diff-evol-tree-search/profile_compute.py at main · ramithuh/diff-evol-tree-search · 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
"""Profile per-epoch computation costs."""
import time
import os
os.environ["XLA_PYTHON_CLIENT_PREALLOCATE"] = "false"

import jax
import jax.numpy as jnp
from jax import jit, vmap
import numpy as np
import optax
from jaxopt import OptaxSolver

from common.setup import get_one_tree_and_seq
from common.tree_func import *

# Simulate 16-leaf, 256 seq, 20 letters, 10 inits
n_leaves = 16
n_ancestors = 15
n_all = 31
seq_length = 256
n_letters = 20
init_count = 10

key = jax.random.PRNGKey(42)
initializer = jax.nn.initializers.kaiming_normal()

# Fake metadata (minimal for JIT)
metadata = {
    'n_all': n_all, 'n_leaves': n_leaves, 'n_ancestors': n_ancestors,
    'seq_length': seq_length, 'n_letters': n_letters, 'n_mutations': 50,
    'args': {'fix_seqs': False, 'fix_tree': False},
    'seq_temp': 0.5, 'lr': 0.1, 'lr_seq': 0.01, 'epochs': 100,
    'tLs': [0, 0.005, 10, 50], 'init_count': init_count,
    'exp_name': None, 'notes': None, 'tags': None, 'project': None,
}
metadata['args']['notes'] = None
metadata['args']['tags'] = None
metadata['args']['project'] = None

# Params
tree_params = {'t': initializer(key + 10, (init_count, n_all - 1, n_ancestors), jnp.float64)}
seq_slices = [initializer(key+i+10, (init_count, seq_length, n_letters), jnp.float64)
              for i in range(n_ancestors)]
seq_params = {'s': jnp.stack(seq_slices, axis=1)}

seqs = jax.random.normal(key, (n_all, seq_length, n_letters), dtype=jnp.float64)
sm = jnp.ones((n_letters, n_letters)) - jnp.identity(n_letters).astype(jnp.float64)

# Objectives
def inner_objective(seq_params, tree_params, data):
    seqs, temp, epoch = data
    return compute_loss_optimized(tree_params, seq_params, seqs, temp, epoch)

def outer_objective(tree_params, seq_params, data):
    seqs, temp, epoch = data
    return compute_loss_optimized(tree_params, seq_params, seqs, temp, epoch)

# Optimizers
seq_optimizer = OptaxSolver(opt=optax.adam(0.01), fun=inner_objective, maxiter=1)
vmap_seq_init = vmap(seq_optimizer.init_state, ({'s': 0}, {'t': 0}, None), 0)
seq_opt_state = vmap_seq_init(seq_params, tree_params, [seqs, metadata['tLs'][0], 0])
jitted_seq_update = jit(vmap(seq_optimizer.update, ({'s': 0}, 0, {'t': 0}, None), 0))

tree_optimizer = OptaxSolver(opt=optax.adam(0.1), fun=outer_objective)
vmap_tree_init = vmap(tree_optimizer.init_state, ({'t': 0}, {'s': 0}, None), 0)
tree_opt_state = vmap_tree_init(tree_params, seq_params, [seqs, metadata['tLs'][0], 0])
jitted_tree_update = jit(vmap(tree_optimizer.update, ({'t': 0}, 0, {'s': 0}, None), 0))

vmap_compute_detailed = jit(vmap(compute_detailed_loss_optimized, ({'t':0}, {'s': 0}, None, None, None, None), 0))

def bench(name, fn, n=20):
    # warmup
    for _ in range(3):
        fn()
    jax.block_until_ready(fn())  # ensure GPU sync
    times = []
    for _ in range(n):
        t0 = time.perf_counter()
        result = fn()
        jax.block_until_ready(result)
        times.append(time.perf_counter() - t0)
    avg = np.mean(times) * 1000
    std = np.std(times) * 1000
    print(f"{name:45s}  {avg:8.2f} ms  (+/- {std:.2f} ms)")

print(f"\n--- Per-epoch compute costs (l={n_leaves}, sl={seq_length}, ic={init_count}) ---\n")
print(f"Device: {jax.devices()[0]}\n")

# Tree update
bench("Tree optimizer update (vmap+jit)",
      lambda: jitted_tree_update(tree_params, tree_opt_state, seq_params, [seqs, metadata['tLs'][0], 0]))

# Seq update
bench("Seq optimizer update (vmap+jit)",
      lambda: jitted_seq_update(seq_params, seq_opt_state, tree_params, [seqs, metadata['tLs'][0], 0]))

# Detailed loss (for logging)
bench("Detailed loss computation (vmap+jit)",
      lambda: vmap_compute_detailed(tree_params, seq_params, seqs, metadata['tLs'][0], sm, 0))

# argmin + get_one_tree_and_seq
cost, _, _, _ = vmap_compute_detailed(tree_params, seq_params, seqs, metadata['tLs'][0], sm, 0)
bench("argmin + get_one_tree_and_seq",
      lambda: get_one_tree_and_seq(tree_params, seq_params, jnp.argmin(cost)))

# Full epoch (tree + seq + cost)
def full_epoch():
    tp, ts_state = jitted_tree_update(tree_params, tree_opt_state, seq_params, [seqs, metadata['tLs'][0], 0])
    sp, ss_state = jitted_seq_update(seq_params, seq_opt_state, tp, [seqs, metadata['tLs'][0], 0])
    c, cs, tf, l = vmap_compute_detailed(tp, sp, seqs, metadata['tLs'][0], sm, 0)
    return c

bench("Full epoch (tree + seq + cost)", full_epoch)

# Scale test: ic=100
print(f"\n--- Same but with ic=100 ---\n")
init_count_100 = 100
tree_params_100 = {'t': initializer(key + 10, (init_count_100, n_all - 1, n_ancestors), jnp.float64)}
seq_slices_100 = [initializer(key+i+10, (init_count_100, seq_length, n_letters), jnp.float64)
                  for i in range(n_ancestors)]
seq_params_100 = {'s': jnp.stack(seq_slices_100, axis=1)}

seq_opt_state_100 = vmap_seq_init(seq_params_100, tree_params_100, [seqs, metadata['tLs'][0], 0])
tree_opt_state_100 = vmap_tree_init(tree_params_100, seq_params_100, [seqs, metadata['tLs'][0], 0])

def full_epoch_100():
    tp, ts_state = jitted_tree_update(tree_params_100, tree_opt_state_100, seq_params_100, [seqs, metadata['tLs'][0], 0])
    sp, ss_state = jitted_seq_update(seq_params_100, seq_opt_state_100, tp, [seqs, metadata['tLs'][0], 0])
    c, cs, tf, l = vmap_compute_detailed(tp, sp, seqs, metadata['tLs'][0], sm, 0)
    return c

bench("Full epoch ic=100 (tree + seq + cost)", full_epoch_100)
print()