aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--Makefile3
-rw-r--r--examples/chunk_file_example.lua53
-rw-r--r--examples/test_nn_lib.lua29
-rw-r--r--io/chunk_file.c22
-rw-r--r--io/chunk_file.h1
-rw-r--r--io/init.lua3
-rw-r--r--io/sgd_buffer.lua2
-rw-r--r--layer/affine.lua10
-rw-r--r--layer/bias.lua4
-rw-r--r--layer/init.lua4
-rw-r--r--layer/sigmoid.lua4
-rw-r--r--layer/softmax_ce.lua4
-rw-r--r--layer/window.lua4
-rw-r--r--matrix/cuda_helper.h4
-rw-r--r--nn/layer_dag.lua39
15 files changed, 153 insertions, 33 deletions
diff --git a/Makefile b/Makefile
index 5c6fa7b..0468f57 100644
--- a/Makefile
+++ b/Makefile
@@ -12,7 +12,8 @@ LUA_LIBS := matrix/init.lua io/init.lua nerv.lua \
nn/init.lua nn/layer_repo.lua nn/param_repo.lua nn/layer_dag.lua \
io/sgd_buffer.lua
INCLUDE := -I build/luajit-2.0/include/luajit-2.0/ -DLUA_USE_APICHECK
-CUDA_BASE := /usr/local/cuda-6.5
+#CUDA_BASE := /usr/local/cuda-6.5
+CUDA_BASE := /usr/local/cuda-5.0
CUDA_INCLUDE := -I $(CUDA_BASE)/include/
INCLUDE += $(CUDA_INCLUDE)
LDFLAGS := -L$(CUDA_BASE)/lib64/ -Wl,-rpath=$(CUDA_BASE)/lib64/ -lcudart -lcublas
diff --git a/examples/chunk_file_example.lua b/examples/chunk_file_example.lua
new file mode 100644
index 0000000..5961c98
--- /dev/null
+++ b/examples/chunk_file_example.lua
@@ -0,0 +1,53 @@
+-- To define a readable and writable chunk, one must define a class with the
+-- following methods: __init(id, global_conf), read(handle), write(handle),
+-- get_info(), set_info(info) and an id attribute. This file demonstrates a
+-- basic chunk implementation which manages the I/O of a matrix
+
+local MatrixChunk = nerv.class("nerv.MatrixChunk")
+
+function MatrixChunk:__init(id, global_conf)
+ self.id = id
+ self.info = {}
+ self.gconf = global_conf
+end
+
+function MatrixChunk:read(handle)
+ -- pass the read handle to the matrix method
+ self.data = nerv.MMatrixFloat.load(handle)
+end
+
+function MatrixChunk:write(handle)
+ -- pass the write handle to the matrix method
+ self.data:save(handle)
+end
+
+function MatrixChunk:get_info()
+ return self.info
+end
+
+function MatrixChunk:set_info(info)
+ self.info = info
+end
+
+function MatrixChunk.create_from_matrix(id, mat)
+ local ins = nerv.MatrixChunk(id)
+ ins.data = mat
+ return ins
+end
+
+mat = nerv.MMatrixFloat(3, 4)
+for i = 0, 2 do
+ for j = 0, 3 do
+ mat[i][j] = i + j
+ end
+end
+
+cd = nerv.MatrixChunk.create_from_matrix("matrix1", mat)
+
+cf = nerv.ChunkFile("test.nerv", "w")
+cf:write_chunk(cd)
+cf:close()
+
+cf2 = nerv.ChunkFile("test.nerv", "r")
+cd2 = cf2:read_chunk("matrix1")
+print(cd2.data)
diff --git a/examples/test_nn_lib.lua b/examples/test_nn_lib.lua
index 04fd7d6..6fdbd67 100644
--- a/examples/test_nn_lib.lua
+++ b/examples/test_nn_lib.lua
@@ -117,7 +117,7 @@ tnet_reader = nerv.TNetReader(gconf,
buffer = nerv.SGDBuffer(gconf,
{
buffer_size = 81920,
- -- randomize = true,
+ randomize = true,
readers = {
{ reader = tnet_reader,
data = {main_scp = 429, ref = 1}}
