, w),
}
}
val_data = dataset.load_voc(batch_size, input_size)
val_data_iter = iter(val_data)
def data_iter_func():
data, label = next(val_data_iter)
return data, label
if False:
sym_file, param_file = load_fname("_darknet53_voc")
sym, params = mx.sym.load(sym_file), nd.load(param_file)
sym, params = spass.sym_quant_prepare(sym, params, inputs_ext)
keys = [
'yolov30_yolooutputv30_expand_dims0',
'yolov30_yolooutputv31_expand_dims0',
'yolov30_yolooutputv32_expand_dims0',
'yolov30_yolooutputv30_tile0',
'yolov30_yolooutputv31_tile0',
'yolov30_yolooutputv32_tile0',
'yolov30_yolooutputv30_broadcast_add1',
'yolov30_yolooutputv31_broadcast_add1',
'yolov30_yolooutputv32_broadcast_add1',
]
base, base_params, base_inputs_ext, top, top_params, top_inputs_ext \
= split_model(sym, params, inputs_ext, keys, logger)
dump_sym, dump_params = load_fname("_darknet53_voc", "mrt.base")
open(dump_sym, "w").write(base.tojson())
nd.save(dump_params, base_params)
dump_sym, dump_params, dump_ext = load_fname("_darknet53_voc",
"mrt.top", True)
open(dump_sym, "w").write(top.tojson())
nd.save(dump_params, top_params)
sim.save_ext(dump_ext, top_inputs_ext)
dump_sym, dump_params = load_fname("_darknet53_voc", "mrt.base")
base, base_params = mx.sym.load(dump_sym), nd.load(dump_params)
dump_sym, dump_params, dump_ext = load_fname("_darknet53_voc", "mrt.top",
True)
top, top_params = mx.sym.load(dump_sym), nd.load(dump_params)
(top_inputs_ext, ) = sim.load_ext(dump_ext)
base_inputs = [mx.sym.var(n) for n in inputs_ext]
base_graph = mx.gluon.nn.SymbolBlock(base, base_inputs)
utils.load_parameters(base_graph, base_params, ctx=ctx)
top_inputs = [mx.sym.var(n) for n in top_inputs_ext]
top_graph = mx.gluon.nn.SymbolBlock(top, top_inputs)
utils.load_parameters(top_graph, top_params, ctx=ctx)
metric = dataset.load_voc_metric()
metric.reset()
def yolov3(data, label):
def net(data):
tmp = base_graph(data.as_in_context(ctx))
outs = top_graph(*tmp)
# print ([o[0][0][:] for o in outs])
return outs
acc = validate_data(net, data, label, metric)
return "{:6.2%}".format(acc)
if False:
mrt = _mrt.MRT(base, base_params, inputs_ext)
for i in range(16):
data, _ = data_iter_func()
mrt.set_data('data', data)
th_dict = mrt.calibrate(ctx=ctx)
_, _, dump_ext = load_fname("_darknet53_voc", "mrt.dict", True)
sim.save_ext(dump_ext, th_dict)
_, _, dump_ext = load_fname("_darknet53_voc", "mrt.dict", True)
(th_dict, ) = sim.load_ext(dump_ext)
if True:
mrt = _mrt.MRT(base, base_params, base_inputs_ext)
mrt.set_th_dict(th_dict)
mrt.set_threshold('data', 2.64)
mrt.set_threshold('yolov30_yolooutputv30_expand_dims0', 1)
mrt.set_threshold('yolov30_yolooutputv31_expand_dims0', 1)
mrt.set_threshold('yolov30_yolooutputv32_expand_dims0', 1)
mrt.set_threshold('yolov30_yolooutputv30_tile0', 416)
mrt.set_threshold('yolov30_yolooutputv31_tile0', 416)
mrt.set_threshold('yolov30_yolooutputv32_tile0', 416)
# mrt.set_fixed('yolov30_yolooutputv30_broadcast_add1')
# mrt.set_fixed('yolov30_yolooutputv31_broadcast_add1')
# mrt.set_fixed('yolov30_yolooutputv32_broadcast_add1')
mrt.set_output_prec(30)
qbase, qbase_params, qbase_inputs_ext = mrt.quantize()
oscales = mrt.get_output_scales()
dump_sym, dump_params, dump_ext = load_fname("_darknet53_voc",
"mrt.quantize", True)
open(dump_sym, "w").write(qbase.tojson())
nd.save(dump_params, qbase_params)
sim.save_ext(dump_ext, qbase_inputs_ext, oscales)
if True:
dump_sym, dump_params, dump_ext = load_fname("_darknet53_voc",
"mrt.quantize", True)
net2_inputs_ext, oscales = sim.load_ext(dump_ext)
inputs = [mx.sym.var(n) for n in net2_inputs_ext]
net2 = utils.load_model(dump_sym, dump_params, inputs, ctx=qctx)
net2_metric = dataset.load_voc_metric()
net2_metric.reset()
def mrt_quantize(data, label):
def net(data):
data = sim.load_real_data(data, 'data', net2_inputs_ext)
outs = net2(data.as_in_context(qctx))
outs = [
o.as_in_context(ctx) / oscales[i]
for i, o in enumerate(outs)
]
# outs = b2_graph(*data)
outs = top_graph(*outs)
return outs
acc = validate_data(net, data, label, net2_metric)
return "{:6.2%}".format(acc)
utils.multi_validate(yolov3,
data_iter_func,
mrt_quantize,
iter_num=iter_num,
logger=logger)
def test_sym_pass(batch_size=10, iter_num=10):
logger = logging.getLogger("log.test.sym.pass")
base_ctx = mx.gpu(1)
ctx = mx.gpu(2)
input_size = 416
h, w = input_size, input_size
inputs_ext = {
'data': {
'shape': (batch_size, 3, h, w),
}
}
val_data = dataset.load_voc(batch_size, input_size)
val_data_iter = iter(val_data)
def data_iter_func():
data, label = next(val_data_iter)
return data, label
sym_file, param_file = load_fname("_darknet53_voc")
sym, params = mx.sym.load(sym_file), nd.load(param_file)
sym, params = spass.sym_quant_prepare(sym, params, inputs_ext)
if False:
th_dict = {}
for i in range(16):
data, _ = data_iter_func()
for k, v in inputs_ext.items():
v['data'] = data
th_dict = calib.sym_calibrate(sym,
params,
inputs_ext,
