1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
|
--author: txh18(Tianxing)
--This recipe is rnnlm with bptt, unfolding for each time instance
--The training framework is the same with Mikolov's rnnlm, Tianxing's XRNN-CPU and Wengong's XRNN-GPU
--It uses DAGLayer to simulate RNNLM unfold
--TODO: the select_linear now accepts a column vector, instead of a row vector
require 'lmptb.lmvocab'
require 'lmptb.lmfeeder'
require 'lmptb.lmutil'
require 'tnn.init'
nerv.include('lmptb/layer/init.lua')
--[[global function rename]]--
printf = nerv.printf
--[[global function rename ends]]--
--global_conf: table
--first_time: bool
--Returns: a ParamRepo
function prepare_parameters(global_conf, first_time)
printf("%s preparing parameters...\n", global_conf.sche_log_pre)
if (first_time) then
ltp_ih = nerv.LinearTransParam("ltp_ih", global_conf)
ltp_ih.trans = global_conf.cumat_type(global_conf.vocab:size() + 1, global_conf.hidden_size) --index 0 is for zero, others correspond to vocab index(starting from 1)
ltp_ih.trans:generate(global_conf.param_random)
ltp_ih.trans[0]:fill(0)
ltp_hh = nerv.LinearTransParam("ltp_hh", global_conf)
ltp_hh.trans = global_conf.cumat_type(global_conf.hidden_size, global_conf.hidden_size)
ltp_hh.trans:generate(global_conf.param_random)
ltp_ho = nerv.LinearTransParam("ltp_ho", global_conf)
ltp_ho.trans = global_conf.cumat_type(global_conf.hidden_size, global_conf.vocab:size())
ltp_ho.trans:generate(global_conf.param_random)
bp_h = nerv.BiasParam("bp_h", global_conf)
bp_h.trans = global_conf.cumat_type(1, global_conf.hidden_size)
bp_h.trans:generate(global_conf.param_random)
bp_o = nerv.BiasParam("bp_o", global_conf)
bp_o.trans = global_conf.cumat_type(1, global_conf.vocab:size())
bp_o.trans:generate(global_conf.param_random)
local f = nerv.ChunkFile(global_conf.param_fn, 'w')
f:write_chunk(ltp_ih)
f:write_chunk(ltp_hh)
f:write_chunk(ltp_ho)
f:write_chunk(bp_h)
f:write_chunk(bp_o)
f:close()
end
local paramRepo = nerv.ParamRepo()
paramRepo:import({global_conf.param_fn}, nil, global_conf)
printf("%s preparing parameters end.\n", global_conf.sche_log_pre)
return paramRepo
end
--global_conf: table
--Returns: nerv.LayerRepo
function prepare_layers(global_conf, paramRepo)
printf("%s preparing layers...\n", global_conf.sche_log_pre)
local recurrentLconfig = {{["bp"] = "bp_h", ["ltp_hh"] = "ltp_hh"}, {["dim_in"] = {global_conf.hidden_size, global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size}, ["break_id"] = global_conf.vocab:get_sen_entry().id, ["independent"] = global_conf.independent, ["clip"] = 10}}
local layers = {
["nerv.IndRecurrentLayer"] = {
["recurrentL1"] = recurrentLconfig,
},
["nerv.SelectLinearLayer"] = {
["selectL1"] = {{["ltp"] = "ltp_ih"}, {["dim_in"] = {1}, ["dim_out"] = {global_conf.hidden_size}, ["vocab"] = global_conf.vocab}},
},
["nerv.SigmoidLayer"] = {
["sigmoidL1"] = {{}, {["dim_in"] = {global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size}}}
},
["nerv.AffineLayer"] = {
["outputL"] = {{["ltp"] = "ltp_ho", ["bp"] = "bp_o"}, {["dim_in"] = {global_conf.hidden_size}, ["dim_out"] = {global_conf.vocab:size()}}},
},
["nerv.SoftmaxCELayer"] = {
["softmaxL"] = {{}, {["dim_in"] = {global_conf.vocab:size(), global_conf.vocab:size()}, ["dim_out"] = {1}}},
},
}
printf("%s adding %d bptt layers...\n", global_conf.sche_log_pre, global_conf.bptt)
for i = 1, global_conf.bptt do
layers["nerv.IndRecurrentLayer"]["recurrentL" .. (i + 1)] = recurrentLconfig
layers["nerv.SigmoidLayer"]["sigmoidL" .. (i + 1)] = {{}, {["dim_in"] = {global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size}}}
layers["nerv.SelectLinearLayer"]["selectL" .. (i + 1)] = {{["ltp"] = "ltp_ih"}, {["dim_in"] = {1}, ["dim_out"] = {global_conf.hidden_size}, ["vocab"] = global_conf.vocab}}
