MCPcopy Create free account
hub / github.com/OpenDriveLab/TCP / __init__

Method __init__

roach/models/torch_layers.py:70–96  ·  view source on GitHub ↗
(self, observation_space, chans=(16, 32, 32, 64, 64), states_neurons=[256],
                 features_dim=256, nblock=2, batch_norm=False, final_relu=True)

Source from the content-addressed store, hash-verified

68
69class ImpalaCNN(nn.Module):
70 def __init__(self, observation_space, chans=(16, 32, 32, 64, 64), states_neurons=[256],
71 features_dim=256, nblock=2, batch_norm=False, final_relu=True):
72 # (16, 32, 32)
73 super().__init__()
74 self.features_dim = features_dim
75 self.final_relu = final_relu
76
77 # image encoder
78 curshape = observation_space['birdview'].shape
79 s = 1 / np.sqrt(len(chans)) # per stack scale
80 self.stacks = nn.ModuleList()
81 for outchan in chans:
82 stack = tu.CnnDownStack(curshape[0], nblock=nblock, outchan=outchan, scale=s, batch_norm=batch_norm)
83 self.stacks.append(stack)
84 curshape = stack.output_shape(curshape)
85
86 # dense after concatenate
87 n_image_latent = tu.intprod(curshape)
88 self.dense = tu.NormedLinear(n_image_latent+states_neurons[-1], features_dim, scale=1.4)
89
90 # state encoder
91 states_neurons = [observation_space['state'].shape[0]] + states_neurons
92 self.state_linear = []
93 for i in range(len(states_neurons)-1):
94 self.state_linear.append(tu.NormedLinear(states_neurons[i], states_neurons[i+1]))
95 self.state_linear.append(nn.ReLU())
96 self.state_linear = nn.Sequential(*self.state_linear)
97
98 def forward(self, birdview, state):
99 # birdview: [b, c, h, w]

Callers

nothing calls this directly

Calls 2

output_shapeMethod · 0.95
__init__Method · 0.45

Tested by

no test coverage detected