From 7e8e54e84c63171e748bbf09516fd517e6821ace Mon Sep 17 00:00:00 2001 From: Gustaf Rydholm Date: Sat, 20 Mar 2021 18:09:06 +0100 Subject: Inital commit for refactoring to lightning --- src/text_recognizer/networks/stn.py | 44 ------------------------------------- 1 file changed, 44 deletions(-) delete mode 100644 src/text_recognizer/networks/stn.py (limited to 'src/text_recognizer/networks/stn.py') diff --git a/src/text_recognizer/networks/stn.py b/src/text_recognizer/networks/stn.py deleted file mode 100644 index e9d216f..0000000 --- a/src/text_recognizer/networks/stn.py +++ /dev/null @@ -1,44 +0,0 @@ -"""Spatial Transformer Network.""" - -from einops.layers.torch import Rearrange -import torch -from torch import nn -from torch import Tensor -import torch.nn.functional as F - - -class SpatialTransformerNetwork(nn.Module): - """A network with differentiable attention. - - Network that learns how to perform spatial transformations on the input image in order to enhance the - geometric invariance of the model. - - # TODO: add arguments to make it more general. - - """ - - def __init__(self) -> None: - super().__init__() - # Initialize the identity transformation and its weights and biases. - linear = nn.Linear(32, 3 * 2) - linear.weight.data.zero_() - linear.bias.data.copy_(torch.tensor([1, 0, 0, 0, 1, 0], dtype=torch.float)) - - self.theta = nn.Sequential( - nn.Conv2d(in_channels=1, out_channels=8, kernel_size=7), - nn.MaxPool2d(kernel_size=2, stride=2), - nn.ReLU(inplace=True), - nn.Conv2d(in_channels=8, out_channels=10, kernel_size=5), - nn.MaxPool2d(kernel_size=2, stride=2), - nn.ReLU(inplace=True), - Rearrange("b c h w -> b (c h w)", h=3, w=3), - nn.Linear(in_features=10 * 3 * 3, out_features=32), - nn.ReLU(inplace=True), - linear, - Rearrange("b (row col) -> b row col", row=2, col=3), - ) - - def forward(self, x: Tensor) -> Tensor: - """The spatial transformation.""" - grid = F.affine_grid(self.theta(x), x.shape) - return F.grid_sample(x, grid, align_corners=False) -- cgit v1.2.3-70-g09d2