old_ths=th_dict,
ctx=ctx)
_, _, dump_ext = load_fname("_darknet53_voc", "dict", True)
sim.save_ext(dump_ext, th_dict)
_, _, dump_ext = load_fname("_darknet53_voc", "dict", True)
(th_dict, ) = sim.load_ext(dump_ext)
inputs = [mx.sym.var(name) for name in inputs_ext]
net1 = mx.gluon.nn.SymbolBlock(sym, inputs)
utils.load_parameters(net1, params, ctx=ctx)
metric = dataset.load_voc_metric()
metric.reset()
def yolov3(data, label):
def net(data):
out = net1(data.as_in_context(ctx))
print([o[0][0][:] for o in out])
return out
acc = validate_data(net, data, label, metric)
return "{:6.2%}".format(acc)
keys = [
'yolov30_yolooutputv30_conv0_fwd',
'yolov30_yolooutputv31_conv0_fwd',
'yolov30_yolooutputv32_conv0_fwd',
]
base, base_params, base_inputs_ext, top, top_params, top_inputs_ext \
= split_model(sym, params, inputs_ext, keys, logger)
dump_sym, dump_params = load_fname("_darknet53_voc", "base")
open(dump_sym, "w").write(base.tojson())
dump_sym, dump_params, dump_ext = load_fname("_darknet53_voc", "top", True)
open(dump_sym, "w").write(top.tojson())
nd.save(dump_params, top_params)
sim.save_ext(dump_ext, top_inputs_ext)
base_inputs = [mx.sym.var(n) for n in base_inputs_ext]
base_graph = mx.gluon.nn.SymbolBlock(base, base_inputs)
utils.load_parameters(base_graph, base_params, ctx=base_ctx)
top_inputs = [mx.sym.var(n) for n in top_inputs_ext]
top_graph = mx.gluon.nn.SymbolBlock(top, top_inputs)
utils.load_parameters(top_graph, top_params, ctx=ctx)
# quantize base graph
if False:
qbase, qbase_params, qbase_prec, base_oscales = calib.sym_simulate(
base, base_params, base_inputs_ext, th_dict)
qbase, qbase_params = calib.sym_realize(qbase, qbase_params,
base_inputs_ext, qbase_prec)
dump_sym, dump_params, dump_ext = load_fname("_darknet53_voc",
"base.quantize", True)
open(dump_sym, "w").write(qbase.tojson())
sim.save_ext(dump_ext, base_inputs_ext, base_oscales)
nd.save(dump_params, qbase_params)
if False:
qb_sym, qb_params, qb_ext = load_fname("_darknet53_voc",
"base.quantize", True)
net2_inputs_ext, base_oscales = sim.load_ext(qb_ext)
net2_inputs = [mx.sym.var(n) for n in net2_inputs_ext]
net2 = utils.load_model(qb_sym, qb_params, net2_inputs, ctx=ctx)
base_metric = dataset.load_voc_metric()
base_metric.reset()
def base_quantize(data, label):
def net(data):
data = sim.load_real_data(data, 'data', net2_inputs_ext)
tmp = list(net2(data.as_in_context(ctx)))
tmp = [t / base_oscales[i] for i, t in enumerate(tmp)]
return top_graph(*tmp)
acc = validate_data(net, data, label, base_metric)
return "{:6.2%}".format(acc)
# quantize top graph
if False:
in_bit, out_bit = 8, 30
outputs_ext = {
'yolov30_yolooutputv30_expand_dims0': {
'threshold': 1,
'type': 'score'
},
'yolov30_yolooutputv31_expand_dims0': {
'threshold': 1,
'type': 'score'
},
'yolov30_yolooutputv32_expand_dims0': {
'threshold': 1,
'type': 'score'
},
'yolov30_yolooutputv30_tile0': {
'threshold': 416,
'type': 'bbox'
},
'yolov30_yolooutputv31_tile0': {
'threshold': 416,
'type': 'bbox'
},
'yolov30_yolooutputv32_tile0': {
'threshold': 416,
'type': 'bbox'
},
'yolov30_yolooutputv30_broadcast_add1': {
'fixed': True,
'type': 'ids'
},
'yolov30_yolooutputv31_broadcast_add1': {
'fixed': True,
'type': 'ids'
},
'yolov30_yolooutputv32_broadcast_add1': {
'fixed': True,
'type': 'ids'
},
}
qsym, qparams, type_ext = anno.mixed_precision(top,
top_params,
top_inputs_ext,
th_dict,
in_bit=in_bit,
out_bit=out_bit,
out_ext=outputs_ext,
runtime="cvm")
out_scales = [type_ext['ids'], type_ext['score'], type_ext['bbox']]
dump_sym, dump_params, dump_ext = load_fname("_darknet53_voc",
"top.quantize", True)
open(dump_sym, "w").write(qsym.tojson())
sim.save_ext(dump_ext, top_inputs_ext, out_scales)
nd.save(dump_params, qparams)
if True:
sym_file, param_file, ext_file = load_fname("_darknet53_voc",
"top.quantize", True)
net3_inputs_ext, net3_scales = sim.load_ext(ext_file)
top_sym = base_graph(mx.sym.Group(base_inputs))
top_names = [c.attr('name') for c in top_sym]
net3_inputs = [mx.sym.var(n) for n in net3_inputs_ext]
net3 = utils.load_model(sym_file, param_file, net3_inputs, ctx=ctx)
top_qmetric = dataset.load_voc_metric()
top_qmetric.reset()
def top_quantize(data, label):
def net(data):
tmp = base_graph(data.as_in_context(base_ctx))
tmp = [t.as_in_context(ctx) for t in tmp]
tmp = [
sim.load_real_data(tmp[i], n, net3_inputs_ext)
for i, n in enumerate(top_names)
]
out = net3(*tmp)
out = [(t / net3_scales[i]) for i, t in enumerate(out)]
print([o[0][0][:] for o in out])
return out
acc = validate_data(net, data, label, top_qmetric)
return "{:6.2%}".format(acc)
# merge quantize model
if False:
qb_sym, qb_params, qb_ext = load_fname("_darknet53_voc",
"base.quantize", True)
qbase, qbase_params = mx.sym.load(qb_sym), nd.load(qb_params)
qbase_inputs_ext, _ = sim.load_ext(qb_ext)
qt_sym, qt_params, qt_ext = load_fname("_darknet53_voc",
"top.quantize", True)
qtop, qtop_params = mx.sym.load(qt_sym), nd.load(qt_params)
