VedantDave/Curiosity
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Replacing Relu with LeakyRelu

Browse Source
This commit is contained in:
Vedant Dave 2023-04-12 18:22:02 +02:00
parent c8fdd11d8c
commit 3e9d8f7a9c
1 changed files with 7 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
DPI/models.py
View File

@ -79,7 +79,7 @@ class ObservationDecoder(nn.Module):
layers.append(nn.ConvTranspose2d(in_channels=self.in_channels[i], out_channels=self.out_channels[i], layers.append(nn.ConvTranspose2d(in_channels=self.in_channels[i], out_channels=self.out_channels[i],
kernel_size=self.kernels[i], stride=2, output_padding=self.output_padding[i])) kernel_size=self.kernels[i], stride=2, output_padding=self.output_padding[i]))
if i!=len(self.kernels)-1: if i!=len(self.kernels)-1:
layers.append(nn.ReLU()) layers.append(nn.LeakyReLU())
self.convtranspose = nn.Sequential(*layers) self.convtranspose = nn.Sequential(*layers)
@ -110,7 +110,7 @@ class Actor(nn.Module):
input_channels = state_size if i == 0 else self.hidden_size input_channels = state_size if i == 0 else self.hidden_size
output_channels = self.hidden_size if i!= self.num_layers-1 else 2*action_size output_channels = self.hidden_size if i!= self.num_layers-1 else 2*action_size
layers.append(nn.Linear(input_channels, output_channels)) layers.append(nn.Linear(input_channels, output_channels))
layers.append(nn.ReLU()) layers.append(nn.LeakyReLU())
self.action_model = nn.Sequential(*layers) self.action_model = nn.Sequential(*layers)
def get_dist(self, mean, std): def get_dist(self, mean, std):
@ -144,7 +144,7 @@ class ValueModel(nn.Module):
input_channels = state_size if i == 0 else self.hidden_size input_channels = state_size if i == 0 else self.hidden_size
output_channels = self.hidden_size if i!= self.num_layers-1 else 1 output_channels = self.hidden_size if i!= self.num_layers-1 else 1
layers.append(nn.Linear(input_channels, output_channels)) layers.append(nn.Linear(input_channels, output_channels))
layers.append(nn.ReLU()) layers.append(nn.LeakyReLU())
self.value_model = nn.Sequential(*layers) self.value_model = nn.Sequential(*layers)
def forward(self, state): def forward(self, state):
@ -158,9 +158,9 @@ class RewardModel(nn.Module):
super().__init__() super().__init__()
self.reward_model = nn.Sequential( self.reward_model = nn.Sequential(
nn.Linear(state_size, hidden_size), nn.Linear(state_size, hidden_size),
nn.ReLU(), nn.LeakyReLU(),
nn.Linear(hidden_size, hidden_size), nn.Linear(hidden_size, hidden_size),
nn.ReLU(), nn.LeakyReLU(),
nn.Linear(hidden_size, 1) nn.Linear(hidden_size, 1)
) )
@ -177,7 +177,7 @@ class TransitionModel(nn.Module):
self.hidden_size = hidden_size self.hidden_size = hidden_size
self.action_size = action_size self.action_size = action_size
self.history_size = history_size self.history_size = history_size
self.act_fn = nn.ReLU() self.act_fn = nn.LeakyReLU()
self.fc_state_action = nn.Linear(state_size + action_size, hidden_size) self.fc_state_action = nn.Linear(state_size + action_size, hidden_size)
self.history_cell = nn.GRUCell(hidden_size + history_size, history_size) self.history_cell = nn.GRUCell(hidden_size + history_size, history_size)
@ -274,7 +274,7 @@ class ProjectionHead(nn.Module):
self.projection_model = nn.Sequential( self.projection_model = nn.Sequential(
nn.Linear(state_size + action_size, hidden_size), nn.Linear(state_size + action_size, hidden_size),
nn.LayerNorm(hidden_size), nn.LayerNorm(hidden_size),
nn.ReLU(), nn.LeakyReLU(),
nn.Linear(hidden_size, hidden_size), nn.Linear(hidden_size, hidden_size),
nn.LayerNorm(hidden_size), nn.LayerNorm(hidden_size),
) )