[breaking] deprecate callbacks=Callbacks() but use a list.

examples/A3C-Gym/train-atari.py

@@ -207,16 +207,16 @@ def get_config():
     return TrainConfig(
         dataflow=dataflow,
         optimizer=tf.train.AdamOptimizer(lr, epsilon=1e-3),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
+        callbacks=[
+            ModelSaver(),
             ScheduledHyperParamSetter('learning_rate', [(80, 0.0003), (120, 0.0001)]),
             ScheduledHyperParamSetter('entropy_beta', [(80, 0.005)]),
             ScheduledHyperParamSetter('explore_factor',
                                       [(80, 2), (100, 3), (120, 4), (140, 5)]),
             master,
             StartProcOrThread(master),
             PeriodicCallback(Evaluator(EVAL_EPISODE, ['state'], ['logits']), 2),
-        ]),
+        ],
         session_config=get_default_sess_config(0.5),
         model=M,
         step_per_epoch=STEP_PER_EPOCH,

examples/CTC-TIMIT/train-timit.py

@@ -96,14 +96,14 @@ def get_config(ds_train, ds_test):
     return TrainConfig(
         dataflow=ds_train,
         optimizer=tf.train.AdamOptimizer(lr, epsilon=1e-3),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
+        callbacks=[
+            ModelSaver(),
             StatMonitorParamSetter('learning_rate', 'error',
                                    lambda x: x * 0.2, 0, 5),
             HumanHyperParamSetter('learning_rate'),
             PeriodicCallback(
                 InferenceRunner(ds_test, [ScalarStats('error')]), 2),
-        ]),
+        ],
         model=Model(),
         step_per_epoch=step_per_epoch,
         max_epoch=70,

examples/Char-RNN/char-rnn.py

@@ -110,10 +110,10 @@ def get_config():
     return TrainConfig(
         dataflow=ds,
         optimizer=tf.train.AdamOptimizer(lr),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
+        callbacks=[
+            ModelSaver(),
             ScheduledHyperParamSetter('learning_rate', [(25, 2e-4)])
-        ]),
+        ],
         model=Model(),
         step_per_epoch=step_per_epoch,
         max_epoch=50,

examples/DeepQNetwork/DQN.py

@@ -177,16 +177,16 @@ def get_config():
     return TrainConfig(
         dataflow=dataset_train,
         optimizer=tf.train.AdamOptimizer(lr, epsilon=1e-3),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
+        callbacks=[
+            ModelSaver(),
             ScheduledHyperParamSetter('learning_rate',
                                       [(150, 4e-4), (250, 1e-4), (350, 5e-5)]),
             RunOp(lambda: M.update_target_param()),
             dataset_train,
             PeriodicCallback(Evaluator(EVAL_EPISODE, ['state'], ['Qvalue']), 3),
             # HumanHyperParamSetter('learning_rate', 'hyper.txt'),
             # HumanHyperParamSetter(ObjAttrParam(dataset_train, 'exploration'), 'hyper.txt'),
-        ]),
+        ],
         # save memory for multiprocess evaluator
         session_config=get_default_sess_config(0.6),
         model=M,

examples/DoReFa-Net/alexnet-dorefa.py

@@ -236,16 +236,16 @@ def get_config():
     return TrainConfig(
         dataflow=data_train,
         optimizer=tf.train.AdamOptimizer(lr, epsilon=1e-5),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
+        callbacks=[
+            ModelSaver(),
             # HumanHyperParamSetter('learning_rate'),
             ScheduledHyperParamSetter(
                 'learning_rate', [(56, 2e-5), (64, 4e-6)]),
             InferenceRunner(data_test,
                             [ScalarStats('cost'),
                              ClassificationError('wrong-top1', 'val-error-top1'),
                              ClassificationError('wrong-top5', 'val-error-top5')])
-        ]),
+        ],
         model=Model(),
         step_per_epoch=10000,
         max_epoch=100,

