fix scripts on new config

example_alexnet.py

@@ -83,7 +83,7 @@ def get_config():
 
     dataset_train = FakeData([(227,227,3), tuple()], 10)
     dataset_train = BatchData(dataset_train, 10)
-    step_per_epoch = 3
+    step_per_epoch = 1
 
     sess_config = get_default_sess_config()
     sess_config.gpu_options.per_process_gpu_memory_fraction = 0.5
@@ -105,12 +105,12 @@ def get_config():
         decay_rate=0.1, staircase=True, name='learning_rate')
     tf.scalar_summary('learning_rate', lr)
 
-    param_dict = np.load('alexnet1.npy').item()
+    param_dict = np.load('alexnet.npy').item()
 
     return TrainConfig(
         dataset=dataset_train,
         optimizer=tf.train.AdamOptimizer(lr),
-        callback=Callbacks([
+        callbacks=Callbacks([
             SummaryWriter(),
             PeriodicSaver(),
             #ValidationError(dataset_test, prefix='test'),
@@ -162,4 +162,4 @@ if __name__ == '__main__':
     #start_train(get_config())
 
     # run alexnet with given model (in npy format)
-    run_test('alexnet.npy')
+    run_test('alexnet-tuned.npy')

scripts/dump_model_params.py

@@ -23,10 +23,9 @@ args = parser.parse_args()
 get_config_func = imp.load_source('config_script', args.config).get_config
 
 with tf.Graph().as_default() as G:
-    global_step_var = tf.Variable(
-        0, trainable=False, name=GLOBAL_STEP_OP_NAME)
+    global_step_var = get_global_step_var()
     config = get_config_func()
-    config['get_model_func'](config['inputs'], is_training=False)
+    config.get_model_func(config.inputs, is_training=False)
     init = sessinit.SaverRestore(args.model)
     sess = tf.Session()
     init.init(sess)

scripts/imgclassify.py

@@ -12,7 +12,7 @@ import imp
 from tensorpack.utils import *
 from tensorpack.utils import sessinit
 from tensorpack.dataflow import *
-from tensorpack.predict import DatasetPredictor
+from tensorpack.predict import PredictConfig, DatasetPredictor
 
 
 parser = argparse.ArgumentParser()
@@ -27,11 +27,14 @@ args = parser.parse_args()
 get_config_func = imp.load_source('config_script', args.config).get_config
 
 with tf.Graph().as_default() as G:
-    global_step_var = tf.Variable(
-        0, trainable=False, name=GLOBAL_STEP_OP_NAME)
-    config = get_config_func()
-    config['session_init'] = sessinit.SaverRestore(args.model)
-    config['output_var'] = 'output:0'
+    train_config = get_config_func()
+    config = PredictConfig(
+        inputs=train_config.inputs,
+        input_dataset_mapping=[train_config.inputs[0]],  # assume first component is image
+        get_model_func=train_config.get_model_func,
+        session_init=sessinit.SaverRestore(args.model),
+        output_var_names=['output:0']
+    )
 
     ds = ImageFromFile(args.images, 3, resize=(227, 227))
     predictor = DatasetPredictor(config, ds, batch=128)
@@ -39,7 +42,7 @@ with tf.Graph().as_default() as G:
 
     if args.output_type == 'label':
         for r in res:
-            print r.argsort()[-top:][::-1]
+            print r[0].argsort(axis=1)[:,-args.top:][:,::-1]
     elif args.output_type == 'label_prob':
         raise NotImplementedError
     elif args.output_type == 'raw':