forked from karpathy/llama2.c
-
Notifications
You must be signed in to change notification settings - Fork 0
/
metrics.py
304 lines (244 loc) · 8.94 KB
/
metrics.py
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
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
"""
Kurtosis/skew metrics deprecated as of 12/12/2023, and are shifting towards eval mode
compute_metric functions in model.py as the standard metric logging and computation
method. Only function still in use is the flatten_dict() utility function.
"""
from re import search
from collections import defaultdict
from functools import partial
from json import dump
from pathlib import Path
from enum import Enum
from typing import List, Dict, Any, DefaultDict, Tuple, Optional
from torch import Tensor, mean, pow
from torch.nn import Module
import torch
import py
import py._io
import py._io.capture
import functools
class InferenceMetric(str, Enum):
"""Statistical, non-visual telemetry metrics
for model performance, original and quantised."""
# Perplexity of an LM forward pass
# Measures model's confusion given a token sequence
# Lower is (usually) better but can overfit to training data
# HT: https://huggingface.co/docs/transformers/perplexity
PERPLEXITY = "perplexity"
# Perplexity rise after static int8 quantisation, lower is better
# Measures how much the quantisation scheme hurt model prediction confidence
# HT: Michael Xie, NUS
PERPLEXITY_QUANT = "perplexity_diff_quantised"
# Max of a distribution, lower is better
# Linear int4/8 scales the weight distribution's min-max to [-128, 127]
# Positive outliers would inflate max,
INF_NORM = "inf_norm"
# Kurtosis / 4th momentum stat of a distributiom
KURTOSIS = "kurtosis"
# Skewness / 3rd momentum stat of a distribution
SKEWNESS = "skewness"
# Mean of a distribution
MEAN = "mean"
# Standard deviation of a distribution
STDDEV = "standard_deviation"
# Variance of a distribution
VARIANCE = "variance"
# Softmax1 sum over sequence keys, per token pos / head / block.
# If sum < 1, the model learnt to focus attention on the +1 bias ghost token.
# If sum << 1, the model strongly relies on this +1 bias for a behavior.
SOFTMAX_SUM = "softmax_sum"
@torch.no_grad()
def kurtosis(x: Tensor, dim: int = 0) -> Tensor:
"""
Batch kurtosis along batch dim.
kurtosis[x] = E[y^2] / (E[y])^2, where y = (x - E[x])^2
excess_kurtosis = kurtosis - 3
"""
# Flatten input x except batch_size dim
x = x.view(x.shape[dim], -1)
y = pow(x - mean(x, dim=1, keepdim=True), 2)
excess_kurtosis = mean(pow(y, 2), dim=1) / pow(mean(y, dim=1), 2) - 3.0
return excess_kurtosis
# TODO: Add skew to hook
@torch.no_grad()
def skewness(x: Tensor, dim: int = 0) -> Tensor:
"""
Batch skewness along batch dim.
skewness[x] = E[y^3] / (E[y^2])^1.5, where y = (x - E[x])
"""
# Flatten input x except batch_size dim
x = x.view(x.shape[dim], -1)
y = x - mean(x, dim=1, keepdim=True)
skewness = mean(pow(y, 3), dim=1) / pow(mean(pow(y, 2), dim=1), 1.5)
return skewness
@torch.no_grad()
def compute_avg_and_std(array: List[float]) -> Dict[str, float]:
avg = sum(array) / len(array)
second_moment = sum([x**2 for x in array]) / len(array)
std = (second_moment - avg**2) ** 0.5
return {"avg": avg, "std": std}
# I/O
def save_results(results: Dict[str, Dict[str, Any]], output_dir: str):
results_dir = Path.cwd() / "results"
results_dir.mkdir(exist_ok=True)
model_name = output_dir.split("/")[1]
filepath = results_dir / f"{model_name}.json"
with filepath.open(mode="w") as fp:
dump(results, fp, indent=4)
def flatten_dict(input_dict, parent_key="", separator="/"):
flattened_dict = {}
for key, value in input_dict.items():
new_key = f"{parent_key}{separator}{key}" if parent_key else key
if isinstance(value, dict):
flattened_dict.update(flatten_dict(value, new_key, separator))
else:
flattened_dict[new_key] = value
return flattened_dict
FLAGS = {
"inspect_attn_mean": False,
}
# Activation Kurtosis Hook
@torch.no_grad()
def save_activations_kurtosis(
activations: DefaultDict,
a: dict,
name: str,
module: Module,
inp: Tuple,
out: Tensor,
) -> None:
"""
PyTorch Forward hook to compute moving average of kurtosis at each forward pass.
