search

ysmiura / ifcc

The code of Improving Factual Completeness and Consistency of Image-to-text Radiology Report Generation
Apache License 2.0
87 stars 19 forks source link

Image features CNN(img) #3

Closed nooralahzadeh closed 3 years ago

nooralahzadeh commented 3 years ago

Hi,

Would you please describe what steps (transform, pad, resize, ...) have you applied to provide CNN image features?

Thanks,

ysmiura commented 3 years ago

A random rotation, a color jittering, a padding to a square resolution, a resizing to 224x224, and a tensor normalization are applied to images. These image transformation processes can be found in clinicgen/data/image2text.py.

nooralahzadeh commented 3 years ago

Thanks for the reply, May I ask you whether this transformation process has been applied in the same manner on the Validation and Test dataset as the Train dataset?

ysmiura commented 3 years ago

The random rotation and the color jittering are only applied to training data.

nooralahzadeh commented 3 years ago

Thanks for the reply, May I ask you about the value for the max-word that you use for the generation? is it as default =128 or =50?

ysmiura commented 3 years ago

For the experiment in the paper, we used max-word=128.