fix accurancy lower then 92

model_zoo/vgg16/src/config.py

@@ -19,7 +19,9 @@ from easydict import EasyDict as edict
 
 cifar_cfg = edict({
     'num_classes': 10,
-    'lr_init': 0.05,
+    'lr_init': 0.01,
+    'lr_max': 0.1,
+    'warmup_epochs': 5,
     'batch_size': 64,
     'epoch_size': 70,
     'momentum': 0.9,

model_zoo/vgg16/train.py

@@ -38,20 +38,25 @@ random.seed(1)
 np.random.seed(1)
 
 
-def lr_steps(global_step, lr_max=None, total_epochs=None, steps_per_epoch=None):
+def lr_steps(global_step, lr_init, lr_max, warmup_epochs, total_epochs, steps_per_epoch):
     """Set learning rate."""
     lr_each_step = []
     total_steps = steps_per_epoch * total_epochs
-    decay_epoch_index = [0.3 * total_steps, 0.6 * total_steps, 0.8 * total_steps]
+    warmup_steps = steps_per_epoch * warmup_epochs
+    if warmup_steps != 0:
+        inc_each_step = (float(lr_max) - float(lr_init)) / float(warmup_steps)
+    else:
+        inc_each_step = 0
     for i in range(total_steps):
-        if i < decay_epoch_index[0]:
-            lr_each_step.append(lr_max)
-        elif i < decay_epoch_index[1]:
-            lr_each_step.append(lr_max * 0.1)
-        elif i < decay_epoch_index[2]:
-            lr_each_step.append(lr_max * 0.01)
+        if i < warmup_steps:
+            lr_value = float(lr_init) + inc_each_step * float(i)
         else:
-            lr_each_step.append(lr_max * 0.001)
+            base = (1.0 - (float(i) - float(warmup_steps)) / (float(total_steps) - float(warmup_steps)))
+            lr_value = float(lr_max) * base * base
+            if lr_value < 0.0:
+                lr_value = 0.0
+        lr_each_step.append(lr_value)
+
     current_step = global_step
     lr_each_step = np.array(lr_each_step).astype(np.float32)
     learning_rate = lr_each_step[current_step:]
@@ -86,7 +91,8 @@ if __name__ == '__main__':
     if args_opt.pre_trained:
         load_param_into_net(net, load_checkpoint(args_opt.pre_trained))
 
-    lr = lr_steps(0, lr_max=cfg.lr_init, total_epochs=cfg.epoch_size, steps_per_epoch=batch_num)
+    lr = lr_steps(0, lr_init=cfg.lr_init, lr_max=cfg.lr_max, warmup_epochs=cfg.warmup_epochs,
+                  total_epochs=cfg.epoch_size, steps_per_epoch=batch_num)
     opt = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), Tensor(lr), cfg.momentum,
                    weight_decay=cfg.weight_decay)
     loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean', is_grad=False)