(self, obs_shape, feature_dim,
channels=[16, 32, 32],
num_layers=2,
num_filters=32,
output_logits=False,
image_channel=3)
| 35 | class PixelEncoder(nn.Module): |
| 36 | """Convolutional encoder of pixels observations.""" |
| 37 | def __init__(self, obs_shape, feature_dim, |
| 38 | channels=[16, 32, 32], |
| 39 | num_layers=2, |
| 40 | num_filters=32, |
| 41 | output_logits=False, |
| 42 | image_channel=3): |
| 43 | super().__init__() |
| 44 | |
| 45 | assert len(obs_shape) == 3 |
| 46 | self.obs_shape = obs_shape |
| 47 | self.feature_dim = feature_dim |
| 48 | self.num_layers = num_layers |
| 49 | self.image_channel = image_channel |
| 50 | self.convs = nn.ModuleList( |
| 51 | [nn.Conv2d(obs_shape[0], num_filters, 3, stride=2)] |
| 52 | ) |
| 53 | for i in range(num_layers - 1): |
| 54 | self.convs.append(nn.Conv2d(num_filters, num_filters, 3, stride=1)) |
| 55 | self.outputs = dict() |
| 56 | |
| 57 | x = torch.randn([32]+list(obs_shape)) |
| 58 | self.out_dim = self.forward_conv(x,flatten=False).shape[-1] |
| 59 | print('conv output dim: ' + str(self.out_dim)) |
| 60 | |
| 61 | self.fc = nn.Linear(num_filters * self.out_dim * self.out_dim, self.feature_dim) |
| 62 | self.ln = nn.LayerNorm(self.feature_dim) |
| 63 | |
| 64 | self.output_logits = output_logits |
| 65 | |
| 66 | def reparameterize(self, mu, logstd): |
| 67 | std = torch.exp(logstd) |
no test coverage detected