Adding encoder

encoder.py

@@ -10,7 +10,7 @@ def tie_weights(src, trg):
 
 OUT_DIM = {2: 39, 4: 35, 6: 31}
 
-
+'''
 class PixelEncoder(nn.Module):
     """Convolutional encoder of pixels observations."""
     def __init__(self, obs_shape, feature_dim, num_layers=2, num_filters=32):
@@ -88,7 +88,92 @@ class PixelEncoder(nn.Module):
             L.log_param('train_encoder/conv%s' % (i + 1), self.convs[i], step)
         L.log_param('train_encoder/fc', self.fc, step)
         L.log_param('train_encoder/ln', self.ln, step)
+'''
 
+class PixelEncoder(nn.Module):
+    """Convolutional encoder of pixels observations."""
+    def __init__(self, obs_shape, feature_dim, num_layers=2, num_filters=32):
+        super().__init__()
+
+        assert len(obs_shape) == 3
+
+        self.feature_dim = feature_dim
+        self.num_layers = num_layers
+
+        self.convs = nn.ModuleList(
+            [nn.Conv2d(obs_shape[0], num_filters, 3, stride=2)]
+        )
+        for i in range(num_layers - 1):
+            self.convs.append(nn.Conv2d(num_filters, num_filters, 3, stride=1))
+
+        out_dim = OUT_DIM[num_layers]
+        self.fc = nn.Linear(num_filters * out_dim * out_dim, self.feature_dim * 2)
+        self.ln = nn.LayerNorm(self.feature_dim * 2)
+
+        self.outputs = dict()
+
+    def reparameterize(self, mu, logstd):
+        std = torch.exp(logstd)
+        eps = torch.randn_like(std)
+        return mu + eps * std
+
+    def forward_conv(self, obs):
+        obs = obs / 255.
+        self.outputs['obs'] = obs
+
+        conv = torch.relu(self.convs[0](obs))
+        self.outputs['conv1'] = conv
+
+        for i in range(1, self.num_layers):
+            conv = torch.relu(self.convs[i](conv))
+            self.outputs['conv%s' % (i + 1)] = conv
+
+        h = conv.view(conv.size(0), -1)
+        return h
+
+    def forward(self, obs, detach=False):
+        h = self.forward_conv(obs)
+
+        if detach:
+            h = h.detach()
+
+        h_fc = self.fc(h)
+        self.outputs['fc'] = h_fc
+
+        h_norm = self.ln(h_fc)
+        self.outputs['ln'] = h_norm
+
+        #out = torch.tanh(h_norm)
+
+        mu, logstd = torch.chunk(h_norm, 2, dim=-1)
+        logstd = torch.tanh(logstd)
+        self.outputs['mu'] = mu
+        self.outputs['logstd'] = logstd
+        self.outputs['std'] = logstd.exp()
+
+        out = self.reparameterize(mu, logstd)
+        self.outputs['tanh'] = out
+        return out
+    
+    def copy_conv_weights_from(self, source):
+        """Tie convolutional layers"""
+        # only tie conv layers
+        for i in range(self.num_layers):
+            tie_weights(src=source.convs[i], trg=self.convs[i])
+
+    def log(self, L, step, log_freq):
+        if step % log_freq != 0:
+            return
+
+        for k, v in self.outputs.items():
+            L.log_histogram('train_encoder/%s_hist' % k, v, step)
+            if len(v.shape) > 2:
+                L.log_image('train_encoder/%s_img' % k, v[0], step)
+
+        for i in range(self.num_layers):
+            L.log_param('train_encoder/conv%s' % (i + 1), self.convs[i], step)
+        L.log_param('train_encoder/fc', self.fc, step)
+        L.log_param('train_encoder/ln', self.ln, step)
 
 class IdentityEncoder(nn.Module):
     def __init__(self, obs_shape, feature_dim, num_layers, num_filters):