add tsn 2020-08-31

dygraph/tsn/README.md

@@ -57,20 +57,20 @@ TSN的训练数据采用UCF101动作识别数据集。数据下载及准备请
 
 1. 多卡训练
 ```bash
-bash multi_gpus_run.sh ./configs/tsn.yaml
+bash multi_gpus_run.sh
 ```
 多卡训练所使用的gpu可以通过如下方式设置：
 - 修改`multi_gpus_run.sh` 中 `export CUDA_VISIBLE_DEVICES=0,1,2,3`（默认为0,1,2,3表示使用0，1，2，3卡号的gpu进行训练）
-- 注意：若修改了batchsize则学习率也要做相应的修改，规则为大batchsize用大lr，即同倍数增长缩小关系。例如，默认batchsize=128，lr=0.001，若batchsize=64，lr=0.0005
+- 注意：多卡训练的参数配置文件为`multi_tsn.yaml`。若修改了batchsize则学习率也要做相应的修改，规则为大batchsize用大lr，即同倍数增长缩小关系。例如，默认四卡batchsize=128，lr=0.001，若batchsize=64，lr=0.0005。
 
 
 2. 单卡训练
 ```bash
-bash single_gpu_run.sh ./configs/tsn.yaml
+bash single_gpu_run.sh
 ```
 单卡训练所使用的gpu可以通过如下方式设置：
-- 修改 `run.sh` 中的 `export CUDA_VISIBLE_DEVICES=0` （表示使用gpu 0 进行模型训练）
-- 注意：若修改了batchsize则学习率也要做相应的修改，规则为大batchsize用大lr，即同倍数增长缩小关系。例如，默认batchsize=128，lr=0.001，若batchsize=64，lr=0.0005
+- 修改 `single_gpu_run.sh` 中的 `export CUDA_VISIBLE_DEVICES=0` （表示使用gpu 0 进行模型训练）
+- 注意：单卡训练的参数配置文件为`single_gpu_run.sh`。若修改了batchsize则学习率也要做相应的修改，规则为大batchsize用大lr，即同倍数增长缩小关系。默认单卡batchsize=64，lr=0.0005；若batchsize=32，lr=0.00025
 ## 模型评估
 
 可通过如下方式进行模型评估:

dygraph/tsn/multi_gpus_run.sh

-configs="./tsn.yaml"
+configs="./multi_tsn.yaml"
 pretrain="" # set pretrain model path if needed
 resume="" # set checkpoints model path if u want to resume training
 save_dir=""

dygraph/tsn/tsn.yaml → dygraph/tsn/multi_tsn.yaml

@@ -15,7 +15,7 @@ TRAIN:
     target_size: 224
     num_reader_threads: 12
     buf_size: 1024
-    batch_size: 256
+    batch_size: 128
     use_gpu: True
     filelist: "./data/dataset/ucf101/ucf101_train_split_1_rawframes.txt"
     learning_rate: 0.001
@@ -23,7 +23,7 @@ TRAIN:
     decay_epochs: [30, 60]
     l2_weight_decay: 1e-4
     momentum: 0.9
-    total_videos: 9738 #239781
+    total_videos: 9738 
 
 VALID:
     short_size: 256

dygraph/tsn/single_gpu_run.sh

-configs="tsn.yaml"
+configs="single_tsn.yaml"
 pretrain="" # set pretrain model path if needed
 resume="" # set checkpoints model path if u want to resume training
 save_dir=""

dygraph/tsn/single_tsn.yaml

+MODEL:
+    name: "TSN"
+    format: "frames" # support for "frames" or "videos"
+    num_classes: 101
+    seg_num: 3
+    seglen: 1
+    image_mean: [0.485, 0.456, 0.406]
+    image_std: [0.229, 0.224, 0.225]
+    num_layers: 50
+    topk: 5
+
+TRAIN:
+    epoch: 80
+    short_size: 256
+    target_size: 224
+    num_reader_threads: 12
+    buf_size: 1024
+    batch_size: 64
+    use_gpu: True
+    filelist: "./data/dataset/ucf101/ucf101_train_split_1_rawframes.txt"
+    learning_rate: 0.0005
+    learning_rate_decay: 0.1
+    decay_epochs: [30, 60]
+    l2_weight_decay: 1e-4
+    momentum: 0.9
+    total_videos: 9738 
+
+VALID:
+    short_size: 256
+    target_size: 224
+    num_reader_threads: 12
+    buf_size: 1024
+    batch_size: 128
+    filelist: "./data/dataset/ucf101/ucf101_val_split_1_rawframes.txt"
+
+TEST:
+    short_size: 256
+    target_size: 224
+    num_reader_threads: 12
+    buf_size: 1024
+    batch_size: 64
+    filelist: "./data/dataset/ucf101/ucf101_val_split_1_rawframes.txt"
\ No newline at end of file

dygraph/tsn/train.py

@@ -181,14 +181,17 @@ def val(epoch, model, cfg, args):
         acc_top1 = fluid.layers.accuracy(input=outputs, label=labels, k=1)
         acc_top5 = fluid.layers.accuracy(input=outputs, label=labels, k=5)
 
-        total_loss += avg_loss.numpy()[0]
+        dy_out = avg_loss.numpy()[0]
+        total_loss += dy_out
         total_acc1 += acc_top1.numpy()[0]
         total_acc5 += acc_top5.numpy()[0]
         total_sample += 1
 
-        print('TEST Epoch {}, iter {}, loss = {}, acc1 {}, acc5 {}'.format(
-            epoch, batch_id,
-            avg_loss.numpy()[0], acc_top1.numpy()[0], acc_top5.numpy()[0]))
+        if batch_id % 5 == 0:
+            print(
+                "TEST Epoch {}, iter {}, loss={:.5f}, acc1 {:.5f}, acc5 {:.5f}".
+                format(epoch, batch_id, total_loss / total_sample, total_acc1 /
+                       total_sample, total_acc5 / total_sample))
 
     print('Finish loss {} , acc1 {} , acc5 {}'.format(
         total_loss / total_sample, total_acc1 / total_sample, total_acc5 /
@@ -297,6 +300,7 @@ def train(args):
                     input=outputs, label=labels, k=1)
                 acc_top5 = fluid.layers.accuracy(
                     input=outputs, label=labels, k=5)
+                dy_out = avg_loss.numpy()[0]
 
                 if use_data_parallel:
                     # (data_parallel step5/6)
@@ -309,16 +313,16 @@ def train(args):
                 optimizer.minimize(avg_loss)
                 video_model.clear_gradients()
 
-                total_loss += avg_loss.numpy()[0]
+                total_loss += dy_out
                 total_acc1 += acc_top1.numpy()[0]
                 total_acc5 += acc_top5.numpy()[0]
                 total_sample += 1
                 train_batch_cost = time.time() - batch_start
                 print(
-                    'TRAIN Epoch: {}, iter: {}, batch_cost: {: .5f} s, reader_cost: {: .5f} s loss={: .6f}, acc1 {: .6f}, acc5 {: .6f} \t'.
+                    'TRAIN Epoch: {}, iter: {}, batch_cost: {:.5f} s, reader_cost: {:.5f} s, loss={:.6f}, acc1 {:.6f}, acc5 {:.6f}  '.
                     format(epoch, batch_id, train_batch_cost, train_reader_cost,
-                           avg_loss.numpy()[0],
-                           acc_top1.numpy()[0], acc_top5.numpy()[0]))
+                           total_loss / total_sample, total_acc1 / total_sample,
+                           total_acc5 / total_sample))
                 batch_start = time.time()
 
             print(
@@ -349,11 +353,11 @@ def train(args):
                 else:
                     if val_acc > best_acc:
                         best_acc = val_acc
-                    if fluid.dygraph.parallel.Env().local_rank == 0:
-                        if not os.path.isdir(args.weights):
-                            os.makedirs(args.weights)
-                        fluid.dygraph.save_dygraph(video_model.state_dict(),
-                                                   args.weights + "/final")
+                        if fluid.dygraph.parallel.Env().local_rank == 0:
+                            if not os.path.isdir(args.weights):
+                                os.makedirs(args.weights)
+                            fluid.dygraph.save_dygraph(video_model.state_dict(),
+                                                       args.weights + "/final")
             else:
                 if fluid.dygraph.parallel.Env().local_rank == 0:
                     if not os.path.isdir(args.weights):