Merge pull request #19 from PaddlePaddle/develop

update

.gitmodules

@@ -7,3 +7,6 @@
 [submodule "fluid/PaddleNLP/Senta"]
 	path = fluid/PaddleNLP/Senta
 	url = https://github.com/baidu/Senta.git
+[submodule "fluid/PaddleNLP/LARK"]
+	path = fluid/PaddleNLP/LARK
+	url = https://github.com/PaddlePaddle/LARK

README.md

@@ -43,6 +43,7 @@ PaddlePaddle 提供了丰富的计算单元，使得用户可以采用模块化
 模型|简介|模型优势|参考论文
 --|:--:|:--:|:--:
 [Transformer](./fluid/PaddleNLP/neural_machine_translation/transformer/README_cn.md)|机器翻译模型|基于self-attention，计算复杂度小，并行度高，容易学习长程依赖，翻译效果更好|[Attention Is All You Need](https://arxiv.org/abs/1706.03762)
+[BERT](https://github.com/PaddlePaddle/LARK/tree/develop/BERT)|语义表示模型|在多个 NLP 任务上取得 SOTA 效果，支持多卡多机训练，支持混合精度训练|[BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805)
 [LAC](https://github.com/baidu/lac/blob/master/README.md)|联合的词法分析模型|能够整体性地完成中文分词、词性标注、专名识别任务|[Chinese Lexical Analysis with Deep Bi-GRU-CRF Network](https://arxiv.org/abs/1807.01882)
 [Senta](https://github.com/baidu/Senta/blob/master/README.md)|情感倾向分析模型集|百度AI开放平台中情感倾向分析模型|-
 [DAM](./fluid/PaddleNLP/deep_attention_matching_net)|语义匹配模型|百度自然语言处理部发表于ACL-2018的工作,用于检索式聊天机器人多轮对话中应答的选择|[Multi-Turn Response Selection for Chatbots with Deep Attention Matching Network](http://aclweb.org/anthology/P18-1103)

fluid/PaddleCV/face_detection/widerface_eval.py

@@ -44,11 +44,15 @@ def infer(args, config):
         shrink, max_shrink = get_shrink(image.size[1], image.size[0])
 
         det0 = detect_face(image, shrink)
-        det1 = flip_test(image, shrink)
-        [det2, det3] = multi_scale_test(image, max_shrink)
-        det4 = multi_scale_test_pyramid(image, max_shrink)
-        det = np.row_stack((det0, det1, det2, det3, det4))
-        dets = bbox_vote(det)
+        if args.use_gpu:
+            det1 = flip_test(image, shrink)
+            [det2, det3] = multi_scale_test(image, max_shrink)
+            det4 = multi_scale_test_pyramid(image, max_shrink)
+            det = np.row_stack((det0, det1, det2, det3, det4))
+            dets = bbox_vote(det)
+        else:
+            # when infer on cpu, use a simple case
+            dets = det0
 
         keep_index = np.where(dets[:, 4] >= args.confs_threshold)[0]
         dets = dets[keep_index, :]

fluid/PaddleCV/gan/cycle_gan/README.md

@@ -74,8 +74,8 @@ env CUDA_VISIBLE_DEVICES=0 python train.py
 
 ```
 env CUDA_VISIBLE_DEVICE=0 python infer.py \
-    --init_model="models/1" --input="./data/inputA/*" \
-    --output="./output"
+    --init_model="checkpoints/1" --input="./data/inputA/*" \
+    --input_style A --output="./output"
 ```
 
 训练150轮的模型预测效果如图2和图3所示：

fluid/PaddleCV/gan/cycle_gan/infer.py

@@ -26,8 +26,10 @@ def infer(args):
     data_shape = [-1, 3, 256, 256]
     input = fluid.layers.data(name='input', shape=data_shape, dtype='float32')
     if args.input_style == "A":
+        model_name = 'g_a'
         fake = build_generator_resnet_9blocks(input, name="g_A")
     elif args.input_style == "B":
+        model_name = 'g_b'
         fake = build_generator_resnet_9blocks(input, name="g_B")
     else:
         raise "Input with style [%s] is not supported." % args.input_style
@@ -37,7 +39,7 @@ def infer(args):
         place = fluid.CUDAPlace(0)
     exe = fluid.Executor(place)
     exe.run(fluid.default_startup_program())
-    fluid.io.load_persistables(exe, args.init_model)
+    fluid.io.load_persistables(exe, args.init_model + "/" + model_name)
 
     if not os.path.exists(args.output):
         os.makedirs(args.output)

fluid/PaddleCV/gan/cycle_gan/layers.py

@@ -3,10 +3,12 @@ import paddle.fluid as fluid
 import numpy as np
 import os
 
-use_cudnn = True
+# cudnn is not better when batch size is 1.
+use_cudnn = False
 if 'ce_mode' in os.environ:
     use_cudnn = False
 
+
 def cal_padding(img_size, stride, filter_size, dilation=1):
     """Calculate padding size."""
     valid_filter_size = dilation * (filter_size - 1) + 1
@@ -18,6 +20,8 @@ def cal_padding(img_size, stride, filter_size, dilation=1):
 
 
 def instance_norm(input, name=None):
+    # TODO(lvmengsi@baidu.com): Check the accuracy when using fluid.layers.layer_norm.
+    # return fluid.layers.layer_norm(input, begin_norm_axis=2) 
     helper = fluid.layer_helper.LayerHelper("instance_norm", **locals())
     dtype = helper.input_dtype()
     epsilon = 1e-5

fluid/PaddleCV/gan/cycle_gan/train.py

@@ -17,7 +17,6 @@ import data_reader
 from utility import add_arguments, print_arguments, ImagePool
 from trainer import *
 
