cma_approach_square_size.py 15 KB
Newer Older
harry1080ti's avatar
harry1080ti committed
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
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
import cma
from multiprocessing import Pool
from os import cpu_count
import time
import path_constant as pc

class cma_approach(object):
    def __init__(self,
        # data path
        path_to_datasrc = "alexnet_data.csv",
        path_to_topology = "alexnet.csv",
        target_col = "Cycles",

        # problem definition
        number_of_partition = 4, max_iteration = 100,
        sigma = 0.5, population_size = 10,

        # constraint
        max_res_unit = 960, initial_res = 0,
        res_step = 1,
        penalty_offest = 10000000000,
        seeding_type="optimised",
        hybird = True
        ):
        self.target_col = target_col
        self.start = time.time()
        self.k = number_of_partition
        self.max_iter = max_iteration
        self.sigma = sigma
        self.max_res_unit = max_res_unit
        self.res_step = res_step
        self.population_size = population_size
        self.penalty_offest = penalty_offest
        self.ending_iter = 0
        self.is_hybird = hybird
        self.data_src = {}

        self.topology_file = path_to_topology
        self.layers = self.parse_topology_file()
        self.parse_data_set_file(path_to_datasrc)

        self.best_layer = number_of_partition * [0]
        self.best_res = number_of_partition * [0]

        self.total_valid_solution = 0
        self.trial = 1
        self.seeding_type = seeding_type

        self.max_res_available = 1920*9

    def parse_topology_file(self):
        layers = []
        with open(pc.TOPOLOGIES_PATH+self.topology_file, 'r') as f:
            next(f)
            for line in f:
                elems = line.strip().split(',')
                layers.append(elems[0])

        for layer in layers:
            self.data_src[layer] = {}
        return layers

    def parse_data_set_file(self, path_to_data_csv):
        first = True
        target_idx = 2
        with open(pc.DATA_SOURCE_PATH+path_to_data_csv, 'r') as f:
            for line in f:
                elems = line.strip().split(',')
                # print(elems)
                if first:
                    for idx, col in enumerate(elems):
                        if self.target_col in col:
                            target_idx = idx
                            break
                    first = False
                else:
                    self.data_src[elems[1]][int(elems[0])] = int(float(elems[target_idx]))

    def regroup_layers(self, sample):
        # #print("DEBUG", sample)
        detail_sample = []
        idx = 0
        for size in sample:
            part = []

            if size == 1:
                part.append(self.layers[idx])
                idx += 1
            else:
                for i in range(0, size):
                    part.append(self.layers[i + idx])
                idx += size

            detail_sample.append(part)

        return detail_sample

    def decode(self, val, max_val):
        return int(val * max_val)

    def encode(self, val, max_val):
        return float(val / max_val)

    def filter_layer(self, layer):
        for idx in range(self.k):
            if layer[idx] <= 0:
                return False

        if sum(layer) != len(self.layers):
            return False

        return True

    def filter_res(self, res):
        # #print(layer, res)
        for idx in range(self.k):
            if res[idx] <= 0:
                return False

        if sum(res) != self.max_res_unit:
            return False

        return True

    def penalty_layer(self, layer):
        penalty_score = self.penalty_offest

        if sum(layer) != len(self.layers):
            penalty_score += self.penalty_offest
        else:
            layer = [abs(val) for val in layer]

        for idx in range(self.k):
            if layer[idx] <= 0:
                penalty_score *= 1.05

        percent_diff = (abs(sum(layer) - len(self.layers)) / len(self.layers))
        penalty_score += percent_diff * self.penalty_offest

        return penalty_score

    # def penalty_res(self, res):
    #     penalty_score = self.penalty_offest

    #     if sum(res) != self.max_res_unit:
    #         penalty_score += self.penalty_offest
    #     else:
    #         res = [abs(val) for val in res]

    #     for idx in range(self.k):
    #         if res[idx] <= 0:
    #             penalty_score *= 1.05

    #     percent_diff = abs(sum(res) - self.max_res_unit) / self.max_res_unit
    #     penalty_score += percent_diff * self.penalty_offest

    #     return penalty_score

    def find_max_latency(self, layer_partition, res_partitions):
        latencies = [0] * len(layer_partition)
        max_latency_idx = 0
        # print(layer_partition)
        # print(res_partitions)
        for idx, part in enumerate(layer_partition):
            res = res_partitions[idx]
            for layer in part:
                latencies[idx] += self.data_src[layer][res]
            if latencies[idx] > latencies[max_latency_idx]:
                max_latency_idx = idx

        return latencies, max_latency_idx

    def eva_hybird_sq(self, layer):
        res = [self.res_step] * self.k
        latencies = []

        while sum([r*r for r in res]) < self.max_res_unit:
            latencies, max_idx = self.find_max_latency(layer, res)
            res[max_idx] += self.res_step