end
local layerRepo = nerv.LayerRepo(layers, paramRepo, global_conf)
printf("%s preparing layers end.\n", global_conf.sche_log_pre)
return layerRepo
end
--global_conf: table
--layerRepo: nerv.LayerRepo
--Returns: a nerv.DAGLayer
function prepare_dagLayer(global_conf, layerRepo)
printf("%s Initing daglayer ...\n", global_conf.sche_log_pre)
--input: input_w, input_w, ... input_w_now, last_activation
local dim_in_t = {}
for i = 1, global_conf.bptt + 1 do dim_in_t[i] = 1 end
dim_in_t[global_conf.bptt + 2] = global_conf.hidden_size
dim_in_t[global_conf.bptt + 3] = global_conf.vocab:size()
--[[ softmax
| \
ouptut i(bptt+3)
|
recurrentL(bptt+1)... recurrentL2-recurrentL1
selectL(bptt+1) selectL2 selectL1
/ | | |
i(bptt+2) i(bptt+1) i2 i1
]]--
local connections_t = {
["selectL1[1]"] = "recurrentL1[1]",
["recurrentL1[1]"] = "sigmoidL1[1]",
["sigmoidL1[1]"] = "outputL[1]",
["outputL[1]"] = "softmaxL[1]",
["softmaxL[1]"] = "<output>[1]"
}
for i = 1, global_conf.bptt, 1 do
connections_t["<input>["..i.."]"] = "selectL"..i.."[1]"
connections_t["selectL"..(i+1).."[1]"] = "recurrentL"..(i+1).."[1]"
connections_t["recurrentL"..(i+1).."[1]"] = "sigmoidL"..(i+1).."[1]"
connections_t["sigmoidL"..(i+1).."[1]"] = "recurrentL"..i.."[2]"
end
connections_t["<input>["..(global_conf.bptt+1).."]"] = "selectL"..(global_conf.bptt+1).."[1]"
connections_t["<input>["..(global_conf.bptt+2).."]"] = "recurrentL"..(global_conf.bptt+1).."[2]"
connections_t["<input>["..(global_conf.bptt+3).."]"] = "softmaxL[2]"
printf("%s printing DAG connections:\n", global_conf.sche_log_pre)
for key, value in pairs(connections_t) do
printf("\t%s->%s\n", key, value)
end
local dagL = nerv.DAGLayerT("dagL", global_conf, {["dim_in"] = dim_in_t, ["dim_out"] = {1}, ["sub_layers"] = layerRepo,
["connections"] = connections_t,
})
dagL:init(global_conf.batch_size)
printf("%s Initing DAGLayer end.\n", global_conf.sche_log_pre)
return dagL
end
--global_conf: table
--dagL: nerv.DAGLayer
--fn: string
--config: table
--Returns: table, result
function propagateFile(global_conf, dagL, fn, config)
printf("%s Begining doing on %s...\n", global_conf.sche_log_pre, fn)
if (config.do_train == true) then printf("%s do_train in config is true.\n", global_conf.sche_log_pre) end
local feeder = nerv.LMFeeder(global_conf, global_conf.batch_size, global_conf.vocab)
feeder:open_file(fn)
local tnow = 1
local token_store = {}
local hidden_store = {}
local sigmoidL_ref = dagL.layers["sigmoidL1"]
local inputL_ref = dagL.layers["selectL1"]
token_store[tnow] = feeder:get_batch()
for i = 1, global_conf.bptt + 1 do
hidden_store[tnow - i] = global_conf.cumat_type(global_conf.batch_size, global_conf.hidden_size)
hidden_store[tnow - i]:fill(0)
token_store[tnow - i] = {}
for j = 1, global_conf.batch_size do token_store[tnow - i][j] = global_conf.vocab.null_token end
end
local dagL_input = {}
for i = 1, global_conf.bptt + 1 do
dagL_input[i] = global_conf.cumat_type(global_conf.batch_size, 1) --changed to row vector, debughtx
end
dagL_input[global_conf.bptt + 2] = global_conf.cumat_type(global_conf.batch_size, global_conf.hidden_size)
dagL_input[global_conf.bptt + 3] = global_conf.cumat_type(global_conf.batch_size, global_conf.vocab:size())
local dagL_output = {global_conf.cumat_type(global_conf.batch_size, 1)}
local dagL_err = {nil} --{global_conf.cumat_type(global_conf.batch_size, 1)}
local dagL_input_err = {}
for i = 1, global_conf.bptt + 1 do
dagL_input_err[i] = nil --global_conf.cumat_type(global_conf.batch_size, global_conf.vocab:size())
end
dagL_input_err[global_conf.bptt + 2] = global_conf.cumat_type(global_conf.batch_size, global_conf.hidden_size)
