modify add demos

tensorflow2fluid/demo/diff_test/diff.py

+import sys
+import math
+val1 = map(float, open(sys.argv[1]).read().strip().split('\n'))
+val2 = map(float, open(sys.argv[2]).read().strip().split('\n'))
+
+if len(val1) != len(val2):
+    raise Exception("Not Same Length")
+
+max_diff = 0
+avg_diff = 0
+for i in range(len(val1)):
+    diff = math.fabs(val1[i] - val2[i])
+    if diff > max_diff:
+        max_diff = diff
+    avg_diff += diff
+avg_diff /= len(val1)
+print("max_diff: {}\tavg_diff: {}".format(max_diff, avg_diff))

tensorflow2fluid/demo/diff_test/resnet_v1_101_infer.py

+# coding:utf-8
+import sys
+sys.path.append("..")
+from paddle_resnet_v1_101.mymodel import KitModel
+import paddle.fluid as fluid
+import numpy
+
+use_cuda = True
+
+
+def model_initialize():
+    # 构建模型结构，并初始化参数
+    result = KitModel()
+    if use_cuda:
+        exe = fluid.Executor(fluid.CUDAPlace(0))
+    else:
+        exe = fluid.Executor(fluid.CPUPlace())
+    exe.run(fluid.default_startup_program())
+
+    # 根据save_var.list列表，加载模型参数
+    var_list = list()
+    global_block = fluid.default_main_program().global_block()
+    with open('../paddle_resnet_v1_101/save_var.list') as f:
+        for line in f:
+            try:
+                # 过滤部分不需要加载的参数（OP配置参数）
+                var = global_block.var(line.strip())
+                var_list.append(var)
+            except:
+                pass
+    fluid.io.load_vars(exe, '../paddle_resnet_v1_101', vars=var_list)
+
+    prog = fluid.default_main_program()
+    return exe, prog, result
+
+
+def test_case(exe, prog, result):
+    # 测试随机数据输入
+    numpy.random.seed(13)
+    img_data = numpy.random.rand(1000, 224, 224, 3)
+    # tf中输入为NHWC，PaddlePaddle则为NCHW，需transpose
+    img_data = numpy.transpose(img_data, (0, 3, 1, 2))
+
+    # input_0为输入数据的张量名，张量名和数据类型须与my_model.py中定义一致
+    for i in range(0, 50):
+        r, = exe.run(
+            fluid.default_main_program(),
+            feed={
+                'input_0':
+                numpy.array(img_data[i * 20:i * 20 + 20], dtype='float32')
+            },
+            fetch_list=[result])
+        r = r.flatten()
+        files = open('fluid_resnet_v1_101.result', 'a+')
+        for i in range(0, r.shape[0]):
+            files.write(str(r[i]) + '\n')
+        files.close()
+
+    # 调用save_inference_model可将模型结构（当前以代码形式保存）和参数均序列化保存
+    # 保存后的模型可使用load_inference_model加载
+    # http://www.paddlepaddle.org/documentation/docs/zh/1.2/api_cn/api_guides/low_level/inference.html#api-guide-inference
+    # fluid.io.save_inference_model("./paddle_model", ["input_0"], [result], exe)
+
+
+if __name__ == "__main__":
+    exe, prog, result = model_initialize()
+    test_case(exe, prog, result)

