caffe2fluid: support new layers conversion

fluid/image_classification/caffe2fluid/README.md

@@ -1,6 +1,15 @@
 ### Caffe2Fluid
 This tool is used to convert a Caffe model to a Fluid model
 
+### Key Features
+1. Convert caffe model to fluid model with codes of defining a network(useful for re-training)
+
+2. Pycaffe is not necessary when just want convert model without do caffe-inference
+
+3. Caffe's customized layers convertion also be supported by extending this tool
+
+4. A bunch of tools in '*examples/imagenet/tools*' are provided to compare the difference
+
 ### HowTo
 1. Prepare caffepb.py in ./proto if your python has no 'pycaffe' module, two options provided here:
     - Generate pycaffe from caffe.proto
@@ -29,10 +38,10 @@ This tool is used to convert a Caffe model to a Fluid model
       ```
 
 3. Use the converted model to infer
-    - See more details in '*examples/imagenet/run.sh*'
+    - See more details in '*examples/imagenet/tools/run.sh*'
 
 4. Compare the inference results with caffe
-    - See more details in '*examples/imagenet/diff.sh*'
+    - See more details in '*examples/imagenet/tools/diff.sh*'
 
 ### How to convert custom layer
 1. Implement your custom layer in a file under '*kaffe/custom_layers*', eg: mylayer.py

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

@@ -1,4 +1,4 @@
-A demo to show converting caffe models on 'imagenet' using caffe2fluid
+A demo to show converting caffe models trained on 'imagenet' using caffe2fluid
 
 ---
 
@@ -10,28 +10,32 @@ A demo to show converting caffe models on 'imagenet' using caffe2fluid
 
 3. Convert the Caffe model to Fluid model
     - generate fluid code and weight file
-    <pre><code>python convert.py alexnet.prototxt \
+        ```python convert.py alexnet.prototxt \
         --caffemodel alexnet.caffemodel \
         --data-output-path alexnet.npy \
         --code-output-path alexnet.py
-    </code></pre>
+        ```
 
     - save weights as fluid model file
-    <pre><code>python alexnet.py alexnet.npy ./fluid_model
-    </code></pre>
+        ```
+        python alexnet.py alexnet.npy ./fluid
+        ```
 
 4. Do inference
-   <pre><code>python infer.py infer ./fluid_mode data/65.jpeg
-</code></pre>
+    ```
+    python infer.py infer ./fluid data/65.jpeg
+    ```
 
 5. convert model and do inference together
-   <pre><code>bash ./run.sh alexnet ./models.caffe/alexnet ./models/alexnet
-</code></pre>
-    The Caffe model is stored in './models.caffe/alexnet/alexnet.prototxt|caffemodel'
-    and the Fluid model will be save in './models/alexnet/alexnet.py|npy'
+   ```
+    bash ./tools/run.sh alexnet ./models.caffe/alexnet ./models/alexnet
+    ```
+    * Assume the Caffe model is stored in '*./models.caffe/alexnet/alexnet.prototxt|caffemodel*'
+    * converted model will be stored as '*./models/alexnet/alexnet.py|npy*'
 
 6. test the difference with caffe's results(need pycaffe installed)
-   <pre><code>bash ./diff.sh resnet
-</code></pre>
-Make sure your caffemodel stored in './models.caffe/resnet'.
-The results will be stored in './results/resnet.paddle|caffe'
+   ```
+    bash ./tools/diff.sh resnet
+    ```
+    * Make sure your caffemodel stored in '*./models.caffe/resnet*'
+    * The results will be stored in '*./results/resnet.paddle|caffe*'

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

@@ -17,8 +17,21 @@ def walk_dir(rootdir):
 def calc_diff(f1, f2):
     import numpy as np
 
