Spaces:
Runtime error
Runtime error
File size: 2,638 Bytes
f39e999 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 |
import torch
from torch import nn
import torch.nn.functional as F
from configs.paths_config import model_paths
class MocoLoss(nn.Module):
def __init__(self):
super(MocoLoss, self).__init__()
print("Loading MOCO model from path: {}".format(model_paths["moco"]))
self.model = self.__load_model()
self.model.cuda()
self.model.eval()
@staticmethod
def __load_model():
import torchvision.models as models
model = models.__dict__["resnet50"]()
# freeze all layers but the last fc
for name, param in model.named_parameters():
if name not in ['fc.weight', 'fc.bias']:
param.requires_grad = False
checkpoint = torch.load(model_paths['moco'], map_location="cpu")
state_dict = checkpoint['state_dict']
# rename moco pre-trained keys
for k in list(state_dict.keys()):
# retain only encoder_q up to before the embedding layer
if k.startswith('module.encoder_q') and not k.startswith('module.encoder_q.fc'):
# remove prefix
state_dict[k[len("module.encoder_q."):]] = state_dict[k]
# delete renamed or unused k
del state_dict[k]
msg = model.load_state_dict(state_dict, strict=False)
assert set(msg.missing_keys) == {"fc.weight", "fc.bias"}
# remove output layer
model = nn.Sequential(*list(model.children())[:-1]).cuda()
return model
def extract_feats(self, x):
x = F.interpolate(x, size=224)
x_feats = self.model(x)
x_feats = nn.functional.normalize(x_feats, dim=1)
x_feats = x_feats.squeeze()
return x_feats
def forward(self, y_hat, y, x):
n_samples = x.shape[0]
x_feats = self.extract_feats(x)
y_feats = self.extract_feats(y)
y_hat_feats = self.extract_feats(y_hat)
y_feats = y_feats.detach()
loss = 0
sim_improvement = 0
sim_logs = []
count = 0
for i in range(n_samples):
diff_target = y_hat_feats[i].dot(y_feats[i])
diff_input = y_hat_feats[i].dot(x_feats[i])
diff_views = y_feats[i].dot(x_feats[i])
sim_logs.append({'diff_target': float(diff_target),
'diff_input': float(diff_input),
'diff_views': float(diff_views)})
loss += 1 - diff_target
sim_diff = float(diff_target) - float(diff_views)
sim_improvement += sim_diff
count += 1
return loss / count, sim_improvement / count, sim_logs
|