Merge pull request #761 from walloollaw/add-examples-for-caffe2fluid

Add examples for caffe2fluid

fluid/image_classification/caffe2fluid/README.md

@@ -2,7 +2,8 @@
 This tool is used to convert a Caffe model to Fluid model
 
 ### Howto
-1, Prepare caffepb.py in ./proto, two options provided
+1, Prepare caffepb.py in ./proto if your python has no 'pycaffe' module, two options provided here:
+
     1) generate it from caffe.proto using protoc
         bash ./proto/compile.sh
 
@@ -12,14 +13,24 @@ This tool is used to convert a Caffe model to Fluid model
 2, Convert the caffe model using 'convert.py' which will generate a python script and a weight(in .npy) file
 
 3, Use the converted model to predict
-    see more detail info in 'tests/lenet/README.md'
+
+    see more detail info in 'examples/xxx'
 
 
-### Supported models
+### Tested models
 - Lenet on mnist dataset
 
 - ResNets:(ResNet-50, ResNet-101, ResNet-152)
-    model addrs:(https://onedrive.live.com/?authkey=%21AAFW2-FVoxeVRck&id=4006CBB8476FF777%2117887&cid=4006CBB8476FF777)
+    model addr: `https://onedrive.live.com/?authkey=%21AAFW2-FVoxeVRck&id=4006CBB8476FF777%2117887&cid=4006CBB8476FF777`_
+
+- GoogleNet:
+    model addr: `https://gist.github.com/jimmie33/7ea9f8ac0da259866b854460f4526034`_
+
+- VGG:
+    model addr: `https://gist.github.com/ksimonyan/211839e770f7b538e2d8`_
+
+- AlexNet:
+    model addr: `https://github.com/BVLC/caffe/tree/master/models/bvlc_alexnet`_
 
 ### Notes
 Some of this code come from here: https://github.com/ethereon/caffe-tensorflow

fluid/image_classification/caffe2fluid/convert.py

@@ -4,8 +4,8 @@ import os
 import sys
 import numpy as np
 import argparse
-from kaffe import KaffeError, print_stderr
 
+from kaffe import KaffeError, print_stderr
 from kaffe.paddle import Transformer
 
 
@@ -47,6 +47,8 @@ def convert(def_path, caffemodel_path, data_output_path, code_output_path,
     except KaffeError as err:
         fatal_error('Error encountered: {}'.format(err))
 
+    return 0
+
 
 def main():
     """ main
@@ -64,9 +66,10 @@ def main():
         help='The phase to convert: test (default) or train')
     args = parser.parse_args()
     validate_arguments(args)
-    convert(args.def_path, args.caffemodel, args.data_output_path,
+    return convert(args.def_path, args.caffemodel, args.data_output_path,
                    args.code_output_path, args.phase)
 
 
 if __name__ == '__main__':
-    main()
+    ret = main()
+    sys.exit(ret)

fluid/image_classification/caffe2fluid/examples/imagenet/README.md

+a demo to show converting caffe models on 'imagenet' using caffe2fluid
+
+---
+
+# How to use
+
+1. prepare python environment
+2. download caffe model to "models.caffe/xxx" which contains "xxx.caffemodel" and "xxx.prototxt"
+3. run the tool
+    eg: bash ./run.sh resnet50 ./models.caffe/resnet50 ./models/resnet50

