Merge pull request #14 from PaddlePaddle/develop

update

fluid/PaddleCV/image_classification/dist_train/README.md

@@ -39,6 +39,8 @@ You can test if distributed training works on a single node before deploying to
 
 ***NOTE: for best performance, we recommend using multi-process mode, see No.3. And together with fp16.***
 
+***NOTE: for nccl2 distributed mode, you must ensure each node train same number of samples, or set skip_unbalanced_data to 1 to do sync training.***
+
 1. simply run `python dist_train.py` to start local training with default configuratioins.
 2. for pserver mode, run `bash run_ps_mode.sh` to start 2 pservers and 2 trainers, these 2 trainers
    will use GPU 0 and 1 to simulate 2 workers.
@@ -90,4 +92,19 @@ The default resnet50 distributed training config is based on this paper: https:/
 
 ### Performance
 
-TBD
+The below figure shows fluid distributed training performances. We did these on a 4-node V100 GPU cluster,
+each has 8 V100 GPU card, with total of 32 GPUs. All modes can reach the "state of the art (choose loss scale carefully when using fp16 mode)" of ResNet50 model with imagenet dataset. The Y axis in the figure shows
+the images/s while the X-axis shows the number of GPUs.
+
+<p align="center">
+<img src="../images/imagenet_dist_performance.png" width=528> <br />
+Performance of Multiple-GPU Training of Resnet50 on Imagenet
+</p>
+
+The second figure shows speed-ups when using multiple GPUs according to the above figure.
+
+<p align="center">
+<img src="../images/imagenet_dist_speedup.png" width=528> <br />
+Speed-ups of Multiple-GPU Training of Resnet50 on Imagenet
+</p>
+

fluid/PaddleCV/rcnn/eval_coco_map.py

@@ -55,7 +55,7 @@ def eval():
         add_conv_body_func=resnet.add_ResNet50_conv4_body,
         add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head,
         use_pyreader=False,
-        is_train=False)
+        mode='val')
     model.build_model(image_shape)
     pred_boxes = model.eval_bbox_out()
     if cfg.MASK_ON:

fluid/PaddleCV/rcnn/infer.py

@@ -35,7 +35,7 @@ def infer():
         add_conv_body_func=resnet.add_ResNet50_conv4_body,
         add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head,
         use_pyreader=False,
-        is_train=False)
+        mode='infer')
     model.build_model(image_shape)
     pred_boxes = model.eval_bbox_out()
     if cfg.MASK_ON:

fluid/PaddleCV/rcnn/models/model_builder.py

@@ -25,12 +25,12 @@ class RCNN(object):
     def __init__(self,
                  add_conv_body_func=None,
                  add_roi_box_head_func=None,
-                 is_train=True,
+                 mode='train',
                  use_pyreader=True,
                  use_random=True):
         self.add_conv_body_func = add_conv_body_func
         self.add_roi_box_head_func = add_roi_box_head_func
-        self.is_train = is_train
+        self.mode = mode
         self.use_pyreader = use_pyreader
         self.use_random = use_random
 
@@ -41,7 +41,7 @@ class RCNN(object):
         self.rpn_heads(body_conv)
         # Fast RCNN
         self.fast_rcnn_heads(body_conv)
-        if not self.is_train:
+        if self.mode != 'train':
             self.eval_bbox()
         # Mask RCNN
         if cfg.MASK_ON:
@@ -115,7 +115,9 @@ class RCNN(object):
                     name='gt_masks', shape=[2], dtype='float32', lod_level=3)
 
     def feeds(self):
-        if not self.is_train:
+        if self.mode == 'infer':
+            return [self.image, self.im_info]
+        if self.mode == 'val':
             return [self.image, self.im_info, self.im_id]
         if not cfg.MASK_ON:
             return [
@@ -213,7 +215,7 @@ class RCNN(object):
         rpn_cls_score_prob = fluid.layers.sigmoid(
             self.rpn_cls_score, name='rpn_cls_score_prob')
 
