02d76709c72932f75ad04b66fe2300f3c18611a2,pyprob/nn/embedding_cnn_2d_5c.py,EmbeddingCNN2D5C,__init__,#EmbeddingCNN2D5C#Any#Any#,8

Before Change

        self._input_shape = input_shape  // expecting 3d: [channels, height, width]
        input_channels = input_shape[0]
        self._output_shape = output_shape
        self._output_dim = util.prod(output_shape)
        self._conv1 = nn.Conv2d(input_channels, 64, 3)
        self._conv2 = nn.Conv2d(64, 64, 3)
        self._conv3 = nn.Conv2d(64, 128, 3)
        self._conv4 = nn.Conv2d(128, 128, 3)
        self._conv5 = nn.Conv2d(128, 128, 3)
        cnn_output_dim = self._forward_cnn(torch.zeros(self._input_shape).unsqueeze(0)).nelement()
        self._lin1 = nn.Linear(cnn_output_dim, self._output_dim)
        self._lin2 = nn.Linear(self._output_dim, self._output_dim)

    def _forward_cnn(self, x):
        x = torch.relu(self._conv1(x))

After Change

    def __init__(self, input_shape, output_shape):
        super().__init__()
        self._input_shape = util.to_size(input_shape)  // expecting 3d: [channels, height, width]
        self._output_shape = util.to_size(output_shape)
        input_channels = self._input_shape[0]
        self._output_dim = util.prod(self._output_shape)
        self._conv1 = nn.Conv2d(input_channels, 64, 3)
        self._conv2 = nn.Conv2d(64, 64, 3)
        self._conv3 = nn.Conv2d(64, 128, 3)
        self._conv4 = nn.Conv2d(128, 128, 3)
        self._conv5 = nn.Conv2d(128, 128, 3)
        cnn_output_dim = self._forward_cnn(torch.zeros(self._input_shape).unsqueeze(0)).nelement()
        self._lin1 = nn.Linear(cnn_output_dim, self._output_dim)
        self._lin2 = nn.Linear(self._output_dim, self._output_dim)

    def _forward_cnn(self, x):
        x = torch.relu(self._conv1(x))

Instances