bh-an / image-captioning

ML models for Image captoining using CNN+LSTM and ResNet+GRU on the Flickr8k dataset

Image-captioning

ResNet+GRU results:

Actual Captions :- 
The two ladies are riding bicycles near the beach .
Two women in summer wear ride beach cruiser tricycles on the concrete near the beach .
Two women on low riding three-wheeled vehicles with baskets .
two women ride their three wheelers .
Two women riding tricycles .

Predicted Caption : Two women ride a three wheelers .
0.5814307369682193

Actual Captions :- 
Two girls arm wrestle as another observes
Two girls arm wrestle while a third girl in a pink shirt and glasses watches .
Two girls arm wrestling , while another looks on .
Two teenage girls arm wrestle while a third girl watches .
Two young girls are arm wrestling in their hotel room while another girl watches .

Predicted Caption : Two girls arm wrestle while a third girl watches .
1.0

Actual Captions :- 
A bunch of dogs are competing in a race .
Five greyhounds are racing on a sand track .
Muzzled greyhounds are racing on the track .
several muzzled greyhound dogs racing around a track
The number 2 dog in the blue vest is in the lead at the dog races .

Predicted Caption : Three muzzled greyhounds race around a track while a dog watches .
0.33180774028439436

CNN+LSTM results:

image-captioning's Issues

Can you improve it please

There is no verification set generator in the code, so there is no verification set loss and training set loss when training the model, and the code for Bleu score evaluation is missing. Can you improve it when you have time?