        # for i in range(0, int(self.max_res_unit/self.res_step - self.k*self.res_step)):
        #     latencies, max_idx = self.find_max_latency(layer, res)
        #     res[max_idx] += self.res_step

        return latencies[max_idx], latencies, res, layer

    # not really in used
    # def evaluate_full_relaxed(self, layer):
    #     seed = []

    #     for i in range(self.k - 1):
    #         seed.append(int(self.max_res_unit/self.k))
    #     seed.append(self.max_res_unit - sum(seed))
    #     # #print(seed)
    #     seed = [self.encode(val, self.max_res_unit) for val in seed[:-1]]

    #     es_res = cma.CMAEvolutionStrategy(seed, \
    #         self.sigma, {'popsize' : self.population_size})
    #     i = 0
    #     while not es_res.stop() and i < self.max_iter:
    #         samples = es_res.ask()
    #         scores = [0] * es_res.popsize
    #         res = [0] * es_res.popsize

    #         for idx, sample in enumerate(samples):
    #             res_assign = [self.decode(val, self.max_res_unit) for val in sample]
    #             res_assign.append(self.max_res_unit - sum(res_assign))
    #             res[idx] = res_assign

    #         for idx, r in enumerate(res):
    #             if self.filter_res(r):
    #                 latencies, max_idx = self.find_max_latency(layer, r)
    #                 scores[idx] = latencies[max_idx]
    #             else:
    #                 scores[idx] = self.penalty_res(r)

    #         # for idx in range(self.population_size):
    #         #     #print(samples[idx], scores[idx])

    #         es_res.tell(samples, scores)
    #         i += 1

    #     res = [self.decode(val, self.max_res_unit) for val in es_res.result[0]]
    #     res.append(self.max_res_unit - sum(res))

    #     if self.filter_res(r):
    #         latencies, max_idx = self.find_max_latency(layer, res)
    #     else:
    #         max_latency = self.penalty_res(r)
    #         latencies = [max_latency]*self.k
    #         max_idx = 0

    #     return latencies[max_idx], latencies, res, layer

    def evaluation_top_level(self, in_val):
        pid, sampling = in_val
        layer = [self.decode(val, len(self.layers)) for val in sampling]
        layer.append(len(self.layers) - sum(layer))
        penalty = 0

        if not self.filter_layer(layer):
            penalty = self.penalty_layer(layer)
            if self.is_hybird:
                return pid, penalty
            else:
                return pid, penalty*4

        layer = self.regroup_layers(layer)
        # if self.is_hybird:
        #     return pid, self.eva_hybird_sq(layer)[0]
        # else:
        #     score, _, res, _ = self.evaluate_full_relaxed(layer)
        #     return pid, score, res

        return pid, self.eva_hybird_sq(layer)[0]

    def run(self):
        self.trial += self.trial
        if (self.seeding_type=="allzeros"):
            self.seed = [0]*(self.k-1)
            self.seed_od = self.seed
        elif (self.seeding_type=="optimised"):
            self.seed = []
            for i in range(self.k - 1):
                self.seed.append(int(len(self.layers)/self.k))
            self.seed.append(len(self.layers) - sum(self.seed))
            self.seed_od = self.seed
            self.seed = [self.encode(val, len(self.layers)) for val in self.seed[:-1]]
        else:
            raise ValueError('Invalid Seeding Strategy')



        self.es = cma.CMAEvolutionStrategy(self.seed, self.sigma, \
            {'popsize' : self.population_size})

        best_overall = self.penalty_offest
        self.i = 0
        while not self.es.stop() and self.i < self.max_iter:
            samples = self.es.ask()
            id_list = [(idx, sample) for idx, sample in enumerate(samples)]
            scores = [0] * self.es.popsize
            invalid_sampling = 0

            res_combintaions = [0] * self.es.popsize

            pool = Pool(processes = 1)#cpu_count() - 4)
            for result in pool.imap_unordered(self.evaluation_top_level, id_list):
                scores[result[0]] = result[1]
                if result[1] >= self.penalty_offest:
                    invalid_sampling += 1
                else:
                    if not self.is_hybird:
                        res_combintaions[result[0]] = result[2]
            pool.close()
            pool.join()

            if not self.is_hybird:
                best_in_iteration = min(scores)

                if best_in_iteration < best_overall and best_in_iteration < self.penalty_offest:
                    best_overall = best_in_iteration
                    self.best_res = res_combintaions[scores.index(min(scores))]

