timprepscius commented on issue #20659:
URL:
https://github.com/apache/incubator-mxnet/issues/20659#issuecomment-947035378
I'm only doing one test now - have to work on other things. Let me know if
this works. It doesn't for me whether on 1.x or master.
I just ran again, using your mx.contrib.onnx.import_to_gluon('onnx_test',
ctx=mx.cpu()), which I didn't realize existed.
Here it is, use complex=False
instantiate like so:
```
if test == 4:
m = DCUnet20(complex=False)
```
input is like so:
```
input = torch.ones((1, 1, 1537, 215))
output = m(input).detach().numpy()
```
if you do end up debugging, I find this useful:
```
np.set_printoptions(threshold=sys.maxsize, precision=None,
floatmode='unique')
```
```
import torch
import torch.nn as nn
def log_(*args):
# print(*args)
pass
class CConv2d(nn.Module):
"""
Class of complex valued convolutional layer
"""
def __init__(self, in_channels, out_channels, kernel_size, stride=1,
padding=0):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.padding = padding
self.stride = stride
self.real_conv = nn.Conv2d(in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=self.kernel_size,
padding=self.padding,
stride=self.stride)
self.im_conv = nn.Conv2d(in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=self.kernel_size,
padding=self.padding,
stride=self.stride)
# Glorot initialization.
nn.init.xavier_uniform_(self.real_conv.weight)
nn.init.xavier_uniform_(self.im_conv.weight)
def forward(self, x):
x_real = x[..., 0]
x_im = x[..., 1]
c_real = self.real_conv(x_real) - self.im_conv(x_im)
c_im = self.im_conv(x_real) + self.real_conv(x_im)
output = torch.stack([c_real, c_im], dim=-1)
return output
# + colab={} colab_type="code" id="GgtxJbSQ5i96"
class CConvTranspose2d(nn.Module):
"""
Class of complex valued dilation convolutional layer
"""
def __init__(self, in_channels, out_channels, kernel_size, stride,
output_padding=0, padding=0):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.output_padding = output_padding
self.padding = padding
self.stride = stride
self.real_convt = nn.ConvTranspose2d(in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=self.kernel_size,
output_padding=self.output_padding,
padding=self.padding,
stride=self.stride)
self.im_convt = nn.ConvTranspose2d(in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=self.kernel_size,
output_padding=self.output_padding,
padding=self.padding,
stride=self.stride)
# Glorot initialization.
nn.init.xavier_uniform_(self.real_convt.weight)
nn.init.xavier_uniform_(self.im_convt.weight)
def forward(self, x):
x_real = x[..., 0]
x_im = x[..., 1]
ct_real = self.real_convt(x_real) - self.im_convt(x_im)
ct_im = self.im_convt(x_real) + self.real_convt(x_im)
output = torch.stack([ct_real, ct_im], dim=-1)
return output
# + colab={} colab_type="code" id="OJSmVrxp5i9-"
class CBatchNorm2d(nn.Module):
"""
Class of complex valued batch normalization layer
"""
def __init__(self, num_features, eps=1e-05, momentum=0.1, affine=True,
track_running_stats=True):
super().__init__()
self.num_features = num_features
self.eps = eps
self.momentum = momentum
self.affine = affine
self.track_running_stats = track_running_stats
self.real_b = nn.BatchNorm2d(num_features=self.num_features,
eps=self.eps, momentum=self.momentum,
affine=self.affine,
track_running_stats=self.track_running_stats)
self.im_b = nn.BatchNorm2d(num_features=self.num_features,
eps=self.eps, momentum=self.momentum,
affine=self.affine,
track_running_stats=self.track_running_stats)
def forward(self, x):
x_real = x[..., 0]
x_im = x[..., 1]
n_real = self.real_b(x_real)
n_im = self.im_b(x_im)
output = torch.stack([n_real, n_im], dim=-1)
return output
# + colab={} colab_type="code" id="N7W37XMO5i-B"
class Encoder(nn.Module):
"""
Class of upsample block
"""
def __init__(self, filter_size=(7,5), stride_size=(2,2), in_channels=1,
out_channels=45, padding=(0,0), complex=True):
super().__init__()
self.filter_size = filter_size
self.stride_size = stride_size
self.in_channels = in_channels
self.out_channels = out_channels
self.padding = padding
if complex:
self.cconv = CConv2d(in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=self.filter_size,
stride=self.stride_size, padding=self.padding)
self.cbn = CBatchNorm2d(num_features=self.out_channels)
else:
self.cconv = nn.Conv2d(in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=self.filter_size,
stride=self.stride_size, padding=self.padding)
self.cbn = nn.BatchNorm2d(num_features=self.out_channels)
self.leaky_relu = nn.LeakyReLU()
def forward(self, x):
conved = self.cconv(x)
normed = self.cbn(conved)
acted = self.leaky_relu(normed)
return acted
# + colab={} colab_type="code" id="fuugYDZs5i-G"
class Decoder(nn.Module):
"""
Class of downsample block
"""
def __init__(self, filter_size=(7,5), stride_size=(2,2), in_channels=1,
out_channels=45,
output_padding=(0,0), padding=(0,0), last_layer=False,
complex=True):
super().__init__()
self.filter_size = filter_size
self.stride_size = stride_size
self.in_channels = in_channels
self.out_channels = out_channels
self.output_padding = output_padding
self.padding = padding
self.complex = complex
self.last_layer = last_layer
if complex:
self.cconvt = CConvTranspose2d(in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=self.filter_size,
stride=self.stride_size, output_padding=self.output_padding,
padding=self.padding)
self.cbn = CBatchNorm2d(num_features=self.out_channels)
else:
self.cconvt = nn.ConvTranspose2d(in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=self.filter_size,
stride=self.stride_size, output_padding=self.output_padding,
padding=self.padding)
self.cbn = nn.BatchNorm2d(num_features=self.out_channels)
self.leaky_relu = nn.LeakyReLU()
def forward(self, x):
conved = self.cconvt(x)
if self.complex:
if not self.last_layer:
normed = self.cbn(conved)
output = self.leaky_relu(normed)
else:
m_phase = conved / (torch.abs(conved) + 1e-8)
m_mag = torch.tanh(torch.abs(conved))
output = m_phase * m_mag
else:
normed = self.cbn(conved)
output = self.leaky_relu(normed)
return output
class DCUnet20(nn.Module):
"""
Deep Complex U-Net class of the model.
