diff --git a/fluid/PaddleRec/tagspace/.run_ce.sh b/fluid/PaddleRec/tagspace/.run_ce.sh new file mode 100755 index 0000000000000000000000000000000000000000..74a0413a846f21cc7eaacd444f527103778be923 --- /dev/null +++ b/fluid/PaddleRec/tagspace/.run_ce.sh @@ -0,0 +1,20 @@ +#!/bin/bash + +export MKL_NUM_THREADS=1 +export OMP_NUM_THREADS=1 + + +export CPU_NUM=1 +export NUM_THREADS=1 + +FLAGS_benchmark=true python train.py --enable_ce --train_dir train_big_data/ --vocab_text_path big_vocab_text.txt --vocab_tag_path big_vocab_tag.txt --model_dir big_model --batch_size 500 | python _ce.py + +cudaid=${tagspace:=0} # use 0-th card as default +export CUDA_VISIBLE_DEVICES=$cudaid + +FLAGS_benchmark=true python train.py --enable_ce --use_cuda 1 --train_dir train_big_data/ --vocab_text_path big_vocab_text.txt --vocab_tag_path big_vocab_tag.txt --model_dir big_model --batch_size 500 --parallel 1 | python _ce.py + +cudaid=${tagspace_4:=0,1,2,3} # use 0-th card as default +export CUDA_VISIBLE_DEVICES=$cudaid + +FLAGS_benchmark=true python train.py --enable_ce --use_cuda 1 --train_dir train_big_data/ --vocab_text_path big_vocab_text.txt --vocab_tag_path big_vocab_tag.txt --model_dir big_model --batch_size 500 --parallel 1 | python _ce.py diff --git a/fluid/PaddleRec/tagspace/__init.py__ b/fluid/PaddleRec/tagspace/__init.py__ new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/fluid/PaddleRec/tagspace/_ce.py b/fluid/PaddleRec/tagspace/_ce.py new file mode 100644 index 0000000000000000000000000000000000000000..b75fa39a2114c9388ffad77be74f52733010aeba --- /dev/null +++ b/fluid/PaddleRec/tagspace/_ce.py @@ -0,0 +1,66 @@ +# this file is only used for continuous evaluation test! + +import os +import sys +sys.path.append(os.environ['ceroot']) +from kpi import CostKpi +from kpi import DurationKpi +from kpi import AccKpi + + +each_pass_duration_cpu1_thread1_kpi = DurationKpi('each_pass_duration_cpu1_thread1', 0.08, 0, actived=True) +train_acc_cpu1_thread1_kpi = AccKpi('train_acc_cpu1_thread1', 0.08, 0) +each_pass_duration_gpu1_kpi = DurationKpi('each_pass_duration_gpu1', 0.08, 0, actived=True) +train_acc_gpu1_kpi = AccKpi('train_acc_gpu1', 0.08, 0) +each_pass_duration_gpu4_kpi = DurationKpi('each_pass_duration_gpu4', 0.08, 0, actived=True) +train_acc_gpu4_kpi = AccKpi('train_acc_gpu4', 0.08, 0) + +tracking_kpis = [ + each_pass_duration_cpu1_thread1_kpi, + train_acc_cpu1_thread1_kpi, + each_pass_duration_gpu1_kpi, + train_acc_gpu1_kpi, + each_pass_duration_gpu4_kpi, + train_acc_gpu4_kpi, + ] + + +def parse_log(log): + ''' + This method should be implemented by model developers. + + The suggestion: + + each line in the log should be key, value, for example: + + " + train_cost\t1.0 + test_cost\t1.0 + train_cost\t1.0 + train_cost\t1.0 + train_acc\t1.2 + " + ''' + for line in log.split('\n'): + fs = line.strip().split('\t') + print(fs) + if len(fs) == 3 and fs[0] == 'kpis': + kpi_name = fs[1] + kpi_value = float(fs[2]) + yield kpi_name, kpi_value + + +def log_to_ce(log): + kpi_tracker = {} + for kpi in tracking_kpis: + kpi_tracker[kpi.name] = kpi + + for (kpi_name, kpi_value) in parse_log(log): + print(kpi_name, kpi_value) + kpi_tracker[kpi_name].add_record(kpi_value) + kpi_tracker[kpi_name].persist() + + +if __name__ == '__main__': + log = sys.stdin.read() + log_to_ce(log) diff --git a/fluid/PaddleRec/tagspace/train.py b/fluid/PaddleRec/tagspace/train.py index 914c824c134a8c60c790dc5473431215924e0dff..419bb1c4b156c148f8bc4bc3a48385b6722f5c68 100644 --- a/fluid/PaddleRec/tagspace/train.py +++ b/fluid/PaddleRec/tagspace/train.py @@ -40,6 +40,10 @@ def parse_args(): '--base_lr', type=float, default=0.01, help='learning rate') parser.add_argument( '--num_devices', type=int, default=1, help='Number of GPU devices') + parser.add_argument( + '--enable_ce', + action='store_true', + help='If set, run the task with continuous evaluation logs.') args = parser.parse_args() return args @@ -51,6 +55,9 @@ def get_cards(args): def train(): """ do training """ args = parse_args() + if args.enable_ce: + fluid.default_startup_program().random_seed = SEED + fluid.default_main_program().random_seed = SEED train_dir = args.train_dir vocab_text_path = args.vocab_text_path vocab_tag_path = args.vocab_tag_path @@ -91,6 +98,7 @@ def train(): model_dir = args.model_dir fetch_list = [avg_cost.name] total_time = 0.0 + ce_info = [] for pass_idx in range(pass_num): epoch_idx = pass_idx + 1 print("epoch_%d start" % epoch_idx) @@ -106,6 +114,7 @@ def train(): "neg_tag": lod_neg_tag }, fetch_list=[avg_cost.name, correct.name]) + ce_info.append(float(np.sum(correct_val)) / (args.num_devices * batch_size)) if batch_id % args.print_batch == 0: print("TRAIN --> pass: {} batch_num: {} avg_cost: {}, acc: {}" .format(pass_idx, (batch_id + 10) * batch_size, @@ -120,9 +129,43 @@ def train(): feed_var_names = ["text", "pos_tag"] fetch_vars = [cos_pos] fluid.io.save_inference_model(save_dir, feed_var_names, fetch_vars, - train_exe) + exe) + # only for ce + if args.enable_ce: + ce_acc = 0 + try: + ce_acc = ce_info[-2] + except: + logger.error("ce info error") + epoch_idx = args.pass_num + device = get_device(args) + if args.use_cuda: + gpu_num = device[1] + print("kpis\teach_pass_duration_gpu%s\t%s" % + (gpu_num, total_time / epoch_idx)) + print("kpis\ttrain_acc_gpu%s\t%s" % + (gpu_num, ce_acc)) + else: + cpu_num = device[1] + threads_num = device[2] + print("kpis\teach_pass_duration_cpu%s_thread%s\t%s" % + (cpu_num, threads_num, total_time / epoch_idx)) + print("kpis\ttrain_acc_cpu%s_thread%s\t%s" % + (cpu_num, threads_num, ce_acc)) + print("finish training") +def get_device(args): + if args.use_cuda: + gpus = os.environ.get("CUDA_VISIBLE_DEVICES", 1) + gpu_num = len(gpus.split(',')) + return "gpu", gpu_num + else: + threads_num = os.environ.get('NUM_THREADS', 1) + cpu_num = os.environ.get('CPU_NUM', 1) + return "cpu", int(cpu_num), int(threads_num) + + if __name__ == "__main__": train()