def test_compile():
x = sym.Variable("x")
y = sym.Variable("y")
z = sym.exp(y + x)
shape = (10, 128)
dtype = tvm.float32
shape_dict = {"x": shape, "y": shape}
def verify(graph, lib):
m = graph_runtime.create(graph, lib, tvm.cpu(0))
# get member functions
set_input, run, get_output = m["set_input"], m["run"], m["get_output"]
na = tvm.nd.array(np.random.uniform(size=shape).astype(dtype))
nb = tvm.nd.array(np.random.uniform(size=shape).astype(dtype))
# set inputs
set_input("x", na)
set_input("y", nb)
# execute
run()
# get outputs
out = tvm.nd.empty(shape, dtype)
get_output(0, out)
np.testing.assert_allclose(out.asnumpy(),
np.exp(na.asnumpy() + nb.asnumpy()))
graph, lib, _ = nnvm.compiler.build(z, "llvm", shape_dict)
assert graph.index.num_nodes == 3
verify(graph, lib)
with nnvm.compiler.build_config(opt_level=0):
graph, lib, _ = nnvm.compiler.build(z, "llvm", shape_dict)
# print(graph.ir())
assert graph.index.num_nodes == 4
verify(graph, lib)
def test_compile_cache():
x = sym.Variable("x")
y = sym.Variable("y")
z = sym.exp(y + x)
shape = (10, 1)
dtype = tvm.float32
shape_dict = {"x": shape, "y": shape}
def verify(graph, lib):
m = graph_runtime.create(graph, lib, tvm.cpu(0))
# get member functions
na = tvm.nd.array(np.random.uniform(size=shape).astype(dtype))
nb = tvm.nd.array(np.random.uniform(size=shape).astype(dtype))
m.run(x=na, y=nb)
# get outputs
out = m.get_output(0, tvm.nd.empty(shape, dtype))
tvm.testing.assert_allclose(out.asnumpy(),
np.exp(na.asnumpy() + nb.asnumpy()))
engine = nnvm.compiler.engine
graph, lib, _ = nnvm.compiler.build(z, "llvm", shape_dict)
inputs = [tvm.placeholder((10, )), tvm.placeholder((10, ))]
gkey = nnvm.compiler.graph_key(nnvm.graph.create(z), inputs, "llvm")
gkey2 = nnvm.compiler.graph_key(nnvm.graph.create(z), inputs + inputs,
"llvm")
gf = engine[gkey]
assert gf is not None
assert engine[gkey2] is None
graph, lib, _ = nnvm.compiler.build(z, "llvm", shape_dict)
assert graph.index.num_nodes == 3
verify(graph, lib)
# Test various set external cache
engine.clear_cache()
engine[gkey] = gf
def test_compile():
x = sym.Variable("x")
y = sym.Variable("y")
z = sym.exp(y + x)
shape = (10, 128)
dtype = tvm.float32
shape_dict = {"x": shape, "y": shape}
def verify(graph, lib):
m = graph_runtime.create(graph, lib, tvm.cpu(0))
# get member functions
set_input, run, get_output = m["set_input"], m["run"], m["get_output"]
na = tvm.nd.array(np.random.uniform(size=shape).astype(dtype))
nb = tvm.nd.array(np.random.uniform(size=shape).astype(dtype))
# set inputs
set_input("x", na)
set_input("y", nb)
# execute
run()
# get outputs
out = tvm.nd.empty(shape, dtype)
get_output(0, out)
tvm.testing.assert_allclose(
out.asnumpy(), np.exp(na.asnumpy() + nb.asnumpy()))
graph, lib, _ = nnvm.compiler.build(z, "llvm", shape_dict)
assert graph.index.num_nodes == 3
verify(graph, lib)
with nnvm.compiler.build_config(opt_level=0):
graph, lib, _ = nnvm.compiler.build(z, "llvm", shape_dict)
# print(graph.ir())
assert graph.index.num_nodes == 4
verify(graph, lib)
def test_compile_cache():
x = sym.Variable("x")
y = sym.Variable("y")
z = sym.exp(y + x)
shape = (10, 1)
dtype = tvm.float32
shape_dict = {"x": shape, "y": shape}
def verify(graph, lib):
m = graph_runtime.create(graph, lib, tvm.cpu(0))
# get member functions
na = tvm.nd.array(np.random.uniform(size=shape).astype(dtype))
nb = tvm.nd.array(np.random.uniform(size=shape).astype(dtype))
m.run(x=na, y=nb)