-        param_obj = cfg.TRAIN if self.is_train else cfg.TEST
+        param_obj = cfg.TRAIN if self.mode == 'train' else cfg.TEST
         pre_nms_top_n = param_obj.rpn_pre_nms_top_n
         post_nms_top_n = param_obj.rpn_post_nms_top_n
         nms_thresh = param_obj.rpn_nms_thresh
@@ -230,7 +232,7 @@ class RCNN(object):
             nms_thresh=nms_thresh,
             min_size=min_size,
             eta=eta)
-        if self.is_train:
+        if self.mode == 'train':
             outs = fluid.layers.generate_proposal_labels(
                 rpn_rois=self.rpn_rois,
                 gt_classes=self.gt_label,
@@ -267,7 +269,7 @@ class RCNN(object):
                 self.mask_int32 = mask_out[2]
 
     def fast_rcnn_heads(self, roi_input):
-        if self.is_train:
+        if self.mode == 'train':
             pool_rois = self.rois
         else:
             pool_rois = self.rpn_rois
@@ -311,7 +313,7 @@ class RCNN(object):
             bias_attr=ParamAttr(
                 name='conv5_mask_b', learning_rate=2., regularizer=L2Decay(0.)))
         act_func = None
-        if not self.is_train:
+        if self.mode != 'train':
             act_func = 'sigmoid'
         mask_fcn_logits = fluid.layers.conv2d(
             input=mask_out,
@@ -325,13 +327,13 @@ class RCNN(object):
                 learning_rate=2.,
                 regularizer=L2Decay(0.)))
 
-        if not self.is_train:
+        if self.mode != 'train':
             mask_fcn_logits = fluid.layers.lod_reset(mask_fcn_logits,
                                                      self.pred_result)
         return mask_fcn_logits
 
     def mask_rcnn_heads(self, mask_input):
-        if self.is_train:
+        if self.mode == 'train':
             conv5 = fluid.layers.gather(self.res5_2_sum,
                                         self.roi_has_mask_int32)
             self.mask_fcn_logits = self.SuffixNet(conv5)

fluid/PaddleCV/rcnn/reader.py

@@ -36,7 +36,9 @@ def roidb_reader(roidb, mode):
     im_height = np.round(roidb['height'] * im_scales)
     im_width = np.round(roidb['width'] * im_scales)
     im_info = np.array([im_height, im_width, im_scales], dtype=np.float32)
-    if mode == 'val' or mode == 'infer':
+    if mode == 'infer':
+        return im, im_info
+    if mode == 'val':
         return im, im_info, im_id
 
     gt_boxes = roidb['gt_boxes'].astype('float32')
@@ -162,8 +164,8 @@ def coco(mode,
             for roidb in roidbs:
                 if cfg.image_name not in roidb['image']:
                     continue
-                im, im_info, im_id = roidb_reader(roidb, mode)
-                batch_out = [(im, im_info, im_id)]
+                im, im_info = roidb_reader(roidb, mode)
+                batch_out = [(im, im_info)]
                 yield batch_out
 
     return reader

fluid/PaddleRec/gru4rec/cluster_train.py

@@ -13,22 +13,26 @@ import net
 
 SEED = 102
 
+
 def parse_args():
     parser = argparse.ArgumentParser("gru4rec benchmark.")
     parser.add_argument(
-        '--train_dir', type=str, default='train_data', help='train file address')
-    parser.add_argument(
-        '--vocab_path', type=str, default='vocab.txt', help='vocab file address')
-    parser.add_argument(
-        '--is_local', type=int, default=1, help='whether local')
+        '--train_dir',
+        type=str,
+        default='train_data',
+        help='train file address')
     parser.add_argument(
-        '--hid_size', type=int, default=100, help='hid size')
+        '--vocab_path',
+        type=str,
+        default='vocab.txt',
+        help='vocab file address')
+    parser.add_argument('--is_local', type=int, default=1, help='whether local')
+    parser.add_argument('--hid_size', type=int, default=100, help='hid size')
     parser.add_argument(
         '--model_dir', type=str, default='model_recall20', help='model dir')
     parser.add_argument(
         '--batch_size', type=int, default=5, help='num of batch size')
-    parser.add_argument(
-        '--pass_num', type=int, default=10, help='num of epoch')
+    parser.add_argument('--pass_num', type=int, default=10, help='num of epoch')
     parser.add_argument(
         '--print_batch', type=int, default=10, help='num of print batch')
     parser.add_argument(
@@ -40,19 +44,33 @@ def parse_args():
     parser.add_argument(
         '--role', type=str, default='pserver', help='trainer or pserver')
     parser.add_argument(
-        '--endpoints', type=str, default='127.0.0.1:6000', help='The pserver endpoints, like: 127.0.0.1:6000, 127.0.0.1:6001')
-    parser.add_argument(
-        '--current_endpoint', type=str, default='127.0.0.1:6000', help='The current_endpoint')
-    parser.add_argument(
-        '--trainer_id', type=int, default=0, help='trainer id ,only trainer_id=0 save model')
-    parser.add_argument(
-        '--trainers', type=int, default=1, help='The num of trianers, (default: 1)')
+        '--endpoints',
+        type=str,
+        default='127.0.0.1:6000',
+        help='The pserver endpoints, like: 127.0.0.1:6000, 127.0.0.1:6001')
+    parser.add_argument(
+        '--current_endpoint',
+        type=str,
+        default='127.0.0.1:6000',
+        help='The current_endpoint')
+    parser.add_argument(
+        '--trainer_id',
+        type=int,
+        default=0,
+        help='trainer id ,only trainer_id=0 save model')
+    parser.add_argument(
+        '--trainers',
+        type=int,
+        default=1,
+        help='The num of trianers, (default: 1)')
     args = parser.parse_args()
     return args
 