tensorflow2fluid/demo/diff_test/resnet_v1_50_infer.py

+# coding:utf-8
+import sys
+sys.path.append("..")
+from paddle_resnet_v1_50.mymodel import KitModel
+import paddle.fluid as fluid
+import numpy
+
+use_cuda = True
+
+
+def model_initialize():
+    # 构建模型结构，并初始化参数
+    result = KitModel()
+    if use_cuda:
+        exe = fluid.Executor(fluid.CUDAPlace(0))
+    else:
+        exe = fluid.Executor(fluid.CPUPlace())
+    exe.run(fluid.default_startup_program())
+
+    # 根据save_var.list列表，加载模型参数
+    var_list = list()
+    global_block = fluid.default_main_program().global_block()
+    with open('../paddle_resnet_v1_50/save_var.list') as f:
+        for line in f:
+            try:
+                # 过滤部分不需要加载的参数（OP配置参数）
+                var = global_block.var(line.strip())
+                var_list.append(var)
+            except:
+                pass
+    fluid.io.load_vars(exe, '../paddle_resnet_v1_50', vars=var_list)
+
+    prog = fluid.default_main_program()
+    return exe, prog, result
+
+
+def test_case(exe, prog, result):
+    # 测试随机数据输入
+    numpy.random.seed(13)
+    img_data = numpy.random.rand(1000, 224, 224, 3)
+    # tf中输入为NHWC，PaddlePaddle则为NCHW，需transpose
+    img_data = numpy.transpose(img_data, (0, 3, 1, 2))
+
+    # input_0为输入数据的张量名，张量名和数据类型须与my_model.py中定义一致
+    for i in range(0, 50):
+        r, = exe.run(
+            fluid.default_main_program(),
+            feed={
+                'input_0':
+                numpy.array(img_data[i * 20:i * 20 + 20], dtype='float32')
+            },
+            fetch_list=[result])
+        r = r.flatten()
+        files = open('fluid_resnet_v1_50.result', 'a+')
+        for i in range(0, r.shape[0]):
+            files.write(str(r[i]) + '\n')
+        files.close()
+
+    # 调用save_inference_model可将模型结构（当前以代码形式保存）和参数均序列化保存
+    # 保存后的模型可使用load_inference_model加载
+    # http://www.paddlepaddle.org/documentation/docs/zh/1.2/api_cn/api_guides/low_level/inference.html#api-guide-inference
+    # fluid.io.save_inference_model("./paddle_model", ["input_0"], [result], exe)
+
+
+if __name__ == "__main__":
+    exe, prog, result = model_initialize()
+    test_case(exe, prog, result)

tensorflow2fluid/demo/diff_test/run.sh

+rm -rf fluid_vgg_19.result
+python vgg_19_infer.py
+echo "paddle fluid vgg_19 model"
+python diff.py fluid_vgg_19.result tf_vgg_19.result
+
+rm -rf fluid_vgg_16.result
+python vgg_16_infer.py
+echo "paddle fluid vgg_16 model"
+python diff.py fluid_vgg_16.result tf_vgg_16.result
+
+rm -rf fluid_resnet_v1_50.result
+python resnet_v1_50_infer.py
+echo "paddle fluid resnet_v1_50 model"
+python diff.py fluid_resnet_v1_50.result tf_resnet_v1_50.result
+
+rm -rf fluid_resnet_v1_101.result
+python resnet_v1_101_infer.py
+echo "paddle fluid resnet_v1_101 model"
+python diff.py fluid_resnet_v1_101.result tf_resnet_v1_101.result

tensorflow2fluid/demo/diff_test/vgg_16_infer.py

+# coding:utf-8
+import sys
+sys.path.append("..")
+from paddle_vgg_16.mymodel import KitModel
+import paddle.fluid as fluid
+import numpy
+
+use_cuda = True
+
+
+def model_initialize():
+    # 构建模型结构，并初始化参数
+    result = KitModel()
+    if use_cuda:
+        exe = fluid.Executor(fluid.CUDAPlace(0))
+    else:
+        exe = fluid.Executor(fluid.CPUPlace())
+    exe.run(fluid.default_startup_program())
+
+    # 根据save_var.list列表，加载模型参数
+    var_list = list()
+    global_block = fluid.default_main_program().global_block()
+    with open('../paddle_vgg_16/save_var.list') as f:
+        for line in f:
+            try:
+                # 过滤部分不需要加载的参数（OP配置参数）
+                var = global_block.var(line.strip())
+                var_list.append(var)
+            except:
+                pass
+    fluid.io.load_vars(exe, '../paddle_vgg_16', vars=var_list)
+
+    prog = fluid.default_main_program()
+    return exe, prog, result
+
+
+def test_case(exe, prog, result):
+    # 测试随机数据输入
+    numpy.random.seed(13)
+    img_data = numpy.random.rand(1000, 224, 224, 3)
+    # tf中输入为NHWC，PaddlePaddle则为NCHW，需transpose
+    img_data = numpy.transpose(img_data, (0, 3, 1, 2))
+
+    # input_0为输入数据的张量名，张量名和数据类型须与my_model.py中定义一致
+    for i in range(0, 50):
+        r, = exe.run(
+            fluid.default_main_program(),
+            feed={
+                'input_0':
+                numpy.array(img_data[i * 20:i * 20 + 20], dtype='float32')
+            },
+            fetch_list=[result])
+        r = r.flatten()
+        files = open('fluid_vgg_16.result', 'a+')
+        for i in range(0, r.shape[0]):
+            files.write(str(r[i]) + '\n')
+        files.close()
+
+    # 调用save_inference_model可将模型结构（当前以代码形式保存）和参数均序列化保存
+    # 保存后的模型可使用load_inference_model加载
+    # http://www.paddlepaddle.org/documentation/docs/zh/1.2/api_cn/api_guides/low_level/inference.html#api-guide-inference
+    # fluid.io.save_inference_model("./paddle_model", ["input_0"], [result], exe)
+
+
+if __name__ == "__main__":
+    exe, prog, result = model_initialize()
+    test_case(exe, prog, result)

