m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge pull request #15830 from wzzju/add_ir_node_encapsulation

Browse Source 
add IrNode&IrVarNode&IrOpNode. test=develop
revert-15774-anakin_subgraph_engine
Zhen Wang 7 years ago committed by GitHub
parent a4b4ecd8b6 548931456c
commit e00c7a2e26
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
5 changed files with 549 additions and 145 deletions
Show Stats Download Patch File Download Diff File

13
paddle/fluid/pybind/ir.cc
Unescape View File

@ -14,6 +14,7 @@
#include "paddle/fluid/pybind/ir.h"
#include <algorithm>
#include <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
@ -116,7 +117,7 @@ void BindNode(py::module *m) {
.def("is_var", &Node::IsVar)
.def("is_ctrl_var", &Node::IsCtrlVar)
.def("clear_inputs", [](Node &self) { self.inputs.clear(); })
.def("inputs_remove",
.def("remove_input",
[](Node &self, int node_id) {
auto pos = std::find_if(
self.inputs.begin(), self.inputs.end(),
@ -125,7 +126,7 @@ void BindNode(py::module *m) {
self.inputs.erase(pos);
}
})
.def("inputs_remove",
.def("remove_input",
[](Node &self, Node &node) {
auto pos =
std::find(self.inputs.begin(), self.inputs.end(), &node);
@ -133,10 +134,10 @@ void BindNode(py::module *m) {
self.inputs.erase(pos);
}
})
.def("inputs_append",
.def("append_input",
[](Node &self, Node &node) { self.inputs.push_back(&node); })
.def("clear_outputs", [](Node &self) { self.outputs.clear(); })
.def("outputs_remove",
.def("remove_output",
[](Node &self, int node_id) {
auto pos = std::find_if(
self.outputs.begin(), self.outputs.end(),
@ -145,7 +146,7 @@ void BindNode(py::module *m) {
self.outputs.erase(pos);
}
})
.def("outputs_remove",
.def("remove_output",
[](Node &self, Node &node) {
auto pos =
std::find(self.outputs.begin(), self.outputs.end(), &node);
@ -153,7 +154,7 @@ void BindNode(py::module *m) {
self.outputs.erase(pos);
}
})
.def("outputs_append",
.def("append_output",
[](Node &self, Node &node) { self.outputs.push_back(&node); })
.def_readwrite("inputs", &Node::inputs)
.def_readwrite("outputs", &Node::outputs);

123
python/paddle/fluid/contrib/slim/quantization/quantization_pass.py
View File

File diff suppressed because it is too large Load Diff

6
python/paddle/fluid/contrib/slim/tests/test_graph.py
Unescape View File

@ -61,16 +61,16 @@ class TestGraph(unittest.TestCase):
opt.minimize(loss)
graph = IrGraph(core.Graph(main.desc), for_test=False)
marked_nodes = set()
for op in graph.all_ops():
for op in graph.all_op_nodes():
if op.name().find('conv2d') > -1:
marked_nodes.add(op)
graph.draw('.', 'residual', marked_nodes)
self.assertFalse(graph.has_circle())
self.assertEqual(graph.graph_num(), 1)
nodes = graph.topology_sort()
self.assertEqual(len(nodes), len(graph.all_ops()))
self.assertEqual(len(nodes), len(graph.all_op_nodes()))
nodes_map = graph.build_adjacency_list()
self.assertEqual(len(nodes_map), len(graph.all_ops()))
self.assertEqual(len(nodes_map), len(graph.all_op_nodes()))
nodes_num = len(graph.all_nodes())
graph.safe_remove_nodes(marked_nodes)
self.assertEqual(len(graph.all_nodes()), nodes_num - len(marked_nodes))

57
python/paddle/fluid/contrib/slim/tests/test_quantization_pass.py
Unescape View File

