def verify_upsampling():
batch = IntIter(list_constraint([1, 4, 8, 16]))
channel = IntIter(list_constraint([1, 3, 4]))
h = IntIter(range_constraint(1, 9, 3))
w = IntIter(range_constraint(1, 9, 3))
dshp = opg.ExtendIter(batch, channel, h, w)
datas = []
for i in range(len(dshp)):
size = np.product(dshp[i])
arr1 = ConstantIter(rand_constraint(-127, 127, size), shape=dshp[i])
arr2 = ConstantIter(rand_constraint(0, 127, size), shape=dshp[i])
datas.extend([arr1, arr2])
data = ConcatIter(*datas)
scale = IntIter(iter_constraint(3), name="scale")
def upsampling(data, scale):
if scale == 0:
raise ValueError("scale must > 0 vs. " + str(scale))
a_np = np.array(data)
b_np = topi.testing.upsampling_python(a_np, scale, "NCHW")
return [b_np]
op_units = opg.OpUnitIter([data, scale], 1)
op_units.eval_data("upsampling", upsampling, is_dump=True)
def verify_dense():
batch = IntIter(list_constraint([1, 4, 8, 16]))
channel = IntIter(list_constraint([1, 3, 32, 64]))
dshp = opg.ExtendIter(batch, channel)
datas = []
for i in range(len(dshp)):
size = np.product(dshp[i])
arr1 = ConstantIter(rand_constraint(-127, 127, size), shape=dshp[i])
arr2 = ConstantIter(rand_constraint(0, 127, size), shape=dshp[i])
arr3 = ConstantIter(rand_constraint(1, 127, size), shape=dshp[i])
datas.extend([arr1, arr2, arr3])
data = ConcatIter(
*datas, ConstantIter(rand_constraint(-127, 127, 112), shape=(4, 4, 7)))
print(len(data))
units = IntIter(list_constraint([1, 32, 64]), name="units")
use_bias = BoolIter(name="use_bias")
op_units = opg.OpUnitIter([data, units, use_bias], 1)
for i in range(len(op_units)):
data, units, use_bias = op_units[i]
data_nd = nd.array(data)
batch, ic = data_nd.shape[:2]
wshp = (units, ic)
wsize = np.product(wshp)
weight = ConstantIter(rand_constraint(-127, 127, wsize), shape=wshp)
weight_nd = nd.array(weight[0])
bshp = (units, )
bias = ConstantIter(rand_constraint(-127, 127, units), shape=bshp)
bias_nd = nd.array(bias[0])
attr = {
'units': units,
'use_bias': use_bias,
}
ins = [data_nd, weight_nd, bias_nd
] if use_bias else [data_nd, weight_nd]
outs, err = None, None
try:
if use_bias:
out = nd.FullyConnected(data_nd, weight_nd, bias_nd, units,
(not use_bias))
else:
out = nd.FullyConnected(data_nd, weight_nd, None, units,
(not use_bias))
outs = [out]
except Exception as e:
err = "Error:\n" + str(e)
print(data_nd.shape, wshp, bshp, attr, outs[0].shape if outs else None,
err.replace("\n", "") if err else None)
opg.dump("dense", attr, ins, outs, err)
def verify_get_valid_counts():
dshp = ConcatIter(PermutationIter(list_constraint([4, 5, 6])),
[[1, 2, 3], [3, 1, 2], [2, 3, 1]])
data_arr = []
for i in range(len(dshp)):
shp = dshp[i]
size = np.product(shp)
arr = ConstantIter(rand_constraint(0, 10, size), shape=shp)
data_arr.append(arr)
data = ConcatIter(*data_arr)
score = IntIter(list_constraint([-1, 0, 1, 5, 7]), name="score_threshold")
def get_valid_counts(data, score_threshold):
np_data = np.array(data)
dshp = np_data.shape
if len(dshp) != 3 or (dshp[2] <= 2):
raise ValueError("data shape error: " + str(dshp))
batch_size, num_anchor, elem_length = dshp
np_out1 = np.zeros(shape=(batch_size, ))
np_out2 = np.zeros(shape=dshp, dtype=INT32)
for i in range(batch_size):
np_out1[i] = 0
inter_idx = 0
for j in range(num_anchor):
score = np_data[i, j, 1]
if score > score_threshold:
for k in range(elem_length):
np_out2[i, inter_idx, k] = np_data[i, j, k]
np_out1[i] += 1
inter_idx += 1
if j >= np_out1[i]:
for k in range(elem_length):
np_out2[i, j, k] = -1.0
return [np_out1, np_out2]
op_units = opg.OpUnitIter([data, score], 1)
op_units.eval_data("get_valid_counts", get_valid_counts, is_dump=True)
def verify_non_max_suppression():
# batch = np.random.randint(low=1, high=10)
# n = np.random.randint(low=10, high=11)
# k = 6
# dshape = (batch, n, k)
# data = tvm.placeholder(dshape, name="data")
# valid_count = tvm.placeholder((dshape[0],), dtype="int32", name="valid_count")
# nms_threshold = np.random.randint(low=1, high=10)
# force_suppress = True if np.random.randint(low=0, high=1) == 1 else False
# nms_topk = np.random.randint(low=1, high=9)
# params = {'iou_threshold':nms_threshold*10, 'coord_start':2, 'score_index':1, 'id_index':0,
# 'force_suppress':force_suppress, 'top_k': nms_topk, 'return_indices':False}
# save_dict(params, case_dir + '/attr.txt')
# np_data = np.random.randint(low=-(2**31-1), high=(2**31-1), size=dshape).astype(data.dtype)
# np_valid_count = np.random.randint(low=1, high=10, size=(batch)).astype(valid_count.dtype)
# device = 'llvm'
# ctx = tvm.context(device, 0)
# with tvm.target.create(device):
# out = non_max_suppression(data, valid_count, -1, nms_threshold, force_suppress, nms_topk, return_indices=False)
# s = topi.generic.schedule_nms(out)
# tvm_data = tvm.nd.array(np_data, ctx)
# tvm_valid_count = tvm.nd.array(np_valid_count, ctx)
# save_txt(tvm_data, case_dir + "/in_0.txt")
# save_txt(tvm_valid_count, case_dir + "/in_1.txt")
# tvm_out = tvm.nd.array(np.zeros(dshape, dtype=data.dtype), ctx)
# f = tvm.build(s, [data, valid_count, out], device)
# f(tvm_data, tvm_valid_count, tvm_out)
# save_txt(tvm_out, case_dir + "/out_0.txt")
batch = IntIter(range_constraint(1, 2))
n = IntIter(rand_constraint(1, 32, 40))
k = IntIter(list_constraint([6]))
dshp = opg.ExtendIter(batch, n, k)
datas = []
for i in range(len(dshp)):
shp = dshp[i]
data = []
for n in range(shp[1]):
elem = rand_constraint(-20, 20, 6)()
elem[0] = random.randint(-1, 10)
data.append(elem)
datas.append([[data]])
data = ConcatIter(*datas)
valid_count = RandomVectorIter(1, 32, 1, 10)
iou = ConcatIter(IntIter(rand_constraint(0, 100, 20)),
IntIter(list_constraint([110])),
name="iou_threshold")
force_suppress = BoolIter(name="force_suppress")
top_k = ConcatIter(IntIter(list_constraint([-1])),
IntIter(rand_constraint(1, 32, 6)),
name="top_k")
id_index = IntIter(list_constraint([0]), name="id_index")
score_index = IntIter(list_constraint([1]), name="score_index")
coord_start = IntIter(list_constraint([2]), name="coord_start")
max_output_size = ConcatIter(IntIter(list_constraint([-1])),
IntIter(rand_constraint(1, 32, 6)),
name="max_output_size")
return_indices = BoolIter(const=False, name="return_indices")
invalid_to_bottom = BoolIter(const=True, name="invalid_to_bottom")
inputs = [
data, valid_count, iou, force_suppress, top_k, id_index, score_index,
coord_start, max_output_size, return_indices, invalid_to_bottom
]
def cstr_func(data, valid_count, *args):
data_np = np.array([data])
count = valid_count[0]
if count > data_np.shape[1]:
return False
return True
op_units = opg.OpUnitIter([
data, valid_count, iou, force_suppress, top_k, id_index, score_index,
coord_start, max_output_size, return_indices, invalid_to_bottom
], 2, [cstr_func])
def non_max_suppression(data, valid_count, iou, force_suppress, top_k,
id_index, score_index, coord_start,
max_output_size, return_indices,
invalid_to_bottom):
device = 'llvm'
ctx = tvm.context(device, 0)
data_np, valid_count_np = np.array(data, dtype="float32"), np.array(
valid_count, dtype="int32")
data_nd, valid_count_nd = tvm.nd.array(data_np, ctx), tvm.nd.array(
valid_count_np, ctx)
dshp = data_nd.shape
data_tvm = tvm.placeholder(dshp, name="data", dtype="float32")
valid_count_tvm = tvm.placeholder((dshp[0], ),
dtype="int32",
name="valid_count")
with tvm.target.create(device):
out = topi.vision.non_max_suppression(data_tvm, valid_count_tvm,
max_output_size, iou / 100,
force_suppress, top_k,
coord_start, score_index,
id_index, return_indices,
invalid_to_bottom)
s = topi.generic.schedule_nms(out)
out_nd = tvm.nd.array(np.zeros(dshp, dtype=data_tvm.dtype), ctx)
f = tvm.build(s, [data_tvm, valid_count_tvm, out], device)
f(data_nd, valid_count_nd, out_nd)
return [out_nd.asnumpy()]
op_units.eval_data("non_max_suppression", non_max_suppression, True)
def verify_max_pool2d():
batch = IntIter(list_constraint([1, 4]))
channel = IntIter(list_constraint([1, 4]))
h = IntIter(range_constraint(1, 9, 3))
w = IntIter(range_constraint(1, 9, 3))
dshp = opg.ExtendIter(batch, channel, h, w)
datas = []
for i in range(len(dshp)):
size = np.product(dshp[i])
arr1 = ConstantIter(rand_constraint(-127, 127, size), shape=dshp[i])
arr2 = ConstantIter(rand_constraint(0, 127, size), shape=dshp[i])
datas.extend([arr1, arr2])
data = ConcatIter(*datas)
print(len(data))
iattr = IntIter(range_constraint(1, 4))
pool_size = ConcatIter(RepeatIter([1, 2, 3], 2), [(2, 3), (3, 1)],
name="pool_size")
strides = ConcatIter(RepeatIter([1, 2], 2), [(1, 2), (2, 1)],
name="strides")
iattr = IntIter(iter_constraint(2))
padding = ConcatIter(VectorIter(iattr, 1),
VectorIter(iattr, 2),
name="padding")
ceil_mode = BoolIter(name="ceil_mode")
def max_pool2d(data, pool_size, strides, padding, ceil_mode):
data_nd = nd.array(data)
pad = padding
if len(padding) == 1:
pad = (padding[0], padding[0])
out = nd.Pooling(data_nd,
pool_size,
pool_type="max",
global_pool=False,
stride=strides,
pad=pad)
data_npy = np.array(data)
ashape = data_npy.shape
n, ic, ih, iw = data_npy.shape
sh, sw = strides
kh, kw = pool_size
bshape = [n, ic, ih, iw]
if len(padding) == 1:
pt = pl = pb = pr = padding[0]
else:
pt = pb = padding[0]
pl = pr = padding[1]
if ceil_mode:
pb += sh - 1
pr += sw - 1
pad_np = np.full((n, ic, ih + pt + pb, iw + pl + pr),
-127).astype(INT32)
# pad_np = np.zeros(shape=(n, ic, ih+pt+pb, iw+pl+pr)).astype(INT32)
no_zero = (range(n), range(ic), (range(pt,
ih + pt)), (range(pl, iw + pl)))
pad_np[np.ix_(*no_zero)] = data_npy
bshape[2] = int(math.floor(float(ashape[2] - kh + pt + pb) / sh) + 1)
bshape[3] = int(math.floor(float(ashape[3] - kw + pl + pr) / sw) + 1)
if pt >= kh or (bshape[2] - 1) * sh - pt >= ashape[2]:
raise ValueError("ceil_mode exceed out of input")
if pl >= kw or (bshape[3] - 1) * sw - pl >= ashape[3]:
raise ValueError("ceil mode exceed out of input")
_, oc, oh, ow = bshape
b_np = np.zeros(shape=(n, oc, oh, ow)).astype(INT32)
for i in range(oh):
for j in range(ow):
b_np[:, :, i, j] = np.max(pad_np[:, :, i * sh:i * sh + kh,
j * sw:j * sw + kw],
axis=(2, 3))
return [b_np]
op_units = opg.OpUnitIter([data, pool_size, strides, padding, ceil_mode],
1)
op_units.eval_data("max_pool2d", max_pool2d, is_dump=True)
def verify_conv2d():
batch = IntIter(list_constraint([1, 4, 8]))
channel = IntIter(list_constraint([1, 3, 4, 8]))
h = IntIter(range_constraint(1, 9, 3))
w = IntIter(range_constraint(1, 9, 3))
dshp = opg.ExtendIter(batch, channel, h, w)
datas = []
for i in range(len(dshp)):
size = np.product(dshp[i])
arr1 = ConstantIter(rand_constraint(-127, 127, size), shape=dshp[i])
arr2 = ConstantIter(rand_constraint(0, 127, size), shape=dshp[i])
datas.extend([arr1, arr2])
data = ConcatIter(*datas)
print(len(data))
num_filter = IntIter(list_constraint([1, 16, 32]), name="num_filter")
kernel = VectorIter(IntIter(list_constraint([1, 2, 3])),
size=2,
name="kernel_size")
strides = RandomVectorIter(1, 4, 2, 1, name="strides")
padding = RandomVectorIter(0, 3, 2, 1, name="padding")
groups = IntIter(list_constraint([1, 2, 4]), name="groups")
use_bias = RandomBoolIter(name="use_bias")
op_units = opg.OpUnitIter(
[data, num_filter, kernel, strides, padding, groups, use_bias], 1)
for i in range(len(op_units)):
data, num_filter, kernel, strides, padding, \
groups, use_bias = op_units[i]
a_np = np.array(data)
batch, ic, h, w = a_np.shape
if groups == 1:
pass
elif random.randint(0, 10) < 5 and ic % groups == 0:
pass
else:
groups = ic
wshp = (num_filter, ic // groups, *kernel)
wsize = np.product(wshp)
weight = ConstantIter(rand_constraint(-127, 127, wsize), shape=wshp)
w_np = np.array(weight[0])
bshp = (num_filter, )
bias = ConstantIter(rand_constraint(-127, 127, num_filter), shape=bshp)
b_np = np.array(bias[0])
rand = random.randint(0, 100)
if rand < 60:
dilation = (1, 1)
elif rand < 70:
dilation = (1, 2)
elif rand < 80:
dilation = (2, 1)
else:
dilation = (2, 2)
attr = {
'channels': num_filter,
'kernel_size': kernel,
'strides': strides,
'padding': padding,
'dilation': dilation,
'groups': groups,
'layout': "NCHW",
'kernel_layout': "OIHW",
'use_bias': use_bias,
}
ins = [a_np, w_np, b_np] if use_bias else [a_np, w_np]
outs, err = None, None
try:
b_nd = nd.array(b_np) if use_bias else None
_ = nd.Convolution(nd.array(a_np),
nd.array(w_np),
b_nd,
kernel,
strides,
dilation,
padding,
num_filter,
groups,
no_bias=(not use_bias))
dw_np = topi.testing.dilate_python(w_np, (1, 1, *dilation))
c_np = topi.testing.conv2d_nchw_python(a_np, dw_np, strides,
padding)
if use_bias:
c_np += b_np.reshape(num_filter, 1, 1)
outs = [c_np]
except Exception as e:
err = "Error:\n" + str(e)
print(a_np.shape, wshp, bshp, attr, outs[0].shape if outs else None,
err.replace("\n", "") if err else None)
opg.dump("conv2d", attr, ins, outs, err)