dagL_input_err[global_conf.bptt + 3] = global_conf.cumat_type(global_conf.batch_size, global_conf.vocab:size())
local result = nerv.LMResult(global_conf, global_conf.vocab)
result:init("rnn")
global_conf.input_word_id = {}
while (1) do
token_store[tnow + 1] = feeder:get_batch() --The next word(to predict)
if (token_store[tnow + 1] == nil) then break end
--dagL:propagate(dagL_input, dagL_output)
for i = 1, global_conf.bptt + 1 do
nerv.LMUtil.set_id(dagL_input[i], token_store[tnow - i + 1], global_conf.vocab)
global_conf.input_word_id["recurrentL"..i] = dagL_input[i] --for IndRecurrent
end
dagL_input[global_conf.bptt + 2]:copy_fromd(hidden_store[tnow - global_conf.bptt - 1])
nerv.LMUtil.set_onehot(dagL_input[global_conf.bptt + 3], token_store[tnow + 1], global_conf.vocab) --for softmax
--local dagL_input = create_dag_input(global_conf, token_store, hidden_store, tnow)
global_conf.timer:tic("dagL-propagate")
dagL:propagate(dagL_input, dagL_output)
global_conf.timer:toc("dagL-propagate")
hidden_store[tnow] = global_conf.cumat_type(global_conf.batch_size, global_conf.hidden_size)
hidden_store[tnow]:copy_fromd(sigmoidL_ref.outputs[1][1])
if (config.do_train == true) then
global_conf.timer:tic("dagL-back_propagate")
dagL:back_propagate(dagL_err, dagL_input_err, dagL_input, dagL_output)
global_conf.timer:toc("dagL-back_propagate")
global_conf.timer:tic("dagL-update")
dagL:update(dagL_err, dagL_input, dagL_output)
global_conf.timer:toc("dagL-update")
inputL_ref.layer.ltp.trans[0]:fill(0) --afraid that this will be updated in select_linear:update
end
for i = 1, global_conf.batch_size, 1 do
if (token_store[tnow + 1][i] ~= global_conf.vocab.null_token) then
result:add("rnn", token_store[tnow + 1][i], math.exp(dagL_output[1][i - 1][0]))
if (config.report_word == true) then
printf("%s %s: <stream %d> <prob %f>\n", global_conf.sche_log_pre, token_store[tnow + 1][i], i, math.exp(dagL_output[1][i - 1][0]))
end
end
if (result["rnn"].cn_w % global_conf.log_w_num == 0) then
printf("%s %d words processed %s.\n", global_conf.sche_log_pre, result["rnn"].cn_w, os.date())
printf("\t%s log prob per sample :%f.\n", global_conf.sche_log_pre, result:logp_sample("rnn"));
--[[
for key, value in pairs(global_conf.timer.rec) do
printf("\t [global_conf.timer]: time spent on %s:%.5fs\n", key, value)
end
]]--
--comment this for debughtx
global_conf.timer:flush()
--nerv.CuMatrix.print_profile()
--nerv.CuMatrix.clear_profile()
end
end
token_store[tnow - 2 - global_conf.bptt] = nil
hidden_store[tnow - 2 - global_conf.bptt] = nil
collectgarbage("collect")
tnow = tnow + 1
end
printf("%s Displaying result:\n", global_conf.sche_log_pre)
printf("%s %s\n", global_conf.sche_log_pre, result:status("rnn"))
printf("%s Doing on %s end.\n", global_conf.sche_log_pre, fn)
return result
end
--returns dagL, paramRepo
function load_net(global_conf)
local paramRepo = prepare_parameters(global_conf, false)
local layerRepo = prepare_layers(global_conf, paramRepo)
local dagL = prepare_dagLayer(global_conf, layerRepo)
return dagL, paramRepo
end
--[[global settings]]--
local set = "ptb"
if (set == "ptb") then
data_dir = "/home/slhome/txh18/workspace/nerv/nerv/nerv/examples/lmptb/PTBdata"
train_fn = data_dir.."/ptb.train.txt"
valid_fn = data_dir.."/ptb.valid.txt"
test_fn = data_dir.."/ptb.test.txt"
work_dir_base = "/home/slhome/txh18/workspace/nerv/lmptb-work"
global_conf = {
lrate = 1, wcost = 1e-6, momentum = 0,
cumat_type = nerv.CuMatrixFloat,
mmat_type = nerv.MMatrixFloat,
hidden_size = 50,
batch_size = 10,
bptt = 6, --train bptt_block's words. could be set to zero
max_iter = 18,
param_random = function() return (math.random() / 5 - 0.1) end,
independent = true,
train_fn = train_fn,
valid_fn = valid_fn,
test_fn = test_fn,
sche_log_pre = "[SCHEDULER]:",