tensorflow2fluid/demo/diff_test/vgg_19_infer.py

+# coding:utf-8
+import sys
+sys.path.append("..")
+from paddle_vgg_19.mymodel import KitModel
+import paddle.fluid as fluid
+import numpy
+
+use_cuda = True
+
+
+def model_initialize():
+    # 构建模型结构，并初始化参数
+    result = KitModel()
+    if use_cuda:
+        exe = fluid.Executor(fluid.CUDAPlace(0))
+    else:
+        exe = fluid.Executor(fluid.CPUPlace())
+    exe.run(fluid.default_startup_program())
+
+    # 根据save_var.list列表，加载模型参数
+    var_list = list()
+    global_block = fluid.default_main_program().global_block()
+    with open('../paddle_vgg_19/save_var.list') as f:
+        for line in f:
+            try:
+                # 过滤部分不需要加载的参数（OP配置参数）
+                var = global_block.var(line.strip())
+                var_list.append(var)
+            except:
+                pass
+    fluid.io.load_vars(exe, '../paddle_vgg_19', vars=var_list)
+
+    prog = fluid.default_main_program()
+    return exe, prog, result
+
+
+def test_case(exe, prog, result):
+    # 测试随机数据输入
+    numpy.random.seed(13)
+    img_data = numpy.random.rand(1000, 224, 224, 3)
+    # tf中输入为NHWC，PaddlePaddle则为NCHW，需transpose
+    img_data = numpy.transpose(img_data, (0, 3, 1, 2))
+
+    # input_0为输入数据的张量名，张量名和数据类型须与my_model.py中定义一致
+    for i in range(0, 50):
+        r, = exe.run(
+            fluid.default_main_program(),
+            feed={
+                'input_0':
+                numpy.array(img_data[i * 20:i * 20 + 20], dtype='float32')
+            },
+            fetch_list=[result])
+        r = r.flatten()
+        files = open('fluid_vgg_19.result', 'a+')
+        for i in range(0, r.shape[0]):
+            files.write(str(r[i]) + '\n')
+        files.close()
+
+    # 调用save_inference_model可将模型结构（当前以代码形式保存）和参数均序列化保存
+    # 保存后的模型可使用load_inference_model加载
+    # http://www.paddlepaddle.org/documentation/docs/zh/1.2/api_cn/api_guides/low_level/inference.html#api-guide-inference
+    # fluid.io.save_inference_model("./paddle_model", ["input_0"], [result], exe)
+
+
+if __name__ == "__main__":
+    exe, prog, result = model_initialize()
+    test_case(exe, prog, result)

tensorflow2fluid/demo/export_to_checkpoint.py

 from tensorflow.contrib.slim.nets import inception
 from tensorflow.contrib.slim.nets import vgg as vgg
 from tensorflow.contrib.slim.nets import resnet_v1 as resnet_v1
-from tensorflow.contrib.framework.python.ops import arg_scope
 import tensorflow.contrib.slim as slim
 import tensorflow as tf
-import numpy
 from six import text_type as _text_type
 
+
 def inception_v3(ckpt_file):
     def get_tuned_variables():
         CHECKPOINT_EXCLUDE_SCOPES = 'InceptionV3/Logits,InceptionV3/AuxLogits'
-        exclusions = [scope.strip() for scope in CHECKPOINT_EXCLUDE_SCOPES.split(',')]
+        exclusions = [
+            scope.strip() for scope in CHECKPOINT_EXCLUDE_SCOPES.split(',')
+        ]
         variables_to_restore = []
-    
         for var in slim.get_model_variables():
             excluded = False
             for exclusion in exclusions:
@@ -24,75 +24,102 @@ def inception_v3(ckpt_file):
         return variables_to_restore
 
     img_size = inception.inception_v3.default_image_size
-    img = tf.placeholder(tf.float32, shape=[None, img_size, img_size, 3], name='inputs')
+    img = tf.placeholder(
+        tf.float32, shape=[None, img_size, img_size, 3], name='inputs')
     with slim.arg_scope(inception.inception_v3_arg_scope()):
-        logits, _ = inception.inception_v3(img, num_classes=1000, is_training=False)
+        logits, _ = inception.inception_v3(
+            img, num_classes=1000, is_training=False)
 
     sess = tf.Session()
     init = tf.global_variables_initializer()
     sess.run(init)
-    load_model = tf.contrib.slim.assign_from_checkpoint_fn(ckpt_file, get_tuned_variables(), ignore_missing_vars=True)
+    load_model = tf.contrib.slim.assign_from_checkpoint_fn(
+        ckpt_file, get_tuned_variables(), ignore_missing_vars=True)
     load_model(sess)
     return sess
 
+
 def resnet_v1_50(ckpt_file):
     img_size = resnet_v1.resnet_v1.default_image_size
-    img = tf.placeholder(tf.float32, shape=[None, img_size, img_size, 3], name='inputs')
+    img = tf.placeholder(
+        tf.float32, shape=[None, img_size, img_size, 3], name='inputs')
     with slim.arg_scope(resnet_v1.resnet_arg_scope()):
-        net, endpoint = resnet_v1.resnet_v1_50(img, num_classes=1000, is_training=False)
+        net, endpoint = resnet_v1.resnet_v1_50(
+            img, num_classes=1000, is_training=False)
 
     sess = tf.Session()
-    load_model = tf.contrib.slim.assign_from_checkpoint_fn(ckpt_file, tf.contrib.slim.get_model_variables("resnet_v1_50"))
+    load_model = tf.contrib.slim.assign_from_checkpoint_fn(
+        ckpt_file, tf.contrib.slim.get_model_variables("resnet_v1_50"))
     load_model(sess)
     return sess
 
+
 def resnet_v1_101(ckpt_file):
     img_size = resnet_v1.resnet_v1.default_image_size
-    img = tf.placeholder(tf.float32, shape=[None, img_size, img_size, 3], name='inputs')
+    img = tf.placeholder(
+        tf.float32, shape=[None, img_size, img_size, 3], name='inputs')
     with slim.arg_scope(resnet_v1.resnet_arg_scope()):
-        net, endpoint = resnet_v1.resnet_v1_101(img, num_classes=1000, is_training=False)
+        net, endpoint = resnet_v1.resnet_v1_101(
+            img, num_classes=1000, is_training=False)
 
     sess = tf.Session()
     sess.run(tf.global_variables_initializer())
-    load_model = tf.contrib.slim.assign_from_checkpoint_fn(ckpt_file, tf.contrib.slim.get_model_variables("resnet_v1_101"))
+    load_model = tf.contrib.slim.assign_from_checkpoint_fn(
+        ckpt_file, tf.contrib.slim.get_model_variables("resnet_v1_101"))
     load_model(sess)
     return sess
 
+
 def vgg_16(ckpt_file):
     img_size = vgg.vgg_16.default_image_size
-    inputs = tf.placeholder(tf.float32, shape=[None, img_size, img_size, 3],
-            name="inputs")
+    inputs = tf.placeholder(
+        tf.float32, shape=[None, img_size, img_size, 3], name="inputs")
     logits, endpoint = vgg.vgg_16(inputs, num_classes=1000, is_training=False)
     sess = tf.Session()
 
