optimize cnnctc learning rate strategy

model_zoo/official/cv/cnnctc/default_config.yaml

@@ -33,8 +33,10 @@ TRAIN_EPOCHS: 3
 run_distribute: False
 PRED_TRAINED: ""
 SAVE_PATH: "./"
-LR: 1e-4
-LR_PARA: 5e-4
+#LR
+base_lr: 0.0005
+warmup_step: 2000
+warmup_ratio: 0.0625
 MOMENTUM: 0.8
 LOSS_SCALE: 8096
 SAVE_CKPT_PER_N_STEP: 2000

model_zoo/official/cv/cnnctc/src/dataset.py

@@ -35,12 +35,14 @@ class NormalizePAD():
     def __call__(self, img):
         # toTensor
         img = np.array(img, dtype=np.float32)
+        # normalize
+        means = [121.58949, 123.93914, 123.418655]
+        stds = [65.70353, 65.142426, 68.61079]
+        img = np.subtract(img, means)
+        img = np.true_divide(img, stds)
+
         img = img.transpose([2, 0, 1])
         img = img.astype(np.float)
-        img = np.true_divide(img, 255)
-        # normalize
-        img = np.subtract(img, 0.5)
-        img = np.true_divide(img, 0.5)
 
         _, _, w = img.shape
         Pad_img = np.zeros(shape=self.max_size, dtype=np.float32)

model_zoo/official/cv/cnnctc/src/lr_schedule.py

@@ -0,0 +1,41 @@
+# Copyright 2020-2021 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+"""lr generator for cnnctc"""
+import math
+
+def linear_warmup_learning_rate(current_step, warmup_steps, base_lr, init_lr):
+    lr_inc = (float(base_lr) - float(init_lr)) / float(warmup_steps)
+    learning_rate = float(init_lr) + lr_inc * current_step
+    return learning_rate
+
+def a_cosine_learning_rate(current_step, base_lr, warmup_steps, decay_steps):
+    base = float(current_step - warmup_steps) / float(decay_steps)
+    learning_rate = (1 + math.cos(base * math.pi)) / 2 * base_lr
+    return learning_rate
+
+def dynamic_lr(config, steps_per_epoch):
+    """dynamic learning rate generator"""
+    base_lr = config.base_lr
+    total_steps = steps_per_epoch * config.TRAIN_EPOCHS
+    warmup_steps = int(config.warmup_step)
+    decay_steps = total_steps - warmup_steps
+    lr = []
+    for i in range(total_steps):
+        if i < warmup_steps:
+            lr.append(linear_warmup_learning_rate(i, warmup_steps, base_lr, base_lr * config.warmup_ratio))
+        else:
+            lr.append(a_cosine_learning_rate(i, base_lr, warmup_steps, decay_steps))
+
+    return lr

model_zoo/official/cv/cnnctc/train.py

@@ -15,7 +15,6 @@
 """cnnctc train"""
 
 
-import ast
 import numpy as np
 import mindspore
 import mindspore.common.dtype as mstype
@@ -30,6 +29,7 @@ from mindspore.train.serialization import load_checkpoint, load_param_into_net
 from src.callback import LossCallBack
 from src.cnn_ctc import CNNCTC_Model, ctc_loss, WithLossCell, CNNCTCTrainOneStepWithLossScaleCell
 from src.dataset import ST_MJ_Generator_batch_fixed_length, ST_MJ_Generator_batch_fixed_length_para
+from src.lr_schedule import dynamic_lr
 from src.model_utils.config import config
 from src.model_utils.device_adapter import get_device_id
 from src.model_utils.moxing_adapter import moxing_wrapper
@@ -88,9 +88,6 @@ def train():
         else:
             ckpt_save_dir = config.SAVE_PATH + "ckpt_standalone/"
 
-    config.LR = ast.literal_eval(config.LR)
-    config.LR_PARA = ast.literal_eval(config.LR_PARA)
-
     ds = dataset_creator(config.run_distribute)
 
     net = CNNCTC_Model(config.NUM_CLASS, config.HIDDEN_SIZE, config.FINAL_FEATURE_WIDTH)
@@ -104,9 +101,11 @@ def train():
         print('train from scratch...')
 
     criterion = ctc_loss()
+    dataset_size = ds.get_dataset_size()
+    lr = Tensor(dynamic_lr(config, dataset_size), mstype.float32)
     opt = mindspore.nn.RMSProp(params=net.trainable_params(),
                                centered=True,
-                               learning_rate=config.LR_PARA,
+                               learning_rate=lr,
                                momentum=config.MOMENTUM,
                                loss_scale=config.LOSS_SCALE)