Mutates specified dict objects with each fwd pass.
"""
m = out.shape[0]
if is_query_key := name[-3:] in (".wq", ".wk") and FLAGS["inspect_attn_mean"]:
bsz, seqlen, _ = out.shape
out = out.detach().clone().view(bsz, seqlen, 6, 48) # num_heads, k/q embed_dim
# print(out, out.shape)
mu = out.mean().item()
mx = out.max().item()
mn = out.min().item()
a[name] = [mu, mx, mn, a[name][-1] + 1 if name in a else 1]
num_params = torch.tensor(out.shape[1:]).prod().item() # exclude batch_size
k = kurtosis(out).sum().item()
n, mu, _ = activations[name]
activations[name] = [
n + m,
(n * mu + k) / (n + m),
num_params,
] # TODO: Weight mean by num params
def activation_hooks(
model: Module, layers_to_save: Optional[List[str]] = None
) -> DefaultDict[str, List[float]]:
"""Registers forward hooks in specified layers.
Parameters
----------
model: PyTorch model
layers_to_save: Module names within ``model`` whose activations we want to save.
If None, save all layers
Returns
-------
activations_dict: dict of lists containing activations of specified layers in
``layers_to_save``.
"""
activations_dict = defaultdict(lambda: [0, 0.0, 0])
a = (
{}
) # Raw attention activations: a = Q @ K.T, where attn = softmax(a/sqrt(d_k)) @ v
hooks = []
for name, module in model.named_modules():
if layers_to_save is None or name in layers_to_save:
unwanted_prefix = "_orig_mod."
hooks += [
module.register_forward_hook(
partial(
save_activations_kurtosis,
activations_dict,
a,
name.replace(unwanted_prefix, ""),
)
)
]
def deregister():
for hook in hooks:
hook.remove()
return activations_dict, deregister, a
CAPTURE_PATTERN = r"feed_forward\.w\d+"
def summarise_layers(
activations_dict: DefaultDict[str, float], num_params: List[int]
) -> Dict[str, Dict[str, float]]:
"""
Compute average and std of activations per layer.
"""
layers = defaultdict(list) # aggregate common layers over blocks
for name, kurtosis in activations_dict.items():
if match := search(CAPTURE_PATTERN, name):
layers[match.group()].append(kurtosis)
summary_per_layer = {
name: compute_avg_and_std(kurtoses) for name, kurtoses in layers.items()
}
mean = (
sum(a * b for a, b in zip(activations_dict.values(), num_params))
/ len(activations_dict)
/ sum(num_params)
)
return {"layers": activations_dict, "summary": summary_per_layer, "mean": mean}
# Analysis
def quantisation_metrics(
saved_activation_kurtosis: Dict[str, List[float]], model: Module
) -> dict:
"""
Compute quantisation proxy metrics post-inference.
Currently: weight and activation kurtosis and skew.
"""
activation_kurtosis = {
name: val[1] for name, val in saved_activation_kurtosis.items()
}
activation_params = [val[2] for val in saved_activation_kurtosis.values()]
weight_kurtosis = {
name.replace("_orig_mod.", ""): kurtosis(param.unsqueeze(0).detach()).item()
for name, param in model.named_parameters()
}
weight_params = [param.numel() for param in model.parameters()]
return {
"activations": summarise_layers(activation_kurtosis, activation_params),
"weights": summarise_layers(weight_kurtosis, weight_params),
}
def get_capture(out, in_):
try:
capture = py.io.StdCaptureFD(out=out, in_=in_)
except:
capture = None
return capture
def reset_capture(capture):
if capture is not None:
capture.reset()
def hide_warnings(function=None, out=True, in_=False):
"""Suppresses C++ warnings in PyTorch underlying methods. Decorate on functions"""
def decorator_hide_warnings(func):
@functools.wraps(func)
def wrapper(*args, **kwargs):
capture = get_capture(out, in_)
result = func(*args, **kwargs)
reset_capture(capture)
return result
return wrapper
if function:
return decorator_hide_warnings(function)
return decorator_hide_warnings
class MovingAverage:
"""dynamic moving average: mean = (mean * i + arr[i]) / (i+1).
accurate to 7 d.p. for 100k elements."""
def __init__(self):
self.mean = torch.tensor(0)
self.n = torch.tensor(0)
def update(self, x):
"""Update moving mean with new value x.
Args:
x (float): new value
"""
self.mean = (self.mean * self.n + x) / (self.n + 1)
self.n += 1
def __call__(self):
return self.mean