fluid/image_classification/caffe2fluid/examples/imagenet/infer.py

+#!/bin/env python
+
+#function:
+#   a demo to show how to use the converted model genereated by caffe2fluid
+#   
+#notes:
+#   only support imagenet data
+
+import os
+import sys
+import inspect
+import numpy as np
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+
+
+def load_data(imgfile, shape):
+    h, w = shape[1:]
+    from PIL import Image
+    im = Image.open(imgfile)
+
+    # The storage order of the loaded image is W(widht),
+    # H(height), C(channel). PaddlePaddle requires
+    # the CHW order, so transpose them.
+    im = im.resize((w, h), Image.ANTIALIAS)
+    im = np.array(im).astype(np.float32)
+    im = im.transpose((2, 0, 1))  # CHW
+    im = im[(2, 1, 0), :, :]  # BGR
+
+    # The mean to be subtracted from each image.
+    # By default, the per-channel ImageNet mean.
+    mean = np.array([104., 117., 124.], dtype=np.float32)
+    mean = mean.reshape([3, 1, 1])
+    im = im - mean
+    return im.reshape([1] + shape)
+
+
+def build_model(net_file, net_name):
+    print('build model with net_file[%s] and net_name[%s]' %
+          (net_file, net_name))
+
+    net_path = os.path.dirname(net_file)
+    module_name = os.path.basename(net_file).rstrip('.py')
+    if net_path not in sys.path:
+        sys.path.insert(0, net_path)
+
+    try:
+        m = __import__(module_name, fromlist=[net_name])
+        MyNet = getattr(m, net_name)
+    except Exception as e:
+        print('failed to load module[%s]' % (module_name))
+        print(e)
+        return None
+
+    input_name = 'data'
+    input_shape = MyNet.input_shapes()[input_name]
+    images = fluid.layers.data(name='image', shape=input_shape, dtype='float32')
+    #label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+
+    net = MyNet({input_name: images})
+    input_shape = MyNet.input_shapes()[input_name]
+    return net, input_shape
+
+
+def dump_results(results, names, root):
+    if os.path.exists(root) is False:
+        os.path.mkdir(root)
+
+    for i in range(len(names)):
+        n = names[i]
+        res = results[i]
+        filename = os.path.join(root, n)
+        np.save(filename + '.npy', res)
+
+
+def infer(net_file, net_name, model_file, imgfile, debug=False):
+    """ do inference using a model which consist 'xxx.py' and 'xxx.npy'
+    """
+    #1, build model
+    net, input_shape = build_model(net_file, net_name)
+    prediction = net.get_output()
+
+    #2, load weights for this model
+    place = fluid.CPUPlace()
+    exe = fluid.Executor(place)
+    startup_program = fluid.default_startup_program()
+    exe.run(startup_program)
+
+    if model_file.find('.npy') > 0:
+        net.load(data_path=model_file, exe=exe, place=place)
+    else:
+        net.load(data_path=model_file, exe=exe)
+
+    #3, test this model
+    test_program = fluid.default_main_program().clone()
+
+    fetch_list_var = []
+    fetch_list_name = []
+    if debug is False:
+        fetch_list_var.append(prediction)
+    else:
+        for k, v in net.layers.items():
+            fetch_list_var.append(v)
+            fetch_list_name.append(k)
+
+    np_images = load_data(imgfile, input_shape)
+    results = exe.run(program=test_program,
+                      feed={'image': np_images},
+                      fetch_list=fetch_list_var)
+
+    if debug is True:
+        dump_path = 'results.layers'
+        dump_results(results, fetch_list_name, dump_path)
+        print('all results dumped to [%s]' % (dump_path))
+    else:
+        result = results[0]
+        print('predicted class:', np.argmax(result))
+
+
+if __name__ == "__main__":
+    """ maybe more convenient to use 'run.sh' to call this tool
+    """
+    net_file = 'models/resnet50/resnet50.py'
+    weight_file = 'models/resnet50/resnet50.npy'
+    imgfile = 'data/65.jpeg'
+    net_name = 'ResNet50'
+
+    argc = len(sys.argv)
+    if argc == 5:
+        net_file = sys.argv[1]
+        weight_file = sys.argv[2]
+        imgfile = sys.argv[3]
+        net_name = sys.argv[4]
+    elif argc > 1:
+        print('usage:')
+        print('\tpython %s [net_file] [weight_file] [imgfile] [net_name]' %
+              (sys.argv[0]))
+        print('\teg:python %s %s %s %s %s' % (sys.argv[0], net_file,
+                                              weight_file, imgfile, net_name))
+        sys.exit(1)
+
+    infer(net_file, net_name, weight_file, imgfile)

fluid/image_classification/caffe2fluid/examples/imagenet/run.sh

+#!/bin/bash
+
+#function:
+#   a tool used to:
+#       1, convert a caffe model
+#       2, do inference using this model
+#
+#usage:
+#   bash run.sh resnet50 ./models.caffe/resnet50 ./models/resnet50
+#
+
+#set -x
+if [[ $# -lt 3 ]];then
+    echo "usage:"
+    echo "  bash $0 [model_name] [cf_model_path] [pd_model_path] [only_convert]"
+    echo "  eg: bash $0 resnet50 ./models.caffe/resnet50 ./models/resnet50"
+    exit 1
+else
+    model_name=$1
+    cf_model_path=$2
+    pd_model_path=$3
+    only_convert=$4
+fi
+
+proto_file=$cf_model_path/${model_name}.prototxt
+caffemodel_file=$cf_model_path/${model_name}.caffemodel
+weight_file=$pd_model_path/${model_name}.npy
+net_file=$pd_model_path/${model_name}.py
+
+if [[ ! -e $proto_file ]];then
+    echo "not found prototxt[$proto_file]"
+    exit 1
+fi
+
+if [[ ! -e $caffemodel_file ]];then
+    echo "not found caffemodel[$caffemodel_file]"
+    exit 1
+fi
+
+if [[ ! -e $pd_model_path ]];then
+    mkdir $pd_model_path
+fi
+
+PYTHON=`which cfpython`
+if [[ -z $PYTHON ]];then
+    PYTHON=`which python`
+fi
+$PYTHON ../../convert.py \
+        $proto_file \
+        --caffemodel $caffemodel_file \
+        --data-output-path $weight_file\
+        --code-output-path $net_file
+
+ret=$?
+if [[ $ret -ne 0 ]];then
+    echo "failed to convert caffe model[$cf_model_path]"
+    exit $ret
+else
+    echo "succeed to convert caffe model[$cf_model_path] to fluid model[$pd_model_path]"
+fi
+
+if [[ -z $only_convert ]];then
+    PYTHON=`which pdpython`
+    if [[ -z $PYTHON ]];then
+        PYTHON=`which python`
+    fi
+    imgfile="data/65.jpeg"
+    net_name=`grep "name" $proto_file | head -n1 | perl -ne 'if(/\"([^\"]+)\"/){ print $1."\n";}'`
+    $PYTHON ./infer.py $net_file $weight_file $imgfile $net_name
+    ret=$?
+fi
+exit $ret

fluid/image_classification/caffe2fluid/examples/mnist/README.md

+a demo to show converting caffe model on 'mnist' using caffe2fluid
+
+---
+
+# How to use
+
+1. prepare python environment
+2. download caffe model to "models.caffe/lenet" which contains "lenet.caffemodel" and "lenet.prototxt"
+3. run the tool
+    eg: bash ./run.sh lenet ./models.caffe/lenet ./models/lenet

fluid/image_classification/caffe2fluid/tests/lenet/predict.py → fluid/image_classification/caffe2fluid/examples/mnist/evaluate.py

@@ -4,12 +4,12 @@
 #   demo to show how to use converted model using caffe2fluid
 #
 
+import sys
+import os
 import numpy as np
 import paddle.v2 as paddle
 import paddle.v2.fluid as fluid
 
-from lenet import LeNet as MyNet
-
 
 def test_model(exe, test_program, fetch_list, test_reader, feeder):
     acc_set = []
@@ -24,10 +24,15 @@ def test_model(exe, test_program, fetch_list, test_reader, feeder):
     return float(acc_val)
 
 
-def main(model_path):
+def evaluate(net_file, model_file):
     """ main
     """
-    print('load fluid model in %s' % (model_path))
+    #1, build model
+    net_path = os.path.dirname(net_file)
+    if net_path not in sys.path:
+        sys.path.insert(0, net_path)
+
+    from lenet import LeNet as MyNet
 
     with_gpu = False
     paddle.init(use_gpu=with_gpu)
@@ -45,10 +50,10 @@ def main(model_path):
     exe.run(fluid.default_startup_program())
 
     #2, load weights
-    if model_path.find('.npy') > 0:
-        net.load(data_path=model_path, exe=exe, place=place)
+    if model_file.find('.npy') > 0:
+        net.load(data_path=model_file, exe=exe, place=place)
     else:
-        net.load(data_path=model_path, exe=exe)
+        net.load(data_path=model_file, exe=exe)
 
     #3, test this model
     test_program = fluid.default_main_program().clone()
@@ -65,10 +70,17 @@ def main(model_path):
 
 
 if __name__ == "__main__":
-    import sys
-    if len(sys.argv) == 2:
-        fluid_model_path = sys.argv[1]
-    else:
-        fluid_model_path = './model.fluid'
-
-    main(fluid_model_path)
+    net_file = 'models/lenet/lenet.py'
+    weight_file = 'models/lenet/lenet.npy'
+
+    argc = len(sys.argv)
+    if argc == 3:
+        net_file = sys.argv[1]
+        weight_file = sys.argv[2]
+    elif argc > 1:
+        print('usage:')
+        print('\tpython %s [net_file] [weight_file]' % (sys.argv[0]))
+        print('\teg:python %s %s %s %s' % (sys.argv[0], net_file, weight_file))
+        sys.exit(1)
+
+    evaluate(net_file, weight_file)

fluid/image_classification/caffe2fluid/examples/mnist/run.sh

+#!/bin/bash
+
+#function:
+#   a tool used to:
+#       1, convert a caffe model
+#       2, do inference using this model
+#
+#usage:
+#   bash run.sh lenet ./models.caffe/lenet ./models/lenet
+#
+
+#set -x
+if [[ $# -lt 3 ]];then
+    echo "usage:"
+    echo "  bash $0 [model_name] [cf_model_path] [pd_model_path] [only_convert]"
+    echo "  eg: bash $0 lenet ./models.caffe/lenet ./models/lenet"
+    exit 1
+else
+    model_name=$1
+    cf_model_path=$2
+    pd_model_path=$3
+    no_eval=$4
+fi
+
+proto_file=$cf_model_path/${model_name}.prototxt
+caffemodel_file=$cf_model_path/${model_name}.caffemodel
+weight_file=$pd_model_path/${model_name}.npy
+net_file=$pd_model_path/${model_name}.py
+
+if [[ ! -e $proto_file ]];then
+    echo "not found prototxt[$proto_file]"
+    exit 1
+fi
+
+if [[ ! -e $caffemodel_file ]];then
+    echo "not found caffemodel[$caffemodel_file]"
+    exit 1
+fi
+
+if [[ ! -e $pd_model_path ]];then
+    mkdir $pd_model_path
+fi
+
+PYTHON=`which cfpython`
+if [[ -z $PYTHON ]];then
+    PYTHON=`which python`
+fi
+$PYTHON ../../convert.py \
+        $proto_file \
+        --caffemodel $caffemodel_file \
+        --data-output-path $weight_file\
+        --code-output-path $net_file
+
+ret=$?
+if [[ $ret -ne 0 ]];then
+    echo "failed to convert caffe model[$cf_model_path]"
+    exit $ret
+else
+    echo "succeed to convert caffe model[$cf_model_path] to fluid model[$pd_model_path]"
+fi
+
+if [[ -z $only_convert ]];then
+    PYTHON=`which pdpython`
+    if [[ -z $PYTHON ]];then
+        PYTHON=`which python`
+    fi
+    net_name=`grep "name" $proto_file | head -n1 | perl -ne 'if(/\"([^\"]+)\"/){ print $1."\n";}'`
+    if [[ $net_name != "LeNet" ]];then
+        echo "only support LeNet"
+        exit 1
+    fi
+    $PYTHON ./evaluate.py $net_file $weight_file
+    ret=$?
+fi
+exit $ret