# get outputs
out = m.get_output(0, tvm.nd.empty(shape, dtype))
tvm.testing.assert_allclose(
out.asnumpy(), np.exp(na.asnumpy() + nb.asnumpy()))
engine = nnvm.compiler.engine
graph, lib, _ = nnvm.compiler.build(z, "llvm", shape_dict)
inputs = [tvm.placeholder((10,)), tvm.placeholder((10,))]
gkey = nnvm.compiler.graph_key(nnvm.graph.create(z), inputs, "llvm")
gkey2 = nnvm.compiler.graph_key(nnvm.graph.create(z), inputs + inputs, "llvm")
gf = engine[gkey]
assert gf is not None
assert engine[gkey2] is None
graph, lib, _ = nnvm.compiler.build(z, "llvm", shape_dict)
assert graph.index.num_nodes == 3
verify(graph, lib)
# Test various set external cache
engine.clear_cache()
engine[gkey] = gf
def test_copy():
x = sym.Variable('x')
z = sym.Variable('z')
y = sym.exp(sym.elemwise_add(x, x, name='add', gpu=2),
name='exp',
gpu=1,
attr={"kk": "1"})
assert y.__copy__().debug_str() == y.debug_str()
def test_unary():
x = sym.Variable('x')
x = sym.exp(x)
x = sym.log(x)
x = sym.sigmoid(x)
x = sym.tanh(x)
x = sym.relu(x)
assert x.list_input_names() == ['x']
def test_unary():
x = sym.Variable('x')
x = sym.exp(x)
x = sym.log(x)
x = sym.sigmoid(x)
x = sym.tanh(x)
x = sym.relu(x)
assert x.list_input_names() == ['x']
def test_graph_gradient():
x0 = sym.Variable('x0')
x1 = sym.Variable('x1')
yg = sym.Variable('yg')
y = sym.exp(sym.mul(x0, x1))
grad_graph = grad(y, [x0], yg)
print("Original graph")
print(y.debug_str())
print("Gradient graph")
print(grad_graph.symbol.debug_str())
def test_run():
x = sym.Variable("x")
y = sym.Variable("y")
z = sym.exp(y + x)
shape = (10, 10)
dtype = tvm.float32
nx = tvm.nd.array(np.random.uniform(size=shape).astype(dtype))
ny = tvm.nd.array(np.random.uniform(size=shape).astype(dtype))
res = _run_graph(z, {"x": nx, "y": ny})
tvm.testing.assert_allclose(
res[0].asnumpy(), np.exp(nx.asnumpy() + ny.asnumpy()))
def test_run():
x = sym.Variable("x")
y = sym.Variable("y")
z = sym.exp(y + x)
shape = (10, 10)
dtype = tvm.float32
nx = tvm.nd.array(np.random.uniform(size=shape).astype(dtype))
ny = tvm.nd.array(np.random.uniform(size=shape).astype(dtype))
res = _run_graph(z, {"x": nx, "y": ny})
np.testing.assert_allclose(res[0].asnumpy(),
np.exp(nx.asnumpy() + ny.asnumpy()))
def test_compose():
x = sym.Variable('x')
z = sym.Variable('z')
y = sym.exp(sym.elemwise_add(x, x, name='add', gpu=2),
name='exp', gpu=1, attr={"kk": "1"})
assert y.list_input_names() == ['x']
assert y.list_output_names() == ["exp_output"]
assert y.list_attr()['gpu'] == '1'
z = y.get_internals()
assert z['add_output'].list_output_names() == ['add_output']
assert y.list_attr(recursive=True)['add$gpu'] == '2'
def test_exp():
x = sym.Variable("x")
y = sym.exp(x)
def forward(x):
return np.exp(x)
def backward(head_grads, x):
return [np.exp(x) * head_grads]
shape = {'x': (1, 3, 32, 32)}
check_function(y, forward, backward, shape=shape)
def test_exp():
x = sym.Variable("x")
y = sym.exp(x)
def forward(x):
return np.exp(x)
def backward(head_grads, x):
return [np.exp(x) * head_grads]
shape = {'x': (1, 3, 32, 32)}
check_function(y, forward, backward, shape=shape)
def test_compose():
x = sym.Variable('x')
z = sym.Variable('z')
y = sym.exp(sym.add(x, x, name='add', gpu=2),
name='exp', gpu=1, attr={"kk": "1"})
assert y.list_inputs() == ['x']
assert y.list_outputs() == ["exp_output"]
assert y.list_attr()['gpu'] == '1'
z = y.get_internals()
assert z['add_output'].list_outputs() == ['add_output']
assert y.list_attr(recursive=True)['add_gpu'] == '2'