examples/DoReFa-Net/svhn-digit-dorefa.py

@@ -163,12 +163,11 @@ def get_config():
     return TrainConfig(
         dataflow=data_train,
         optimizer=tf.train.AdamOptimizer(lr, epsilon=1e-5),
-        callbacks=Callbacks([
-            StatPrinter(),
+        callbacks=[
             ModelSaver(),
             InferenceRunner(data_test,
                             [ScalarStats('cost'), ClassificationError()])
-        ]),
+        ],
         model=Model(),
         step_per_epoch=step_per_epoch,
         max_epoch=200,

examples/GAN/DCGAN-CelebA.py

@@ -109,9 +109,7 @@ def get_config():
     return TrainConfig(
         dataflow=dataset,
         optimizer=tf.train.AdamOptimizer(lr, beta1=0.5, epsilon=1e-3),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
-        ]),
+        callbacks=[ModelSaver()],
         session_config=get_default_sess_config(0.5),
         model=Model(),
         step_per_epoch=300,

examples/GAN/GAN.py

@@ -102,7 +102,6 @@ def run_step(self):
 
 
 class RandomZData(DataFlow):
-
     def __init__(self, shape):
         super(RandomZData, self).__init__()
         self.shape = shape

examples/GAN/Image2Image.py

@@ -168,10 +168,10 @@ def get_config():
     return TrainConfig(
         dataflow=dataset,
         optimizer=tf.train.AdamOptimizer(lr, beta1=0.5, epsilon=1e-3),
-        callbacks=Callbacks([
-            StatPrinter(), PeriodicCallback(ModelSaver(), 3),
+        callbacks=[
+            PeriodicCallback(ModelSaver(), 3),
             ScheduledHyperParamSetter('learning_rate', [(200, 1e-4)])
-        ]),
+        ],
         model=Model(),
         step_per_epoch=dataset.size(),
         max_epoch=300,

examples/GAN/InfoGAN-mnist.py

@@ -158,9 +158,7 @@ def get_config():
     return TrainConfig(
         dataflow=dataset,
         optimizer=tf.train.AdamOptimizer(lr, beta1=0.5, epsilon=1e-6),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
-        ]),
+        callbacks=[ModelSaver()],
         session_config=get_default_sess_config(0.5),
         model=Model(),
         step_per_epoch=500,

examples/HED/hed.py

@@ -173,13 +173,13 @@ def get_config():
     return TrainConfig(
         dataflow=dataset_train,
         optimizer=tf.train.AdamOptimizer(lr, epsilon=1e-3),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
+        callbacks=[
+            ModelSaver(),
             ScheduledHyperParamSetter('learning_rate', [(30, 6e-6), (45, 1e-6), (60, 8e-7)]),
             HumanHyperParamSetter('learning_rate'),
             InferenceRunner(dataset_val,
                             BinaryClassificationStats('prediction', 'edgemap4d'))
-        ]),
+        ],
         model=Model(),
         step_per_epoch=step_per_epoch,
         max_epoch=100,

examples/Inception/inception-bn.py

@@ -160,17 +160,16 @@ def get_config():
     return TrainConfig(
         dataflow=dataset_train,
         optimizer=tf.train.MomentumOptimizer(lr, 0.9),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
+        callbacks=[
+            ModelSaver(),
             InferenceRunner(dataset_val, [
                 ClassificationError('wrong-top1', 'val-top1-error'),
                 ClassificationError('wrong-top5', 'val-top5-error')]),
-            # HumanHyperParamSetter('learning_rate', 'hyper-googlenet.txt')
             ScheduledHyperParamSetter('learning_rate',
                                       [(8, 0.03), (14, 0.02), (17, 5e-3),
                                        (19, 3e-3), (24, 1e-3), (26, 2e-4),
                                        (30, 5e-5)])
-        ]),
+        ],
         session_config=get_default_sess_config(0.99),
         model=Model(),
         step_per_epoch=step_per_epoch,

