Commit
·
89bc030
1
Parent(s):
cc1972d
adding first pass dataset loading script
Browse files- pythia_training_metrics.py +159 -0
pythia_training_metrics.py
ADDED
@@ -0,0 +1,159 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import datasets
|
2 |
+
import pickle
|
3 |
+
|
4 |
+
|
5 |
+
_DESCRIPTION = """\
|
6 |
+
Dataset for storing training metrics of pythia models
|
7 |
+
"""
|
8 |
+
|
9 |
+
class PythiaTrainingMetrics(datasets.GeneratorBasedBuilder):
|
10 |
+
|
11 |
+
MODEL_SIZES = [
|
12 |
+
"70m",
|
13 |
+
"160m",
|
14 |
+
"410m",
|
15 |
+
"1b",
|
16 |
+
"1.4b",
|
17 |
+
"2.8b",
|
18 |
+
"6.9b"
|
19 |
+
]
|
20 |
+
|
21 |
+
_GRADIENTS_DESCRIPTION = """\
|
22 |
+
Dataset for storing gradients of pythia models
|
23 |
+
"""
|
24 |
+
|
25 |
+
_WEIGHTS_DESCRIPTION = """\
|
26 |
+
Dataset for storing weights of pythia models
|
27 |
+
"""
|
28 |
+
|
29 |
+
_WEIGHTS_MINI_DESCRIPTION = """\
|
30 |
+
Dataset for storing weights of pythia models (minimizes the amount of gradients per
|
31 |
+
checkpoint to only 2)
|
32 |
+
"""
|
33 |
+
|
34 |
+
_ACTIVATIONS_DESCRIPTION = """\
|
35 |
+
Dataset for storing activations of pythia models
|
36 |
+
"""
|
37 |
+
|
38 |
+
BUILDER_CONFIGS = [
|
39 |
+
datasets.BuilderConfig(
|
40 |
+
name="gradients",
|
41 |
+
description=_WEIGHTS_DESCRIPTION,
|
42 |
+
version="1.0.0",
|
43 |
+
),
|
44 |
+
datasets.BuilderConfig(
|
45 |
+
name="gradients_mini",
|
46 |
+
description=_WEIGHTS_MINI_DESCRIPTION,
|
47 |
+
version="1.0.0",
|
48 |
+
),
|
49 |
+
datasets.BuilderConfig(
|
50 |
+
name="activations ",
|
51 |
+
description=_ACTIVATIONS_DESCRIPTION,
|
52 |
+
version="1.0.0",
|
53 |
+
),
|
54 |
+
datasets.BuilderConfig(
|
55 |
+
name="weights",
|
56 |
+
description=_WEIGHTS_DESCRIPTION,
|
57 |
+
version="1.0.0",
|
58 |
+
),
|
59 |
+
datasets.BuilderConfig(
|
60 |
+
name="all",
|
61 |
+
description="All the metrics",
|
62 |
+
version="1.0.0",
|
63 |
+
)
|
64 |
+
]
|
65 |
+
|
66 |
+
def _info(self):
|
67 |
+
"""
|
68 |
+
TODO: Got to figure out how to represent the features etc.
|
69 |
+
|
70 |
+
how do we do this if each feature is dependent on the model size?
|
71 |
+
"""
|
72 |
+
|
73 |
+
features_dict = {
|
74 |
+
"checkpoint_step": datasets.Value('int32'),
|
75 |
+
"layer_name": datasets.Value('string'),
|
76 |
+
}
|
77 |
+
|
78 |
+
if self.config.name in ["activations", "weights"]:
|
79 |
+
features_dict['data'] = datasets.Sequence(datasets.Value('float32'))
|
80 |
+
elif self.config_name in ["gradients", "gradients_mini"]:
|
81 |
+
features_dict['gradient_step'] = datasets.Value('int32')
|
82 |
+
features_dict['gradient'] = datasets.Sequence(datasets.Value('float32'))
|
83 |
+
|
84 |
+
features = datasets.Features(features_dict)
|
85 |
+
|
86 |
+
return datasets.DatasetInfo(
|
87 |
+
description=_DESCRIPTION,
|
88 |
+
features=features,
|
89 |
+
)
|
90 |
+
|
91 |
+
|
92 |
+
def _split_generators(self, dl_manager: datasets.DownloadManager):
|
93 |
+
"""
|
94 |
+
Returns data for different splits - we define a split as a model size.
|
95 |
+
"""
|
96 |
+
|
97 |
+
model_size_to_fp = { model_size: [] for model_size in self.MODEL_SIZES }
|
98 |
+
|
99 |
+
checkpoint_steps = [0, 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1000, ]
|
100 |
+
checkpoint_steps.extend([3000 + (i * 10000) for i in range(0, 15)])
|
101 |
+
|
102 |
+
def get_gradient_step(step: int):
|
103 |
+
"""
|
104 |
+
Return a list of the gradient steps that are used at a given checkpoint step.
|
105 |
+
"""
|
106 |
+
return list(range(max(0, step-5), min(step+6, 143_000)))
|
107 |
+
|
108 |
+
for model_size in self.MODEL_SIZES:
|
109 |
+
for checkpoint_step in checkpoint_steps:
|
110 |
+
|
111 |
+
directory_path = f"./models/{model_size}/checkpoint_{checkpoint_step}"
|
112 |
+
|
113 |
+
if self.config.name == "activations":
|
114 |
+
model_size_to_fp[model_size].append(f"{directory_path}/checkpoint_activations.pickle")
|
115 |
+
elif self.config_name == "weights":
|
116 |
+
model_size_to_fp[model_size].append(f"{directory_path}/checkpoint_weights.pickle")
|
117 |
+
elif self.config_name == "gradients":
|
118 |
+
for gradient_step in get_gradient_step(checkpoint_step):
|
119 |
+
model_size_to_fp[model_size].append(f"{directory_path}/checkpoint_gradients_{gradient_step}.pickle")
|
120 |
+
elif self.config_name == "gradients_mini":
|
121 |
+
for gradient_step in get_gradient_step(checkpoint_step)[:2]:
|
122 |
+
model_size_to_fp[model_size].append(f"{directory_path}/checkpoint_gradients_mini_{gradient_step}.pickle")
|
123 |
+
|
124 |
+
downloaded_files = dl_manager.download_and_extract(model_size_to_fp)
|
125 |
+
|
126 |
+
return [
|
127 |
+
datasets.SplitGenerator(
|
128 |
+
name=datasets.Split.TRAIN,
|
129 |
+
gen_kwargs={
|
130 |
+
"filepaths": downloaded_fps
|
131 |
+
}
|
132 |
+
) for downloaded_fps in downloaded_files.values()
|
133 |
+
]
|
134 |
+
|
135 |
+
def _generate_examples(self, filepaths):
|
136 |
+
|
137 |
+
# the filepaths should be a list of filepaths
|
138 |
+
if isinstance(filepaths, str):
|
139 |
+
filepaths = [filepaths]
|
140 |
+
|
141 |
+
global_idx = 0 # the unique identifier for the example
|
142 |
+
|
143 |
+
for filepath in filepaths:
|
144 |
+
with open(filepath, encoding="utf-8") as f:
|
145 |
+
data = pickle.load(f)
|
146 |
+
|
147 |
+
# extract checkpoint step from the filepath
|
148 |
+
checkpoint_step = int(filepath.split("/")[1].split("_")[-1])
|
149 |
+
|
150 |
+
if self.config.name in ["activations", "weights"]:
|
151 |
+
for layer_name, layer_data in data.items():
|
152 |
+
for data in layer_data:
|
153 |
+
yield global_idx, {"checkpoint_step": checkpoint_step, "layer_name": layer_name, "data": data}
|
154 |
+
global_idx += 1
|
155 |
+
elif self.config.name in ["gradients", "gradients_mini"]:
|
156 |
+
for layer_name, layer_data in data.items():
|
157 |
+
for gradient_step, gradient in layer_data.items():
|
158 |
+
yield global_idx, {"checkpoint_step": checkpoint_step, "layer_name": layer_name, "gradient_step": gradient_step, "gradient": gradient}
|
159 |
+
global_idx += 1
|