fluid/image_classification/caffe2fluid/kaffe/caffe/resolver.py

@@ -54,7 +54,6 @@ def show_fallback_warning():
     WARNING: PyCaffe not found!
     Falling back to a pure protocol buffer implementation.
     * Conversions will be drastically slower.
-    * This backend is UNTESTED!
 ------------------------------------------------------------
 
 '''

fluid/image_classification/caffe2fluid/kaffe/graph.py

@@ -175,6 +175,7 @@ class GraphBuilder(object):
         kind = NodeKind.map_raw_kind(layer.type)
         if kind is None:
             raise KaffeError('Unknown layer type encountered: %s' % layer.type)
+
         # We want to use the layer's top names (the "output" names), rather than the
         # name attribute, which is more of readability thing than a functional one.
         # Other layers will refer to a node by its "top name".
@@ -235,6 +236,7 @@ class GraphBuilder(object):
                 node.add_parent(parent_node)
             if len(layer.top) > 1:
                 raise KaffeError('Multiple top nodes are not supported.')
+
             for output_name in layer.top:
                 if output_name == layer.name:
                     # Output is named the same as the node. No further action required.

fluid/image_classification/caffe2fluid/kaffe/layers.py

@@ -51,16 +51,78 @@ LAYER_DESCRIPTORS = {
     'Threshold': shape_identity,
 }
 
-LAYER_TYPES = LAYER_DESCRIPTORS.keys()
+# layer types in 'V1LayerParameter'
+# (v1layertype name, enum value, mapped to layer type)
+v1_layertypes = [
+    ('ABSVAL', 35),
+    ('ACCURACY', 1),
+    ('ARGMAX', 30),
+    ('BNLL', 2),
+    ('CONCAT', 3),
+    ('CONVOLUTION', 4),
+    ('DATA', 5),
+    ('DECONVOLUTION', 39),
+    ('DROPOUT', 6),
+    ('ELTWISE', 25),
+    ('EXP', 38),
+    ('FLATTEN', 8),
+    ('IM2COL', 11),
+    ('INNERPRODUCT', 14),
+    ('LRN', 15),
+    ('MEMORYDATA', 29),
+    ('MULTINOMIALLOGISTICLOSS', 16),
+    ('MVN', 34),
+    ('POOLING', 17),
+    ('POWER', 26),
+    ('RELU', 18),
+    ('SIGMOID', 19),
+    ('SIGMOIDCROSSENTROPYLOSS', 27),
+    ('SILENCE', 36),
+    ('SOFTMAX', 20),
+    ('SPLIT', 22),
+    ('SLICE', 33),
+    ('TANH', 23),
+    ('WINDOWDATA', 24),
+    ('THRESHOLD', 31),
+]
 
+LAYER_TYPES = LAYER_DESCRIPTORS.keys()
 LayerType = type('LayerType', (), {t: t for t in LAYER_TYPES})
 
+#map the layer name in V1 to standard name
+V1_LAYER_MAP = {'_not_init_': True}
+
+
+def get_v1_layer_map():
+    global V1_LAYER_MAP
+    if '_not_init_' not in V1_LAYER_MAP:
+        return V1_LAYER_MAP
+    else:
+        del V1_LAYER_MAP['_not_init_']
+
+    name2layer = {}
+    for n in LAYER_TYPES:
+        name2layer[n.upper()] = n
+
+    for l in v1_layertypes:
+        n, v = l
+        if n in name2layer and v not in V1_LAYER_MAP:
+            V1_LAYER_MAP[v] = name2layer[n]
+        else:
+            raise KaffeError('not found v1 layer type %s' % n)
+    return V1_LAYER_MAP
+
 
 class NodeKind(LayerType):
     @staticmethod
     def map_raw_kind(kind):
         if kind in LAYER_TYPES:
             return kind
+
+        v1_layers = get_v1_layer_map()
+        if kind in v1_layers:
+            return v1_layers[kind]
+        else:
             return None
 