-    d1 = np.load(f1).flatten()
-    d2 = np.load(f2).flatten()
+    d1 = np.load(f1)
+    d2 = np.load(f2)
+
+    print d1.shape
+    print d2.shape
+    #print d1[0, 0, 0:10, 0:10]
+    #print d2[0, 0, 0:10, 0:10]
+    #d1 = d1[:, :, 1:-2, 1:-2]
+    #d2 = d2[:, :, 1:-2, 1:-2]
+
+    d1 = d1.flatten()
+    d2 = d2.flatten()
+
+    #print d1[:10]
+    #print d2[:10]
 
     d1_num = reduce(lambda x, y: x * y, d1.shape)
     d2_num = reduce(lambda x, y: x * y, d2.shape)
@@ -36,15 +49,16 @@ def calc_diff(f1, f2):
         return -1.0, -1.0
 
 
-def compare(path1, path2):
+def compare(path1, path2, no_exception):
     def diff(f1, f2):
         max_df, sq_df = calc_diff(f1, f2)
-        print('compare %s <=> %s with result[max_df:%.4e, sq_df:%.4e]' %
-              (f1, f2, max_df, sq_df))
-        assert (max_df < 1e-5), \
-                'max_df is too large with value[%.6e]' % (max_df)
-        assert (sq_df < 1e-10), \
-                'sq_df is too large with value[%.6e]' % (sq_df)
+        print('[max_df:%.4e, sq_df:%.4e] when compare %s <=> %s' %
+              (max_df, sq_df, os.path.basename(f1), os.path.basename(f2)))
+        if no_exception is False:
+            assert (max_df < 1e-5), \
+                    'max_df is too large with value[%.6e]' % (max_df)
+            assert (sq_df < 1e-10), \
+                    'sq_df is too large with value[%.6e]' % (sq_df)
 
     if os.path.exists(path1) is False:
         print('not found %s' % (path1))
@@ -73,13 +87,17 @@ def diff(f1, f2):
     if len(sys.argv) == 1:
         path1 = 'lenet.tf/results'
         path2 = 'lenet.paddle/results'
-    elif len(sys.argv) == 3:
+    elif len(sys.argv) >= 3:
         path1 = sys.argv[1]
         path2 = sys.argv[2]
+        if len(sys.argv) == 4:
+            no_exception = True
+        else:
+            no_exception = False
     else:
         print('usage:')
         print(' %s [path1] [path2]' % (sys.argv[0]))
         exit(1)
 
-    print('compare inner result in %s %s' % (path1, path2))
-    exit(compare(path1, path2))
+    #print('compare inner result in %s %s' % (path1, path2))
+    exit(compare(path1, path2, no_exception))

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

@@ -213,7 +213,6 @@ def caffe_infer(prototxt, caffemodel, datafile):
     results = []
     names = []
     for k, v in net.blobs.items():
-        k = k.rstrip('_output')
         k = k.replace('/', '_')
         names.append(k)
         results.append(v.data.copy())
@@ -259,7 +258,7 @@ def caffe_infer(prototxt, caffemodel, datafile):
             print('usage:')
             print('\tpython %s dump [net_file] [weight_file] [datafile] [net_name]' \
                     % (sys.argv[0]))
-            print('\teg:python dump %s %s %s %s %s' % (sys.argv[0],\
+            print('\teg:python %s dump %s %s %s %s' % (sys.argv[0],\
                 net_file, weight_file, datafile, net_name))
             sys.exit(1)
 

fluid/image_classification/caffe2fluid/examples/imagenet/tools/cmp.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+
+#
+#function:
+#   a tool used to compare the results produced by paddle and caffe
+#
+
+if [[ $# -lt 2 ]];then
+    echo "usage:"
+    echo "  bash $0 [model_name] [param_name] [caffe_name]"
+    exit 1
+fi
+
+model_name=$1
+param_name=$2
+paddle_file="./results/${model_name}.paddle/${param_name}.npy"
+if [[ $# -eq 3 ]];then
+    caffe_file="./results/${model_name}.caffe/${3}.npy"
+else
+    caffe_file="./results/${model_name}.caffe/${2}.npy"
+fi
+python ./compare.py $paddle_file $caffe_file

