Why divide 0.07

[vvthai10]

I don't understand why you divide right here
text_probs = image_features.unsqueeze(1) @ text_features.permute(0, 2, 1) text_probs = text_probs[:, 0, ...]/0.07 text_probs = text_probs.squeeze()
and another here
logit_scale = self.logit_scale.exp() # nn.Parameter(torch.ones([]) * np.log(1 / 0.07)) logits_per_image = logit_scale * image_features @ text_features.t() logits_per_text = logits_per_image.t()

And in another paper , they use multi with 100, example
for layer in range(len(det_patch_tokens)): det_patch_tokens[layer] = det_patch_tokens[layer] / det_patch_tokens[layer].norm(dim=-1, keepdim=True) anomaly_map = (100.0 * det_patch_tokens[layer] @ text_features) anomaly_map = torch.softmax(anomaly_map, dim=-1)[:, :, 1] anomaly_score = torch.mean(anomaly_map, dim=-1) det_loss += loss_bce(anomaly_score, image_label)

[zqhang]

It is the temperature ratio of the original CLIP.

[vvthai10]

But i see in https://github.com/mlfoundations/open_clip#usage, they use 100, you think it make result different?

[zqhang]

i dont think it will affect much

…

---- Replied Message ---- | From | ***@***.***> | | Date | 05/24/2024 14:30 | | To | ***@***.***> | | Cc | Qihang ***@***.***>***@***.***> | | Subject | Re: [zqhang/AnomalyCLIP] Why divide 0.07 (Issue #15) | But i see in https://github.com/mlfoundations/open_clip#usage, they use 100, you think it make result different? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[vvthai10]

Another question, i want use Linear layer to get local feature of images, You think when train text by your idea and with adapter linear. It can be better

[zqhang]

you can have a try

[vvthai10]

I am trying but i don't have make high ACC, I just high in first epoch and after it reduce and I don't know why :(

[vvthai10]

I train with dataset visa, I train 11 epoch, but loss and image_loss is 3.7960, 0.5325. I feel something wrong. I setup by your setting. Can you share me about loss and image_loss when u train?

…

On Fri, May 24, 2024 at 3:18 PM Qihang Zhou ***@***.***> wrote: you can have a try — Reply to this email directly, view it on GitHub <#15 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AR65VJJZ5CQCRLN2CUACCLTZD3ZVRAVCNFSM6AAAAABIFKQWISVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMRYHA4DINJQHE> . You are receiving this because you authored the thread.Message ID: ***@***.***>