-    load_model = tf.contrib.slim.assign_from_checkpoint_fn(ckpt_file,
-            tf.contrib.slim.get_model_variables("vgg_16"))
+    load_model = tf.contrib.slim.assign_from_checkpoint_fn(
+        ckpt_file, tf.contrib.slim.get_model_variables("vgg_16"))
     load_model(sess)
     return sess
 
+
 def vgg_19(ckpt_file):
     img_size = vgg.vgg_19.default_image_size
-    inputs = tf.placeholder(tf.float32, shape=[None, img_size, img_size, 3],
-            name="inputs")
+    inputs = tf.placeholder(
+        tf.float32, shape=[None, img_size, img_size, 3], name="inputs")
     logits, endpoint = vgg.vgg_19(inputs, num_classes=1000, is_training=False)
     sess = tf.Session()
-    load_model = tf.contrib.slim.assign_from_checkpoint_fn(ckpt_file,
-            tf.contrib.slim.get_model_variables("vgg_19"))
+    load_model = tf.contrib.slim.assign_from_checkpoint_fn(
+        ckpt_file, tf.contrib.slim.get_model_variables("vgg_19"))
     load_model(sess)
     return sess
 
+
 def save_checkpoint(sess, save_dir):
     saver = tf.train.Saver()
-    saver.save(sess, save_dir+"/model")
+    saver.save(sess, save_dir + "/model")
+
 
 def get_parser():
     import argparse
     parser = argparse.ArgumentParser()
-    parser.add_argument("--model", "-m", type=_text_type, default=None, help="inception_v3/resnet_v1_50/resnet_v1_101/vgg_16/vgg_19")
-    parser.add_argument("--ckpt_file", "-c", type=_text_type, default=None, help="parameters ckpt file")
-    parser.add_argument("--save_dir", "-s", type=_text_type, default=None, help="model path")
+    parser.add_argument(
+        "--model",
+        "-m",
+        type=_text_type,
+        default=None,
+        help="inception_v3/resnet_v1_50/resnet_v1_101/vgg_16/vgg_19")
+    parser.add_argument(
+        "--ckpt_file",
+        "-c",
+        type=_text_type,
+        default=None,
+        help="parameters ckpt file")
+    parser.add_argument(
+        "--save_dir", "-s", type=_text_type, default=None, help="model path")
     return parser
 
+
 if __name__ == "__main__":
     parser = get_parser()
     args = parser.parse_args()
@@ -110,5 +137,7 @@ if __name__ == "__main__":
     elif args.model == "vgg_19":
         sess = vgg_19(args.ckpt_file)
     else:
-        raise Exception("Only support inception_v3/resnet_v1_50/resnet_v1_101/vgg_16/vgg_19")
+        raise Exception(
+            "Only support inception_v3/resnet_v1_50/resnet_v1_101/vgg_16/vgg_19"
+        )
     save_checkpoint(sess, args.save_dir)

tensorflow2fluid/demo/resnet_v1_101.sh

+export CUDA_VISIBLE_DEVICES=-1
+
+wget http://download.tensorflow.org/models/resnet_v1_101_2016_08_28.tar.gz
+tar xzvf resnet_v1_101_2016_08_28.tar.gz
+python export_to_checkpoint.py --model resnet_v1_101 --ckpt_file resnet_v1_101.ckpt --save_dir resnet_v1_101_checkpoint
+rm resnet_v1_101_2016_08_28.tar.gz resnet_v1_101.ckpt
+
+tf2fluid --meta_file resnet_v1_101_checkpoint/model.meta \
+         --ckpt_dir resnet_v1_101_checkpoint \
+         --in_nodes inputs \
+         --input_shape None,224,224,3 \
+         --output_nodes resnet_v1_101/predictions/Softmax \
+         --save_dir paddle_resnet_v1_101