examples/Inception/inceptionv3.py

@@ -268,8 +268,8 @@ def get_config():
     return TrainConfig(
         dataflow=dataset_train,
         optimizer=tf.train.AdamOptimizer(lr, epsilon=1e-3),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
+        callbacks=[
+            ModelSaver(),
             InferenceRunner(dataset_val, [
                 ClassificationError('wrong-top1', 'val-error-top1'),
                 ClassificationError('wrong-top5', 'val-error-top5')]),
@@ -278,7 +278,7 @@ def get_config():
                                        (17, 0.003), (22, 1e-3), (36, 2e-4),
                                        (41, 8e-5), (48, 1e-5), (53, 2e-6)]),
             HumanHyperParamSetter('learning_rate')
-        ]),
+        ],
         session_config=get_default_sess_config(0.9),
         model=Model(),
         step_per_epoch=5000,

examples/PennTreebank/PTB-LSTM.py

@@ -126,8 +126,8 @@ def get_config():
         data=train_data,
         model=M,
         optimizer=tf.train.GradientDescentOptimizer(lr),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
+        callbacks=[
+            ModelSaver(),
             HyperParamSetterWithFunc(
                 'learning_rate',
                 lambda e, x: x * 0.80 if e > 6 else x),
@@ -139,7 +139,7 @@ def get_config():
                     'validation_perplexity',
                     np.exp(self.trainer.stat_holder.get_stat_now('validation_cost') / SEQ_LEN))),
             RunOp(lambda: M.reset_lstm_state()),
-        ]),
+        ],
         max_epoch=70,
     )
 

examples/ResNet/cifar10-resnet.py

@@ -141,13 +141,13 @@ def get_config():
     return TrainConfig(
         dataflow=dataset_train,
         optimizer=tf.train.MomentumOptimizer(lr, 0.9),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
+        callbacks=[
+            ModelSaver(),
             InferenceRunner(dataset_test,
                             [ScalarStats('cost'), ClassificationError()]),
             ScheduledHyperParamSetter('learning_rate',
                                       [(1, 0.1), (82, 0.01), (123, 0.001), (300, 0.0002)])
-        ]),
+        ],
         model=Model(n=NUM_UNITS),
         step_per_epoch=step_per_epoch,
         max_epoch=400,

examples/ResNet/imagenet-resnet.py

@@ -189,15 +189,15 @@ def get_config():
     return TrainConfig(
         dataflow=dataset_train,
         optimizer=tf.train.MomentumOptimizer(lr, 0.9, use_nesterov=True),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
+        callbacks=[
+            ModelSaver(),
             InferenceRunner(dataset_val, [
                 ClassificationError('wrong-top1', 'val-error-top1'),
                 ClassificationError('wrong-top5', 'val-error-top5')]),
             ScheduledHyperParamSetter('learning_rate',
                                       [(30, 1e-2), (60, 1e-3), (85, 1e-4), (95, 1e-5)]),
             HumanHyperParamSetter('learning_rate'),
-        ]),
+        ],
         model=Model(),
         step_per_epoch=5000,
         max_epoch=110,

examples/ResNet/svhn-resnet.py

@@ -70,14 +70,13 @@ def get_config():
     return TrainConfig(
         dataflow=dataset_train,
         optimizer=tf.train.MomentumOptimizer(lr, 0.9),
-        callbacks=Callbacks([
-            StatPrinter(),
+        callbacks=[
             ModelSaver(),
             InferenceRunner(dataset_test,
                             [ScalarStats('cost'), ClassificationError()]),
             ScheduledHyperParamSetter('learning_rate',
                                       [(1, 0.1), (20, 0.01), (28, 0.001), (50, 0.0001)])
-        ]),
+        ],
         model=Model(n=18),
         step_per_epoch=step_per_epoch,
         max_epoch=500,

examples/SimilarityLearning/mnist-embeddings.py

@@ -141,11 +141,10 @@ def get_config(model):
         dataflow=dataset,
         model=model(),
         optimizer=tf.train.GradientDescentOptimizer(lr),
-        callbacks=Callbacks([
-            StatPrinter(),
+        callbacks=[
             ModelSaver(),
             ScheduledHyperParamSetter('learning_rate', [(10, 1e-5), (20, 1e-6)])
-        ]),
+        ],
         step_per_epoch=step_per_epoch,
         max_epoch=20,
     )

