!3227 modify some network scripts

Merge pull request !3227 from anzhengqi/modify-networks

official/cv/resnet/infer.py

@@ -61,7 +61,7 @@ def infer_net():
 
     # create dataset
     dataset = create_dataset(dataset_path=config.data_path, do_train=False, batch_size=config.batch_size,
-                             target=target)
+                             image_size=config.eval_image_size, target=target)
     step_size = dataset.get_dataset_size()
 
     # define net

official/cv/resnet/src/dataset_infer.py

@@ -89,7 +89,8 @@ class ImgDataset:
         return len(self.data)
 
 
-def create_dataset(dataset_path, do_train, repeat_num=1, batch_size=32, target="Ascend", distribute=False):
+def create_dataset(dataset_path, do_train, repeat_num=1, batch_size=32, image_size=224,
+                   target="Ascend", distribute=False):
     """
     create a train or eval imagenet2012 dataset for resnet50
 
@@ -123,7 +124,6 @@ def create_dataset(dataset_path, do_train, repeat_num=1, batch_size=32, target="
                                        num_parallel_workers=8, shuffle=True,
                                        num_shards=device_num, shard_id=rank_id)
 
-    image_size = 224
     mean = [0.485 * 255, 0.456 * 255, 0.406 * 255]
     std = [0.229 * 255, 0.224 * 255, 0.225 * 255]
 
@@ -163,7 +163,8 @@ def create_dataset(dataset_path, do_train, repeat_num=1, batch_size=32, target="
     return data_set
 
 
-def create_dataset2(dataset_path, do_train, repeat_num=1, batch_size=32, target="Ascend", distribute=False):
+def create_dataset2(dataset_path, do_train, repeat_num=1, batch_size=32, image_size=224,
+                    target="Ascend", distribute=False):
     """
     create a train or eval imagenet2012 dataset for resnet101
     Args:
@@ -195,7 +196,6 @@ def create_dataset2(dataset_path, do_train, repeat_num=1, batch_size=32, target=
         data_set = ds.GeneratorDataset(source=dataset_generator, column_names=["label", "image", "filename"],
                                        num_parallel_workers=8, shuffle=True,
                                        num_shards=device_num, shard_id=rank_id)
-    image_size = 224
     mean = [0.475 * 255, 0.451 * 255, 0.392 * 255]
     std = [0.275 * 255, 0.267 * 255, 0.278 * 255]
 
@@ -232,7 +232,8 @@ def create_dataset2(dataset_path, do_train, repeat_num=1, batch_size=32, target=
     return data_set
 
 
-def create_dataset3(dataset_path, do_train, repeat_num=1, batch_size=32, target="Ascend", distribute=False):
+def create_dataset3(dataset_path, do_train, repeat_num=1, batch_size=32, image_size=224,
+                    target="Ascend", distribute=False):
     """
     create a train or eval imagenet2012 dataset for se-resnet50
 
@@ -264,7 +265,6 @@ def create_dataset3(dataset_path, do_train, repeat_num=1, batch_size=32, target=
         data_set = ds.GeneratorDataset(source=dataset_generator, column_names=["label", "image", "filename"],
                                        num_parallel_workers=8, shuffle=True,
                                        num_shards=device_num, shard_id=rank_id)
-    image_size = 224
     mean = [123.68, 116.78, 103.94]
     std = [1.0, 1.0, 1.0]
 

official/nlp/textcnn/mr_config.yaml

@@ -18,6 +18,7 @@ pre_trained: False
 num_classes: 2
 batch_size: 64
 epoch_size: 4
+sink_size: -1
 weight_decay: 3e-5
 keep_checkpoint_max: 1
 checkpoint_path: "./checkpoint/"

official/nlp/textcnn/mr_config_cpu.yaml

@@ -18,6 +18,7 @@ pre_trained: False
 num_classes: 2
 batch_size: 64
 epoch_size: 4
+sink_size: -1
 weight_decay: 3e-5
 keep_checkpoint_max: 1
 checkpoint_path: "./checkpoint/"

official/nlp/textcnn/src/textcnn.py

@@ -14,11 +14,40 @@
 # ============================================================================
 """TextCNN"""
 
+import os
+
 import mindspore.ops as ops
 import mindspore.nn as nn
 from mindspore import Tensor
 from mindspore.nn.cell import Cell
 import mindspore
+from mindspore.train.callback import Callback
+
+
+class EvalCallback(Callback):
+    """
+    Evaluation per epoch, and save the best accuracy checkpoint.
+    """
+    def __init__(self, model, eval_ds, begin_eval_epoch=1, save_path="./"):
+        self.model = model
+        self.eval_ds = eval_ds
+        self.begin_eval_epoch = begin_eval_epoch
+        self.best_acc = 0
+        self.save_path = save_path
+
+    def epoch_end(self, run_context):
+        """
+        evaluate at epoch end.
+        """
+        cb_params = run_context.original_args()
+        cur_epoch = cb_params.cur_epoch_num
+        if cur_epoch >= self.begin_eval_epoch:
+            res = self.model.eval(self.eval_ds)
+            acc = res["acc"]
+            if acc > self.best_acc:
+                self.best_acc = acc
+                mindspore.save_checkpoint(cb_params.train_network, os.path.join(self.save_path, "best_acc.ckpt"))
+                print("the best epoch is", cur_epoch, "best acc is", self.best_acc)
 
 
 class SoftmaxCrossEntropyExpand(Cell):

official/nlp/textcnn/sst2_config.yaml

@@ -18,8 +18,9 @@ pre_trained: False
 num_classes: 2
 batch_size: 64
 epoch_size: 4
+sink_size: 200
 weight_decay: 3e-5
-keep_checkpoint_max: 1
+keep_checkpoint_max: 10
 checkpoint_path: "./checkpoint/"
 checkpoint_file_path: "train_textcnn-4_149.ckpt"
 word_len: 51

official/nlp/textcnn/subj_config.yaml

@@ -18,6 +18,7 @@ pre_trained: False
 num_classes: 2
 batch_size: 64
 epoch_size: 4
+sink_size: -1
 weight_decay: 3e-5
 keep_checkpoint_max: 1
 checkpoint_path: "./checkpoint/"

official/nlp/textcnn/train.py

@@ -30,7 +30,7 @@ from model_utils.moxing_adapter import moxing_wrapper
 from model_utils.device_adapter import get_device_id, get_rank_id
 from model_utils.config import config
 from src.textcnn import TextCNN
-from src.textcnn import SoftmaxCrossEntropyExpand
+from src.textcnn import SoftmaxCrossEntropyExpand, EvalCallback
 from src.dataset import MovieReview, SST2, Subjectivity
 
 set_seed(1)
@@ -55,7 +55,10 @@ def train_net():
         instance = SST2(root_dir=config.data_path, maxlen=config.word_len, split=0.9)
 
     dataset = instance.create_train_dataset(batch_size=config.batch_size, epoch_size=config.epoch_size)
+    eval_dataset = instance.create_test_dataset(batch_size=config.batch_size)
     batch_num = dataset.get_dataset_size()
+    if config.sink_size == -1:
+        config.sink_size = batch_num
 
     base_lr = float(config.base_lr)
     learning_rate = []
@@ -79,13 +82,19 @@ def train_net():
 
     model = Model(net, loss_fn=loss, optimizer=opt, metrics={'acc': Accuracy()})
 
-    config_ck = CheckpointConfig(save_checkpoint_steps=int(config.epoch_size * batch_num / 2),
+    config_ck = CheckpointConfig(save_checkpoint_steps=config.sink_size,
                                  keep_checkpoint_max=config.keep_checkpoint_max)
     time_cb = TimeMonitor(data_size=batch_num)
     ckpt_save_dir = os.path.join(config.output_path, config.checkpoint_path)
     ckpoint_cb = ModelCheckpoint(prefix="train_textcnn", directory=ckpt_save_dir, config=config_ck)
     loss_cb = LossMonitor()
-    model.train(config.epoch_size, dataset, callbacks=[time_cb, ckpoint_cb, loss_cb])
+    eval_callback = EvalCallback(model, eval_dataset, save_path=ckpt_save_dir)
+    if config.device_target == "CPU":
+        model.train(config.epoch_size, dataset, callbacks=[time_cb, ckpoint_cb, loss_cb])
+    else:
+        epoch_count = config.epoch_size * batch_num // config.sink_size
+        model.train(epoch_count, dataset, callbacks=[time_cb, ckpoint_cb, loss_cb, eval_callback],
+                    sink_size=config.sink_size)
     print("train success")
 
 

official/nlp/transformer/README.md

@@ -508,3 +508,7 @@ Refer to the [ModelZoo FAQ](https://gitee.com/mindspore/models#FAQ) for some com
 
   **A**: At the end stage of training, the model accuracy usually drifts irregularly. Because we have to use a third-party perl scripts for evaluation, we can't find the best checkpoint as soon as the training process finished.
   You can try to evaluate the last several checkpoints to find the best one.
+
+- **Q: why the shape match error such as "For 'Add', x.shape and y.shape need to broadcast." occurs?**
+
+  **A**: because all of the parameters is supported by the dataset in readme. if users use new datasets, please modify parameters in the same time.

official/nlp/transformer/README_CN.md

@@ -481,3 +481,7 @@ train.py已经设置了一些种子，避免数据集轮换和权重初始化的
 - **Q: 为什么我最后的checkpoint的精度不好？**
 
   **A**： 因为我们需要使用一个第三方的perl脚本来进行验证，所以我们没办法在训练的时候就获取最优checkpoint。你可以尝试对最后的多个checkpoint进行验证，从中获取最好的一个。
+
+- **Q: 为什么报错类似＂For 'Add', x.shape and y.shape need to broadcast.＂，Shape不匹配的问题？**
+
+  **A**: 因为已有的配置参数是根据readme中提供的数据集配置的，用户更换数据集后，需要根据参数定义重新配置相关参数．