2_training/hamiltonian/train_ham.py

#!/usr/bin/env python
"""DeepH-E3 Hamiltonian training for diamond — with live monitoring.

Uses our own 1_data_prepare data.  Training runs continuously (no
chunking) to keep LR-scheduler state intact.  Every CHECK_INTERVAL
epochs the val_loss is compared to reference milestones; if it lags
more than STALL_RATIO× the reference, training is stopped for
investigation.

Usage:
    python train_ham.py [params.json]             # train on our data
    python train_ham.py [params.json] --ref-graph # diagnostic: use reference graph PKL

Reference milestones (from diamond-qe reference training):
    Epoch  300: val_loss = 1.76e-05   LR = 2.00e-03
    Epoch  600: val_loss = 8.40e-06   LR = 2.00e-03
    Epoch  900: val_loss = 3.92e-06   LR = 1.00e-03  (first LR drop)
    Epoch 1200: val_loss = 2.63e-06   LR = 2.50e-04
    Epoch 1500: val_loss = 2.40e-06   LR = 1.25e-04
    Epoch 1800: val_loss = 2.03e-06   LR = 3.13e-05
    Epoch 2017: val_loss = 1.98e-06   LR = 1.56e-05  (best model)
"""
import glob
import json
import os
import re
import subprocess
import sys
import time

SCRIPT_DIR     = os.path.dirname(os.path.abspath(__file__))
DATA_DIR       = os.path.abspath(
    os.path.join(SCRIPT_DIR, '..', '..', '1_data_prepare', 'data'))
PARAMS_DEFAULT = os.path.abspath(
    os.path.join(SCRIPT_DIR, '..', '..', '1_data_prepare', 'params.json'))

DATASET_DIR  = os.path.join(SCRIPT_DIR, 'dataset')
GRAPH_DIR    = os.path.join(SCRIPT_DIR, '..', 'graph')
GRAPH_PKL    = os.path.join(GRAPH_DIR,
                             'HGraph-diamond_qe_e3-rFromDFT-edge=Aij.pkl')
REF_GRAPH_PKL = (
    '/home/apolyukhin/scripts/ml/diamond-qe/deeph-e3/graph/'
    'HGraph-diamond_qe_e3-rFromDFT-edge=Aij.pkl'
)
RESULTS_DIR  = os.path.join(SCRIPT_DIR, 'results')
INI_PATH     = os.path.join(SCRIPT_DIR, 'train.ini')
LAUNCHER     = os.path.join(SCRIPT_DIR, '_launcher.py')
TRAIN_LOG    = os.path.join(SCRIPT_DIR, 'train.log')

# Reference val_loss milestones
REF_MILESTONES = {
    300:  1.76e-5,
    600:  8.40e-6,
    900:  3.92e-6,
    1200: 2.63e-6,
    1500: 2.40e-6,
    1800: 2.03e-6,
}

CHECK_INTERVAL = 300    # check every this many epochs
STALL_RATIO    = 5.0    # stop if val_loss > STALL_RATIO × reference milestone
POLL_SECONDS   = 30     # how often to poll the log file


def load_params(path=None):
    with open(path or PARAMS_DEFAULT) as f:
        return json.load(f)


def build_dataset_links():
    """Create dataset/ with numbered dirs, each symlinking files from aohamiltonian/."""
    os.makedirs(DATASET_DIR, exist_ok=True)
    disp_dirs = sorted(glob.glob(os.path.join(DATA_DIR, 'disp-*')))
    n = 0
    for disp_dir in disp_dirs:
        ao_dir = os.path.abspath(
            os.path.join(disp_dir, 'reconstruction', 'aohamiltonian'))
        if not os.path.exists(os.path.join(ao_dir, 'hamiltonians.h5')):
            print(f'  WARNING: no hamiltonians.h5 in {ao_dir}, skipping')
            continue
        dest = os.path.join(DATASET_DIR, f'{n:02d}')
        os.makedirs(dest, exist_ok=True)
        for fname in os.listdir(ao_dir):
            link = os.path.join(dest, fname)
            if not os.path.exists(link):
                os.symlink(os.path.join(ao_dir, fname), link)
        n += 1
    print(f'  {n} dataset dirs ready in {DATASET_DIR}')
    return n


def write_train_ini(params, use_ref_graph=False, num_epoch=None, results_dir=None):
    """Write train.ini with reference hyperparameters and our own data."""
    t = params.get('hamiltonian', {})
    save_dir = results_dir or RESULTS_DIR