fluid/image_classification/caffe2fluid/examples/imagenet/tools/cmp_layers.sh

@@ -0,0 +1,39 @@
+#!/bin/bash
+
+#function:
+#   a tool used to compare all layers' results
+#
+
+if [[ $# -ne 1 ]];then
+    echo "usage:"
+    echo "  bash $0 [model_name]"
+    echo "  eg:bash $0 alexnet"
+    exit 1
+fi
+
+model_name=$1
+prototxt="models.caffe/$model_name/${model_name}.prototxt"
+layers=$(cat $prototxt | perl -ne 'if(/^\s+name\s*:\s*\"([^\"]+)/){print $1."\n";}')
+
+for i in $layers;do
+    cf_npy="results/${model_name}.caffe/${i}.npy"
+    pd_npy="results/${model_name}.paddle/${i}.npy"
+
+    if [[ ! -e $cf_npy ]];then
+        echo "caffe's result not exist[$cf_npy]"
+        continue
+    fi
+
+    if [[ ! -e $pd_npy ]];then
+        echo "paddle's result not exist[$pd_npy]"
+        continue
+    fi
+
+    python compare.py $cf_npy $pd_npy no_exception
+    if [[ $? -eq 0 ]];then
+        echo "succeed to compare layer[$i]"
+    else
+        echo "failed to compare layer[$i]"
+    fi
+
+done

fluid/image_classification/caffe2fluid/examples/imagenet/diff.sh → fluid/image_classification/caffe2fluid/examples/imagenet/tools/diff.sh

@@ -36,7 +36,7 @@ model_caffemodel="models.caffe/${model_name}/${model_name}.caffemodel"
 paddle_results="$results_root/${model_name}.paddle"
 rm -rf $paddle_results
 rm -rf "results.paddle"
-bash run.sh $model_name ./models.caffe/$model_name ./models/$model_name
+bash ./tools/run.sh $model_name ./models.caffe/$model_name ./models/$model_name
 if [[ $? -ne 0 ]] || [[ ! -e "results.paddle" ]];then
     echo "not found paddle's results, maybe failed to convert"
     exit 1

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

@@ -6,7 +6,7 @@
 #       2, do inference(only in fluid) using this model
 #
 #usage:
-#   bash run.sh resnet50 ./models.caffe/resnet50 ./models/resnet50
+#   cd caffe2fluid/examples/imagenet && bash run.sh resnet50 ./models.caffe/resnet50 ./models/resnet50
 #
 
 #set -x

fluid/image_classification/caffe2fluid/kaffe/custom_layers/__init__.py

@@ -7,13 +7,14 @@
 import axpy
 import flatten
 import argmax
+import reshape
 
 #custom layer import ends
 
 custom_layers = get_registered_layers()
 
 
-def set_args(f, params):
+def set_args(f, params, node=None):
     """ set args for function 'f' using the parameters in node.layer.parameters
 
     Args:
@@ -24,19 +25,15 @@ def set_args(f, params):
         arg_names (list): a list of argument names
         kwargs (dict): a dict contains needed arguments
     """
+    from ..protobuf_to_dict import protobuf_to_dict
+
     argc = f.__code__.co_argcount
     arg_list = f.__code__.co_varnames[0:argc]
 
     kwargs = {}
     for arg_name in arg_list:
-        try:
-            v = getattr(params, arg_name, None)
-        except Exception as e:
-            #maybe failed to extract caffe's parameters
-            v = None
-
-        if v is not None:
-            kwargs[arg_name] = v
+        if arg_name in params:
+            kwargs[arg_name] = params[arg_name]
 
     return arg_list, kwargs
 
@@ -54,7 +51,7 @@ def compute_output_shape(kind, node):
 
     parents = node.parents
     inputs = [list(p.output_shape) for p in parents]
-    arg_names, kwargs = set_args(shape_func, node.layer.parameters)
+    arg_names, kwargs = set_args(shape_func, node.params)
 
     if len(inputs) == 1:
         inputs = inputs[0]
@@ -80,7 +77,7 @@ def make_node(template, kind, node):
     layer_func = custom_layers[kind]['layer']
 
     #construct arguments needed by custom layer function from node's parameters
-    arg_names, kwargs = set_args(layer_func, node.layer.parameters)
+    arg_names, kwargs = set_args(layer_func, node.params, node)
 
     return template('custom_layer', kind, **kwargs)