Ce sentiment classification (#2389)

* add ce for models_ce

* add ce for sentiment_classification

PaddleNLP/sentiment_classification/.run_ce.sh

+#! /bin/bash
+export FLAGS_enable_parallel_graph=1
+export FLAGS_sync_nccl_allreduce=1
+export FLAGS_fraction_of_gpu_memory_to_use=0.95
+export CPU_NUM=1
+
+
+# run_train on train.tsv and do_val on test.tsv
+train() {
+    python -u run_classifier.py \
+        --task_name 'senta' \
+        --use_cuda true \
+        --do_train true \
+        --do_val true \
+        --do_infer false \
+        --batch_size 16 \
+        --data_dir ./senta_data/ \
+        --vocab_path ./senta_data/word_dict.txt \
+        --checkpoints ./save_models \
+        --save_steps 500 \
+        --validation_steps 50 \
+        --epoch 2 \
+        --senta_config_path ./senta_config.json \
+        --skip_steps 10 \
+        --random_seed 0 \
+        --enable_ce
+}
+
+
+export CUDA_VISIBLE_DEVICES=0
+train | grep "dev evaluation" | grep "ave loss" | tail -1 | awk '{print "kpis\ttrain_loss_senta_card1\t"$5"\nkpis\ttrain_acc_senta_card1\t"$8"\nkpis\teach_step_duration_senta_card1\t"$11}' | tr -d "," | python _ce.py
+sleep 20
+
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+train | grep "dev evaluation" | grep "ave loss" | tail -1 | awk '{print "kpis\ttrain_loss_senta_card4\t"$5"\nkpis\ttrain_acc_senta_card4\t"$8"\nkpis\teach_step_duration_senta_card4\t"$11}' | tr -d "," | python _ce.py

PaddleNLP/sentiment_classification/_ce.py

+# 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_step_duration_senta_card1 = DurationKpi('each_step_duration_senta_card1', 0.01, 0, actived=False)
+train_loss_senta_card1 = CostKpi('train_loss_senta_card1', 0.05, 0, actived=True)
+train_acc_senta_card1 = AccKpi('train_acc_senta_card1', 0.02, 0, actived=True)
+each_step_duration_senta_card4 = DurationKpi('each_step_duration_senta_card4', 0.01, 0, actived=False)
+train_loss_senta_card4 = CostKpi('train_loss_senta_card4', 0.05, 0, actived=True)
+train_acc_senta_card4 = AccKpi('train_acc_senta_card4', 0.02, 0, actived=True)
+
+tracking_kpis = [
+        each_step_duration_senta_card1,
+        train_loss_senta_card1,
+        train_acc_senta_card1,
+        each_step_duration_senta_card4,
+        train_loss_senta_card4,
+        train_acc_senta_card4,
+]
+
+
+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)

PaddleNLP/sentiment_classification/run_classifier.py

@@ -58,6 +58,7 @@ run_type_g.add_arg("task_name", str, None,
 run_type_g.add_arg("do_train", bool, True, "Whether to perform training.")
 run_type_g.add_arg("do_val", bool, True, "Whether to perform evaluation.")
 run_type_g.add_arg("do_infer", bool, True, "Whether to perform inference.")
+parser.add_argument('--enable_ce', action='store_true', help='If set, run the task with continuous evaluation logs.')
 
 args = parser.parse_args()
 # yapf: enable.
@@ -199,6 +200,8 @@ def main(args):
         print("Max train steps: %d" % max_train_steps)
 
         train_program = fluid.Program()
+        if args.enable_ce and args.random_seed is not None:
+            train_program.random_seed = args.random_seed
 
         with fluid.program_guard(train_program, startup_prog):
             with fluid.unique_name.guard():
@@ -364,6 +367,7 @@ def main(args):
         inference(exe, infer_prog, infer_pyreader,
             [prop.name], "infer")
 
+
 if __name__ == "__main__":
     print_arguments(args)
     main(args)