_, out_scales = sim.load_ext(qt_ext)
maps = dict(
zip([c.attr('name') for c in qbase],
[c.attr('name') for c in base]))
qsym, qparams = merge_model(qbase, qbase_params, qbase_inputs_ext,
qtop, qtop_params, maps)
sym_file, param_file, ext_file = load_fname("_darknet53_voc",
"all.quantize", True)
open(sym_file, "w").write(qsym.tojson())
nd.save(param_file, qparams)
sim.save_ext(ext_file, qbase_inputs_ext, out_scales)
if False:
sym_file, param_file, ext_file = load_fname("_darknet53_voc",
"all.quantize", True)
net4_inputs_ext, net4_scales = sim.load_ext(ext_file)
net4_inputs = [mx.sym.var(n) for n in net4_inputs_ext]
net4 = utils.load_model(sym_file, param_file, net4_inputs, ctx=ctx)
all_qmetric = dataset.load_voc_metric()
all_qmetric.reset()
def all_quantize(data, label):
def net(data):
data = sim.load_real_data(data, 'data', net4_inputs_ext)
out = net4(data.as_in_context(ctx))
out = [(t / net4_scales[i]) for i, t in enumerate(out)]
return out
acc = validate_data(net, data, label, all_qmetric)
return "{:6.2%}".format(acc)
if False:
sym_file, param_file, ext_file = load_fname("_darknet53_voc",
"all.quantize", True)
net4_inputs_ext, net4_scales = sim.load_ext(ext_file)
datadir = "/data/voc/data/"
for i in range(50):
countdir = datadir + "/" + str(i)
os.makedirs(countdir, exist_ok=True)
data, label = data_iter_func()
data = sim.load_real_data(data, 'data', net4_inputs_ext)
np.save(countdir + "/data.npy", data.asnumpy().astype('int8'))
np.save(countdir + "/label.npy", label.asnumpy())
# data = sim.load_real_data(data, 'data', net4_inputs_ext)
# np.save("/tmp/yolo/data", data.asnumpy().astype('int8'))
# out = net4(data.as_in_context(ctx))
# for i, o in enumerate(out):
# np.save("/tmp/yolo/result"+str(i), o.asnumpy().astype('int32'))
exit()
utils.multi_validate(
yolov3,
data_iter_func,
top_quantize,
# base_quantize, # top_quantize, all_quantize,
iter_num=iter_num,
logger=logger)
def test_sym_pass(batch_size=10, iter_num=10):
logger = logging.getLogger("log.test.sym.pass")
calib_ctx = mx.gpu(0)
ctx = [mx.gpu(int(i)) for i in "1,2,3,4".split(',') if i.strip()]
inputs_ext = { 'data': {
'shape': (batch_size, 3, 224, 224),
} }
inputs = [mx.sym.var(name) for name in inputs_ext]
logger.info("load dataset, symbol and parameters")
data_iter = ds.load_imagenet_rec(batch_size)
def data_iter_func():
data = data_iter.next()
return data.data[0], data.label[0]
for i in range(10):
if i == 3:
break
data, _ = data_iter_func()
data_iter.reset()
version = "19"
net1 = utils.load_model(*load_fname(version), inputs, ctx=ctx)
acc_top1 = mx.metric.Accuracy()
acc_top5 = mx.metric.TopKAccuracy(5)
acc_top1.reset()
acc_top5.reset()
def vgg(data, label):
data = gluon.utils.split_and_load(data, ctx_list=ctx, batch_axis=0, even_split=False)
res = [net1.forward(d) for d in data]
res = nd.concatenate(res)
acc_top1.update(label, res)
_, top1 = acc_top1.get()
acc_top5.update(label, res)
_, top5 = acc_top5.get()
return "top1={:6.2%} top5={:6.2%}".format(top1, top5)
sym_fname, param_fname = load_fname(version)
print(sym_fname, param_fname)
exit()
sym, params = mx.sym.load(sym_fname), nd.load(param_fname)
sym, params = spass.sym_quant_prepare(sym, params, inputs_ext)
if True:
mrt = _mrt.MRT(sym, params, inputs_ext)
mrt.set_data('data', data)
mrt.calibrate(ctx=calib_ctx)
mrt.set_output_prec(8)
qsym, qparams, inputs_ext = mrt.quantize()
dump_sym, dump_params, dump_ext = load_fname(version, "sym.quantize", True)
sim.save_ext(dump_ext, inputs_ext)
nd.save(dump_params, qparams)
open(dump_sym, "w").write(qsym.tojson())
dump_sym, dump_params, dump_ext = load_fname(version, "sym.quantize", True)
(inputs_ext,) = sim.load_ext(dump_ext)
net3 = utils.load_model(dump_sym, dump_params, inputs, ctx=ctx)
qacc_top1 = mx.metric.Accuracy()
qacc_top5 = mx.metric.TopKAccuracy(5)
qacc_top1.reset()
qacc_top5.reset()
def cvm_quantize(data, label):
data = sim.load_real_data(data, 'data', inputs_ext)
data = gluon.utils.split_and_load(data, ctx_list=ctx, batch_axis=0, even_split=False)
res = [net3.forward(d) for d in data]
res = nd.concatenate(res)
qacc_top1.update(label, res)
_, top1 = qacc_top1.get()
qacc_top5.update(label, res)
_, top5 = qacc_top5.get()
return "top1={:6.2%} top5={:6.2%}".format(top1, top5)
utils.multi_validate(vgg, data_iter_func,
cvm_quantize,
iter_num=iter_num, logger=logger)
def test_mx_quantize(batch_size=10, iter_num=10):
logger = logging.getLogger("log.test.mx.quantize")
ctx = [mx.gpu(int(i)) for i in "1,3".split(',') if i.strip()]
inputs_ext = { 'data': {
'shape': (batch_size, 3, 224, 224),
}}
inputs = [mx.sym.var(n) for n in inputs_ext]
data_iter = ds.load_imagenet_rec(batch_size)
def data_iter_func():
data = data_iter.next()
return data.data[0], data.label[0]
data, _ = data_iter_func()
net1 = utils.load_model(*load_fname(version), inputs, ctx=ctx)
acc_top1 = mx.metric.Accuracy()
acc_top5 = mx.metric.TopKAccuracy(5)
