wagner-niklas / CAGE_expression_inference

Project to infere emotional expressions and benchmark datasets by Niklas Wagner, Felix Mätzler, Samed R. Vossberg, Helen Schneider and Svetlana Pavlitska.

MIT License

11 stars 2 forks source link

using AffectNet8_Swin_VA without the pretrained model from AffectNet8_Swin_Combined, the validation ccc only is 0.51 #4

Closed tempbrucefu closed 1 week ago

tempbrucefu commented 3 weeks ago

not using the pretrained model from AffectNet8_Swin_Combined, the codes are as following on AffectNet8_Swin_VA,
MODEL = models.swin_v2_t(weights="DEFAULT")

MODEL.head = torch.nn.Linear(in_features=768, out_features=10, bias=True)

# MODEL.to(DEVICE)
# MODEL.load_state_dict(torch.load("../AffectNet8_Swin_Combined/model.pt"))
MODEL.head = torch.nn.Linear(in_features=768, out_features=2, bias=True)

The performance on NUM_EPOCHS = 25 only has 0.51

tempbrucefu commented 3 weeks ago

for AffectNet8_Maxvit_VA, it is close to the report result, I got 0.75 of ccc after using the default NUM_EPOCHS = 20

wagner-niklas commented 3 weeks ago

Hi @tempbrucefu, thanks for sharing your information. Your metholody of not using a fine-tuning approach and your insights are very interesting. Without a dedicated pre-trained model, we also noticed quite unstable gradients for only VA. As known from literature, keep in mind that such training could require way more epochs than 25 to achieve similar results and especially different architectures operate different.

Best regards