Support conversion of recognize_digits models #19

kuke · 2018-04-12T03:29:02Z

Resolve #18

I add this pull request immediately because I find that the fit_a_line model is too simple. And the parameters we passed parameters to the node maker are not enough to handle complex operators such as conv2d, batch norm_op covered by the recognize_digits_conv model. So I redesigned the interface and hope that it would be convenient for everyone to the implementation of complex models.

…ts_conv model

varunarora · 2018-04-12T22:24:46Z

Aha! Yes, tf2onxx had to the same. Can't just do a passthrough of things. But I think this was inevitable, so I am glad we have gotten to it this way.

varunarora

Really enjoying this..

varunarora · 2018-04-12T22:30:53Z

convert.py

+
+                    # Append some customized arguments of the node maker here
+                    if op.type == 'conv2d':
+                        kernel_shape = fluid.executor.fetch_var(


Since this is all static, we could make fetch_var available inside the ops module, so we don't have to such a conditional here

Agree. We should make convert.py more clean. So I change to pass operator and scope as arguments.

varunarora · 2018-04-12T22:34:48Z

fluid_onnx/ops.py

@@ -281,6 +301,25 @@ def pad_op():
    pass


+def pool2d_op(inputs, attrs, outputs):
+    if attrs['global_pooling'] is False:


kuke · 2018-04-13T05:30:53Z

Verified the conversion's correctness

python validate.py --fluid_model ./extras/recognize_digits_conv.inference.model/ --onnx_model ./recognize_digits_conv.onnx

-----------  Configuration Arguments -----------
a: 0.0
b: 1.0
batch_size: 10
expected_decimal: 5
fluid_model: ./extras/recognize_digits_conv.inference.model/
onnx_model: ./recognize_digits_conv.onnx
------------------------------------------------
Inference results for fluid model:
[array([[0.03526556, 0.00531864, 0.4901957 , 0.03146124, 0.21433197,
        0.01664129, 0.09860657, 0.0232774 , 0.0488808 , 0.03602083],
       [0.08501429, 0.0067768 , 0.44258946, 0.0571987 , 0.14585115,
        0.03440049, 0.11217487, 0.02188424, 0.04691784, 0.04719224],
       [0.08955405, 0.00735261, 0.28424332, 0.02858522, 0.18232903,
        0.02950039, 0.13921854, 0.03707843, 0.08816215, 0.11397629],
       [0.0623313 , 0.00577816, 0.51687336, 0.03530243, 0.15085357,
        0.02701959, 0.07948152, 0.01347689, 0.04822892, 0.06065419],
       [0.09505039, 0.00879493, 0.27774006, 0.01832659, 0.13748896,
        0.01448278, 0.25627935, 0.03942731, 0.0796967 , 0.07271288],
       [0.11718861, 0.00460373, 0.526473  , 0.02148582, 0.10427924,
        0.01576571, 0.08279876, 0.02664964, 0.05605122, 0.04470441],
       [0.0657331 , 0.00556582, 0.27806717, 0.04050628, 0.25586566,
        0.02690179, 0.13714735, 0.08382507, 0.0529239 , 0.05346391],
       [0.10572249, 0.00510338, 0.25388333, 0.05231774, 0.2349722 ,
        0.02573268, 0.15960246, 0.04436921, 0.06631914, 0.05197732],
       [0.07063236, 0.00825579, 0.39233148, 0.0192927 , 0.2183245 ,
        0.01888983, 0.14734125, 0.03661045, 0.04431242, 0.04400918],
       [0.05180368, 0.00526981, 0.5286077 , 0.02330228, 0.17934334,
        0.01972133, 0.06492393, 0.02322024, 0.04617354, 0.05763409]],
      dtype=float32)]