"""
def __init__(self, complex=False):
super().__init__()
self.complex = complex
# for istft
self.set_size(model_complexity=int(45//1.414), input_channels=1,
model_depth=20)
self.encoders = []
self.model_length = 20 // 2
for i in range(self.model_length):
module = Encoder(in_channels=self.enc_channels[i],
out_channels=self.enc_channels[i + 1],
filter_size=self.enc_kernel_sizes[i],
stride_size=self.enc_strides[i], padding=self.enc_paddings[i], complex=complex)
self.add_module("encoder{}".format(i), module)
self.encoders.append(module)
self.decoders = []
for i in range(self.model_length):
if i != self.model_length - 1:
module = Decoder(in_channels=self.dec_channels[i] +
self.enc_channels[self.model_length - i], out_channels=self.dec_channels[i +
1],
filter_size=self.dec_kernel_sizes[i],
stride_size=self.dec_strides[i], padding=self.dec_paddings[i],
output_padding=self.dec_output_padding[i],
complex=complex)
else:
module = Decoder(in_channels=self.dec_channels[i] +
self.enc_channels[self.model_length - i], out_channels=self.dec_channels[i +
1],
filter_size=self.dec_kernel_sizes[i],
stride_size=self.dec_strides[i], padding=self.dec_paddings[i],
output_padding=self.dec_output_padding[i],
last_layer=True, complex=complex)
self.add_module("decoder{}".format(i), module)
self.decoders.append(module)
def forward(self, x):
log_('x : ', x.shape)
orig_x = x
xs = []
for i, encoder in enumerate(self.encoders):
xs.append(x)
x = encoder(x)
log_('Encoder : ', x.shape)
p = x
for i, decoder in enumerate(self.decoders):
p = decoder(p)
if i == self.model_length - 1:
break
log_('Decoder : ', p.shape)
p = torch.cat([p, xs[self.model_length - 1 - i]], dim=1)
# u9 - the mask
mask = p
log_('mask : ', mask.shape)
output = mask * orig_x
output = torch.squeeze(output, 1)
return output
def set_size(self, model_complexity, model_depth=20, input_channels=1):
if model_depth == 20:
self.enc_channels = [input_channels,
model_complexity,
model_complexity,
model_complexity * 2,
model_complexity * 2,
model_complexity * 2,
model_complexity * 2,
model_complexity * 2,
model_complexity * 2,
model_complexity * 2,
128]
self.enc_kernel_sizes = [(7, 1),
(1, 7),
(6, 4),
(7, 5),
(5, 3),
(5, 3),
(5, 3),
(5, 3),
(5, 3),
(5, 3)]
self.enc_strides = [(1, 1),
(1, 1),
(2, 2),
(2, 1),
(2, 2),
(2, 1),
(2, 2),
(2, 1),
(2, 2),
(2, 1)]
self.enc_paddings = [(3, 0),
(0, 3),
(0, 0),
(0, 0),
(0, 0),
(0, 0),
(0, 0),
(0, 0),
(0, 0),
(0, 0)]
self.dec_channels = [0,
model_complexity * 2,
model_complexity * 2,
model_complexity * 2,
model_complexity * 2,
model_complexity * 2,
model_complexity * 2,
model_complexity * 2,
model_complexity,
model_complexity,
1]
self.dec_kernel_sizes = [(6, 3),
(6, 3),
(6, 3),
(6, 4),
(6, 3),
(6, 4),
(8, 5),
(7, 5),
(1, 7),
(7, 1)]
self.dec_strides = [(2, 1), #
(2, 2), #
(2, 1), #
(2, 2), #
(2, 1), #
(2, 2), #
(2, 1), #
(2, 2), #
(1, 1),
(1, 1)]
self.dec_paddings = [(0, 0),
(0, 0),
(0, 0),
(0, 0),
(0, 0),
(0, 0),
(0, 0),
(0, 0),
(0, 3),
(3, 0)]
self.dec_output_padding = [(0,0),
(0,0),
(0,0),
(0,0),
(0,0),
(0,0),
(0,0),
(0,0),
(0,0),
(0,0)]
else:
raise ValueError("Unknown model depth : {}".format(model_depth))
```
--
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: issues-unsubscr...@mxnet.apache.org
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@mxnet.apache.org
For additional commands, e-mail: issues-h...@mxnet.apache.org