+
 def get_cards(args):
     return args.num_devices
 
+
 def train():
     """ do training """
     args = parse_args()
@@ -67,7 +85,8 @@ def train():
         buffer_size=1000, word_freq_threshold=0, is_train=True)
 
     # Train program
-    src_wordseq, dst_wordseq, avg_cost, acc = net.network(vocab_size=vocab_size, hid_size=hid_size)
+    src_wordseq, dst_wordseq, avg_cost, acc = net.all_vocab_network(
+        vocab_size=vocab_size, hid_size=hid_size)
 
     # Optimization to minimize lost
     sgd_optimizer = fluid.optimizer.SGD(learning_rate=args.base_lr)
@@ -97,8 +116,10 @@ def train():
                 lod_dst_wordseq = utils.to_lodtensor([dat[1] for dat in data],
                                                      place)
                 ret_avg_cost = exe.run(main_program,
-                                            feed={  "src_wordseq": lod_src_wordseq,
-                                                    "dst_wordseq": lod_dst_wordseq},
+                                       feed={
+                                           "src_wordseq": lod_src_wordseq,
+                                           "dst_wordseq": lod_dst_wordseq
+                                       },
                                        fetch_list=fetch_list)
                 avg_ppl = np.exp(ret_avg_cost[0])
                 newest_ppl = np.mean(avg_ppl)
@@ -113,7 +134,8 @@ def train():
             feed_var_names = ["src_wordseq", "dst_wordseq"]
             fetch_vars = [avg_cost, acc]
             if args.trainer_id == 0:
-                fluid.io.save_inference_model(save_dir, feed_var_names, fetch_vars, exe)
+                fluid.io.save_inference_model(save_dir, feed_var_names,
+                                              fetch_vars, exe)
                 print("model saved in %s" % save_dir)
         print("finish training")
 
@@ -123,7 +145,8 @@ def train():
     else:
         print("run distribute training")
         t = fluid.DistributeTranspiler()
-        t.transpile(args.trainer_id, pservers=args.endpoints, trainers=args.trainers)
+        t.transpile(
+            args.trainer_id, pservers=args.endpoints, trainers=args.trainers)
         if args.role == "pserver":
             print("run psever")
             pserver_prog = t.get_pserver_program(args.current_endpoint)
@@ -136,5 +159,6 @@ def train():
             print("run trainer")
             train_loop(t.get_trainer_program())
 
+
 if __name__ == "__main__":
     train()

fluid/PaddleRec/gru4rec/infer.py

@@ -11,6 +11,7 @@ import paddle
 
 import utils
 
+
 def parse_args():
     parser = argparse.ArgumentParser("gru4rec benchmark.")
     parser.add_argument(
@@ -22,12 +23,15 @@ def parse_args():
     parser.add_argument(
         '--model_dir', type=str, default='model_recall20', help='model dir')
     parser.add_argument(
-        '--use_cuda', type=int, default='1', help='whether use cuda')  
+        '--use_cuda', type=int, default='0', help='whether use cuda')
     parser.add_argument(
         '--batch_size', type=int, default='5', help='batch_size')
+    parser.add_argument(
+        '--vocab_path', type=str, default='vocab.txt', help='vocab file')
     args = parser.parse_args()
     return args
 
+
 def infer(test_reader, use_cuda, model_path):
     """ inference function """
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
@@ -72,11 +76,16 @@ if __name__ == "__main__":
     test_dir = args.test_dir
     model_dir = args.model_dir
     batch_size = args.batch_size
+    vocab_path = args.vocab_path
     use_cuda = True if args.use_cuda else False
-    print("start index: ", start_index, " last_index:" ,last_index)
+    print("start index: ", start_index, " last_index:", last_index)
     vocab_size, test_reader = utils.prepare_data(
-        test_dir, "", batch_size=batch_size,
-        buffer_size=1000, word_freq_threshold=0, is_train=False)
+        test_dir,
+        vocab_path,
+        batch_size=batch_size,
+        buffer_size=1000,
+        word_freq_threshold=0,
+        is_train=False)
 
     for epoch in range(start_index, last_index + 1):
         epoch_path = model_dir + "/epoch_" + str(epoch)