zotanika commented on a change in pull request #8260:
URL: https://github.com/apache/tvm/pull/8260#discussion_r672730636
##########
File path: python/tvm/relay/frontend/caffe.py
##########
@@ -511,19 +511,72 @@ def convert_deconv(self, op):
if weight:
kh, kw = params["kernel_size"]
weight_shape = [-1, conv_params.num_output, kh, kw]
- weight_value = np.asarray(weight.data, np.float32)
+ if not weight.data:
+ if conv_params.weight_filler:
+ _filler = conv_params.weight_filler.value
+ weight_value = np.full(weight.shape.dim, _filler,
np.float32)
+ else:
+ raise tvm.error.OpAttributeInvalid("At least weight_filler
must be given")
+ else:
+ weight_value = np.asarray(weight.data, np.float32)
weight_value = np.reshape(weight_value, weight_shape)
else:
- raise Exception("No weight value of layer {} in
caffemodel".format(op.name))
+ raise tvm.error.OpAttributeRequired(
+ "No weight value of layer {} in caffemodel".format(op.name)
+ )
weight_expr = self.exp_tab.new_const(weight_value, dtype="float32")
in_expr = self.exp_tab.get_expr(inputs[0])
- out = _op.nn.conv2d_transpose(data=in_expr, weight=weight_expr,
**params)
- if bias:
+ groups = params["groups"]
+ channels = params["channels"]
+
+ if bias:
bias_value = np.asarray(bias.data, np.float32)
bias_expr = self.exp_tab.new_const(bias_value, dtype="float32")
- out = _op.nn.bias_add(out, bias_expr)
+
+ if groups > channels:
+ raise tvm.error.OpAttributeInvalid(
+ "Groups cannot be larger than the number of input channels"
+ )
+
+ if groups == channels:
Review comment:
Dear @makihiro ,Thank you for your valuable comment.
`groups==1` is the only case that the original author intended to handle, no
matter how many channels are given in this case. The routine at line 544 are
actually falling back to the original implementation of back-end of Deconv op.
For the necessity of handling multiple groups in Caffe Deconv op, you may
refer to several implementations of image-enhancement networks including
upscaling layers, such as
[demosaicnet_caffe](https://github.com/mgharbi/demosaicnet_caffe).
##########
File path: python/tvm/relay/frontend/caffe.py
##########
@@ -511,19 +511,72 @@ def convert_deconv(self, op):
if weight:
kh, kw = params["kernel_size"]
weight_shape = [-1, conv_params.num_output, kh, kw]
- weight_value = np.asarray(weight.data, np.float32)
+ if not weight.data:
+ if conv_params.weight_filler:
+ _filler = conv_params.weight_filler.value
+ weight_value = np.full(weight.shape.dim, _filler,
np.float32)
+ else:
+ raise tvm.error.OpAttributeInvalid("At least weight_filler
must be given")
+ else:
+ weight_value = np.asarray(weight.data, np.float32)
weight_value = np.reshape(weight_value, weight_shape)
else:
- raise Exception("No weight value of layer {} in
caffemodel".format(op.name))
+ raise tvm.error.OpAttributeRequired(
+ "No weight value of layer {} in caffemodel".format(op.name)
+ )
weight_expr = self.exp_tab.new_const(weight_value, dtype="float32")
in_expr = self.exp_tab.get_expr(inputs[0])
- out = _op.nn.conv2d_transpose(data=in_expr, weight=weight_expr,
**params)
- if bias:
+ groups = params["groups"]
+ channels = params["channels"]
+
+ if bias:
bias_value = np.asarray(bias.data, np.float32)
bias_expr = self.exp_tab.new_const(bias_value, dtype="float32")
- out = _op.nn.bias_add(out, bias_expr)
+
+ if groups > channels:
+ raise tvm.error.OpAttributeInvalid(
+ "Groups cannot be larger than the number of input channels"
+ )
+
+ if groups == channels:
Review comment:
Dear @mshr-h, Thank you for your valuable comment.
`groups==1` is the only case that the original author intended to handle, no
matter how many channels are given in this case. The routine at line 544 are
actually falling back to the original implementation of back-end of Deconv op.
For the necessity of handling multiple groups in Caffe Deconv op, you may
refer to several implementations of image-enhancement networks including
upscaling layers, such as
[demosaicnet_caffe](https://github.com/mgharbi/demosaicnet_caffe).
##########
File path: tests/python/frontend/caffe/test_forward.py
##########
@@ -452,6 +452,33 @@ def test_forward_Deconvolution():
bias_filler=dict(type="xavier"),
),
)
+ _test_deconvolution(
+ data,
+ convolution_param=dict(
+ num_output=16,
+ bias_term=False,
+ pad=0,
+ kernel_size=2,
+ stride=2,
+ dilation=1,
+ group=16,
+ weight_filler=dict(type="xavier"),
+ bias_filler=dict(type="xavier"),
+ ),
+ )
+ data = np.random.rand(1, 100, 32, 32).astype(np.float32)
+ _test_deconvolution(
+ data,
+ convolution_param=dict(
+ num_output=100,
+ bias_term=False,
+ pad=0,
+ kernel_size=2,
+ stride=2,
+ dilation=1,
+ group=100,
Review comment:
Dear @lixiaoquan , Thank you for your suggestion. Fixed the test case.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