@@ -128,9 +128,12 @@ sm = sublayer_repo:get_layer("softmax_ce0")
main = layer_repo:get_layer("main")
main:init(gconf.batch_size)
gconf.cnt = 0
+-- data = buffer:get_data()
+-- input = {data.main_scp, data.ref}
+-- while true do
for data in buffer.get_data, buffer do
- if gconf.cnt == 1000 then break end
- gconf.cnt = gconf.cnt + 1
+-- if gconf.cnt == 100 then break end
+-- gconf.cnt = gconf.cnt + 1
input = {data.main_scp, data.ref}
output = {}
@@ -141,11 +144,21 @@ for data in buffer.get_data, buffer do
main:back_propagate(err_output, err_input, input, output)
main:update(err_input, input, output)
- nerv.utils.printf("cross entropy: %.8f\n", sm.total_ce)
- nerv.utils.printf("correct: %d\n", sm.total_correct)
- nerv.utils.printf("frames: %d\n", sm.total_frames)
- nerv.utils.printf("err/frm: %.8f\n", sm.total_ce / sm.total_frames)
- nerv.utils.printf("accuracy: %.8f\n", sm.total_correct / sm.total_frames)
+-- nerv.utils.printf("cross entropy: %.8f\n", sm.total_ce)
+-- nerv.utils.printf("correct: %d\n", sm.total_correct)
+-- nerv.utils.printf("frames: %d\n", sm.total_frames)
+-- nerv.utils.printf("err/frm: %.8f\n", sm.total_ce / sm.total_frames)
+-- nerv.utils.printf("accuracy: %.8f\n", sm.total_correct / sm.total_frames)
collectgarbage("collect")
end
+nerv.utils.printf("cross entropy: %.8f\n", sm.total_ce)
+nerv.utils.printf("correct: %d\n", sm.total_correct)
+nerv.utils.printf("accuracy: %.3f%%\n", sm.total_correct / sm.total_frames * 100)
+nerv.utils.printf("writing back...\n")
+cf = nerv.ChunkFile("output.nerv", "w")
+for i, p in ipairs(main:get_params()) do
+ print(p)
+ cf:write_chunk(p)
+end
+cf:close()
nerv.Matrix.print_profile()
diff --git a/io/chunk_file.c b/io/chunk_file.c
index 4e987b7..aa7dd1c 100644
--- a/io/chunk_file.c
+++ b/io/chunk_file.c
@@ -10,6 +10,11 @@
do { \
if ((exp) != (ret)) INVALID_FORMAT_ERROR(fn); \
} while (0)
+#define CHECK_FILE_OPEN(pfh) \
+ do { \
+ if ((pfh)->closed) \
+ nerv_error(L, "operations on a closed file"); \
+ } while (0)
const char *nerv_chunk_file_tname = "nerv.ChunkFile";
const char *nerv_chunk_file_handle_tname = "nerv.ChunkFileHandle";
@@ -109,6 +114,7 @@ int nerv_chunk_file_open_write(lua_State *L, const char *fn) {
if (!fp) nerv_error(L, "Error while opening chunk file: %s", fn);
lfp = (ChunkFileHandle *)malloc(sizeof(ChunkFileHandle));
lfp->fp = fp;
+ lfp->closed = 0;
luaT_pushudata(L, lfp, nerv_chunk_file_handle_tname);
lua_setfield(L, -2, "handle");
luaT_pushmetatable(L, nerv_chunk_file_tname);
@@ -174,6 +180,7 @@ int nerv_chunk_file_open_read(lua_State *L, const char *fn) {
lua_setfield(L, -2, "metadata");
lfp = (ChunkFileHandle *)malloc(sizeof(ChunkFileHandle));
lfp->fp = fp;
+ lfp->closed = 0;
luaT_pushudata(L, lfp, nerv_chunk_file_handle_tname);
lua_setfield(L, -2, "handle");
luaT_pushmetatable(L, nerv_chunk_file_tname);
@@ -215,6 +222,7 @@ int nerv_chunk_file_write_chunkdata(lua_State *L) {
const char *metadata_str = lua_tolstring(L, 2, NULL);
lua_getfield(L, 1, "handle");