Inference results for ONNX model:
Outputs(_0=array([[0.03526554, 0.00531865, 0.49019575, 0.03146119, 0.21433191,
        0.0166413 , 0.09860668, 0.02327737, 0.04888073, 0.03602083],
       [0.08501417, 0.00677679, 0.44258943, 0.05719874, 0.14585117,
        0.03440046, 0.11217497, 0.02188419, 0.04691781, 0.04719226],
       [0.08955389, 0.00735262, 0.28424323, 0.02858525, 0.18232921,
        0.0295004 , 0.13921838, 0.03707845, 0.08816212, 0.11397647],
       [0.06233123, 0.00577816, 0.51687336, 0.03530247, 0.15085363,
        0.02701959, 0.07948162, 0.01347688, 0.0482289 , 0.06065418],
       [0.09505048, 0.00879495, 0.27774018, 0.0183266 , 0.1374888 ,
        0.01448278, 0.25627935, 0.03942732, 0.07969675, 0.07271287],
       [0.11718868, 0.00460375, 0.52647275, 0.02148584, 0.10427925,
        0.0157657 , 0.08279874, 0.02664967, 0.05605123, 0.04470445],
       [0.06573316, 0.00556582, 0.27806693, 0.04050625, 0.25586566,
        0.02690177, 0.13714729, 0.08382516, 0.05292396, 0.05346398],
       [0.10572256, 0.00510339, 0.25388315, 0.05231774, 0.23497239,
        0.02573269, 0.15960251, 0.04436918, 0.06631912, 0.05197728],
       [0.07063232, 0.00825581, 0.39233148, 0.01929271, 0.21832474,
        0.01888983, 0.14734115, 0.03661038, 0.04431238, 0.0440092 ],
       [0.05180358, 0.00526981, 0.5286082 , 0.02330226, 0.17934316,
        0.01972132, 0.06492393, 0.02322023, 0.0461735 , 0.05763406]],
      dtype=float32))


The exported model achieves 5-decimal precision.

kuke

Updated and this pr is ready for review

kuke · 2018-04-13T05:34:55Z

convert.py

+
+                    # Append some customized arguments of the node maker here
+                    if op.type == 'conv2d':
+                        kernel_shape = fluid.executor.fetch_var(


Agree. We should make convert.py more clean. So I change to pass operator and scope as arguments.

kuke

Some ideas about why I make the change.

kuke · 2018-04-13T05:38:46Z

fluid/utils.py

+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+


I add a new directory fluid here to put some modules related to Fluid but independent of ONNX.

kuke · 2018-04-13T05:41:11Z

fluid_onnx/ops.py

@@ -224,8 +252,63 @@ def lppool_op():
    pass


-def matmul_op(inputs, attrs, outputs):
-    return make_node('MatMul', inputs=inputs, outputs=outputs)
+def mul_op(operator, scope):


As we can see, sometimes we can't implement the operator in Fluid by only operator in ONNX. So it would be better to make the function name (mul_op here) consistent with the op's name in Fluid ('mul').

kuke · 2018-04-13T05:43:30Z

fluid_onnx/ops.py

+    inputs, attrs, outputs = get_op_io_info(operator)
+
+    # Flatten input(X) and input(Y) into 2-D matries
+    x_flat_out = [inputs['X'][0] + '@flatten_0']


Please pay attention to the intermediate tensors between nodes. They should be named uniquely in the ONNX graph.

kuke · 2018-04-13T08:34:00Z

README.md


 - [fit_a_line](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_fit_a_line.py)
- [machine_translation](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_machine_translation.py)
+- [recognize_digits](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_recognize_digits.py)


Another model recognize_digits_mlp also has been verified.

sidgoyal78

Thanks, I think this looks pretty good to me.

kuke · 2018-04-16T01:36:29Z

@sidgoyal78 Thanks!