-
 parser = argparse.ArgumentParser(description=__doc__)
 add_arg = functools.partial(add_arguments, argparser=parser)
 # yapf: disable
@@ -36,7 +35,7 @@ add_arg('run_ce',            bool,  False,       "Whether to run for model ce.")
 def train(args):
 
     max_images_num = data_reader.max_images_num()
-    shuffle=True
+    shuffle = True
     if args.run_ce:
         np.random.seed(10)
         fluid.default_startup_program().random_seed = 90
@@ -66,9 +65,11 @@ def train(args):
     exe.run(fluid.default_startup_program())
     A_pool = ImagePool()
     B_pool = ImagePool()
-    
-    A_reader = paddle.batch(data_reader.a_reader(shuffle=shuffle), args.batch_size)()
-    B_reader = paddle.batch(data_reader.b_reader(shuffle=shuffle), args.batch_size)()
+
+    A_reader = paddle.batch(
+        data_reader.a_reader(shuffle=shuffle), args.batch_size)()
+    B_reader = paddle.batch(
+        data_reader.b_reader(shuffle=shuffle), args.batch_size)()
     if not args.run_ce:
         A_test_reader = data_reader.a_test_reader()
         B_test_reader = data_reader.b_test_reader()
@@ -119,13 +120,13 @@ def train(args):
         if not os.path.exists(out_path):
             os.makedirs(out_path)
         fluid.io.save_persistables(
-            exe, out_path + "/g_a", main_program=g_A_trainer.program, filename="params")
+            exe, out_path + "/g_a", main_program=g_A_trainer.program)
         fluid.io.save_persistables(
-            exe, out_path + "/g_b", main_program=g_B_trainer.program, filename="params")
+            exe, out_path + "/g_b", main_program=g_B_trainer.program)
         fluid.io.save_persistables(
-            exe, out_path + "/d_a", main_program=d_A_trainer.program, filename="params")
+            exe, out_path + "/d_a", main_program=d_A_trainer.program)
         fluid.io.save_persistables(
-            exe, out_path + "/d_b", main_program=d_B_trainer.program, filename="params")
+            exe, out_path + "/d_b", main_program=d_B_trainer.program)
         print("saved checkpoint to {}".format(out_path))
         sys.stdout.flush()
 
@@ -144,8 +145,21 @@ def train(args):
 
     if args.init_model:
         init_model()
-    losses=[[], []]
+    losses = [[], []]
     t_time = 0
+
+    g_A_trainer_program = fluid.CompiledProgram(
+        g_A_trainer.program).with_data_parallel(
+            loss_name=g_A_trainer.g_loss_A.name)
+    g_B_trainer_program = fluid.CompiledProgram(
+        g_B_trainer.program).with_data_parallel(
+            loss_name=g_B_trainer.g_loss_B.name)
+    d_B_trainer_program = fluid.CompiledProgram(
+        d_B_trainer.program).with_data_parallel(
+            loss_name=d_B_trainer.d_loss_B.name)
+    d_A_trainer_program = fluid.CompiledProgram(
+        d_A_trainer.program).with_data_parallel(
+            loss_name=d_A_trainer.d_loss_A.name)
     for epoch in range(args.epoch):
         batch_id = 0
         for i in range(max_images_num):
@@ -158,7 +172,7 @@ def train(args):
             s_time = time.time()
             # optimize the g_A network
             g_A_loss, fake_B_tmp = exe.run(
-                g_A_trainer.program,
+                g_A_trainer_program,
                 fetch_list=[g_A_trainer.g_loss_A, g_A_trainer.fake_B],
                 feed={"input_A": tensor_A,
                       "input_B": tensor_B})
@@ -167,14 +181,14 @@ def train(args):
 
             # optimize the d_B network
             d_B_loss = exe.run(
-                d_B_trainer.program,
+                d_B_trainer_program,
                 fetch_list=[d_B_trainer.d_loss_B],
                 feed={"input_B": tensor_B,
                       "fake_pool_B": fake_pool_B})[0]
 
             # optimize the g_B network
             g_B_loss, fake_A_tmp = exe.run(
-                g_B_trainer.program,
+                g_B_trainer_program,
                 fetch_list=[g_B_trainer.g_loss_B, g_B_trainer.fake_A],
                 feed={"input_A": tensor_A,
                       "input_B": tensor_B})
@@ -183,16 +197,16 @@ def train(args):
 
             # optimize the d_A network
             d_A_loss = exe.run(
-                d_A_trainer.program,
+                d_A_trainer_program,
                 fetch_list=[d_A_trainer.d_loss_A],
                 feed={"input_A": tensor_A,
                       "fake_pool_A": fake_pool_A})[0]
             batch_time = time.time() - s_time
             t_time += batch_time
-            print("epoch{}; batch{}; g_A_loss: {}; d_B_loss: {}; g_B_loss: {}; d_A_loss: {}; "
-                  "Batch_time_cost: {:.2f}".format(
-                epoch, batch_id, g_A_loss[0], d_B_loss[0], g_B_loss[0],
-                d_A_loss[0], batch_time))
+            print(
+                "epoch{}; batch{}; g_A_loss: {}; d_B_loss: {}; g_B_loss: {}; d_A_loss: {}; "
+                "Batch_time_cost: {:.2f}".format(epoch, batch_id, g_A_loss[
+                    0], d_B_loss[0], g_B_loss[0], d_A_loss[0], batch_time))
             losses[0].append(g_A_loss[0])
             losses[1].append(d_A_loss[0])
             sys.stdout.flush()

LARK @ 8dbdf489

+Subproject commit 8dbdf4892a9c22a39a20537fd8584b760f41d963

fluid/PaddleRec/tagspace/.run_ce.sh

+#!/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

fluid/PaddleRec/tagspace/_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_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)

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()