    if use_ref_graph:
        graph_dir_val      = REF_GRAPH_PKL
        processed_data_val = ''
        save_graph_val     = ''
        print(f'  Using REFERENCE graph PKL: {graph_dir_val}')
    elif os.path.exists(GRAPH_PKL):
        graph_dir_val      = os.path.abspath(GRAPH_PKL)
        processed_data_val = ''
        save_graph_val     = ''
        print(f'  Using cached graph: {graph_dir_val}')
    else:
        graph_dir_val      = ''
        processed_data_val = DATASET_DIR
        save_graph_val     = os.path.abspath(GRAPH_DIR)
        print(f'  Graph not found — will build from {DATASET_DIR}')

    ini = f"""; DeepH-E3 training config — generated by train_ham.py

[basic]
device = {t.get('device', 'cuda')}
dtype = float
save_dir = {save_dir}
additional_folder_name = diamond_e3
simplified_output = False
seed = 42
checkpoint_dir =
use_new_hypp = True

[data]
graph_dir = {graph_dir_val}
DFT_data_dir =
processed_data_dir = {processed_data_val}
save_graph_dir = {save_graph_val}
target_data = hamiltonian
dataset_name = diamond_qe_e3
get_overlap = False

[train]
num_epoch = {num_epoch or t.get('num_epoch', 3000)}
batch_size = 1
extra_validation = []
extra_val_test_only = True

train_size = {t.get('train_size', 30)}
val_size   = {t.get('val_size', 10)}
test_size  = 10

min_lr = 1e-5

[hyperparameters]
learning_rate = {t.get('learning_rate', 0.002)}
Adam_betas = (0.9, 0.999)
scheduler_type = 1
scheduler_params = (factor=0.5, cooldown=40, patience=120, threshold=0.05)
revert_decay_patience = 20
revert_decay_rate = 0.8

[target]
target = hamiltonian
target_blocks_type = all
target_blocks =
selected_element_pairs =
convert_net_out = False

[network]
cutoff_radius = {t.get('cutoff_radius', 7.4)}
only_ij = False
spherical_harmonics_lmax = {t.get('lmax', 4)}
spherical_basis_irreps =
irreps_embed = {t.get('irreps_embed', '64x0e')}
irreps_mid = {t.get('irreps_mid', '64x0e+32x1o+16x2e+8x3o+8x4e')}
num_blocks = {t.get('num_blocks', 3)}
ignore_parity = False

irreps_embed_node =
irreps_edge_init =
irreps_mid_node =
irreps_post_node =
irreps_out_node =
irreps_mid_edge =
irreps_post_edge =
out_irreps =
"""
    with open(INI_PATH, 'w') as f:
        f.write(ini)
    print(f'  Written {INI_PATH}')


def write_launcher(params):
    t = params.get('hamiltonian', {})
    deeph_e3_dir = t.get('deeph_e3_dir', '/home/apolyukhin/Development/DeepH-E3')
    with open(LAUNCHER, 'w') as f:
        f.write(f"""import sys, torch
torch.serialization.add_safe_globals([slice])
try:
    from torch_geometric.data.data import DataEdgeAttr, DataTensorAttr
    from torch_geometric.data.storage import GlobalStorage
    torch.serialization.add_safe_globals([DataEdgeAttr, DataTensorAttr, GlobalStorage])
except ImportError:
    pass
sys.path.insert(0, '{deeph_e3_dir}')
from deephe3 import DeepHE3Kernel
kernel = DeepHE3Kernel()
kernel.train('{INI_PATH}')
""")


def parse_train_log(log_path):
    """Return {epoch: val_loss} from training log."""
    result = {}
    try:
        with open(log_path) as f:
            for line in f:
                m = re.search(r'Epoch #(\d+)\s+\|.*Val loss:\s+([\d.e+\-]+)', line)
                if m:
                    result[int(m.group(1))] = float(m.group(2))
    except (FileNotFoundError, OSError):
        pass
    return result


def monitor_loop(proc, log_path=None):
    """Poll log every POLL_SECONDS; report at CHECK_INTERVAL boundaries.

    Returns True if stall was detected (proc already terminated), else False.
    """
    log_path = log_path or TRAIN_LOG
    last_reported = 0

    while proc.poll() is None:
        time.sleep(POLL_SECONDS)
        history = parse_train_log(log_path)
        if not history:
            continue

        latest = max(history)
        boundary = (latest // CHECK_INTERVAL) * CHECK_INTERVAL
        if boundary > last_reported and boundary > 0:
            last_reported = boundary
            # Best val_loss up to this boundary
            window = {ep: v for ep, v in history.items() if ep <= boundary + 20}
            if not window:
                continue
            val = min(window.values())
            ref = REF_MILESTONES.get(boundary)
            if ref:
                ratio = val / ref
                flag = 'OK' if ratio < STALL_RATIO else 'STALLED'
                print(f'  [epoch ~{boundary:4d}] val={val:.3e}  '
                      f'ref={ref:.3e}  ratio={ratio:.1f}x  [{flag}]',
                      flush=True)
                if ratio > STALL_RATIO:
                    print(f'\nStall detected ({ratio:.1f}x above reference). '
                          f'Terminating training.')
                    proc.terminate()
                    return True
            else:
                print(f'  [epoch ~{boundary:4d}] val={val:.3e}', flush=True)

    return False


def main():
    use_ref_graph = '--ref-graph' in sys.argv
    params_path   = next((a for a in sys.argv[1:] if not a.startswith('--')), None)
    params = load_params(params_path)

    t = params.get('hamiltonian', {})
    conda_env  = t.get('conda_env', 'deeph')
    conda_base = params['paths']['conda_base']

    # Diagnostic mode: use reference graph, separate results dir and log
    if use_ref_graph:
        results_dir = os.path.join(SCRIPT_DIR, 'results_ref')
        train_log   = os.path.join(SCRIPT_DIR, 'train_ref.log')
        num_epoch   = 500
        if not os.path.exists(REF_GRAPH_PKL):
            print(f'ERROR: Reference graph not found:\n  {REF_GRAPH_PKL}')
            sys.exit(1)
        print('=== DIAGNOSTIC: using reference graph PKL (500 epochs) ===')
    else:
        results_dir = RESULTS_DIR
        train_log   = TRAIN_LOG
        num_epoch   = None

    os.makedirs(results_dir, exist_ok=True)
    os.makedirs(os.path.abspath(GRAPH_DIR), exist_ok=True)

    # Build dataset symlinks only if graph PKL is missing (non-ref mode)
    if not use_ref_graph and not os.path.exists(GRAPH_PKL):
        print('Step 1: Building dataset symlinks...')
        n = build_dataset_links()
        if n == 0:
            print('ERROR: No aohamiltonian data found. Run reconstruct.py first.')
            sys.exit(1)
    else:
        print('Step 1: Skipped (graph PKL available).')

    print('Step 2: Writing train.ini...')
    write_train_ini(params, use_ref_graph=use_ref_graph,
                    num_epoch=num_epoch, results_dir=results_dir)
    write_launcher(params)

    print('\nReference milestones:')
    for ep, vl in sorted(REF_MILESTONES.items()):
        print(f'  Epoch {ep:5d}: val_loss = {vl:.2e}')
    print(f'\nStall threshold : {STALL_RATIO}× reference')
    print(f'Training log    : {train_log}')
    print(f'Conda env       : {conda_env}')
    print(f'\nStep 3: Launching training...\n', flush=True)

    activate = (f'source {conda_base}/etc/profile.d/conda.sh'
                f' && conda activate {conda_env}')

    with open(train_log, 'w') as log_f:
        proc = subprocess.Popen(
            ['bash', '-c', f'{activate} && python {LAUNCHER}'],
            stdout=log_f,
            stderr=subprocess.STDOUT,
        )

    try:
        stalled = monitor_loop(proc, train_log)
    except KeyboardInterrupt:
        print('\nInterrupted — stopping training.')
        proc.terminate()
        stalled = False

    rc = proc.wait()

    # Final summary
    history = parse_train_log(train_log)
    if history:
        best_ep  = min(history, key=history.get)
        best_val = history[best_ep]
        last_ep  = max(history)
        last_val = history[last_ep]
        print(f'\nTraining finished (exit code {rc}).')
        print(f'  Last  epoch {last_ep:5d}: val_loss = {last_val:.3e}')
        print(f'  Best  epoch {best_ep:5d}: val_loss = {best_val:.3e}')
        print(f'  Reference best (epoch 2017): 1.98e-06')
        print(f'  Ratio to reference best: {best_val / 1.98e-6:.2f}×')
    else:
        print(f'\nTraining finished (exit code {rc}). No log data parsed.')

    if stalled:
        print('\nInvestigation hints:')
        print('  - Check train.log for LR at stall point.')
        print('  - If val_loss is flat, verify data matches reference:')
        print('    diff sizes: ls -lh 1_data_prepare/data/disp-01/'
              'reconstruction/aohamiltonian/hamiltonians.h5')
        print('    vs reference: /home/apolyukhin/scripts/ml/diamond-qe/'
              'disp-01/reconstruction/aohamiltonian/hamiltonians.h5')


if __name__ == '__main__':
    main()