def test_exp():
x = sym.Variable("x")
y = sym.exp(x)
def forward(x):
return np.exp(x)
def backward(x):
return np.exp(x)
dtype = "float32"
dshape = (1, 3, 32, 32)
inputs = {'x': (dshape, x)}
helper(y, inputs, dtype, forward, backward)
def test_fusible_network():
""" The network is as following:
data
|
exp
/ \
sqrt log
\ /
b_add
|
tanh
"""
batch_size = 1
data_shape = (batch_size, 3, 224, 224)
data = symbol.Variable('data', shape=data_shape, dtype="float32")
shape_dict = {"data": data_shape}
params = {}
params["data"] = np.random.uniform(-1, 1,
size=data_shape).astype("float32")
exp = symbol.exp(data, name='exp')
sqrt = symbol.sqrt(exp, name='sqrt')
log = symbol.log(exp, name='log')
ret = sqrt + log
ret = symbol.tanh(ret)
# Fuse log and broadcast_add.
check_annotated_graph(ret, ['exp', 'log', 'broadcast_add'], 8,
shape_dict,
params)
# Fuse log, broadcast_add, and tanh
check_annotated_graph(ret, ['exp', 'sqrt', 'none', 'elemwise_add'], 6,
shape_dict, params)
# No operator will be fused.
check_annotated_graph(ret, ['log', 'sqrt', 'none', 'tanh'], 11,
shape_dict, params)
# All operators will be fused.
check_annotated_graph(ret, [''], 2, shape_dict, params)
# All operators will be fused since all of them are annotated to the
# same device.
check_annotated_graph(ret,
['exp', 'sqrt', 'broadcast_add', 'none', 'log',
'tanh'], 2, shape_dict, params)
# Fuse exp, sqrt, log, and boradcast_add
check_annotated_graph(ret, ['tanh'], 4, shape_dict, params)
def test_exp():
x = sym.Variable("x")
y = sym.exp(x)
dtype = "float32"
dshape = (1, 3, 32, 32)
oshape = dshape
for target, ctx in ctx_list():
graph, lib, _ = nnvm.compiler.build(y, target, {"x": dshape})
m = graph_runtime.create(graph, lib, ctx)
data = np.random.uniform(size=dshape).astype(dtype)
m.run(x=data)
out = m.get_output(0, tvm.nd.empty(oshape, dtype))
y_np = np.exp(data)
np.testing.assert_allclose(out.asnumpy(), y_np, atol=1e-5, rtol=1e-5)
def test_exp():
x = sym.Variable("x")
y = sym.exp(x)
def forward(x):
return np.exp(x)
def backward(x):
return np.exp(x)
dtype = "float32"
dshape = (1, 3, 32, 32)
inputs = {'x': (dshape, x)}
helper(y, inputs, dtype, forward, backward)
def test_exp():
x = sym.Variable("x")
y = sym.exp(x)
def forward(x):
return np.exp(x)
def backward(head_grads, x):
return [np.exp(x) * head_grads]
dtype = "float32"
dshape = (1, 3, 32, 32)
inputs = [('x', dshape, x)]
helper(y, inputs, dtype, forward, backward)
def test_place_device():
x = sym.Variable('x', device_group="stage1")
y = sym.add(x, x, name='add1')
y = sym.cast(y, dtype=1, name="cast1")
z = sym.add(y, y, device_group="stage2", name="add2")
z = sym.add(z, sym.exp(y, device_group="stage2"), name="add3")
g = graph.create(z)
g._set_json_attr("device_group_attr_key", "device_group")
g._set_json_attr("device_assign_map", {"stage1": 0, "stage2" : 1}, "dict_str_int")
g._set_json_attr("device_copy_op", "cross_device_copy")
g = g.apply("PlaceDevice")
jgraph = json.loads(g.apply('SaveJSON').json_attr('json'))
jnodes = jgraph['nodes']
jnode_row_ptr = jgraph['node_row_ptr']
nindex = {n['name']: i for i, n in enumerate(jnodes)}
assert g.json_attr('device')[jnode_row_ptr[nindex["add2"]]] == 1
assert g.json_attr('device')[jnode_row_ptr[nindex["add3"]]] == 1
assert g.json_attr('device')[jnode_row_ptr[nindex["cast1"]]] == 0
def test_place_device():
x = sym.Variable('x', device_group="stage1")
y = sym.add(x, x, name='add1')
y = sym.cast(y, dtype=1, name="cast1")
