diff --git a/fluid/PaddleRec/gru4rec/.run_ce.sh b/fluid/PaddleRec/gru4rec/.run_ce.sh new file mode 100755 index 0000000000000000000000000000000000000000..902b49f1335ea0f527dab4c4826d36ead04d7680 --- /dev/null +++ b/fluid/PaddleRec/gru4rec/.run_ce.sh @@ -0,0 +1,22 @@ +#!/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 --train_dir train_big_data --vocab_path vocab_big.txt --use_cuda 0 --batch_size 500 --model_dir model_output --pass_num 2 --enable_ce --step_num 10 | python _ce.py + + +cudaid=${gru4rec:=0} # use 0-th card as default +export CUDA_VISIBLE_DEVICES=$cudaid + +FLAGS_benchmark=true python train.py --train_dir train_big_data --vocab_path vocab_big.txt --use_cuda 1 --batch_size 500 --model_dir model_output --pass_num 2 --enable_ce --step_num 1000 | python _ce.py + + +cudaid=${gru4rec_4:=0,1,2,3} # use 0-th card as default +export CUDA_VISIBLE_DEVICES=$cudaid + +FLAGS_benchmark=true python train.py --train_dir train_big_data --vocab_path vocab_big.txt --use_cuda 1 --parallel 1 --num_devices 2 --batch_size 500 --model_dir model_output --pass_num 2 --enable_ce --step_num 1000 | python _ce.py diff --git a/fluid/PaddleRec/gru4rec/__init__.py b/fluid/PaddleRec/gru4rec/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/fluid/PaddleRec/gru4rec/_ce.py b/fluid/PaddleRec/gru4rec/_ce.py new file mode 100644 index 0000000000000000000000000000000000000000..90ddbf787a32e819c76394d44c5f9a41e3b08685 --- /dev/null +++ b/fluid/PaddleRec/gru4rec/_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_ppl_cpu1_thread1_kpi = CostKpi('train_ppl_cpu1_thread1', 0.08, 0) +each_pass_duration_gpu1_kpi = DurationKpi('each_pass_duration_gpu1', 0.08, 0, actived=True) +train_ppl_gpu1_kpi = CostKpi('train_ppl_gpu1', 0.08, 0) +each_pass_duration_gpu4_kpi = DurationKpi('each_pass_duration_gpu4', 0.08, 0, actived=True) +train_ppl_gpu4_kpi = CostKpi('train_ppl_gpu4', 0.08, 0) + +tracking_kpis = [ + each_pass_duration_cpu1_thread1_kpi, + train_ppl_cpu1_thread1_kpi, + each_pass_duration_gpu1_kpi, + train_ppl_gpu1_kpi, + each_pass_duration_gpu4_kpi, + train_ppl_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/gru4rec/train.py b/fluid/PaddleRec/gru4rec/train.py index 568f497085ec1574962b0c5a86906a4f07209cd6..b43926b69eaf002380a261a0689be91ec3f6ff90 100644 --- a/fluid/PaddleRec/gru4rec/train.py +++ b/fluid/PaddleRec/gru4rec/train.py @@ -40,6 +40,12 @@ 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( + '--step_num', type=int, default=1000, help='Number of steps') + 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 +57,9 @@ 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 hid_size = args.hid_size train_dir = args.train_dir vocab_path = args.vocab_path @@ -84,6 +93,7 @@ def train(): model_dir = args.model_dir fetch_list = [avg_cost.name] + ce_info = [] total_time = 0.0 for pass_idx in six.moves.xrange(pass_num): epoch_idx = pass_idx + 1 @@ -105,8 +115,11 @@ def train(): fetch_list=fetch_list) avg_ppl = np.exp(ret_avg_cost[0]) newest_ppl = np.mean(avg_ppl) + ce_info.append(newest_ppl) if i % args.print_batch == 0: print("step:%d ppl:%.3f" % (i, newest_ppl)) + if args.enable_ce and i > args.step_num: + break t1 = time.time() total_time += t1 - t0 @@ -117,8 +130,43 @@ def train(): fetch_vars = [avg_cost, acc] fluid.io.save_inference_model(save_dir, feed_var_names, fetch_vars, exe) print("model saved in %s" % save_dir) + + # only for ce + if args.enable_ce: + ce_ppl = 0 + try: + ce_ppl = ce_info[-2] + except: + print("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_ppl_gpu%s\t%s" % + (gpu_num, ce_ppl)) + 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_ppl_cpu%s_thread%s\t%s" % + (cpu_num, threads_num, ce_ppl)) + 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()