Other aspects

During our preliminary investigation, we found that the network complexity and the value of lambda are critical for training a model effectively. Therefore, we validate all our models using the following values of lambda and network complexity measured using the minimum number of filters.

We varied the value of lambda as below-

1. 10

2. 50

3. 100

The minimum number of filters used in any model were varied as-

1. 32

2. 64

3. 128

In the literature on CNNs, we have observed that many research papers suggest using Upsampling whereas others suggest using Fractional Strided Convolution. Fractional Strided convolution learns its own upsampling parameters during the training. Upsampling is faster on the other hand as it uses a pre-determined strategy for upsampling.

Therefore, fractional strided convolution performs better since the model learns a specific upsampling for the dataset. On the other hand, Upsampling is faster as it uses a pre-determined strategy. In certain cases where dataset follows some interpolation pattern then upsampling can be beneficial in terms of saving time.

We test both of these parameters and report our results.

We want to test the impact of depth of a network on learning in case of Image Colorization. Hence, we will use a deeper Unet architecture with 15 layers and a shallow architecture with 10 layers.

We will report the closeness score on both these architectures when trained using different hyperparameters.