Vision Models in Google Colab
Budget: $15 – $25 USD
Buget:
$150
Steps:
1)
Find 5 vision models that does semantic segmentation.
Could be from:
https://huggingface.co/models?pipeline_tag=image-segmentation&sort=trending
https://paperswithcode.com/task/semantic-segmentation
github.com, or other sources.
Important: models should be recent (from the last 9 years: 2015-2024), well known and having an article with results. Because we are going to benchmark them and get new results. Make sure all found models can be executed and re-trained in the free version of Google Colab using GPU environment.
2)
Create a free version of Google Colab notebook using GPU environment. In notebook do the following:
2.1) change each model's architecture found at step 1, by appending a small model like this:
input_layer = custom_input_layer()
hidden = keras.layers.Dense(10, activation='gelu')(input_layer)
hidden = keras.layers.Concatenate()([input_layer, hidden])
hidden = keras.layers.Dense(10, activation='gelu')(hidden)
2.2) original models found at step 1, should preserve its original weights and parameters, because we just extend the original models by merging them with a part of another model. Make sure the secondary model preserve its weights as well.
2.3) The secondary model could be appended at the begging of found model or at the end of found model. Make a function that have a option to allow us to chose how to append the model (begging, end).
2.4) When extending the found model, the resulting model should have connected custom_input_layer of secondary model with input of found model and secondary final layer should be done connected using layer concatenation
3)
Train again the models on their original datasets which previously were trained from their original paper. We should compare the results of original model and modified model. In notebook do the following:
3.1) Train each modified model from step 2 on their original dataset
3.2) Collect all their results in similar format as original model's paper
$150
Steps:
1)
Find 5 vision models that does semantic segmentation.
Could be from:
https://huggingface.co/models?pipeline_tag=image-segmentation&sort=trending
https://paperswithcode.com/task/semantic-segmentation
github.com, or other sources.
Important: models should be recent (from the last 9 years: 2015-2024), well known and having an article with results. Because we are going to benchmark them and get new results. Make sure all found models can be executed and re-trained in the free version of Google Colab using GPU environment.
2)
Create a free version of Google Colab notebook using GPU environment. In notebook do the following:
2.1) change each model's architecture found at step 1, by appending a small model like this:
input_layer = custom_input_layer()
hidden = keras.layers.Dense(10, activation='gelu')(input_layer)
hidden = keras.layers.Concatenate()([input_layer, hidden])
hidden = keras.layers.Dense(10, activation='gelu')(hidden)
2.2) original models found at step 1, should preserve its original weights and parameters, because we just extend the original models by merging them with a part of another model. Make sure the secondary model preserve its weights as well.
2.3) The secondary model could be appended at the begging of found model or at the end of found model. Make a function that have a option to allow us to chose how to append the model (begging, end).
2.4) When extending the found model, the resulting model should have connected custom_input_layer of secondary model with input of found model and secondary final layer should be done connected using layer concatenation
3)
Train again the models on their original datasets which previously were trained from their original paper. We should compare the results of original model and modified model. In notebook do the following:
3.1) Train each modified model from step 2 on their original dataset
3.2) Collect all their results in similar format as original model's paper