pfh = luaT_checkudata(L, -1, nerv_chunk_file_handle_tname);
+ CHECK_FILE_OPEN(pfh);
start = ftello(pfh->fp);
write_chunk_header_plain(pfh->fp, 0, &status); /* fill zeros */
CHECK_WRITE(status);
@@ -245,6 +253,7 @@ int nerv_chunk_file_get_chunkdata(lua_State *L) {
lua_getfield(L, 1, "handle");
pfh = luaT_checkudata(L, -1, nerv_chunk_file_handle_tname);
+ CHECK_FILE_OPEN(pfh);
lua_pop(L, 1); /* pop handle */
lua_getfield(L, 1, "metadata");
/* now stack: self, k, metadata */
@@ -260,10 +269,20 @@ int nerv_chunk_file_get_chunkdata(lua_State *L) {
return 1;
}
+int nerv_chunk_file_close(lua_State *L) {
+ ChunkFileHandle *pfh;
+ lua_getfield(L, 1, "handle");
+ pfh = luaT_checkudata(L, -1, nerv_chunk_file_handle_tname);
+ CHECK_FILE_OPEN(pfh);
+ fclose(pfh->fp);
+ pfh->closed = 1;
+ return 0;
+}
+
int nerv_chunk_file_handle_destroy(lua_State *L) {
ChunkFileHandle *pfh = luaT_checkudata(L, 1,
nerv_chunk_file_handle_tname);
- fclose(pfh->fp);
+ if (!pfh->closed) fclose(pfh->fp);
free(pfh);
return 0;
}
@@ -285,6 +304,7 @@ static int nerv_chunk_data_destroy(lua_State *L) {
static const luaL_Reg nerv_chunk_file_methods[] = {
{"get_chunkdata", nerv_chunk_file_get_chunkdata},
{"_write_chunkdata", nerv_chunk_file_write_chunkdata},
+ {"close", nerv_chunk_file_close},
{"__init", nerv_chunk_file___init},
{NULL, NULL}
};
diff --git a/io/chunk_file.h b/io/chunk_file.h
index 9ece117..9bae59d 100644
--- a/io/chunk_file.h
+++ b/io/chunk_file.h
@@ -8,6 +8,7 @@ extern const char *nerv_chunk_data_tname;
typedef struct ChunkFileHandle {
FILE *fp;
+ int closed;
} ChunkFileHandle;
typedef struct ChunkInfo {
diff --git a/io/init.lua b/io/init.lua
index c151804..b722a81 100644
--- a/io/init.lua
+++ b/io/init.lua
@@ -18,6 +18,9 @@ function nerv.ChunkFile:write_chunk(chunk)
end
function nerv.ChunkFile:read_chunk(id, global_conf)
+ if self.metadata == nil then
+ nerv.error("wrong file opening mode")
+ end
local metadata = self.metadata[id]
if metadata == nil then
nerv.error("chunk with id %s does not exist", id)
diff --git a/io/sgd_buffer.lua b/io/sgd_buffer.lua
index bf72744..381b863 100644
--- a/io/sgd_buffer.lua
+++ b/io/sgd_buffer.lua
@@ -15,7 +15,7 @@ function SGDBuffer:__init(global_conf, buffer_conf)
local buffs = {}
for id, width in pairs(reader_spec.data) do
buffs[id] = {data = global_conf.mmat_type(self.buffer_size, width),
- leftover = {},
+ leftover = nil,
width = width}
end
table.insert(self.readers, {buffs = buffs,
diff --git a/layer/affine.lua b/layer/affine.lua
index 59a0e91..2cd7acb 100644
--- a/layer/affine.lua
+++ b/layer/affine.lua
@@ -41,7 +41,7 @@ function AffineLayer:init()
self.bc:fill(0)
end
-function nerv.AffineLayer:update(bp_err, input, output)
+function AffineLayer:update(bp_err, input, output)
local ltp = self.ltp.trans
local bp = self.bp.trans
local ltc = self.ltc
@@ -60,13 +60,17 @@ function nerv.AffineLayer:update(bp_err, input, output)
ltp:add(ltp, ltp, 1.0, -gconf.lrate * gconf.wcost)
end
-function nerv.AffineLayer:propagate(input, output)
