ConvTranspose bias tensor dimensions are not aligned with spec

[talc23]

Describe the bug

according to spec B.shape==[C] where C is the input's number of channels.
however model building is failing unless B.shape==[M] where M is the output's number of channels.

I think that maybe B should be accumulated and then broadcasted across Conv2DBackpropInput output. right now B is added as is to Conv2DBackpropInput output

To Reproduce

import onnx.helper as helper
from onnx import TensorProto
import numpy as np
from onnx_tf.backend import prepare

N=1
Cx=2
Hx=3
Wx=3
xDims = [N, Cx, Hx, Wx]

Cy=3
Hy=5
Wy=5
yDims = [N, Cy, Hy, Wy]

Ww=3
Hw=3
wDims = [Cx, Cy, Hw, Ww]

bDims = [Cx]

graph_proto = helper.make_graph(
[
helper.make_node("ConvTranspose", ["X", "W", "B"], ["Y"], output_shape=yDims,
pads=[1,1,1,1], strides=[2,2]),
],
"ConvTranspose1",
[
helper.make_tensor_value_info('X' , TensorProto.FLOAT, xDims),
helper.make_tensor_value_info('W' , TensorProto.FLOAT, wDims),
helper.make_tensor_value_info('B' , TensorProto.FLOAT, bDims),
],
[
helper.make_tensor_value_info('Y', TensorProto.FLOAT, yDims),
]
)
#graph_proto.initializer.extend([_sample_float_tensor()])
#graph_proto.initializer[0].name = 'W1'
model = helper.make_model(graph_proto)
f = open('convtranspose_test_model.pb', "wb")
f.write(model.SerializeToString())
f.close()

np.random.seed(0)
#create the input and output file:
#input = np.arange(1, 4, dtype=np.float32);
x = None
w = None
b = None

varType = 'float'
if varType == 'float':
x = np.random.random(np.prod(xDims)).reshape(xDims)
w = np.random.random(np.prod(wDims)).reshape(wDims)
b = np.random.random(np.prod(bDims)).reshape(bDims)
elif varType == 'int':
x = np.random.randint(-5,5,np.prod(xDims)).reshape(xDims)
w = np.random.randint(-5,5,np.prod(wDims)).reshape(wDims)
b = np.random.randint(-5,5,np.prod(bDims)).reshape(bDims)
elif varType == 'ones':
x = np.ones(np.prod(xDims)).reshape(xDims)
w = np.ones(np.prod(wDims)).reshape(wDims)
w[:,1,:,:]*=2
b = np.zeros(np.prod(bDims)).reshape(bDims)

np.savez('convtranspose_input_data', X=x, W=w, B=b)
#np.savez('convtranspose_output_data', outputs=output)

print("\ngraph proto:\n")
print(graph_proto)

tf_rep = prepare(model) # Import the ONNX model to Tensorflow
y = tf_rep.run([x,w,b])

failure:
File "c:\Users\tchercke\Documents\Repos\aipg_inference-dl\modules\framework_importer\scripts\simple_convtranspose_model_gen.py", line 84, in
tf_rep = prepare(model) # Import the ONNX model to Tensorflow
ValueError: Dimensions must be equal, but are 3 and 2 for 'Add' (op: 'Add') with input shapes: [1,5,5,3], [2].

Python, ONNX, ONNX-TF, Tensorflow version

This section can be obtained by running get_version.py from util folder.
Python version:
3.6.6 |Anaconda, Inc.| (default, Jun 28 2018, 11:27:44) [MSC v.1900 64 bit (AMD64)]
ONNX version:
1.3.0
ONNX-TF version:
1.2.0
Tensorflow version:
1.11.0

[fumihwh]

@talc23
The spec maybe incorrect.
Pytorch: https://pytorch.org/docs/stable/nn.html#torch.nn.ConvTranspose2d
Tensorflow: https://github.com/tensorflow/tensorflow/blob/f43d458a318d4d97298710654f1692f6e8364f82/tensorflow/python/keras/layers/convolutional.py#L839
Bias should have shape (M) (say feature maps or filters)

[talc23]

I'll open an issue in ONNX repo.

[fumihwh]

@talc23
Already opened.

[talc23]

I'll close mine

[fumihwh]

@talc23
onnx/onnx#1566
FYI, merged.