log_w_num = 1000, --give a message when log_w_num words have been processed
timer = nerv.Timer()
}
global_conf.work_dir = work_dir_base.."/h"..global_conf.hidden_size.."bp"..global_conf.bptt.."slr"..global_conf.lrate --..os.date("_%bD%dH%H") --comment this for testing
global_conf.param_fn = global_conf.work_dir.."/params"
elseif (set == "test") then
train_fn = "/slfs1/users/txh18/workspace/nerv-project/some-text"
valid_fn = "/slfs1/users/txh18/workspace/nerv-project/some-text"
test_fn = "/slfs1/users/txh18/workspace/nerv-project/some-text"
work_dir = "/slfs1/users/txh18/workspace/nerv-project/lmptb-work-play"
global_conf = {
lrate = 0.1, wcost = 1e-6, momentum = 0,
cumat_type = nerv.CuMatrixFloat,
mmat_type = nerv.MMatrixFloat,
hidden_size = 5,
batch_size = 1,
bptt = 0, --train bptt_block's words. could be set to zero
max_iter = 15,
param_random = function() return (math.random() / 5 - 0.1) end,
independent = true,
train_fn = train_fn,
valid_fn = valid_fn,
test_fn = test_fn,
work_dir = work_dir,
param_fn = work_dir .. "/params",
sche_log_pre = "[SCHEDULER]:",
log_w_num = 80000, --give a message when log_w_num words have been processed
timer = nerv.Timer()
}
end
local vocab = nerv.LMVocab()
global_conf["vocab"] = vocab
printf("%s printing global_conf...\n", global_conf.sche_log_pre)
for key, value in pairs(global_conf) do
printf("\t%s=%s\n", key, value)
end
printf("%s wait 3 seconds...\n", global_conf.sche_log_pre)
nerv.LMUtil.wait(3)
printf("%s creating work_dir...\n", global_conf.sche_log_pre)
os.execute("mkdir -p "..global_conf.work_dir)
scheduler = " printf(\"===INITIAL VALIDATION===\\n\") \
dagL, paramRepo = load_net(global_conf) \
printf(\"===INITIAL VALIDATION===\\n\") \
local result = propagateFile(global_conf, dagL, global_conf.valid_fn, {do_train = false, report_word = false}) \
ppl_rec = {} \
lr_rec = {} \
ppl_rec[0] = result:ppl_net(\"rnn\") ppl_last = ppl_rec[0] \
lr_rec[0] = 0 \
printf(\"\\n\") \
local lr_half = false \
for iter = 1, global_conf.max_iter, 1 do \
printf(\"===ITERATION %d LR %f===\\n\", iter, global_conf.lrate) \
global_conf.sche_log_pre = \"[SCHEDULER ITER\"..iter..\" LR\"..global_conf.lrate..\"]:\" \
dagL, paramRepo = load_net(global_conf) \
propagateFile(global_conf, dagL, global_conf.train_fn, {do_train = true, report_word = false}) \
printf(\"===VALIDATION %d===\\n\", iter) \
local result = propagateFile(global_conf, dagL, global_conf.valid_fn, {do_train = false, report_word = false}) \
ppl_rec[iter] = result:ppl_net(\"rnn\") \
lr_rec[iter] = global_conf.lrate \
if (ppl_last / ppl_rec[iter] < 1.03 or lr_half == true) then \
global_conf.lrate = (global_conf.lrate / 2) \
lr_half = true \
end \
if (ppl_rec[iter] < ppl_last) then \
printf(\"%s saving net to file %s...\\n\", global_conf.sche_log_pre, global_conf.param_fn) \
paramRepo:export(global_conf.param_fn, nil) \
ppl_last = ppl_rec[iter] \
else \
printf(\"%s PPL did not improve, rejected...\\n\", global_conf.sche_log_pre) \
end \
printf(\"\\n\") \
nerv.LMUtil.wait(2) \
end \
printf(\"===VALIDATION PPL record===\\n\") \
for i = 0, #ppl_rec do printf(\"<ITER%d LR%.5f: %.3f> \", i, lr_rec[i], ppl_rec[i]) end \
printf(\"\\n\") \
printf(\"===FINAL TEST===\\n\") \
global_conf.sche_log_pre = \"[SCHEDULER FINAL_TEST]:\" \
dagL, _ = load_net(global_conf) \
propagateFile(global_conf, dagL, global_conf.test_fn, {do_train = false, report_word = false})"
printf("%s printing schedule:\n", global_conf.sche_log_pre)
printf("%s\n", scheduler)
printf("%s wait 3 seconds...\n", global_conf.sche_log_pre)
nerv.LMUtil.wait(3)
--[[global settings end]]--
global_conf.vocab:build_file(global_conf.train_fn)
prepare_parameters(global_conf, true)
assert(loadstring(scheduler))()
|