[Dygraph]add fp16 and fleet api (#2185)

* add fp16 and fleet api

* add instructions for fleet api and fp16

* modify example of fleet api

* modify docs of fleet api

dygraph/docs/tutorials/GETTING_STARTED_cn.md

@@ -22,6 +22,8 @@ PaddleDetection在[tools](https://github.com/PaddlePaddle/PaddleDetection/tree/m
 | :----------------------: | :------------: | :---------------: | :--------------: | :-----------------: |
 |          -c              |      ALL       |  指定配置文件  |  None  |  **必选**，例如-c configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml |
 |        --eval            |     train      |  是否边训练边测试  |  False  |  可选，如需指定，直接`--eval`即可 |
+|      --fleet         |       train     |  是否使用fleet API训练  |  False  |  可以使用--fleet来指定使用fleet API进行多机训练  |
+|      --fp16        |       train     |  是否开启混合精度训练  |  False  |  可以使用--fp16来指定使用混合精度训练  |
 |          -o              |      ALL       |  设置或更改配置文件里的参数内容  |  None  |  可选，例如：`-o use_gpu=False`  |
 |       --slim_config             |     ALL      |  模型压缩策略配置文件  |  None  |  可选，例如`--slim_config configs/slim/prune/yolov3_prune_l1_norm.yml`  |
 |       --output_dir       |      infer/export_model     |  预测后结果或导出模型保存路径  |  `./output`  |  可选，例如`--output_dir=output`  |
@@ -29,7 +31,6 @@ PaddleDetection在[tools](https://github.com/PaddlePaddle/PaddleDetection/tree/m
 |      --infer_dir         |       infer     |  用于预测的图片文件夹路径  |  None  |  可选  |
 |      --infer_img         |       infer     |  用于预测的图片路径  |  None  |  可选，`--infer_img`和`--infer_dir`必须至少设置一个  |
 
-
 ### 训练
 
 - 单卡训练
@@ -45,6 +46,20 @@ export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
 python -m paddle.distributed.launch --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.yml
 ```
 
+- 混合精度训练
+
+```bash
+export CUDA_VISIBLE_DEVICES=0
+python tools/train.py -c configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.yml --fp16
+```
+
+- fleet API训练
+
+```bash
+# fleet API用于多机训练，启动方式与单机多卡训练方式基本一致，只不过需要使用--ips指定ip列表以及--fleet开启多机训练
+python -m paddle.distributed.launch --ips="xx.xx.xx.xx,yy.yy.yy.yy" --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.yml --fleet
+```
+
 - 边训练边评估
 ```bash
 python tools/train.py -c configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.yml --eval

dygraph/ppdet/engine/env.py

@@ -21,9 +21,13 @@ import random
 import numpy as np
 
 import paddle
-from paddle.distributed import ParallelEnv
+from paddle.distributed import ParallelEnv, fleet
 
-__all__ = ['init_parallel_env', 'set_random_seed']
+__all__ = ['init_parallel_env', 'set_random_seed', 'init_fleet_env']
+
+
+def init_fleet_env():
+    fleet.init(is_collective=True)
 
 
 def init_parallel_env():

dygraph/ppdet/engine/trainer.py

@@ -24,7 +24,8 @@ import numpy as np
 from PIL import Image
 
 import paddle
-from paddle.distributed import ParallelEnv
+from paddle.distributed import ParallelEnv, fleet
+from paddle import amp
 from paddle.static import InputSpec
 
 from ppdet.core.workspace import create
@@ -174,10 +175,17 @@ class Trainer(object):
             self.load_weights(self.cfg.pretrain_weights)
 
         model = self.model
-        if self._nranks > 1:
+        if self.cfg.fleet:
+            model = fleet.distributed_model(model)
+            self.optimizer = fleet.distributed_optimizer(
+                self.optimizer).user_defined_optimizer
+        elif self._nranks > 1:
             model = paddle.DataParallel(self.model)
-        else:
-            model = self.model
+
+        # initial fp16
+        if self.cfg.fp16:
+            scaler = amp.GradScaler(
+                enable=self.cfg.use_gpu, init_loss_scaling=1024)
 
         self.status.update({
             'epoch_id': self.start_epoch,
@@ -203,13 +211,25 @@ class Trainer(object):
                 self.status['step_id'] = step_id
                 self._compose_callback.on_step_begin(self.status)
 
-                # model forward
-                outputs = model(data)
-                loss = outputs['loss']
+                if self.cfg.fp16:
+                    with amp.auto_cast(enable=self.cfg.use_gpu):
+                        # model forward
+                        outputs = model(data)
+                        loss = outputs['loss']
+
+                    # model backward
+                    scaled_loss = scaler.scale(loss)
+                    scaled_loss.backward()
+                    # in dygraph mode, optimizer.minimize is equal to optimizer.step
+                    scaler.minimize(self.optimizer, scaled_loss)
+                else:
+                    # model forward
+                    outputs = model(data)
+                    loss = outputs['loss']
+                    # model backward
+                    loss.backward()
+                    self.optimizer.step()
 
-                # model backward
-                loss.backward()
-                self.optimizer.step()
                 curr_lr = self.optimizer.get_lr()
                 self.lr.step()
                 self.optimizer.clear_grad()

dygraph/tools/train.py

@@ -33,7 +33,7 @@ from paddle.distributed import ParallelEnv
 
 from ppdet.core.workspace import load_config, merge_config, create
 from ppdet.utils.checkpoint import load_weight, load_pretrain_weight
-from ppdet.engine import Trainer, init_parallel_env, set_random_seed
+from ppdet.engine import Trainer, init_parallel_env, set_random_seed, init_fleet_env
 
 import ppdet.utils.cli as cli
 import ppdet.utils.check as check
@@ -65,13 +65,24 @@ def parse_args():
         default=False,
         help="If set True, enable continuous evaluation job."
         "This flag is only used for internal test.")
+    parser.add_argument(
+        "--fp16",
+        action='store_true',
+        default=False,
+        help="Enable mixed precision training.")
+    parser.add_argument(
+        "--fleet", action='store_true', default=False, help="Use fleet or not")
     args = parser.parse_args()
     return args
 
 
 def run(FLAGS, cfg):
-    # init parallel environment if nranks > 1
-    init_parallel_env()
+    # init fleet environment
+    if cfg.fleet:
+        init_fleet_env()
+    else:
+        # init parallel environment if nranks > 1
+        init_parallel_env()
 
     if FLAGS.enable_ce:
         set_random_seed(0)
@@ -91,6 +102,8 @@ def main():
     FLAGS = parse_args()
 
     cfg = load_config(FLAGS.config)
+    cfg['fp16'] = FLAGS.fp16
+    cfg['fleet'] = FLAGS.fleet
     merge_config(FLAGS.opt)
     if FLAGS.slim_config:
         slim_cfg = load_config(FLAGS.slim_config)