+function AffineLayer:propagate(input, output)
-- apply linear transform
output[1]:mul(input[1], self.ltp.trans, 1.0, 0.0, 'N', 'N')
-- add bias
output[1]:add_row(self.bp.trans, 1.0)
end
-function nerv.AffineLayer:back_propagate(next_bp_err, bp_err, input, output)
+function AffineLayer:back_propagate(next_bp_err, bp_err, input, output)
next_bp_err[1]:mul(bp_err[1], self.ltp.trans, 1.0, 0.0, 'N', 'T')
end
+
+function AffineLayer:get_params()
+ return {self.ltp, self.bp}
+end
diff --git a/layer/bias.lua b/layer/bias.lua
index 6ddfe11..8cd326b 100644
--- a/layer/bias.lua
+++ b/layer/bias.lua
@@ -22,3 +22,7 @@ function BiasLayer:propagate(input, output)
output[1]:copy_fromd(input[1])
output[1]:add_row(self.bias.trans, 1.0)
end
+
+function BiasLayer:get_params()
+ return {self.bias}
+end
diff --git a/layer/init.lua b/layer/init.lua
index 38bcd7f..3011f8e 100644
--- a/layer/init.lua
+++ b/layer/init.lua
@@ -58,6 +58,10 @@ function Layer:check_dim_len(len_in, len_out)
end
end
+function Layer:get_params()
+ nerv.error_method_not_implemented()
+end
+
function Layer:get_dim()
return self.dim_in, self.dim_out
end
diff --git a/layer/sigmoid.lua b/layer/sigmoid.lua
index 220b7af..dd10fb9 100644
--- a/layer/sigmoid.lua
+++ b/layer/sigmoid.lua
@@ -25,3 +25,7 @@ end
function SigmoidLayer:back_propagate(next_bp_err, bp_err, input, output)
next_bp_err[1]:sigmoid_grad(bp_err[1], output[1])
end
+
+function SigmoidLayer:get_params()
+ return {}
+end
diff --git a/layer/softmax_ce.lua b/layer/softmax_ce.lua
index cd57010..79e859e 100644
--- a/layer/softmax_ce.lua
+++ b/layer/softmax_ce.lua
@@ -50,3 +50,7 @@ function SoftmaxCELayer:back_propagate(next_bp_err, bp_err, input, output)
end
next_bp_err[1]:add(self.soutput, label, 1.0, -1.0)
end
+
+function SoftmaxCELayer:get_params()
+ return {}
+end
diff --git a/layer/window.lua b/layer/window.lua
index 8e9e761..b381c9b 100644
--- a/layer/window.lua
+++ b/layer/window.lua
@@ -22,3 +22,7 @@ function WindowLayer:propagate(input, output)
output[1]:copy_fromd(input[1])
output[1]:scale_row(self.window.trans)
end
+
+function WindowLayer:get_params()
+ return {self.window}
+end
diff --git a/matrix/cuda_helper.h b/matrix/cuda_helper.h
index 88619fd..d6effdb 100644
--- a/matrix/cuda_helper.h
+++ b/matrix/cuda_helper.h
@@ -52,10 +52,10 @@ static const char *cublasGetErrorString(cublasStatus_t err) {
return "CUBLAS_STATUS_EXECUTION_FAILED";
case CUBLAS_STATUS_INTERNAL_ERROR:
return "CUBLAS_STATUS_INTERNAL_ERROR";
- case CUBLAS_STATUS_NOT_SUPPORTED:
+/* case CUBLAS_STATUS_NOT_SUPPORTED:
return "CUBLAS_STATUS_NOT_SUPPORTED";
case CUBLAS_STATUS_LICENSE_ERROR:
- return "CUBLAS_STATUS_LICENSE_ERROR";
+ return "CUBLAS_STATUS_LICENSE_ERROR"; */
}
return "<unknown>";
}
diff --git a/nn/layer_dag.lua b/nn/layer_dag.lua
index 4ee829e..2dda7c9 100644
--- a/nn/layer_dag.lua
+++ b/nn/layer_dag.lua
@@ -38,7 +38,7 @@ local function discover(id, layers, layer_repo)
return ref
end
-function nerv.DAGLayer:__init(id, global_conf, layer_conf)
