cifar10.json

{
    "common_description": "Comparison of models with 1, 2, and 3 EPSes. One of the models with 1 EPS has very small stds initialization, the initialization of others are very similar to each other - init_epses_composition_unit_{empirical,theoretical}_output_std. More EPSes require smaller learning rate, otherwise they don't converge. One of the models has l-2 regularization, but in my experience it's not important.",
    "experiments": [
        {
            "rel_dir": "cifar10/28x28_grayscale/1_eps_k=4_q=4_very_small_stds/2020-05-15T21:07:44",
            "name": "One EPS, very small phi",
            "description": "EPS(K=4,Q=4), Initialized with very small stds, lr=3e-3, no l-2 regularization"
        },
        {
            "rel_dir": "cifar10/28x28_grayscale/1_eps_k=4_q=4_init_epses_composition_unit_empirical_output_std/2020-05-15T21:47:54",
            "name": "One EPS",
            "description": "EPS(K=4,Q=4), init_epses_composition_unit_empirical_output_std, lr=3e-3, no l-2 regularization"
        },
        {
            "rel_dir": "cifar10/28x28_grayscale/2_epses/2020-05-16T21:59:34",
            "name": "2 EPSes, x_0.σ=0.4617, output_of_linear_without_bias.σ=2.813e3",
            "description": "EPS(K=4,Q=4)+EPS(K=3,Q=6), init_epses_composition_unit_theoretical_output_std, lr=1.5e-4, no l-2 regularization"
        },
        {
            "rel_dir": "cifar10/28x28_grayscale/2_epses/2020-05-17T13:28:37",
            "name": "2 EPSes, x_0.σ=0.4263, output_of_linear_without_bias.σ=6.81e-2",
            "description": "EPS(K=4,Q=4)+EPS(K=3,Q=6), init_epses_composition_unit_theoretical_output_std, lr=1.5e-4, no l-2 regularization"
        },        
        {
            "rel_dir": "cifar10/28x28_grayscale/2_epses/2020-05-17T13:22:23",
            "name": "2 EPSes, x_0.σ=0.395, output_of_linear_without_bias.σ=1.197e-6",
            "description": "EPS(K=4,Q=4)+EPS(K=3,Q=6), init_epses_composition_unit_theoretical_output_std, lr=1.5e-4, no l-2 regularization"
        },
        {
            "rel_dir": "cifar10/28x28_grayscale/2_epses/2020-05-17T13:57:49",
            "name": "2 EPSes, x_0.σ=0.381, output_of_linear_without_bias.σ=6.65e-9",
            "description": "EPS(K=4,Q=4)+EPS(K=3,Q=6), init_epses_composition_unit_theoretical_output_std, lr=1.5e-4, no l-2 regularization"
        },
        {
            "rel_dir": "cifar10/28x28_grayscale/2_epses/2020-05-17T15:27:31",
            "name": "2 EPSes, x_0.σ=0.3692, output_of_linear_without_bias.σ=6.876e-11",
            "description": "EPS(K=4,Q=4)+EPS(K=3,Q=6), init_epses_composition_unit_theoretical_output_std, lr=1.5e-4, no l-2 regularization"
        },
        {
            "rel_dir": "cifar10/28x28_grayscale/2_epses/2020-05-17T15:05:49",
            "name": "2 EPSes, x_0.σ=σ=0.3542, output_of_linear_without_bias.σ=1.834-13",
            "description": "EPS(K=4,Q=4)+EPS(K=3,Q=6), init_epses_composition_unit_theoretical_output_std, lr=1.5e-4, no l-2 regularization"
        },
        {
            "rel_dir": "cifar10/28x28_grayscale/2_epses/2020-05-17T14:19:47",
            "name": "2 EPSes, x_0.σ=σ=0.3275. output_of_linear_without_bias.σ=2.214e-18, didn't work.",
            "description": "EPS(K=4,Q=4)+EPS(K=3,Q=6), init_epses_composition_unit_theoretical_output_std, lr=1.5e-4, no l-2 regularization. During training, the EPS's parameters didn't change."
        }        
    ]
}