examples/SpatialTransformer/mnist-addition.py

@@ -155,12 +155,12 @@ def get_config():
     return TrainConfig(
         dataflow=dataset_train,
         optimizer=tf.train.AdamOptimizer(lr, epsilon=1e-3),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
+        callbacks=[
+            ModelSaver(),
             InferenceRunner(dataset_test,
                             [ScalarStats('cost'), ClassificationError()]),
             ScheduledHyperParamSetter('learning_rate', [(200, 1e-4)])
-        ]),
+        ],
         session_config=get_default_sess_config(0.5),
         model=Model(),
         step_per_epoch=step_per_epoch,

examples/cifar-convnet.py

@@ -122,12 +122,12 @@ def lr_func(lr):
     return TrainConfig(
         dataflow=dataset_train,
         optimizer=tf.train.AdamOptimizer(lr, epsilon=1e-3),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
+        callbacks=[
+            ModelSaver(),
             InferenceRunner(dataset_test, ClassificationError()),
             StatMonitorParamSetter('learning_rate', 'val_error', lr_func,
                                    threshold=0.001, last_k=10),
-        ]),
+        ],
         session_config=sess_config,
         model=Model(cifar_classnum),
         step_per_epoch=step_per_epoch,

examples/mnist-convnet.py

@@ -140,14 +140,13 @@ def get_config():
     return TrainConfig(
         dataflow=dataset_train,  # the DataFlow instance for training
         optimizer=tf.train.AdamOptimizer(lr),
-        callbacks=Callbacks([
-            StatPrinter(),  # print statistics in terminal after every epoch
+        callbacks=[
             ModelSaver(),   # save the model after every epoch
             InferenceRunner(    # run inference(for validation) after every epoch
                 dataset_test,   # the DataFlow instance used for validation
                 # Calculate both the cost and the error for this DataFlow
                 [ScalarStats('cross_entropy_loss'), ClassificationError('incorrect')]),
-        ]),
+        ],
         model=Model(),
         step_per_epoch=step_per_epoch,
         max_epoch=100,

examples/svhn-digit-convnet.py

@@ -101,11 +101,11 @@ def get_config():
     return TrainConfig(
         dataflow=data_train,
         optimizer=tf.train.AdamOptimizer(lr),
-        callbacks=Callbacks([
-            StatPrinter(), ModelSaver(),
+        callbacks=[
+            ModelSaver(),
             InferenceRunner(data_test,
                             [ScalarStats('cost'), ClassificationError()])
-        ]),
+        ],
         model=Model(),
         step_per_epoch=step_per_epoch,
         max_epoch=350,

tensorpack/callbacks/group.py

@@ -59,13 +59,20 @@ def __init__(self, cbs):
         for cb in cbs:
             assert isinstance(cb, Callback), cb.__class__
         # move "StatPrinter" to the last
-        for cb in cbs:
+        # TODO don't need to manually move in the future.
+        found = False
+        for idx, cb in enumerate(cbs):
             if isinstance(cb, StatPrinter):
+                if found:
+                    raise ValueError("Callbacks cannot contain two StatPrinter!")
                 sp = cb
                 cbs.remove(sp)
                 cbs.append(sp)
-                break
-        else:
+                if idx != len(cbs) - 1:
+                    logger.warn("StatPrinter should appear as the last element of callbacks! "
+                                "This is now fixed automatically, but may not work in the future.")
+                found = True
+        if not found:
             raise ValueError("Callbacks must contain StatPrinter for stat and writer to work properly!")
 
         self.cbs = cbs

tensorpack/train/config.py

@@ -79,9 +79,9 @@ def assert_type(v, tp):
 
         if isinstance(callbacks, Callbacks):
             # keep quiet now because I haven't determined the final API yet.
-            # logger.warn("[Deprecated] API of TrainConfig(callbacks=) has changed!")
-            # logger.warn("[Deprecated] Please change the option 'callbacks=' to a list of "
-                        # "callbacks without StatPrinter().")
+            logger.warn("[Deprecated] API of TrainConfig(callbacks=) has changed!")
+            logger.warn("[Deprecated] Please change the argument 'callbacks=' to a *list* of "
+                        "callbacks without StatPrinter().")
             callbacks = callbacks.cbs[:-1]  # the last one is StatPrinter()
         assert_type(callbacks, list)
         if extra_callbacks is None: