logasja/instagram-ginza

Prompt

input

Prompt

target

Prompt

output

Training logs here

Model Description

This model uses a modified vnet for 2D input/output implemented here with the following configuration.

{
    "activation": "ReLU",
    "batch_norm": false,
    "filter_num": [
        64,
        128,
        256,
        512,
        512
    ],
    "n_labels": 3,
    "output_activation": "tanh",
    "pool": false,
    "res_num_ini": 1,
    "res_num_max": 3,
    "unpool": false
}

{
    "alpha": 0.0001,
    "batch": 64,
    "epochs": 500,
    "epsilon": 1,
    "input": "(256, 256)",
    "losses": {
        "FEAT_ArcFace": {
            "d": "cosine_similarity",
            "f": "ArcFace",
            "name": "FEAT_ArcFace",
            "reduction": "sum_over_batch_size",
            "threshold": 0.68,
            "weight": 0.1
        },
        "TopIQ": {
            "full_ref": true,
            "lower_better": false,
            "name": "TopIQ",
            "reduction": "sum_over_batch_size",
            "score_range": "~0, ~1",
            "weight": 0.9
        },
        "mean_squared_error": {
            "name": "mean_squared_error",
            "reduction": "sum_over_batch_size",
            "weight": 0.1
        }
    },
    "mixed_precision": true,
    "optimizer": {
        "amsgrad": false,
        "beta_1": 0.9,
        "beta_2": 0.999,
        "clipnorm": null,
        "clipvalue": null,
        "ema_momentum": 0.99,
        "ema_overwrite_frequency": null,
        "epsilon": 1e-07,
        "global_clipnorm": null,
        "gradient_accumulation_steps": null,
        "learning_rate": 9.999999747378752e-05,
        "loss_scale_factor": null,
        "name": "adamw",
        "use_ema": false,
        "weight_decay": 0.004
    },
    "seed": "BIIIIIGSTRETCH",
    "testing": 0.01,
    "training": 0.99
}

Model Architecture Plot