+function DAGLayer:__init(id, global_conf, layer_conf)
local layers = {}
local inputs = {}
local outputs = {}
@@ -131,7 +131,7 @@ function nerv.DAGLayer:__init(id, global_conf, layer_conf)
self.gconf = global_conf
end
-function nerv.DAGLayer:init(batch_size) -- topology sort
+function DAGLayer:init(batch_size) -- topology sort
for i, conn in ipairs(self.parsed_conn) do
local _, output_dim
local ref_from, port_from, ref_to, port_to
@@ -174,7 +174,7 @@ function nerv.DAGLayer:init(batch_size) -- topology sort
end
end
-function nerv.DAGLayer:set_inputs(input)
+function DAGLayer:set_inputs(input)
for i = 1, #self.dim_in do
local layer = self.inputs[i][1]
local port = self.inputs[i][2]
@@ -182,7 +182,7 @@ function nerv.DAGLayer:set_inputs(input)
end
end
-function nerv.DAGLayer:set_outputs(output)
+function DAGLayer:set_outputs(output)
for i = 1, #self.dim_out do
local layer = self.outputs[i][1]
local port = self.outputs[i][2]
@@ -190,7 +190,7 @@ function nerv.DAGLayer:set_outputs(output)
end
end
-function nerv.DAGLayer:set_err_inputs(bp_err)
+function DAGLayer:set_err_inputs(bp_err)
for i = 1, #self.dim_out do
local layer = self.outputs[i][1]
local port = self.outputs[i][2]
@@ -198,7 +198,7 @@ function nerv.DAGLayer:set_err_inputs(bp_err)
end
end
-function nerv.DAGLayer:set_err_outputs(next_bp_err)
+function DAGLayer:set_err_outputs(next_bp_err)
for i = 1, #self.dim_in do
local layer = self.inputs[i][1]
local port = self.inputs[i][2]
@@ -206,30 +206,28 @@ function nerv.DAGLayer:set_err_outputs(next_bp_err)
end
end
-function nerv.DAGLayer:update(bp_err, input, output)
+function DAGLayer:update(bp_err, input, output)
self:set_err_inputs(bp_err)
self:set_inputs(input)
self:set_outputs(output)
+ -- print("update")
for id, ref in pairs(self.queue) do
+ -- print(ref.layer.id)
ref.layer:update(ref.err_inputs, ref.inputs, ref.outputs)
end
end
-function nerv.DAGLayer:propagate(input, output)
+function DAGLayer:propagate(input, output)
self:set_inputs(input)
self:set_outputs(output)
for i = 1, #self.queue do
local ref = self.queue[i]
- --[[
- print(ref.inputs[1])
- print(ref.outputs[1])
- print(#ref.inputs, #ref.outputs)
- --]]
+ -- print(ref.layer.id)
ref.layer:propagate(ref.inputs, ref.outputs)
end
end
-function nerv.DAGLayer:back_propagate(next_bp_err, bp_err, input, output)
+function DAGLayer:back_propagate(next_bp_err, bp_err, input, output)
self:set_err_outputs(next_bp_err)
self:set_err_inputs(bp_err)
self:set_inputs(input)
@@ -238,8 +236,15 @@ function nerv.DAGLayer:back_propagate(next_bp_err, bp_err, input, output)
local ref = self.queue[i]
-- print(ref.layer.id)
ref.layer:back_propagate(ref.err_outputs, ref.err_inputs, ref.inputs, ref.outputs)
- -- if #ref.err_outputs > 0 then
- -- print(ref.err_outputs[1])
- -- end
end
end
+
+function DAGLayer:get_params()
+ local res = {}
+ for id, ref in pairs(self.queue) do
+ for i, p in ipairs(ref.layer:get_params()) do
+ table.insert(res, p)
+ end
+ end
+ return res
+end
Powered by cgit, Themed by Ted Yin.