zero shot classification results

[gefend]

I tried using your script for zero shot classification together with the pretrained weights (both resnet18 and resnet50). The calssification results I got are very random (accuracy 17-22% for each class). Maybe there is an aditional step needed or the weights I downloaded are not the trained weights?

[marshuang80]

To help you debug this, can you please provide the following information?

• What dataset are you using?
• How are you preprocessing the data?
• What are your classification tasks?
• How are you generating the class prompts?

[gefend]

1. I'm using the Chexpert 5X200 dataset
2. I have used the "Zeroshot classification for CheXpert5x200" script from your Readme file without additional preproccesing.
   The only change I made in the "Zeroshot classification for CheXpert5x200" was the generation of the chexpert 5X200 data set using your function preprocess_chexpert_5X200_data from gloria.datasets.preprocess_datasets .
3. The classification task is the same as your classification task on chexpert 5X200, classification of each image into on of the 5 classes.
4. As part of using your script I am generating the class_prompts using your function generate_chexpert_class_prompts
   Thank you for your answer!

[marshuang80]

Got it, thanks for the info. May I ask how you are computing the results? Using different random seeds I was still able to get an accuracy of 60+

labels = df[gloria.constants.CHEXPERT_COMPETITION_TASKS].to_numpy().argmax(axis=1)
pred = similarities[gloria.constants.CHEXPERT_COMPETITION_TASKS].to_numpy().argmax(axis=1)
acc = len(labels[labels == pred]) / len(labels)
print(acc) # 0.607

[gefend]

The following is the code I used, I tried to replace the results calculation to the one you suggested but I get the same results. The only thing that I think I'm doing different than you is that I take only 200 images every run because of my GPU memory capacity.

import torch
import gloria
import pandas as pd
from gloria.datasets.preprocess_datasets import preprocess_chexpert_5x200_data

df = preprocess_chexpert_5x200_data()
df = df[0:200]
# load model
device = "cuda" if torch.cuda.is_available() else "cpu"
gloria_model = gloria.load_gloria(device=device)

cls_prompts = gloria.generate_chexpert_class_prompts()

# process input images and class prompts
processed_txt = gloria_model.process_class_prompts(cls_prompts, device)
processed_imgs = gloria_model.process_img(df['Path'].tolist(), device)

# zero-shot classification on 1000 images
similarities = gloria.zero_shot_classification(
    gloria_model, processed_imgs, processed_txt)

labels = df[gloria.constants.CHEXPERT_COMPETITION_TASKS].to_numpy().argmax(axis=1)
pred = similarities[gloria.constants.CHEXPERT_COMPETITION_TASKS].to_numpy().argmax(axis=1)
acc = len(labels[labels == pred]) / len(labels) #0.17

[baileyyeah0326]

Hi, I have the same issue with you. Have you figured it out?