     @staticmethod

fluid/image_classification/caffe2fluid/kaffe/paddle/network.py

@@ -27,6 +27,9 @@ def layer(op):
         self.layers[name] = layer_output
         # This output is now the input for the next layer.
         self.feed(layer_output)
+        #print('output shape of %s:' % (name))
+        #print layer_output.shape
+
         # Return self for chained calls.
         return self
 
@@ -158,41 +161,64 @@ class Network(object):
         output = fluid.layers.relu(x=input)
         return output
 
-    @layer
-    def max_pool(self, input, k_h, k_w, s_h, s_w, name, padding=None):
-        if padding is None:
-            padding = [0, 0]
+    def _adjust_pad_if_needed(self, i_hw, k_hw, s_hw, p_hw):
+        #adjust the padding if needed
+        i_h, i_w = i_hw
+        k_h, k_w = k_hw
+        s_h, s_w = s_hw
+        p_h, p_w = p_hw
+
+        def is_consistent(i, k, s, p):
+            o = i + 2 * p - k
+            if o % s == 0:
+                return True
+            else:
+                return False
 
+        real_p_h = 0
+        real_p_w = 0
+        if is_consistent(i_h, k_h, s_h, p_h) is False:
+            real_p_h = int(k_h / 2)
+
+        if is_consistent(i_w, k_w, s_w, p_w) is False:
+            real_p_w = int(k_w / 2)
+
+        return [real_p_h, real_p_w]
+
+    def pool(self, pool_type, input, k_h, k_w, s_h, s_w, name, padding):
         # Get the number of channels in the input
-        h_i, w_i = input.shape[2:]
-        fluid = import_fluid()
-        output = fluid.layers.pool2d(
-            input=input,
-            pool_size=[k_h, k_w],
-            pool_stride=[s_h, s_w],
-            pool_padding=padding,
-            pool_type='max')
-        return output
+        in_hw = input.shape[2:]
+        k_hw = [k_h, k_w]
+        s_hw = [s_h, s_w]
 
-    @layer
-    def avg_pool(self, input, k_h, k_w, s_h, s_w, name, padding=None):
         if padding is None:
-            padding = [0, 0]
+            #fix bug about the difference between conv and pool
+            #more info: https://github.com/BVLC/caffe/issues/1318
+            padding = self._adjust_pad_if_needed(in_hw, k_hw, s_hw, [0, 0])
 
-        # Get the number of channels in the input
-        h_i, w_i = input.shape[2:]
         fluid = import_fluid()
         output = fluid.layers.pool2d(
             input=input,
-            pool_size=[k_h, k_w],
-            pool_stride=[s_h, s_w],
+            pool_size=k_hw,
+            pool_stride=s_hw,
             pool_padding=padding,
-            pool_type='avg')
+            pool_type=pool_type)
         return output
 
+    @layer
+    def max_pool(self, input, k_h, k_w, s_h, s_w, name, padding=None):
+        return self.pool('max', input, k_h, k_w, s_h, s_w, name, padding)
+
+    @layer
+    def avg_pool(self, input, k_h, k_w, s_h, s_w, name, padding=None):
+        return self.pool('avg', input, k_h, k_w, s_h, s_w, name, padding)
+
     @layer
     def lrn(self, input, radius, alpha, beta, name, bias=1.0):
-        raise Exception('lrn() not implemented yet')
+        fluid = import_fluid()
+        output = fluid.layers.lrn(input=input, \
+                n=radius, k=bias, alpha=alpha, beta=beta, name=name)
+        return output
 
     @layer
     def concat(self, inputs, axis, name):
@@ -228,7 +254,7 @@ class Network(object):
     @layer
     def softmax(self, input, name):
         fluid = import_fluid()
-        output = fluid.layers.softmax(x=input, name=name)
+        output = fluid.layers.softmax(input)
         return output
 
     @layer
@@ -256,5 +282,8 @@ class Network(object):
         return output
 
     @layer
-    def dropout(self, input, keep_prob, name):
-        raise Exception('dropout() not implemented yet')
+    def dropout(self, input, drop_prob, name, is_test=True):
+        fluid = import_fluid()
+        output = fluid.layers.dropout(
+            input, dropout_prob=drop_prob, is_test=is_test, name=name)
+        return output

fluid/image_classification/caffe2fluid/kaffe/paddle/transformer.py

@@ -132,8 +132,7 @@ class TensorFlowMapper(NodeMapper):
         # just scales by alpha (as does Krizhevsky's paper).
         # We'll account for that here.
         alpha = params.alpha / float(params.local_size)
-        return TensorFlowNode('lrn',
-                              int(params.local_size / 2), alpha, params.beta)
+        return TensorFlowNode('lrn', params.local_size, alpha, params.beta)
 
     def map_concat(self, node):
         return TensorFlowNode('concat', node.parameters.axis)
@@ -191,22 +190,33 @@ class TensorFlowEmitter(object):
     def emit_setup_def(self):
         return self.statement('def setup(self):')
 
