testing and loading script

.gitignore

 *.gz
+*.npy
 train_log
 
 # Byte-compiled / optimized / DLL files

example_mnist.py

@@ -62,7 +62,7 @@ def get_model(inputs, is_training):
     # fc will have activation summary by default. disable this for the output layer
     logits = FullyConnected('fc1', l, out_dim=10,
                          summary_activation=False, nl=tf.identity)
-    prob = tf.nn.softmax(logits, name='output')
+    prob = tf.nn.softmax(logits, name='prob')
 
     y = one_hot(label, 10)
     cost = tf.nn.softmax_cross_entropy_with_logits(logits, y)

scripts/dump_model_params.py

+#!/usr/bin/env python2
+# -*- coding: UTF-8 -*-
+# File: dump_model_params.py
+# Author: Yuxin Wu <ppwwyyxx@gmail.com>
+
+import argparse
+import cv2
+import tensorflow as tf
+import imp
+
+from tensorpack.utils import *
+from tensorpack.utils import sessinit
+from tensorpack.dataflow import *
+from tensorpack.predict import DatasetPredictor
+
+
+parser = argparse.ArgumentParser()
+parser.add_argument(dest='config')
+parser.add_argument(dest='model')
+parser.add_argument(dest='output')
+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['get_model_func'](config['inputs'], is_training=False)
+    init = sessinit.SaverRestore(args.model)
+    sess = tf.Session()
+    init.init(sess)
+    with sess.as_default():
+        sessinit.dump_session_params(args.output)
+
+
+

scripts/imgclassify.py

+#!/usr/bin/env python2
+# -*- coding: UTF-8 -*-
+# File: imgclassify.py
+# Author: Yuxin Wu <ppwwyyxx@gmail.com>
+
+
+import argparse
+import cv2
+import tensorflow as tf
+import imp
+
+from tensorpack.utils import *
+from tensorpack.utils import sessinit
+from tensorpack.dataflow import *
+from tensorpack.predict import DatasetPredictor
+
+
+parser = argparse.ArgumentParser()
+parser.add_argument(dest='config')
+parser.add_argument(dest='model')
+parser.add_argument(dest='images', nargs='+')
+parser.add_argument('--output_type', default='label',
+                    choices=['label', 'label-prob', 'raw'])
+parser.add_argument('--top', default=1, type=int)
+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'
+
+    ds = ImageFromFile(args.images, 3, resize=(227, 227))
+    predictor = DatasetPredictor(config, ds, batch=128)
+    res = predictor.get_all_result()
+
+    if args.output_type == 'label':
+        for r in res:
+            print r.argsort()[-top:][::-1]
+    elif args.output_type == 'label_prob':
+        raise NotImplementedError
+    elif args.output_type == 'raw':
+        print res

tensorpack/dataflow/common.py

@@ -17,6 +17,7 @@ class BatchData(DataFlow):
             if set, might return a data point of a different shape
         """
         self.ds = ds
+        if not remainder:
             assert batch_size <= ds.size()
         self.batch_size = batch_size
         self.remainder = remainder
@@ -85,7 +86,6 @@ class FakeData(DataFlow):
         for _ in xrange(self._size):
             yield [np.random.random(k) for k in self.shapes]
 
-
 class MapData(DataFlow):
     """ Apply a function to the given index in the datapoint"""
     def __init__(self, ds, func, index=0):

tensorpack/dataflow/image.py

+#!/usr/bin/env python2
+# -*- coding: UTF-8 -*-
+# File: image.py
+# Author: Yuxin Wu <ppwwyyxx@gmail.com>
+
+import numpy as np
+import cv2
+from .base import DataFlow
+
+__all__ = ['ImageFromFile']
+
+class ImageFromFile(DataFlow):
+    """ generate rgb images from files """
+    def __init__(self, files, channel, resize=None):
+        """ files: list of file path
+            channel: 1 or 3 channel
+            resize: a (w, h) tuple. If given, will force a resize
+        """
+        self.files = files
+        self.channel = int(channel)
+        self.resize = resize
+
+    def size(self):
+        return len(self.files)
+
+    def get_data(self):
+        for f in self.files:
+            im = cv2.imread(
+                f, cv2.IMREAD_GRAYSCALE if self.channel == 1 else cv2.IMREAD_COLOR)
+            if self.channel == 3:
+                im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
+            if self.resize is not None:
+                im = cv2.resize(im, self.resize)
+            yield (im,)
+

tensorpack/predict.py

 #!/usr/bin/env python2
 # -*- coding: UTF-8 -*-
-# File: infer.py
+# File: predict.py
 # Author: Yuxin Wu <ppwwyyxx@gmail.com>
 
 import tensorflow as tf
@@ -11,7 +11,7 @@ import numpy as np
 from utils import *
 from utils.modelutils import describe_model
 from utils import logger
-from dataflow import DataFlow
+from dataflow import DataFlow, BatchData
 
 def get_predict_func(config):
     """
@@ -27,6 +27,10 @@ def get_predict_func(config):
 
     sess_init = config['session_init']
 
+    # Provide this if only specific output is needed.
+    # by default will evaluate all outputs as well as cost
+    output_var_name = config.get('output_var', None)
+
     # input/output variables
     input_vars = config['inputs']
     get_model_func = config['get_model_func']
@@ -38,18 +42,30 @@ def get_predict_func(config):
     sess_init.init(sess)
 
     def run_input(dp):
+        # TODO if input and dp not aligned?
         feed = dict(zip(input_vars, dp))
-        results = sess.run(
-            [cost_var] + output_vars, feed_dict=feed)
+        if output_var_name is not None:
+            fetches = tf.get_default_graph().get_tensor_by_name(output_var_name)
+            results = sess.run(fetches, feed_dict=feed)
+            return results[0]
+        else:
+            fetches = [cost_var] + output_vars
+            results = sess.run(fetches, feed_dict=feed)
             cost = results[0]
             outputs = results[1:]
             return cost, outputs
     return run_input
 
 class DatasetPredictor(object):
-    def __init__(self, predict_config, dataset):
+    def __init__(self, predict_config, dataset, batch=0):
+        """
+        A predictor with the given predict_config, run on the given dataset
+        if batch is larger than zero, the dataset will be batched
+        """
         assert isinstance(dataset, DataFlow)
         self.ds = dataset
+        if batch > 0:
+            self.ds = BatchData(self.ds, batch, remainder=True)
         self.predict_func = get_predict_func(predict_config)
 
     def get_result(self):

tensorpack/utils/sessinit.py

@@ -4,6 +4,7 @@
 # Author: Yuxin Wu <ppwwyyxx@gmail.com>
 
 from abc import abstractmethod
+import numpy as np
 import tensorflow as tf
 
 from . import logger
@@ -24,7 +25,7 @@ class SaverRestore(SessionInit):
         saver = tf.train.Saver()
         saver.restore(sess, self.path)
         logger.info(
-            "Restore checkpoint from {}".format(ckpt.model_checkpoint_path))
+            "Restore checkpoint from {}".format(self.path))
 
     def set_path(self, model_path):
         self.path = model_path
@@ -44,3 +45,12 @@ class ParamRestore(SessionInit):
                 continue
             logger.info("Restoring param {}".format(name))
             sess.run(var.assign(value))
+
+def dump_session_params(path):
+    var = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
+    result = {}
+    for v in var:
+        result[v.name] = v.eval()
+    logger.info("Params to save to {}:".format(path))
+    logger.info(str(result.keys()))
+    np.save(path, result)