acc_top1.reset()
acc_top5.reset()
def mobilenet(data, label):
data = gluon.utils.split_and_load(data, ctx_list=ctx, batch_axis=0, even_split=False)
res = [net1.forward(d) for d in data]
res = nd.concatenate(res)
acc_top1.update(label, res)
_, top1 = acc_top1.get()
acc_top5.update(label, res)
_, top5 = acc_top5.get()
return "top1={:6.2%} top5={:6.2%}".format(top1, top5)
calib_ctx = mx.gpu(1)
sym_fname, param_fname = load_fname(version)
sym, params = mx.sym.load(sym_fname), nd.load(param_fname)
sym, params = spass.sym_quant_prepare(sym, params, inputs_ext)
if True:
if True:
mrt = _mrt.MRT(sym, params, inputs_ext)
mrt.set_data('data', data)
mrt.calibrate()
# [ 0.0008745864 0.03330660510427334 ] 0.6670066884888368 0.7753906
# mrt.set_threshold("mobilenet0_dense0_weight", 0.67)
# # [ -0.0036011334 0.054821780899052534 ] 1.100036751338784 1.4626989
# mrt.set_threshold("mobilenet0_conv24_batchnorm24_fwd_weight", 1.1)
# # [ 0.013243316 1.7543557133786065 ] 70.18747185088569 94.66275
# mrt.set_threshold("mobilenet0_conv23_batchnorm23_fwd_weight", 35.10)
# # [ -0.0016149869 0.05713169649243355 ] 1.1442489167675376 1.7122083
# mrt.set_threshold("mobilenet0_conv20_batchnorm20_fwd_weight", 1.144)
# # [ -0.0015804865 0.04523811489343643 ] 0.9063427844084799 1.0745146
# mrt.set_threshold("mobilenet0_conv16_batchnorm16_fwd_weight", 0.90)
# # [ 0.4315614 2.447332109723772 ] 49.37820360490254 63.959927
# mrt.set_threshold("mobilenet0_conv2_batchnorm2_fwd", 49.37)
# # [ 0.9770754 1.3392452512468611 ] 27.761980422905516 40.729546
# mrt.set_threshold("mobilenet0_relu2_fwd", 27.76)
# [ 1.0975745 1.0489919010632773 ] 22.077412493692915 23.784576
# mrt.set_threshold("mobilenet0_relu4_fwd", 22.08)
# # [ 0.9885562 2.360489403014386 ] 48.19834426651407 69.22121
# mrt.set_threshold("mobilenet0_conv5_batchnorm5_fwd", 48.2)
# # [ 0.7895588 1.0544661745870065 ] 21.878882319617176 30.95745
# mrt.set_threshold("mobilenet0_relu17_fwd", 21.88)
# # [ 0.8717863 1.0887600296120434 ] 22.646986888608513 28.265652
# mrt.set_threshold("mobilenet0_relu19_fwd", 22.65)
# # [ 0.35124516 0.6501711574631898 ] 13.354668314135012 20.770807
# mrt.set_threshold("mobilenet0_relu20_fwd", 13.35)
# # [ 0.9378179 1.110470714216975 ] 23.147232155910086 27.886068
# mrt.set_threshold("mobilenet0_relu21_fwd", 23.15)
# # [ 0.36263302 0.6352599878026505 ] 13.067832775738754 17.18809
# mrt.set_threshold("mobilenet0_relu22_fwd", 13.07)
# # [ 0.19875833 0.49999100821358816 ] 10.198578498193196 16.625143
# mrt.set_threshold("mobilenet0_relu24_fwd", 10.2)
# # [ 0.32357717 1.6308352606637138 ] 65.55698759215218 75.84912
# mrt.set_threshold("mobilenet0_conv25_batchnorm25_fwd", 32.94)
# # [ 0.36793178 1.512995992388044 ] 30.62785163096019 49.464615
# mrt.set_threshold("mobilenet0_relu26_fwd", 30.63)
# # [ 18.028658 38.61970520019531 ] 790.4227619171143 805.51886
# mrt.set_threshold("sum0", 790.423)
mrt.set_output_prec(8)
qsym, qparams, inputs_ext = mrt.quantize()
else:
inputs_ext['data']['data'] = data
th_dict = calib.sym_calibrate(sym, params, inputs_ext, ctx=calib_ctx)
qsym, qparams, precs, _ = calib.sym_simulate(sym, params, inputs_ext, th_dict)
qsym, qparams = calib.sym_realize(qsym, qparams, inputs_ext, precs)
dump_sym, dump_params, dump_ext = load_fname(version, "sym.quantize", True)
sim.save_ext(dump_ext, inputs_ext)
nd.save(dump_params, qparams)
open(dump_sym, "w").write(qsym.tojson())
dump_sym, dump_params = load_fname(version, "nnvm.compile")
nnvm_sym, nnvm_params = spass.mxnet_to_nnvm(qsym, qparams, inputs_ext)
spass.cvm_build(nnvm_sym, nnvm_params, inputs_ext, dump_sym, dump_params)
dump_sym, dump_params, dump_ext = load_fname(version, "sym.quantize", True)
(inputs_ext,) = sim.load_ext(dump_ext)
net2 = utils.load_model(dump_sym, dump_params, inputs, ctx=ctx)
qacc_top1 = mx.metric.Accuracy()
qacc_top5 = mx.metric.TopKAccuracy(5)
qacc_top1.reset()
qacc_top5.reset()
def cvm_quantize(data, label):
data = sim.load_real_data(data, 'data', inputs_ext)
data = gluon.utils.split_and_load(data, ctx_list=ctx, batch_axis=0, even_split=False)
res = [net2.forward(d) for d in data]
res = nd.concatenate(res)
qacc_top1.update(label, res)
_, top1 = qacc_top1.get()
qacc_top5.update(label, res)
_, top5 = qacc_top5.get()
return "top1={:6.2%} top5={:6.2%}".format(top1, top5)
utils.multi_validate(mobilenet, data_iter_func,
cvm_quantize,
iter_num=iter_num, logger=logger)
def validate_model(sym_path, prm_path, ctx, num_channel=3,
input_size=224, batch_size=16, iter_num=10,
ds_name='imagenet', from_scratch=0, lambd=None,
dump_model=False, input_shape=None):
from gluon_zoo import save_model
flag = [False]*from_scratch + [True]*(2-from_scratch)
model_name, _ = path.splitext(path.basename(sym_path))
model_dir = path.dirname(sym_path)
input_shape = input_shape if input_shape else \
(batch_size, num_channel, input_size, input_size)
logger = logging.getLogger("log.validate.%s"%model_name)
if not path.exists(sym_path) or not path.exists(prm_path):
save_model(model_name)
model = Model.load(sym_path, prm_path)
model.prepare(input_shape)
# model = init(model, input_shape)
print(tpass.collect_op_names(model.symbol, model.params))
data_iter_func = ds.data_iter(ds_name, batch_size, input_size=input_size)
data, _ = data_iter_func()
# prepare
mrt = model.get_mrt()
# mrt = MRT(model)
# calibrate
mrt.set_data(data)
prefix = path.join(model_dir, model_name+'.mrt.dict')
_, _, dump_ext = utils.extend_fname(prefix, True)
if flag[0]:
th_dict = mrt.calibrate(lambd=lambd)
sim.save_ext(dump_ext, th_dict)
else:
(th_dict,) = sim.load_ext(dump_ext)
mrt.set_th_dict(th_dict)
mrt.set_input_prec(8)
mrt.set_output_prec(8)
if flag[1]:
mrt.quantize()
mrt.save(model_name+".mrt.quantize", datadir=model_dir)
else:
mrt = MRT.load(model_name+".mrt.quantize", datadir=model_dir)
# dump model
if dump_model:
datadir = "/data/ryt"
model_name = model_name + "_tfm"
dump_shape = (1, num_channel, input_size, input_size)
mrt.current_model.to_cvm(
model_name, datadir=datadir, input_shape=input_shape)
data = data[0].reshape(dump_shape)
data = sim.load_real_data(
data.astype("float64"), 'data', mrt.get_inputs_ext())
np.save(datadir+"/"+model_name+"/data.npy", data.astype('int8').asnumpy())
sys.exit(0)
# validate
org_model = load_model(Model.load(sym_path, prm_path), ctx)
cvm_quantize = load_model(
mrt.current_model, ctx,
inputs_qext=mrt.get_inputs_ext())
utils.multi_validate(org_model, data_iter_func, cvm_quantize,
iter_num=iter_num,
logger=logging.getLogger('mrt.validate'))
logger.info("test %s finished.", model_name)
def test_mrt_quant(batch_size=1, iter_num=10, from_scratchi=0):
logger = logging.getLogger("log.test.mrt.quantize")
flag = [False] * from_scratch + [True] * (4 - from_scratch)
ctx = mx.gpu(1)
qctx = mx.gpu(3)
input_size = 512
input_shape = (batch_size, 3, input_size, input_size)
# define data iter function, get:
# get_iter_func
val_data = dataset.load_voc(batch_size, input_size)
val_data_iter = iter(val_data)
def data_iter_func():
data, label = next(val_data_iter)
return data, label
# split model, get:
# base, base_params, top, top_params, top_inputs_ext
base, base_params, top, top_params, top_inputs_ext = \
None, None, None, None, None
if flag[0]:
sym_file, param_file = load_fname()
sym, params = mx.sym.load(sym_file), nd.load(param_file)
# mrt = MRT(sym, params, input_shape)
sym, params = tfm.init(sym, params, input_shape)
keys = [
"ssd0_multiperclassdecoder0_zeros_like0",
# "ssd0_multiperclassdecoder0_concat0",
# "ssd0_multiperclassdecoder0__mulscalar0",
"ssd0_multiperclassdecoder0_slice_axis0",
# "ssd0_multiperclassdecoder0_zeros_like1",
"ssd0_normalizedboxcenterdecoder0_concat0",
]
base, base_params, top, top_params, top_inputs_ext \
= split_model(sym, params, {'data': input_shape}, keys)
dump_sym, dump_params = load_fname("mrt.base")
open(dump_sym, "w").write(base.tojson())
nd.save(dump_params, base_params)
dump_sym, dump_params, dump_ext = load_fname("mrt.top", True)
open(dump_sym, "w").write(top.tojson())
nd.save(dump_params, top_params)
sim.save_ext(dump_ext, top_inputs_ext)
else:
dump_sym, dump_params = load_fname("mrt.base")
base, base_params = mx.sym.load(dump_sym), nd.load(dump_params)
dump_sym, dump_params, dump_ext = load_fname("mrt.top", True)
top, top_params = mx.sym.load(dump_sym), nd.load(dump_params)
(top_inputs_ext, ) = sim.load_ext(dump_ext)
base_graph = mx.gluon.nn.SymbolBlock(base, [mx.sym.var('data')])
nbase_params = convert_params_dtype(base_params,
src_dtypes="float64",
dest_dtype="float32")
utils.load_parameters(base_graph, nbase_params, ctx=ctx)
top_graph = mx.gluon.nn.SymbolBlock(
top, [mx.sym.var(n) for n in top_inputs_ext])
ntop_params = convert_params_dtype(top_params,
src_dtypes="float64",
dest_dtype="float32")
utils.load_parameters(top_graph, ntop_params, ctx=ctx)
# calibrate split model, get:
# th_dict
th_dict = None
if flag[1]:
mrt = MRT(base, base_params, input_shape)
for i in range(1):
data, _ = data_iter_func()
mrt.set_data(data)
th_dict = mrt.calibrate(ctx=ctx)
mrt.save("mrt.dict")
else:
mrt = MRT.load("mrt.dict")
# quantize split model, get:
# qbase, qbase_params, qbase_inputs_ext, oscales, maps
qbase, qbase_params, qbase_inputs_ext, oscales, maps = \
None, None, None, None, None
if flag[2]:
# mrt = MRT(base, base_params, input_shape)
# mrt.set_th_dict(th_dict)
# mrt.set_threshold('data', 2.64)
# mrt.set_fixed("ssd0_multiperclassdecoder0_concat0")
# mrt.set_fixed("ssd0_multiperclassdecoder0__mulscalar0")
# mrt.set_fixed("ssd0_multiperclassdecoder0_zeros_like1")
mrt.set_threshold("ssd0_multiperclassdecoder0_slice_axis0", 1)
# mrt.set_threshold("ssd0_normalizedboxcenterdecoder0_concat0", 512)
mrt.set_output_prec(30)
qbase, qbase_params, qbase_inputs_ext = mrt.quantize()
oscales = mrt.get_output_scales()
maps = mrt.get_maps()
dump_sym, dump_params, dump_ext = load_fname("mrt.quantize", True)
open(dump_sym, "w").write(qbase.tojson())
nd.save(dump_params, qbase_params)
sim.save_ext(dump_ext, qbase_inputs_ext, oscales, maps)
else:
qb_sym, qb_params, qb_ext = load_fname("mrt.quantize", True)
qbase, qbase_params = mx.sym.load(qb_sym), nd.load(qb_params)
qbase_inputs_ext, oscales, maps = sim.load_ext(qb_ext)
# merge quantized split model, get:
# qsym, qparams, oscales2
qsym, qparams = None, None
if flag[3]:
name_maps = {
"ssd0_slice_axis41": "ssd0_multiperclassdecoder0_zeros_like0",
"ssd0_slice_axis42": "ssd0_multiperclassdecoder0_slice_axis0",
"ssd0_slice_axis43": "ssd0_normalizedboxcenterdecoder0_concat0",
}
oscales_dict = dict(zip([c.attr('name') for c in base], oscales))
oscales2 = [oscales_dict[name_maps[c.attr('name')]] for c in top]
def box_nms(node, params, graph):
name, op_name = node.attr('name'), node.attr('op_name')
childs, attr = sutils.sym_iter(
node.get_children()), node.list_attr()
if op_name == '_greater_scalar':
valid_thresh = sutils.get_attr(attr, 'scalar', 0)
attr['scalar'] = int(valid_thresh * oscales[1])
node = sutils.get_mxnet_op(op_name)(*childs, **attr, name=name)
elif op_name == '_contrib_box_nms':
valid_thresh = sutils.get_attr(attr, 'valid_thresh', 0)
attr['valid_thresh'] = int(valid_thresh * oscales[1])
node = sutils.get_mxnet_op(op_name)(*childs, **attr, name=name)
return node
qsym, qparams = merge_model(qbase, qbase_params, top, top_params, maps,
box_nms)
sym_file, param_file, ext_file = load_fname("mrt.all.quantize", True)
open(sym_file, "w").write(qsym.tojson())
nd.save(param_file, qparams)
sim.save_ext(ext_file, qbase_inputs_ext, oscales2)
else:
dump_sym, dump_params, dump_ext = load_fname("mrt.all.quantize", True)
qsym, qparams = mx.sym.load(dump_sym), nd.load(dump_params)
_, oscales2 = sim.load_ext(dump_ext)
if False:
dump_shape = (1, 3, input_size, input_size)
compile_to_cvm(qsym,
qparams,
"ssd_tfm",
datadir="/data/ryt",
input_shape=dump_shape)
exit()
metric = dataset.load_voc_metric()
metric.reset()
def yolov3(data, label):
def net(data):
tmp = base_graph(data.as_in_context(ctx))
outs = top_graph(*tmp)
return outs
acc = validate_data(net, data, label, metric)
return "{:6.2%}".format(acc)
net2 = mx.gluon.nn.SymbolBlock(qsym,
[mx.sym.var(n) for n in qbase_inputs_ext])
nqparams = convert_params_dtype(qparams,
src_dtypes="float64",
dest_dtype="float32")
utils.load_parameters(net2, nqparams, ctx=qctx)
net2_metric = dataset.load_voc_metric()
net2_metric.reset()
def mrt_quantize(data, label):
def net(data):
data = sim.load_real_data(data, 'data', qbase_inputs_ext)
outs = net2(data.as_in_context(qctx))
outs = [o.as_in_context(ctx) / oscales2[i] \
for i, o in enumerate(outs)]
return outs
acc = validate_data(net, data, label, net2_metric)
return "{:6.2%}".format(acc)
utils.multi_validate(yolov3,
data_iter_func,
mrt_quantize,
iter_num=iter_num,
logger=logger)
def test_sym_pass(batch_size=10, iter_num=10):
logger = logging.getLogger("log.test.sym.pass")
version = ""
sym_fname, param_fname = load_fname(version)
sym, params = mx.sym.load(sym_fname), nd.load(param_fname)
params = {k.split(':')[1]: v for k, v in params.items()}
calib_ctx = mx.gpu(2)
ctx = [mx.gpu(int(i)) for i in "1,2,3,4,5,6,7".split(',') if i.strip()]
inputs_ext = {
'data': {
'shape': (batch_size, 3, 224, 224),
}
}
inputs = [mx.sym.var(name) for name in inputs_ext]
logger.info("load dataset, symbol and parameters")
order = sutils.topo_sort(sym)
for op_head in order:
if op_head.attr('name') == 'classifier':
break
sym = op_head
net = mx.gluon.nn.SymbolBlock(sym, inputs)
load_parameters(net, params, ctx=ctx)
data_iter = ds.load_imagenet_rec(batch_size)
def data_iter_func():
data = data_iter.next()
return data.data[0], data.label[0]
for i in range(10):
if i == 3:
break
data, _ = data_iter_func()
data_iter.reset()
acc_top1 = mx.metric.Accuracy()
acc_top5 = mx.metric.TopKAccuracy(5)
acc_top1.reset()
acc_top5.reset()
def resnet(data, label):
data = gluon.utils.split_and_load(data,
ctx_list=ctx,
batch_axis=0,
even_split=False)
res = [net.forward(d) for d in data]
res = nd.concatenate(res)
acc_top1.update(label, res)
_, top1 = acc_top1.get()
acc_top5.update(label, res)
_, top5 = acc_top5.get()
return "top1={:6.2%} top5={:6.2%}".format(top1, top5)
sym, params = spass.sym_quant_prepare(sym, params, inputs_ext)
qsym, qparams, precs, _ = calib.sym_simulate(sym, params, inputs_ext, data,
calib_ctx)
qsym, qparams = calib.sym_realize(qsym, qparams, inputs_ext, precs, "cvm")
dump_sym, dump_params, dump_ext = load_fname(version, "sym.quantize", True)
sim.save_ext(dump_ext, inputs_ext)
nd.save(dump_params, qparams)
open(dump_sym, "w").write(qsym.tojson())
dump_sym, dump_params, dump_ext = load_fname(version, "sym.quantize", True)
(inputs_ext, ) = sim.load_ext(dump_ext)
net3 = utils.load_model(dump_sym, dump_params, inputs, ctx=ctx)
qacc_top1 = mx.metric.Accuracy()
qacc_top5 = mx.metric.TopKAccuracy(5)
qacc_top1.reset()
qacc_top5.reset()
def cvm_quantize(data, label):
data = sim.load_real_data(data, 'data', inputs_ext)
data = gluon.utils.split_and_load(data,
ctx_list=ctx,
batch_axis=0,
even_split=False)
res = [net3.forward(d) for d in data]
res = nd.concatenate(res)
qacc_top1.update(label, res)
_, top1 = qacc_top1.get()
qacc_top5.update(label, res)
_, top5 = qacc_top5.get()
return "top1={:6.2%} top5={:6.2%}".format(top1, top5)
utils.multi_validate(resnet,
data_iter_func,
cvm_quantize,
iter_num=iter_num,
logger=logger)
def test_sym_pass(batch_size=10, iter_num=10, quantize=True):
logger = logging.getLogger("log.test.sym.pass")
calib_ctx = mx.gpu(2)
ctx = [mx.gpu(int(i)) for i in "1,2,3,4".split(',') if i.strip()]
input_size = 299
version = "v3"
h, w = input_size, input_size
inputs_ext = {
'data': {
'shape': (batch_size, 3, h, w),
}
}
inputs = [mx.sym.var(name) for name in inputs_ext]
logger.info("load dataset, symbol and parameters")
data_iter = ds.load_imagenet_rec(batch_size, input_size)
def data_iter_func():
data = data_iter.next()
return data.data[0], data.label[0]
net1 = utils.load_model(*load_fname(version), inputs, ctx=ctx)
acc_top1 = mx.metric.Accuracy()
acc_top5 = mx.metric.TopKAccuracy(5)
acc_top1.reset()
acc_top5.reset()
def inception_v3(data, label):
data = gluon.utils.split_and_load(data,
ctx_list=ctx,
batch_axis=0,
even_split=False)
res = [net1.forward(d) for d in data]
res = nd.concatenate(res)
acc_top1.update(label, res)
_, top1 = acc_top1.get()
acc_top5.update(label, res)
_, top5 = acc_top5.get()
return "top1={:6.2%} top5={:6.2%}".format(top1, top5)
if quantize:
sym_file, param_file = load_fname(version)
sym, params = mx.sym.load(sym_file), nd.load(param_file)
sym, params = spass.sym_quant_prepare(sym, params, inputs_ext)
data, _ = data_iter_func()
if True:
dump_sym, dump_params, dump_ext = load_fname(version, "mrt", True)
mrt = _mrt.MRT(sym, params, inputs_ext)
mrt.set_data('data', data)
mrt.calibrate(ctx=calib_ctx)
mrt.set_output_prec(8)
qsym, qparams, inputs_ext = mrt.quantize()
else:
dump_sym, dump_params, dump_ext = load_fname(
version, "sym.quantize", True)
inputs_ext['data']['data'] = data
th_dict = calib.sym_calibrate(sym,
params,
inputs_ext,
ctx=calib_ctx)
qsym, qparams, precs, _ = calib.sym_simulate(
sym, params, inputs_ext, th_dict)
qsym, qparams = calib.sym_realize(qsym, qparams, inputs_ext, precs)
sim.save_ext(dump_ext, inputs_ext)
nd.save(dump_params, qparams)
open(dump_sym, "w").write(qsym.tojson())
dump_sym, dump_params, dump_ext = load_fname(version, "mrt", True)
(inputs_ext, ) = sim.load_ext(dump_ext)
net2 = utils.load_model(dump_sym, dump_params, inputs, ctx=ctx)
qacc_top1 = mx.metric.Accuracy()
qacc_top5 = mx.metric.TopKAccuracy(5)
qacc_top1.reset()
qacc_top5.reset()
def cvm_quantize(data, label):
data = sim.load_real_data(data, 'data', inputs_ext)
data = gluon.utils.split_and_load(data,
ctx_list=ctx,
batch_axis=0,
even_split=False)
res = [net2.forward(d) for d in data]
res = nd.concatenate(res)
qacc_top1.update(label, res)
_, top1 = qacc_top1.get()
qacc_top5.update(label, res)
_, top5 = qacc_top5.get()
return "top1={:6.2%} top5={:6.2%}".format(top1, top5)
utils.multi_validate(inception_v3,
data_iter_func,
cvm_quantize,
iter_num=iter_num,
logger=logger)
_, top5 = acc_top5.get()
return "top1={:6.2%} top5={:6.2%}".format(top1, top5)
# sym, params = mx.sym.load(sym_file), nd.load(param_file)
# sym, params = spass.sym_quant_prepare(sym, params, inputs_ext)
# qsym, qparams, precs, _ = calib.sym_simulate(sym, params, inputs_ext, data, ctx)
# qsym, qparams = calib.sym_realize(qsym, qparams, inputs_ext, precs, "cvm")
# dump_sym, dump_params, dump_ext = load_fname("", "sym.quantize", True)
# sim.save_ext(dump_ext, inputs_ext)
# nd.save(dump_params, qparams)
# open(dump_sym, "w").write(qsym.tojson())
dump_sym, dump_params, dump_ext = load_fname("", "sym.quantize", True)
sym, params = mx.sym.load(dump_sym), nd.load(dump_params)
(inputs_ext, ) = sim.load_ext(dump_ext)
if True:
_mrt.std_dump(sym, params, inputs_ext, data, "alexnet", is_mxnet=True)
exit()
inputs = [mx.sym.var(n) for n in inputs_ext]
net2 = utils.load_model(dump_sym, dump_params, inputs, ctx=ctx)
qacc_top1 = mx.metric.Accuracy()
qacc_top5 = mx.metric.TopKAccuracy(5)
qacc_top1.reset()
qacc_top5.reset()
def cvm_quantize(data, label):
data = sim.load_real_data(data, 'data', inputs_ext)
data = gluon.utils.split_and_load(data,
ctx_list=ctx,
def test_mrt_quant(batch_size=1, iter_num=10):
logger = logging.getLogger("log.test.mrt.quantize")
ctx = mx.gpu(1)
qctx = mx.gpu(3)
input_size = 512
h, w = input_size, input_size
inputs_ext = {
'data': {
'shape': (batch_size, 3, h, w),
}
}
val_data = dataset.load_voc(batch_size, input_size)
val_data_iter = iter(val_data)
def data_iter_func():
data, label = next(val_data_iter)
return data, label
sym_file, param_file = load_fname()
sym, params = mx.sym.load(sym_file), nd.load(param_file)
sym, params = spass.sym_quant_prepare(sym, params, inputs_ext)
keys = [
"ssd0_multiperclassdecoder0_concat0",
"ssd0_multiperclassdecoder0__mulscalar0",
"ssd0_multiperclassdecoder0_slice_axis0",
"ssd0_multiperclassdecoder0_zeros_like1",
"ssd0_normalizedboxcenterdecoder0_concat0",
]
base, base_params, base_inputs_ext, top, top_params, top_inputs_ext \
= _mrt.split_model(sym, params, inputs_ext, keys)
dump_sym, dump_params = load_fname("mrt.base")
open(dump_sym, "w").write(base.tojson())
nd.save(dump_params, base_params)
dump_sym, dump_params, dump_ext = load_fname("mrt.top", True)
open(dump_sym, "w").write(top.tojson())
nd.save(dump_params, top_params)
sim.save_ext(dump_ext, top_inputs_ext)
dump_sym, dump_params = load_fname("mrt.base")
base, base_params = mx.sym.load(dump_sym), nd.load(dump_params)
dump_sym, dump_params, dump_ext = load_fname("mrt.top", True)
top, top_params = mx.sym.load(dump_sym), nd.load(dump_params)
(top_inputs_ext, ) = sim.load_ext(dump_ext)
base_inputs = [mx.sym.var(n) for n in inputs_ext]
base_graph = mx.gluon.nn.SymbolBlock(base, base_inputs)
utils.load_parameters(base_graph, base_params, ctx=ctx)
top_inputs = [mx.sym.var(n) for n in top_inputs_ext]
top_graph = mx.gluon.nn.SymbolBlock(top, top_inputs)
utils.load_parameters(top_graph, top_params, ctx=ctx)
metric = dataset.load_voc_metric()
metric.reset()
def yolov3(data, label):
def net(data):
tmp = base_graph(data.as_in_context(ctx))
outs = top_graph(*tmp)
return outs
acc = validate_data(net, data, label, metric)
return "{:6.2%}".format(acc)
# utils.multi_validate(yolov3, data_iter_func,
# iter_num=iter_num, logger=logger)
# exit()
if True:
mrt = _mrt.MRT(base, base_params, inputs_ext)
for i in range(16):
data, _ = data_iter_func()
mrt.set_data('data', data)
th_dict = mrt.calibrate(ctx=ctx)
_, _, dump_ext = load_fname("mrt.dict", True)
sim.save_ext(dump_ext, th_dict)
_, _, dump_ext = load_fname("mrt.dict", True)
(th_dict, ) = sim.load_ext(dump_ext)
if True:
mrt = _mrt.MRT(base, base_params, base_inputs_ext)
mrt.set_th_dict(th_dict)
mrt.set_threshold('data', 2.64)
mrt.set_fixed("ssd0_multiperclassdecoder0_concat0")
mrt.set_fixed("ssd0_multiperclassdecoder0__mulscalar0")
mrt.set_fixed("ssd0_multiperclassdecoder0_zeros_like1")
mrt.set_threshold("ssd0_multiperclassdecoder0_slice_axis0", 1)
# mrt.set_threshold("ssd0_normalizedboxcenterdecoder0_concat0", 512)
mrt.set_output_prec(30)
qbase, qbase_params, qbase_inputs_ext = mrt.quantize()
oscales = mrt.get_output_scales()
maps = mrt.get_maps()
dump_sym, dump_params, dump_ext = load_fname("mrt.quantize", True)
open(dump_sym, "w").write(qbase.tojson())
nd.save(dump_params, qbase_params)
sim.save_ext(dump_ext, qbase_inputs_ext, oscales, maps)
# merge quantize model
if True:
qb_sym, qb_params, qb_ext = load_fname("mrt.quantize", True)
qbase, qbase_params = mx.sym.load(qb_sym), nd.load(qb_params)
qbase_inputs_ext, oscales, maps = sim.load_ext(qb_ext)
name_maps = {
"ssd0_slice_axis41": "ssd0_multiperclassdecoder0_concat0",
"ssd0_slice_axis42": "ssd0_multiperclassdecoder0_slice_axis0",
"ssd0_slice_axis43": "ssd0_normalizedboxcenterdecoder0_concat0",
}
oscales_dict = dict(zip([c.attr('name') for c in base], oscales))
oscales = [oscales_dict[name_maps[c.attr('name')]] for c in top]
def box_nms(node, params, graph):
name, op_name = node.attr('name'), node.attr('op_name')
childs, attr = sutils.sym_iter(
node.get_children()), node.list_attr()
if op_name == '_greater_scalar':
valid_thresh = sutils.get_attr(attr, 'scalar', 0)
attr['scalar'] = int(valid_thresh * oscales[1])
node = sutils.get_mxnet_op(op_name)(*childs, **attr, name=name)
elif op_name == '_contrib_box_nms':
valid_thresh = sutils.get_attr(attr, 'valid_thresh', 0)
attr['valid_thresh'] = int(valid_thresh * oscales[1])
node = sutils.get_mxnet_op(op_name)(*childs, **attr, name=name)
return node
qsym, qparams = _mrt.merge_model(qbase, qbase_params, top, top_params,
maps, box_nms)
sym_file, param_file, ext_file = load_fname("mrt.all.quantize", True)
open(sym_file, "w").write(qsym.tojson())
nd.save(param_file, qparams)
sim.save_ext(ext_file, qbase_inputs_ext, oscales)
if True:
dump_sym, dump_params, dump_ext = load_fname("mrt.all.quantize", True)
net2_inputs_ext, oscales = sim.load_ext(dump_ext)
inputs = [mx.sym.var(n) for n in net2_inputs_ext]
net2 = utils.load_model(dump_sym, dump_params, inputs, ctx=qctx)
net2_metric = dataset.load_voc_metric()
net2_metric.reset()
def mrt_quantize(data, label):
def net(data):
data = sim.load_real_data(data, 'data', net2_inputs_ext)
outs = net2(data.as_in_context(qctx))
outs = [
o.as_in_context(ctx) / oscales[i]
for i, o in enumerate(outs)
]
return outs
acc = validate_data(net, data, label, net2_metric)
return "{:6.2%}".format(acc)
utils.multi_validate(yolov3,
data_iter_func,
mrt_quantize,
iter_num=iter_num,
logger=logger)