-    def emit_convert_def(self, input_nodes):
-        def data_layer_def(name, shape, dtype=None):
-            if dtype is None:
-                dtype = 'float32'
+    def emit_shape_def(self, input_nodes):
+        self.outdent()
+        func_def = self.statement('@classmethod')
+        func_def += self.statement('def input_shapes(cls):')
+        self.indent()
 
-            layer_var = name + '_layer'
-            shape = [str(s) for s in shape[1:]]
-            layer_def = '%s = fluid.layers.data(name="%s", shape=[%s], dtype="%s")'\
-                    % (layer_var, name, ','.join(shape), dtype)
-            return layer_var, layer_def
+        input_shapes = {}
+        for n in input_nodes:
+            name = n.name
+            output_shape = n.output_shape
+            shape = [str(s) for s in output_shape[1:]]
+            input_shapes[name] = ', '.join(shape)
+        input_shapes = ['"%s": [%s]' % (n, l) for n, l in input_shapes.items()]
+        shape_str = ','.join(input_shapes)
+        func_def += self.statement('return {%s}' % (shape_str))
+        return '\n\n' + func_def
 
+    def emit_convert_def(self, input_nodes):
         codes = []
         inputs = {}
+        codes.append('shapes = cls.input_shapes()')
         for n in input_nodes:
             name = n.name
-            layer_var, layer_def = data_layer_def(n.name, n.output_shape)
+            layer_var = name + '_layer'
+            layer_def = '%s = fluid.layers.data(name="%s", shape=shapes["%s"],'\
+                    ' dtype="float32")' % (layer_var, name, name)
+            #layer_var, layer_def = data_layer_def(n.name, n.output_shape)
             codes.append(layer_def)
             inputs[name] = layer_var
 
@@ -229,7 +239,7 @@ class TensorFlowEmitter(object):
         func_def += self.statement('import paddle.v2.fluid as fluid')
         for l in codes:
             func_def += self.statement(l)
-        return '\n\n' + func_def
+        return '\n' + func_def
 
     def emit_main_def(self, name):
         if name is None:
@@ -273,6 +283,7 @@ class TensorFlowEmitter(object):
                 b += self.emit_node(node)
             blocks.append(b[:-1])
         s = s + '\n\n'.join(blocks)
+        s += self.emit_shape_def(input_nodes)
         s += self.emit_convert_def(input_nodes)
         s += self.emit_main_def(name)
         return s

fluid/image_classification/caffe2fluid/tests/lenet/README.md

-### Convert lenet model from caffe format into paddle format(fluid api)
-
-### Howto
-1, Prepare your caffepb.py
-
-2, Download a lenet caffe-model
-    lenet_iter_10000.caffemodel
-        download address: https://github.com/ethereon/caffe-tensorflow/raw/master/examples/mnist/lenet_iter_10000.caffemodel
-        md5: cbec75c1c374b6c1981c4a1eb024ae01  
-
-    lenet.prototxt
-        download address: https://raw.githubusercontent.com/BVLC/caffe/master/examples/mnist/lenet.prototxt
-        md5: 27384af843338ab90b00c8d1c81de7d5
-
-
-2, Convert this model(make sure caffepb.py is ready in ../../proto)
-    convert to npy format
-        bash ./convert.sh lenet.prototxt lenet.caffemodel lenet.py lenet.npy
-
-    save to fluid format(optional)
-        bash ./convert.sh lenet.prototxt lenet.caffemodel lenet.py lenet.npy && python ./lenet.py ./lenet.npy ./fluid.model
-
-4, Use this new model(paddle installed in this python)
-    use fluid format
-        python ./predict.py ./fluid.model
-
-    use npy format
-        python ./predict.py ./lenet.npy

fluid/image_classification/caffe2fluid/tests/lenet/convert.sh

-#!/bin/bash
-
-#function:
-#   convert a caffe model
-#   eg:
-#       bash ./convert.sh ./model.caffe/lenet.prototxt ./model.caffe/lenet.caffemodel lenet.py lenet.npy
-
-if [[ $# -ne 4 ]];then
-    echo "usage:"
-    echo "  bash $0 [PROTOTXT] [CAFFEMODEL] [PY_NAME] [WEIGHT_NAME]"
-    echo "  eg: bash $0 lenet.prototxt lenet.caffemodel lenet.py lenet.npy"
-    exit 1
-fi
-
-WORK_ROOT=$(dirname `readlink -f ${BASH_SOURCE[0]}`)
-if [[ -z $PYTHON ]];then
-    PYTHON=`which python`
-fi
-
-PROTOTXT=$1
-CAFFEMODEL=$2
-PY_NAME=$3
-WEIGHT_NAME=$4
-CONVERTER_PY="$WORK_ROOT/../../convert.py"
-
-$PYTHON $CONVERTER_PY $PROTOTXT --caffemodel $CAFFEMODEL --code-output-path=$PY_NAME --data-output-path=$WEIGHT_NAME
-ret=$?
-if [[ $ret -eq 0 ]];then
-    echo "succeed to convert caffe model[$CAFFEMODEL, $PROTOTXT] to paddle model[$PY_NAME, $WEIGHT_NAME]"
-else
-    echo "failed to convert caffe model[$CAFFEMODEL, $PROTOTXT]"
-fi
-exit $ret