            ##print(str(self.i) + ":", \
            #    "Sigma:",round(self.es.sigma, 4), \
            #    "|| Valid sampling percentage:", \
            #        (self.population_size - invalid_sampling) /self.population_size *100)
            ##print("invalid sampling", invalid_sampling)
            self.valid_sampling_percentage = (self.population_size - invalid_sampling) /self.population_size *100
            self.total_valid_solution += self.population_size - invalid_sampling
            self.samples = samples
            self.scores = scores
            self.es.tell(samples, scores)

            self.end = time.time()
            self.best_layer = [self.decode(val, len(self.layers)) for val in self.es.result[0]]
            self.best_layer.append(len(self.layers) - sum(self.best_layer))
            self.report()
            self.i += 1

        self.ending_iter = self.i

    def report(self):
        ##print(self.i, self.es.sigma)
        max_latency = 0
        layer = []
        res = []
        latencies = []

        if self.is_hybird:
            if not self.filter_layer(self.best_layer):
                ##print("RESULT NOT VALID")
                ##print("Layer:", self.best_layer, "sum: ", sum(self.best_layer))
                #print(self.penalty_layer(self.best_layer))

                with open(pc.RESULT_CSV_PATH+'cma_logmore_sq.csv', 'a') as csvFile:
                    writer = csv.writer(csvFile, delimiter=',', lineterminator="\n")
                    writer.writerow([self.target_col,self.i,self.k, self.topology_file, 0,0, 0, 0, 0, 0, 0, layer, res, self.end-self.start, self.es.sigma, self.seed_od,self.valid_sampling_percentage, self.trial, self.population_size, self.max_res_unit, self.seeding_type])
                csvFile.close

                return False

            layer = self.regroup_layers(self.best_layer)
            max_latency, latencies, res, layers = self.eva_hybird_sq(layer)
        else:
            if not self.filter_res(self.best_res) and not self.filter_layer(self.best_layer):
                #print("RESULT NOT VALID")
                #print("Layer:", self.best_layer, "sum: ", sum(self.best_layer))
                #print("Res:", self.best_res, "sum: ", sum(self.best_res))
                return False

            layer = self.regroup_layers(self.best_layer)
            res = self.best_res
            latencies, max_idx = self.find_max_latency(layer, self.best_res)
            max_latency = latencies[max_idx]

        # generate data for mapping the full array
        full_latency, full_max_idx = self.find_max_latency([self.layers], [129]*len(self.layers))

        # PLEASE UNCOMMENT OUT THIS PART IF YOU NOT USING THE BASH SCRIPT WE HAVE PROVIDED 
        print("================================= RESULT =================================")
        print("Solution: (out of", self.total_valid_solution, "solutions)")
        print(layer)
        print("Res mapping:")
        print(res)
        print("Latency for each partition: ")
        print(latencies)
        print("Final Latency:", max_latency*self.k, "|| Throught put:", 1/max_latency)
        print("==========================================================================")
        print("Map to full array (", self.max_res_unit, ")")
        print("Final Latency:", full_latency[full_max_idx], "|| Throught put:", 1/full_latency[full_max_idx])
        print("==========================================================================")
        print("Throughtput Ratio:", (1/max_latency)/(1/full_latency[full_max_idx]))
        print("Latency increase:", (max_latency*self.k)/full_latency[full_max_idx])

        with open(pc.RESULT_CSV_PATH+'cma_logmore_sq.csv', 'a') as csvFile:
            writer = csv.writer(csvFile, delimiter=',', lineterminator="\n")
            writer.writerow([self.target_col,self.i,self.k, self.topology_file, 1,(1/max_latency), max_latency*self.k, 1/full_latency[full_max_idx], full_latency[full_max_idx], (1/max_latency)/(1/full_latency[full_max_idx]), (max_latency*self.k)/full_latency[full_max_idx], layer, res, self.end-self.start, self.es.sigma, self.seed_od,self.valid_sampling_percentage, self.trial, self.population_size, self.max_res_unit, self.seeding_type])
        csvFile.close
        return True

if __name__ == "__main__":
    import csv
    import sys

    topology = sys.argv[1]
    k = int(sys.argv[2])
    population_size = int(sys.argv[3])
    max_res_unit = int(sys.argv[4])
    seeding_type = sys.argv[5]
    target_col = sys.argv[6]

    es_hybird = cma_approach(
        path_to_datasrc = str(topology)+"_square_mem_bound.csv",
        path_to_topology = str(topology)+".csv",
        target_col = str(target_col),

        number_of_partition = k, max_iteration = 10000,
        sigma = 0.5, population_size = population_size,

        max_res_unit = max_res_unit, initial_res = 0,
        res_step = 3,
        penalty_offest = 100000000000,
        seeding_type = seeding_type,
        hybird = True
    )

    trials = 1
    #print("======== HYBRID ======== ( k:", k, "trials:", trials, ")")
    es_hybird.run()
    while not es_hybird.report() and trials < 20:
        #print("======== HYBRID ======== ( k:", k, "trials:", trials, ")")
        es_hybird.run()
        trials += 1

    k += 1
    #print("convergence takes", trials, "trials")