@ -130,15 +130,16 @@ class TestQuantizationTransformPass(unittest.TestCase):
loss = linear_fc(3)
opt = fluid.optimizer.Adam(learning_rate=0.001)
opt.minimize(loss)
exe = fluid.Executor(fluid.CPUPlace())
place = fluid.CPUPlace()
exe = fluid.Executor(place)
graph = IrGraph(core.Graph(main.desc), for_test=False)
transform_pass = QuantizationTransformPass(
scope=fluid.global_scope(),
program_exe=exe,
place=place,
activation_quantize_type=quant_type)
transform_pass.apply(graph)
marked_nodes = set()
for op in graph.all_ops():
for op in graph.all_op_nodes():
if op.name().find('quantize') > -1:
marked_nodes.add(op)
graph.draw('.', 'quantize_fc_' + quant_type, marked_nodes)
@ -146,7 +147,7 @@ class TestQuantizationTransformPass(unittest.TestCase):
self.check_program(transform_pass, program)
val_graph = IrGraph(core.Graph(program.desc), for_test=False)
val_marked_nodes = set()
for op in val_graph.all_ops():
for op in val_graph.all_op_nodes():
if op.name().find('quantize') > -1:
val_marked_nodes.add(op)
val_graph.draw('.', 'val_fc_' + quant_type, val_marked_nodes)
@ -166,15 +167,16 @@ class TestQuantizationTransformPass(unittest.TestCase):
loss = residual_block(2)
opt = fluid.optimizer.Adam(learning_rate=0.001)
opt.minimize(loss)
exe = fluid.Executor(fluid.CPUPlace())
place = fluid.CPUPlace()
exe = fluid.Executor(place)
graph = IrGraph(core.Graph(main.desc), for_test=False)
transform_pass = QuantizationTransformPass(
scope=fluid.global_scope(),
program_exe=exe,
place=place,
activation_quantize_type=quant_type)
transform_pass.apply(graph)
marked_nodes = set()
for op in graph.all_ops():
for op in graph.all_op_nodes():
if op.name().find('quantize') > -1:
marked_nodes.add(op)
graph.draw('.', 'quantize_residual_' + quant_type, marked_nodes)
@ -182,7 +184,7 @@ class TestQuantizationTransformPass(unittest.TestCase):
self.check_program(transform_pass, program)
val_graph = IrGraph(core.Graph(program.desc), for_test=False)
val_marked_nodes = set()
for op in val_graph.all_ops():
for op in val_graph.all_op_nodes():
if op.name().find('quantize') > -1:
val_marked_nodes.add(op)
val_graph.draw('.', 'val_residual_' + quant_type, val_marked_nodes)
@ -231,17 +233,17 @@ class TestQuantizationFreezePass(unittest.TestCase):
with fluid.scope_guard(scope):
exe.run(startup)
transform_pass = QuantizationTransformPass(
scope=scope, program_exe=exe, activation_quantize_type=quant_type)
scope=scope, place=place, activation_quantize_type=quant_type)
transform_pass.apply(main_graph)
transform_pass.apply(test_graph)
dev_name = '_gpu_' if use_cuda else '_cpu_'
marked_nodes = set()
for op in main_graph.all_ops():
for op in main_graph.all_op_nodes():
if op.name().find('quantize') > -1:
marked_nodes.add(op)
main_graph.draw('.', 'main' + dev_name + quant_type, marked_nodes)
marked_nodes = set()
for op in test_graph.all_ops():
for op in test_graph.all_op_nodes():
if op.name().find('quantize') > -1:
marked_nodes.add(op)
test_graph.draw('.', 'test' + dev_name + quant_type, marked_nodes)
@ -251,11 +253,6 @@ class TestQuantizationFreezePass(unittest.TestCase):
iters = 5
batch_size = 8
#train_exe = fluid.ParallelExecutor(
