Ayushman72
commited on
Commit
·
578a261
1
Parent(s):
2f9fa75
Revert "Update README.md"
Browse filesThis reverts commit 7657166733c3abeb525aa0258af4270329374542.
README.md
CHANGED
|
@@ -1,10 +1,59 @@
|
|
| 1 |
---
|
| 2 |
pipeline_tag: text-to-image
|
| 3 |
tags:
|
| 4 |
-
-
|
| 5 |
-
-
|
|
|
|
|
|
|
|
|
|
| 6 |
---
|
|
|
|
| 7 |
|
| 8 |
-
This model
|
| 9 |
-
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
pipeline_tag: text-to-image
|
| 3 |
tags:
|
| 4 |
+
- image captioing
|
| 5 |
+
- vit
|
| 6 |
+
- gpt
|
| 7 |
+
- gpt2
|
| 8 |
+
- torch
|
| 9 |
---
|
| 10 |
+
# Image Captioning using ViT and GPT2 architecture
|
| 11 |
|
| 12 |
+
This is my attempt to make a transformer model which takes image as the input and provides a caption for the image
|
| 13 |
+
|
| 14 |
+
## Model Architecture
|
| 15 |
+
It comprises of 12 ViT encoder and 12 GPT2 decoders
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
## Training
|
| 20 |
+
The model was trained on the dataset Flickr30k which comprises of 30k images and 5 captions for each image
|
| 21 |
+
The model was trained for 8 epochs (which took 10hrs on kaggle's P100 GPU)
|
| 22 |
+
|
| 23 |
+
## Results
|
| 24 |
+
The model acieved a BLEU-4 score of 0.2115, CIDEr score of 0.4, METEOR score of 0.25, and SPICE score of 0.19 on the Flickr8k dataset
|
| 25 |
+
|
| 26 |
+
These are the loss curves.
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## Predictions
|
| 33 |
+
To predict your own images download the models.py, predict.py and the requirements.txt and then run the following commands->
|
| 34 |
+
|
| 35 |
+
`pip install -r requirements.txt`
|
| 36 |
+
|
| 37 |
+
`python predict.py`
|
| 38 |
+
|
| 39 |
+
*Predicting for the first time will take time as it has to download the model weights (1GB)*
|
| 40 |
+
|
| 41 |
+
Here are a few examples of the prediction done on the Validation dataset
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
As we can see these are not the most amazing predictions. The performance could be improved by training it further and using an even bigger dataset like MS COCO (500k captioned images)
|
| 54 |
+
|
| 55 |
+
## FAQ
|
| 56 |
+
|
| 57 |
+
Check the [full notebook](./imagecaptioning.ipynb) or [Kaggle](https://www.kaggle.com/code/ayushman72/imagecaptioning)
|
| 58 |
+
|
| 59 |
+
Download the [weights](https://drive.google.com/file/d/1X51wAI7Bsnrhd2Pa4WUoHIXvvhIcRH7Y/view?usp=drive_link) of the model
|