fluid/image_classification/caffe2fluid/tests/lenet/lenet.py

-### generated by caffe2fluid, your net is in class "LeNet" ###
-
-import math
-import os
-import numpy as np
-
-
-def import_fluid():
-    import paddle.v2.fluid as fluid
-    return fluid
-
-
-def layer(op):
-    '''Decorator for composable network layers.'''
-
-    def layer_decorated(self, *args, **kwargs):
-        # Automatically set a name if not provided.
-        name = kwargs.setdefault('name', self.get_unique_name(op.__name__))
-        # Figure out the layer inputs.
-        if len(self.terminals) == 0:
-            raise RuntimeError('No input variables found for layer %s.' % name)
-        elif len(self.terminals) == 1:
-            layer_input = self.terminals[0]
-        else:
-            layer_input = list(self.terminals)
-        # Perform the operation and get the output.
-        layer_output = op(self, layer_input, *args, **kwargs)
-        # Add to layer LUT.
-        self.layers[name] = layer_output
-        # This output is now the input for the next layer.
-        self.feed(layer_output)
-        # Return self for chained calls.
-        return self
-
-    return layer_decorated
-
-
-class Network(object):
-    def __init__(self, inputs, trainable=True):
-        # The input nodes for this network
-        self.inputs = inputs
-        # The current list of terminal nodes
-        self.terminals = []
-        # Mapping from layer names to layers
-        self.layers = dict(inputs)
-        # If true, the resulting variables are set as trainable
-        self.trainable = trainable
-        # Switch variable for dropout
-        self.paddle_env = None
-        self.setup()
-
-    def setup(self):
-        '''Construct the network. '''
-        raise NotImplementedError('Must be implemented by the subclass.')
-
-    def load(self, data_path, exe=None, place=None, ignore_missing=False):
-        '''Load network weights.
-        data_path: The path to the numpy-serialized network weights
-        ignore_missing: If true, serialized weights for missing layers are ignored.
-        '''
-        fluid = import_fluid()
-        #load fluid mode directly
-        if os.path.isdir(data_path):
-            assert (exe is not None), \
-                'must provide a executor to load fluid model'
-            fluid.io.load_persistables_if_exist(executor=exe, dirname=data_path)
-            return True
-
-        #load model from a npy file
-        if exe is None or place is None:
-            if self.paddle_env is None:
-                place = fluid.CPUPlace()
-                exe = fluid.Executor(place)
-                self.paddle_env = {'place': place, 'exe': exe}
-                exe = exe.run(fluid.default_startup_program())
-            else:
-                place = self.paddle_env['place']
-                exe = self.paddle_env['exe']
-
-        data_dict = np.load(data_path).item()
-        for op_name in data_dict:
-            layer = self.layers[op_name]
-            for param_name, data in data_dict[op_name].iteritems():
-                try:
-                    name = '%s_%s' % (op_name, param_name)
-                    v = fluid.global_scope().find_var(name)
-                    w = v.get_tensor()
-                    w.set(data, place)
-                except ValueError:
-                    if not ignore_missing:
-                        raise
-        return True
-
-    def feed(self, *args):
-        '''Set the input(s) for the next operation by replacing the terminal nodes.
-        The arguments can be either layer names or the actual layers.
-        '''
-        assert len(args) != 0
-        self.terminals = []
-        for fed_layer in args:
-            if isinstance(fed_layer, basestring):
-                try:
-                    fed_layer = self.layers[fed_layer]
-                except KeyError:
-                    raise KeyError('Unknown layer name fed: %s' % fed_layer)
-            self.terminals.append(fed_layer)
-        return self
-
-    def get_output(self):
-        '''Returns the current network output.'''
-        return self.terminals[-1]
-
-    def get_unique_name(self, prefix):
-        '''Returns an index-suffixed unique name for the given prefix.
-        This is used for auto-generating layer names based on the type-prefix.
-        '''
-        ident = sum(t.startswith(prefix) for t, _ in self.layers.items()) + 1
-        return '%s_%d' % (prefix, ident)
-
-    @layer
-    def conv(self,
-             input,
-             k_h,
-             k_w,
-             c_o,
-             s_h,
-             s_w,
-             name,
-             relu=True,
-             padding=None,
-             group=1,
-             biased=True):
-        if padding is None:
-            padding = [0, 0]
-
-        # Get the number of channels in the input
-        c_i, h_i, w_i = input.shape[1:]
-
-        # Verify that the grouping parameter is valid
-        assert c_i % group == 0
-        assert c_o % group == 0