Yibing Liu added 4 commits April 11, 2018 20:23

Pass op's parameters into node_maker

48876b4

Add conv2d pool2d batch_norm relu & softmax to support recognize_digi…

446dca8

…ts_conv model

Fix conv_op batchnorm_op and add_op in runtime

82d1f81

Merge branch 'develop' of upstream into convert_conv

d6a60e8

kuke changed the title ~~Support conversion of recognize_digits_conv model~~ [WIP] Support conversion of recognize_digits_conv model Apr 12, 2018

kuke mentioned this pull request Apr 12, 2018

Revise the design again for final review #6

Closed

varunarora reviewed Apr 12, 2018

View reviewed changes

Yibing Liu added 3 commits April 12, 2018 21:31

Verified the recognize_digits_conv model conversion after fixing mul_op

0035259

Pass operator & scope to node maker to hide more details

55f140c

Fix padding in conv2d_op

233df53

kuke requested review from pkuyym and sidgoyal78 April 13, 2018 05:32

kuke commented Apr 13, 2018

View reviewed changes

kuke changed the title ~~[WIP] Support conversion of recognize_digits_conv model~~ Support conversion of recognize_digits_conv model Apr 13, 2018

kuke commented Apr 13, 2018

View reviewed changes

Yibing Liu added 2 commits April 13, 2018 00:29

Add recognize_digits into supported models in README

b577b05

Add tanh_op to support recognize_digits_mlp

467775a

kuke commented Apr 13, 2018

View reviewed changes

kuke changed the title ~~Support conversion of recognize_digits_conv model~~ Support conversion of recognize_digits model Apr 13, 2018

kuke changed the title ~~Support conversion of recognize_digits model~~ Support conversion of recognize_digits models Apr 13, 2018

Hide parameter node's making

557e61b

kuke mentioned this pull request Apr 14, 2018

Some important feedforward models to be supported at first stage #24

Closed

7 tasks

sidgoyal78 reviewed Apr 15, 2018

View reviewed changes

sidgoyal78 approved these changes Apr 16, 2018

View reviewed changes

kuke merged commit 57c845a into PaddlePaddle:develop Apr 16, 2018

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Support conversion of recognize_digits models #19

Support conversion of recognize_digits models #19

kuke commented Apr 12, 2018 •

edited

Loading

varunarora commented Apr 12, 2018

varunarora left a comment

varunarora Apr 12, 2018

kuke Apr 13, 2018

varunarora Apr 12, 2018

kuke commented Apr 13, 2018 •

edited

Loading

kuke left a comment

kuke Apr 13, 2018

kuke left a comment

kuke Apr 13, 2018

kuke Apr 13, 2018

kuke Apr 13, 2018 •

edited

Loading

kuke Apr 13, 2018

sidgoyal78 left a comment

kuke commented Apr 16, 2018

Support conversion of recognize_digits models #19

Support conversion of recognize_digits models #19

Conversation

kuke commented Apr 12, 2018 • edited Loading

varunarora commented Apr 12, 2018

varunarora left a comment

Choose a reason for hiding this comment

varunarora Apr 12, 2018

Choose a reason for hiding this comment

kuke Apr 13, 2018

Choose a reason for hiding this comment

varunarora Apr 12, 2018

Choose a reason for hiding this comment

kuke commented Apr 13, 2018 • edited Loading

kuke left a comment

Choose a reason for hiding this comment

kuke Apr 13, 2018

Choose a reason for hiding this comment

kuke left a comment

Choose a reason for hiding this comment

kuke Apr 13, 2018

Choose a reason for hiding this comment

kuke Apr 13, 2018

Choose a reason for hiding this comment

kuke Apr 13, 2018 • edited Loading

Choose a reason for hiding this comment

kuke Apr 13, 2018

Choose a reason for hiding this comment

sidgoyal78 left a comment

Choose a reason for hiding this comment

kuke commented Apr 16, 2018

kuke commented Apr 12, 2018 •

edited

Loading

kuke commented Apr 13, 2018 •

edited

Loading

kuke Apr 13, 2018 •

edited

Loading