tensorflow2fluid/demo/resnet_v1_50.sh

+export CUDA_VISIBLE_DEVICES=-1
+
+wget http://download.tensorflow.org/models/resnet_v1_50_2016_08_28.tar.gz
+tar xzvf resnet_v1_50_2016_08_28.tar.gz
+python export_to_checkpoint.py --model resnet_v1_50 --ckpt_file resnet_v1_50.ckpt --save_dir resnet_v1_50_checkpoint
+rm resnet_v1_50_2016_08_28.tar.gz resnet_v1_50.ckpt
+
+tf2fluid --meta_file resnet_v1_50_checkpoint/model.meta \
+         --ckpt_dir resnet_v1_50_checkpoint \
+         --in_nodes inputs \
+         --input_shape None,224,224,3 \
+         --output_nodes resnet_v1_50/predictions/Softmax \
+         --save_dir paddle_resnet_v1_50

tensorflow2fluid/demo/vgg_16.sh

+export CUDA_VISIBLE_DEVICES=-1
+
+wget http://download.tensorflow.org/models/vgg_16_2016_08_28.tar.gz
+tar xzvf vgg_16_2016_08_28.tar.gz
+python export_to_checkpoint.py --model vgg_16 --ckpt_file vgg_16.ckpt --save_dir vgg_16_checkpoint
+rm vgg_16_2016_08_28.tar.gz vgg_16.ckpt
+
+tf2fluid --meta_file vgg_16_checkpoint/model.meta \
+         --ckpt_dir vgg_16_checkpoint \
+         --in_nodes inputs \
+         --input_shape None,224,224,3 \
+         --output_nodes vgg_16/fc8/squeezed \
+         --save_dir paddle_vgg_16

tensorflow2fluid/demo/vgg_19.sh

+export CUDA_VISIBLE_DEVICES=-1
+
+wget http://download.tensorflow.org/models/vgg_19_2016_08_28.tar.gz
+tar xzvf vgg_19_2016_08_28.tar.gz
+python export_to_checkpoint.py --model vgg_19 --ckpt_file vgg_19.ckpt --save_dir vgg_19_checkpoint
+rm vgg_19_2016_08_28.tar.gz vgg_19.ckpt
+
+tf2fluid --meta_file vgg_19_checkpoint/model.meta \
+         --ckpt_dir vgg_19_checkpoint \
+         --in_nodes inputs \
+         --input_shape None,224,224,3 \
+         --output_nodes vgg_19/fc8/squeezed \
+         --save_dir paddle_vgg_19

tensorflow2fluid/src/paddle_emitter.py

@@ -413,11 +413,7 @@ class PaddleEmitter(object):
         data1 = node.inputs[0]
         data2 = node.inputs[1]
         axis = self.get_axis(data1, data2)
-        code = list()
-        code.append("# {}, {}, {}".format(node.layer_name, data1.layer_name,
-                                          data2.layer_name))
-        code.append("{} = layers.elementwise_add({}, {}, axis={})".format(
-            node.output_name, data1.ref_name, data2.ref_name, axis))
+        code = "{} = layers.elementwise_add({}, {}, axis={})".format(node.output_name, data1.ref_name, data2.ref_name, axis)
         return code
 
     def emit_mean(self, node):

tensorflow2fluid/src/tensorflow_parser.py

@@ -41,6 +41,7 @@ class TensorflowCkptParser(object):
 
             graph_def, ver = tensorflow.get_default_graph()._as_graph_def(
                 add_shapes=True)
+#            self.sess = sess
 
         if in_nodes is not None and input_shape is not None:
             graph_def = strip_unused_lib.strip_unused(
@@ -62,6 +63,11 @@ class TensorflowPbParser(object):
         tensorflow.reset_default_graph()
         original_graph_def = tensorflow.GraphDef()
         original_graph_def.ParseFromString(serialized)
+        
+#        tensorflow.import_graph_def(origin_graph_def, name="")
+#        self.sess = tensorflow.Session(graph=tf.get_default_graph())
+#        self.sess.run(tensorflow.global_variables_initializer())
+
         original_graph_def = strip_unused_lib.strip_unused(
             input_graph_def=original_graph_def,
             input_node_names=in_nodes,
@@ -84,7 +90,6 @@ class TensorflowPbParser(object):
             if in_type_list[in_nodes[i]] == 1 or in_type_list[
                     in_nodes[i]] == 0:
                 dtype = tensorflow.float32
-                print(input_shape[i])
                 x = tensorflow.placeholder(dtype, shape=input_shape[i])
             elif in_type_list[in_nodes[i]] == 3:
                 dtype = tensorflow.int32