-
-        fluid = import_fluid()
-        prefix = name + '_'
-        output = fluid.layers.conv2d(
-            input=input,
-            filter_size=[k_h, k_w],
-            num_filters=c_o,
-            stride=[s_h, s_w],
-            padding=padding,
-            groups=group,
-            param_attr=fluid.ParamAttr(name=prefix + "weights"),
-            bias_attr=fluid.ParamAttr(name=prefix + "biases"),
-            act="relu" if relu is True else None)
-        return output
-
-    @layer
-    def relu(self, input, name):
-        fluid = import_fluid()
-        output = fluid.layers.relu(x=input)
-        return output
-
-    @layer
-    def max_pool(self, input, k_h, k_w, s_h, s_w, name, padding=None):
-        if padding is None:
-            padding = [0, 0]
-
-        # Get the number of channels in the input
-        h_i, w_i = input.shape[2:]
-        fluid = import_fluid()
-        output = fluid.layers.pool2d(
-            input=input,
-            pool_size=[k_h, k_w],
-            pool_stride=[s_h, s_w],
-            pool_padding=padding,
-            pool_type='max')
-        return output
-
-    @layer
-    def avg_pool(self, input, k_h, k_w, s_h, s_w, name, padding=None):
-        if padding is None:
-            padding = [0, 0]
-
-        # Get the number of channels in the input
-        h_i, w_i = input.shape[2:]
-        fluid = import_fluid()
-        output = fluid.layers.pool2d(
-            input=input,
-            pool_size=[k_h, k_w],
-            pool_stride=[s_h, s_w],
-            pool_padding=padding,
-            pool_type='avg')
-        return output
-
-    @layer
-    def lrn(self, input, radius, alpha, beta, name, bias=1.0):
-        raise Exception('lrn() not implemented yet')
-
-    @layer
-    def concat(self, inputs, axis, name):
-        fluid = import_fluid()
-        output = fluid.layers.concat(input=inputs, axis=axis)
-        return output
-
-    @layer
-    def add(self, inputs, name):
-        fluid = import_fluid()
-        output = inputs[0]
-        for i in inputs[1:]:
-            output = fluid.layers.elementwise_add(x=output, y=i)
-        return output
-
-    @layer
-    def fc(self, input, num_out, name, relu=True, act=None):
-        fluid = import_fluid()
-
-        if act is None:
-            act = 'relu' if relu is True else None
-
-        prefix = name + '_'
-        output = fluid.layers.fc(
-            name=name,
-            input=input,
-            size=num_out,
-            act=act,
-            param_attr=fluid.ParamAttr(name=prefix + 'weights'),
-            bias_attr=fluid.ParamAttr(name=prefix + 'biases'))
-        return output
-
-    @layer
-    def softmax(self, input, name):
-        fluid = import_fluid()
-        output = fluid.layers.softmax(x=input, name=name)
-        return output
-
-    @layer
-    def batch_normalization(self, input, name, scale_offset=True, relu=False):
-        # NOTE: Currently, only inference is supported
-        fluid = import_fluid()
-        prefix = name + '_'
-        param_attr = None if scale_offset is False else fluid.ParamAttr(
-            name=prefix + 'scale')
-        bias_attr = None if scale_offset is False else fluid.ParamAttr(
-            name=prefix + 'offset')
-        mean_name = prefix + 'mean'
-        variance_name = prefix + 'variance'
-        output = fluid.layers.batch_norm(
-            name=name,
-            input=input,
-            is_test=True,
-            param_attr=param_attr,
-            bias_attr=bias_attr,
-            moving_mean_name=mean_name,
-            moving_variance_name=variance_name,
-            epsilon=1e-5,
-            act='relu' if relu is True else None)
-
-        return output
-
-    @layer
-    def dropout(self, input, keep_prob, name):
-        raise Exception('dropout() not implemented yet')
-
-
-class LeNet(Network):
-    def setup(self):
-        self.feed('data')
-        self.conv(5, 5, 20, 1, 1, relu=False, name='conv1')
-        self.max_pool(2, 2, 2, 2, name='pool1')
-        self.conv(5, 5, 50, 1, 1, relu=False, name='conv2')
-        self.max_pool(2, 2, 2, 2, name='pool2')
-        self.fc(500, name='ip1')
-        self.fc(10, relu=False, name='ip2')
-        self.softmax(name='prob')
-
-    @classmethod
-    def convert(cls, npy_model, fluid_path):
-        import paddle.v2.fluid as fluid
-        data_layer = fluid.layers.data(
-            name="data", shape=[1, 28, 28], dtype="float32")
-        feed_data = {"data": data_layer}
-        net = cls(feed_data)
-        place = fluid.CPUPlace()
-        exe = fluid.Executor(place)
-        exe.run(fluid.default_startup_program())
-        net.load(data_path=npy_model, exe=exe, place=place)
-        fluid.io.save_persistables(executor=exe, dirname=fluid_path)
-
-
-if __name__ == "__main__":
-    #usage: python xxxnet.py xxx.npy ./model
-
-    import sys
-    npy_weight = sys.argv[1]
-    fluid_model = sys.argv[2]
-    LeNet.convert(npy_weight, fluid_model)
-    exit(0)