# main_program=quantized_main_program,
# use_cuda=bool(use_cuda),
# loss_name=loss.name,
# scope=scope)
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.mnist.train(), buf_size=500),
@ -269,9 +266,7 @@ class TestQuantizationFreezePass(unittest.TestCase):
loss_v = exe.run(program=quantized_main_program,
feed=feeder.feed(data),
fetch_list=[loss])
#loss_v = train_exe.run(feed=feeder.feed(data),
# fetch_list=[loss.name])
#print('{}: {}'.format('loss' + dev_name + quant_type, loss_v))
print('{}: {}'.format('loss' + dev_name + quant_type, loss_v))
test_data = next(test_reader())
with fluid.program_guard(quantized_test_program):
@ -287,7 +282,7 @@ class TestQuantizationFreezePass(unittest.TestCase):
freeze_pass = QuantizationFreezePass(scope=scope, place=place)
freeze_pass.apply(test_graph)
marked_nodes = set()
for op in test_graph.all_ops():
for op in test_graph.all_op_nodes():
if op.name().find('quantize') > -1:
marked_nodes.add(op)
test_graph.draw('.', 'test_freeze' + dev_name + quant_type,
@ -299,21 +294,21 @@ class TestQuantizationFreezePass(unittest.TestCase):
feed=feeder.feed(test_data),
fetch_list=[loss])
self.assertAlmostEqual(test_loss1, test_loss2, delta=5e-3)
#print('{}: {}'.format('test_loss1' + dev_name + quant_type, test_loss1))
#print('{}: {}'.format('test_loss2' + dev_name + quant_type, test_loss2))
print('{}: {}'.format('test_loss1' + dev_name + quant_type, test_loss1))
print('{}: {}'.format('test_loss2' + dev_name + quant_type, test_loss2))
w_freeze = np.array(scope.find_var('conv2d_1.w_0').get_tensor())
# Maybe failed, this is due to the calculation precision
# self.assertAlmostEqual(np.sum(w_freeze), np.sum(w_quant))
#print('{}: {}'.format('w_freeze' + dev_name + quant_type,
# np.sum(w_freeze)))
#print('{}: {}'.format('w_quant' + dev_name + quant_type,
# np.sum(w_quant)))
print('{}: {}'.format('w_freeze' + dev_name + quant_type,
np.sum(w_freeze)))
print('{}: {}'.format('w_quant' + dev_name + quant_type,
np.sum(w_quant)))
# Convert parameter to 8-bit.
convert_int8_pass = ConvertToInt8Pass(scope=scope, place=place)
convert_int8_pass.apply(test_graph)
marked_nodes = set()
for op in test_graph.all_ops():
for op in test_graph.all_op_nodes():
if op.name().find('quantize') > -1:
marked_nodes.add(op)
test_graph.draw('.', 'test_int8' + dev_name + quant_type, marked_nodes)
@ -330,14 +325,14 @@ class TestQuantizationFreezePass(unittest.TestCase):
w_8bit = np.array(scope.find_var('conv2d_1.w_0.int8').get_tensor())
self.assertEqual(w_8bit.dtype, np.int8)
self.assertEqual(np.sum(w_8bit), np.sum(w_freeze))
#print('{}: {}'.format('w_8bit' + dev_name + quant_type, np.sum(w_8bit)))
#print('{}: {}'.format('w_freeze' + dev_name + quant_type,
# np.sum(w_freeze)))
print('{}: {}'.format('w_8bit' + dev_name + quant_type, np.sum(w_8bit)))
print('{}: {}'.format('w_freeze' + dev_name + quant_type,
np.sum(w_freeze)))
mobile_pass = TransformForMobilePass()
mobile_pass.apply(test_graph)
marked_nodes = set()
for op in test_graph.all_ops():
for op in test_graph.all_op_nodes():
if op.name().find('quantize') > -1:
marked_nodes.add(op)
test_graph.draw('.', 'test_mobile' + dev_name + quant_type,

495
python/paddle/fluid/framework.py
View File

File diff suppressed because it is too large Load Diff
Write Preview
Loading…
Cancel
Save