A few days ago, I saved a model using torch.save(model, PATH) and uploaded to gdrive. Now, when I load the model, I am getting different results. Before saving, I was getting a Chamfer Distance of 0.0017 (A metric I am using for training, the lower it is, the better) but now around 0.0350 if I use model.eval() before running and 0.0200 if I use model.train().
Here’s my code:
class Transform(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv1d(128*3+128,512,1)
self.conv2 = nn.Conv1d(512,512,1)
self.conv3 = nn.Conv1d(512,512,1)
self.conv4 = nn.Conv1d(512,1024,1)
self.bn1 = nn.BatchNorm1d(512)
self.bn2 = nn.BatchNorm1d(512)
self.bn3 = nn.BatchNorm1d(512)
# self.bn4 = nn.BatchNorm1d(512)
def forward(self, inp_global):
xb = F.relu(self.bn1(self.conv1(inp_global)))
xb = F.relu(self.bn2(self.conv2(xb)))
xb = F.relu(self.bn3(self.conv3(xb)))
xb = self.conv4(xb)
xb = nn.MaxPool1d(xb.size(-1))(xb)
output = nn.Flatten(1)(xb)
return output
class GlobalEncode(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv1d(3,64,1)
self.conv2 = nn.Conv1d(64,64,1)
self.conv3 = nn.Conv1d(64,128,1)
self.conv4 = nn.Conv1d(128,128,1)
self.bn1 = nn.BatchNorm1d(64)
self.bn2 = nn.BatchNorm1d(64)
self.bn3 = nn.BatchNorm1d(128)
# self.bn4 = nn.BatchNorm1d(512)
def forward(self, input):
xb = F.relu(self.bn1(self.conv1(input)))
xb = F.relu(self.bn2(self.conv2(xb)))
xb = F.relu(self.bn3(self.conv3(xb)))
xb = self.conv4(xb)
xb = nn.MaxPool1d(xb.size(-1))(xb)
return nn.Flatten(1)(xb)
class PointCloud(nn.Module):
def __init__(self):
super().__init__()
self.encode = Transform()
self.global_encode = GlobalEncode()
self.nb_heads = 8
self.conv1 = nn.Conv1d(in_channels=1024*self.nb_heads,
out_channels=1024*self.nb_heads, kernel_size=1,
groups=self.nb_heads)
self.conv2 = nn.Conv1d(in_channels=1024*self.nb_heads,
out_channels=1024*self.nb_heads, kernel_size=1,
groups=self.nb_heads)
self.conv3 = nn.Conv1d(in_channels=1024*self.nb_heads,
out_channels=256*3*self.nb_heads, kernel_size=1,
groups=self.nb_heads)
self.conv4 = nn.Conv1d(in_channels=256*3*self.nb_heads,
out_channels=256*3*self.nb_heads, kernel_size=1,
groups=self.nb_heads)
self.bn1 = nn.BatchNorm1d(1024*self.nb_heads)
self.bn2 = nn.BatchNorm1d(1024*self.nb_heads)
self.bn3 = nn.BatchNorm1d(256*3*self.nb_heads)
self.dp = nn.Dropout(p=0.2)
self.fc1 = nn.Linear(1024, 512)
self.fc2 = nn.Linear(512, 256)
self.fc3 = nn.Linear(256, 128)
self.fc4 = nn.Linear(128, 8)
self.dropout = nn.Dropout(p=0.2)
def forward(self, input, input_knn):
inp_global = self.global_encode(input).unsqueeze(1)
inp_global = inp_global.repeat(1,2048,1)
xb = torch.cat((input_knn, inp_global), dim = 2).transpose(1,2)
enc = self.encode(xb)
xb = enc.repeat(1,8).unsqueeze(2)
xb = F.relu(self.conv1(xb))
xb = F.relu(self.conv2(xb))
xb = F.relu(self.conv3(xb))
xb = self.conv4(xb)
output = xb.reshape(-1, 2048,3)
xb = self.dropout(F.relu(self.fc1(enc)))
xb = self.dropout(F.relu(self.fc2(xb)))
xb = F.relu(self.fc3(xb))
labels = self.fc4(xb)
return output, labels
, moreover even while using model.eval(), I am getting different results if I shuffle the dataset while validation. Could it be because of BatchNorm or Dropout?