MORN question~

[JingLiJJ]

您好。我想问一下，在MORN里
图片过了CNN出来的size为什么是1311，而不是paper上写的2311呢？
还有，

MORAN_v2/models/morn.py

Lines 66 to 69 in 1eb6988

	offsets_grid = nn.functional.grid_sample(offsets_pool, grid)
	offsets_grid = offsets_grid.permute(0, 2, 3, 1).contiguous()
	offsets_x = torch.cat([grid_x, grid_y + offsets_grid], 3)
	x_rectified = nn.functional.grid_sample(x, offsets_x)

为什么只对y轴进行了矫正呢？还是我理解错了？

[Canjie-Luo]

您好，MORAN v1用的是两个维度的，即xy，但是我们发现cropped text基本没有水平位移的问题，所以MORAN v2去掉了x维度，只在竖直方向上纠正字符，能应用于大多数场景，同时也更好端到端训练，毕竟解空间变小了。

[JingLiJJ]

感谢回答！那我想问一下这个模型可以做不定长训练么？ Get Outlook for iOS<https://aka.ms/o0ukef>

…

________________________________ From: Canjie Luo <notifications@github.com> Sent: Thursday, May 2, 2019 11:51:52 AM To: Canjie-Luo/MORAN_v2 Cc: JingLiJJ; Author Subject: Re: [Canjie-Luo/MORAN_v2] MORN question~ (#69) 您好，MORAN v1用的是两个维度的，即xy，但是我们发现cropped text基本没有水平位移的问题，所以MORAN v2去掉了x维度，只在竖直方向上纠正字符，能应用于大多数场景，同时也更好端到端训练，毕竟解空间变小了。 ― You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#69 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AH3MWVDCQN5LBYUJRMNIP4DPTJQNRANCNFSM4HI4AFLQ>.

[Canjie-Luo]

您说的不定长指的是？

[JingLiJJ]

在训练的时候固定图片的高度，然后等比例缩放长度。像CRNN一样 发送自 Windows 10 版邮件<https://go.microsoft.com/fwlink/?LinkId=550986>应用

…

________________________________ 发件人: Canjie Luo <notifications@github.com> 发送时间: Monday, May 6, 2019 2:39:17 PM 收件人: Canjie-Luo/MORAN_v2 抄送: JingLiJJ; Author 主题: Re: [Canjie-Luo/MORAN_v2] MORN question~ (#69) 您说的不定长指的是？ ― You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#69 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AH3MWVGI5AYRK6J3WYNFNKDPT7HBLANCNFSM4HI4AFLQ>.

[Canjie-Luo]

这个是可以实现的，您可以按照CRNN写一个alignCollate（dataset.py中），或者在MORN中随机缩放target_W。

[JingLiJJ]

好哒。谢谢！ Get Outlook for iOS<https://aka.ms/o0ukef>

…

________________________________ From: Canjie Luo <notifications@github.com> Sent: Tuesday, May 7, 2019 10:42:10 PM To: Canjie-Luo/MORAN_v2 Cc: JingLiJJ; Author Subject: Re: [Canjie-Luo/MORAN_v2] MORN question~ (#69) 这个是可以实现的，您可以按照CRNN写一个alignCollate（dataset.py中），或者在MORN中随机缩放target_W。 ― You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#69 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AH3MWVEMWALCA7LFKG5UB3TPUGIMFANCNFSM4HI4AFLQ>.

[tanhui2975]

请问第80行的代码应该是x_rectified = nn.functional.grid_sample(x, offsets_x)
还是x_rectified = nn.functional.grid_sample(x_rectified, offsets_x) ？ @Canjie-Luo

MORAN_v2/models/morn.py

Lines 66 to 80 in 1eb6988

	offsets_grid = nn.functional.grid_sample(offsets_pool, grid)
	offsets_grid = offsets_grid.permute(0, 2, 3, 1).contiguous()
	offsets_x = torch.cat([grid_x, grid_y + offsets_grid], 3)
	x_rectified = nn.functional.grid_sample(x, offsets_x)
	for iteration in range(enhance):
	offsets = self.cnn(x_rectified)
	offsets_posi = nn.functional.relu(offsets, inplace=False)
	offsets_nega = nn.functional.relu(-offsets, inplace=False)
	offsets_pool = self.pool(offsets_posi) - self.pool(offsets_nega)
	offsets_grid += nn.functional.grid_sample(offsets_pool, grid).permute(0, 2, 3, 1).contiguous()
	offsets_x = torch.cat([grid_x, grid_y + offsets_grid], 3)
	x_rectified = nn.functional.grid_sample(x, offsets_x)

[Canjie-Luo]

您好，原代码是对的，在纠正后的图像上预测offset，但是在原图x上采样。 tanhui2975 <notifications@github.com> 于2019年7月18日周四 下午5:17写道：

…

请问第80行的代码应该是x_rectified = nn.functional.grid_sample(x, offsets_x) 还是x_rectified = nn.functional.grid_sample(x_rectified, offsets_x) ？ @Canjie-Luo <https://github.com/Canjie-Luo> https://github.com/Canjie-Luo/MORAN_v2/blob/1eb698848025900aa4638c5ca6e2605cb4681f58/models/morn.py#L66-L80 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#69?email_source=notifications&email_token=AHVELTHRGW62UUMUEAU37ZDQAAYLRA5CNFSM4HI4AFL2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD2H3T6I#issuecomment-512735737>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AHVELTATK57VDYFE4NZQPL3QAAYLRANCNFSM4HI4AFLQ> .