ainstein/run.py at main · edhirst/ainstein · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
import tensorflow as tf

tfk = tf.keras
tfk.backend.set_floatx("float64")
import numpy as np
import yaml

import wandb
from helper_functions import argument_parser, wandb_helper
from network.ball import BallNetwork
from sampling.ball import BallSample, CubeSample


# Main body function for performing the metric training
def main(hyperparameters_file, runtime_args, wandb_id=None):
    ###########################################################################
    ### Training & Logging set-up ###
    # Load the hyperparameters YAML file
    with open(hyperparameters_file, "r") as file:
        args = yaml.safe_load(file)

    for arg, arg_val in args.items():
        if arg in runtime_args:
            args[arg] = runtime_args[arg]

    # Check if restoring WandB
    if wandb_id is not None:
        args, x_train = wandb_helper.restore_wandb(args, wandb_id)
    else:
        # Initialise the training and val batches
        x_train, x_val = None, None

    # Check and set seeds for reproducibility
    rng = np.random.default_rng()
    # ...for NumPy
    if args["np_seed"] is None:
        args["np_seed"] = int(rng.integers(2**32 - 2))
    # ...for TensorFlow
    if args["tf_seed"] is None:
        args["tf_seed"] = int(rng.integers(2**32 - 2))

    # Make sure the config things are ints
    args["np_seed"] = int(args["np_seed"])
    args["tf_seed"] = int(args["tf_seed"])
    np.random.seed(args["np_seed"])
    tf.random.set_seed(args["tf_seed"])
    tfk.utils.set_random_seed(args["tf_seed"])

    # Print some random characters to check seed applied correctly
    print("TF random key: ", tf.random.uniform(shape=[6]))
    print("NP random key: ", np.random.randint(1, np.iinfo(np.int32).max, size=6))

    # Start a WeightsandBiases session, and allow resuming from checkpoint
    wandb.init(
        project="Ainstein_ball",
        entity="logml",
        config=args,
        id=wandb_id,
        resume="allow",
    )

    # Allow WandB to control the hyperparameters for sweeps (amounts to
    # over-writing the hyperparameters file with new values).
    hp = wandb.config

    # Add run identifiers for saving tracability
    run_name = wandb.run.name or ""
    run_id = wandb.run.id or 42
    hp["run_identifiers"] = (run_name, run_id)

    ###########################################################################
    ### Data set-up ###
    # Create training and validation samples
    if wandb_id is None:
        # Ball patch sampling
        if hp["ball"]:
            train_sample = BallSample(
                hp.num_samples,
                dimension=hp.dim,
                patch_width=hp["patch_width"],
                density_power=hp["density_power"],
            )
            if hp["validate"]:
                val_sample = BallSample(
                    hp.num_val_samples,
                    dimension=hp.dim,
                    patch_width=hp["patch_width"],
                    density_power=hp["density_power"],
                )
        # Cube patch sampling (full functionality unlikely entirely compatible at present)
        else:
            assert hp["n_patches"] == 1, (
                "Cube sampling only suitable for local geometries where don't need the ball structure for patching (set n_patches = 1)"
            )
            train_sample = CubeSample(
                hp.num_samples,
                dimension=hp.dim,
                width=hp["patch_width"],
                density_power=hp["density_power"],
            )
            if hp["validate"]:
                val_sample = CubeSample(
                    hp.num_val_samples,
                    dimension=hp.dim,
                    width=hp["patch_width"],
                    density_power=hp["density_power"],
                )
    # If wandb_id is not None
    else:
        train_sample = x_train
        if hp["validate"]:
            val_sample = x_val

    # Convert to tf objects
    train_sample_tf = tf.convert_to_tensor(train_sample, dtype=tf.dtypes.float64)
    val_sample_tf = None
    if hp["validate"]:
        val_sample_tf = tf.convert_to_tensor(val_sample, dtype=tf.dtypes.float64)

    ###########################################################################
    ### Run ML ###
    # Instantiate the network
    network = BallNetwork(hp=hp, print_losses=hp.print_losses)

    # Train!
    loss_hist = network.train(
        x_train=train_sample_tf, validate=hp["validate"], x_val=val_sample_tf
    )

    # Close the WandB session
    wandb.finish()

    return loss_hist, train_sample_tf, val_sample_tf


###############################################################################
if __name__ == "__main__":
    # Extract the runtime args
    args = argument_parser.get_args()

    # Extract any specified WandB id passed to the run
    wandb_id = args.wandb_id

    # Create a dict of the training arguments
    runtime_args = argument_parser.prune_none_args(args)

    # Perform the training
    lh, train_data, val_data = main(args.hyperparams, runtime_args, wandb_id)