z = sym.add(y, y, device_group="stage2", name="add2")
z = sym.add(z, sym.exp(y, device_group="stage2"), name="add3")
g = graph.create(z)
g._set_json_attr("device_group_attr_key", "device_group")
g._set_json_attr("device_assign_map", {"stage1": 0, "stage2" : 1}, "dict_str_int")
g._set_json_attr("device_copy_op", "cross_device_copy")
g = g.apply("PlaceDevice")
jgraph = json.loads(g.apply('SaveJSON').json_attr('json'))
jnodes = jgraph['nodes']
jnode_row_ptr = jgraph['node_row_ptr']
nindex = {n['name']: i for i, n in enumerate(jnodes)}
assert g.json_attr('device')[jnode_row_ptr[nindex["add2"]]] == 1
assert g.json_attr('device')[jnode_row_ptr[nindex["add3"]]] == 1
assert g.json_attr('device')[jnode_row_ptr[nindex["cast1"]]] == 0
def test_rpc_executor():
host = "localhost"
port = 9021
server = rpc.Server(host, port, use_popen=True)
time.sleep(1)
x = sym.Variable("x")
y = sym.Variable("y")
z = sym.exp(y + x)
shape = (10, 128)
dtype = tvm.float32
shape_dict = {"x": shape, "y": shape}
tmp = util.tempdir()
lib_name = tmp.relpath("net.o")
graph, lib, _ = nnvm.compiler.build(z, "llvm", shape_dict)
# save module
lib.save(lib_name)
remote = rpc.connect(host, port)
remote.upload(lib_name)
ctx = remote.cpu(0)
# load remote
rlib = remote.load_module("net.o")
# Create remotemodule
m = graph_runtime.create(graph, rlib, remote.cpu(0))
# get member functions
set_input, run, get_output = m["set_input"], m["run"], m["get_output"]
na = tvm.nd.array(np.ones(shape).astype(dtype), ctx)
nb = tvm.nd.array(np.ones(shape).astype(dtype), ctx)
# set inputs
set_input("x", na)
set_input("y", nb)
# execute
run()
# get outputs
out = tvm.nd.empty(shape, dtype, ctx)
get_output(0, out)
tvm.testing.assert_allclose(out.asnumpy(),
np.exp(na.asnumpy() + nb.asnumpy()))
server.terminate()
def test_rpc_executor():
host = "localhost"
port = 9021
server = rpc.Server(host, port, use_popen=True)
time.sleep(1)
x = sym.Variable("x")
y = sym.Variable("y")
z = sym.exp(y + x)
shape = (10, 128)
dtype = tvm.float32
shape_dict = {"x": shape, "y": shape}
tmp = util.tempdir()
lib_name = tmp.relpath("net.o")
graph, lib, _ = nnvm.compiler.build(z, "llvm", shape_dict)
# save module
lib.save(lib_name)
remote = rpc.connect(host, port)
remote.upload(lib_name)
ctx = remote.cpu(0)
# load remote
rlib = remote.load_module("net.o")
# Create remotemodule
m = graph_runtime.create(graph, rlib, remote.cpu(0))
# get member functions
set_input, run, get_output = m["set_input"], m["run"], m["get_output"]
na = tvm.nd.array(np.ones(shape).astype(dtype), ctx)
nb = tvm.nd.array(np.ones(shape).astype(dtype), ctx)
# set inputs
set_input("x", na)
set_input("y", nb)
# execute
run()
# get outputs
out = tvm.nd.empty(shape, dtype, ctx)
get_output(0, out)
tvm.testing.assert_allclose(
out.asnumpy(), np.exp(na.asnumpy() + nb.asnumpy()))
server.terminate()
def test_fusible_network(device, target):
R""" The network is as following:
data
|
exp
/ \
sqrt log
\ /
b_add
|
tanh
"""
if not tvm.module.enabled(device):
print("Skip test because %s is not enabled." % device)
return
batch_size = 1
data_shape = (batch_size, 3, 224, 224)
data = symbol.Variable('data', shape=data_shape, dtype="float32")
shape_dict = {"data": data_shape}
params = {}
params["data"] = np.random.uniform(-1, 1,
size=data_shape).astype("float32")
exp = symbol.exp(data, name='exp')
sqrt = symbol.sqrt(exp, name='sqrt')
log = symbol.log(exp, name='log')
ret = sqrt + log
ret = symbol.tanh(ret)
fallback_device = tvm.context("cpu")
target = {"cpu": "llvm", device: target}
# Fuse log and broadcast_add.
op_name_device = {
"exp": "cpu",
"log": "cpu",
"broadcast_add": "cpu",
"sqrt": device,
"elemwise_add": device,
"tanh": device
}
check_annotated_graph(ret, target, op_name_device, 8, fallback_device,
shape_dict, params)
# Fuse log, broadcast_add, and tanh
op_name_device = {
"exp": "cpu",
"log": device,
"broadcast_add": device,
"sqrt": "cpu",
"elemwise_add": "cpu",
"tanh": device
}
check_annotated_graph(ret, target, op_name_device, 6, fallback_device,
shape_dict, params)
# No operator will be fused.
op_name_device = {
"exp": device,
"log": "cpu",
"broadcast_add": device,
"sqrt": "cpu",
"elemwise_add": device,
"tanh": "cpu"
}
check_annotated_graph(ret, target, op_name_device, 11, fallback_device,
shape_dict, params)
# All operators will be fused.
op_name_device = {
"exp": device,
"log": device,
"broadcast_add": device,
"sqrt": device,
"elemwise_add": device,
"tanh": device
}
check_annotated_graph(ret, target, op_name_device, 2, fallback_device,
shape_dict, params)
# All operators will be fused since all of them are annotated to the
# same device.
op_name_device = {
"exp": "cpu",
"log": "cpu",
"broadcast_add": "cpu",
"sqrt": "cpu",
"elemwise_add": "cpu",
"tanh": "cpu"
}
check_annotated_graph(ret, target, op_name_device, 2, fallback_device,
shape_dict, params)
# Fuse exp, sqrt, log, and boradcast_add
op_name_device = {
"exp": device,
"log": device,
"broadcast_add": device,
"sqrt": device,
"elemwise_add": device,
"tanh": "cpu"
}
check_annotated_graph(ret, target, op_name_device, 4, fallback_device,
shape_dict, params)
def test_flatten():
x = sym.Variable("x", shape=(10, 20, 10))
y = sym.flatten(x) * 2
y = sym.exp(y, name="y")
sdict = infer_shape(y)
assert(sdict["y"][0] == [10, 200])
def test_copy():
x = sym.Variable('x')
z = sym.Variable('z')
y = sym.exp(sym.elemwise_add(x, x, name='add', gpu=2),
name='exp', gpu=1, attr={"kk": "1"})
assert y.__copy__().debug_str() == y.debug_str()
def test_op_name():
x = sym.Variable('x')
y = sym.exp(x)
op_name = y.attr("op_name")
op_func = sym.__dict__[op_name]
z = op_func(x)
def test_op_name():
x = sym.Variable('x')
y = sym.exp(x)
op_name = y.attr("op_name")
op_func = sym.__dict__[op_name]
z = op_func(x)
def elu(data):
return -0.5 * sym.relu(1 - sym.exp(data)) + sym.relu(data)
def test_flatten():
x = sym.Variable("x", shape=(10, 20, 10))
y = sym.flatten(x) * 2
y = sym.exp(y, name="y")
sdict = infer_shape(y)
assert(sdict["y"][0] == [10, 200])
def elu(data):
return -0.5 * sym.relu(1 - sym.exp(data)) + sym.relu(data)
