rdiehlmartinez commited on
Commit
2178182
·
1 Parent(s): a378726

tested working version of weight subconfig

Browse files
Files changed (1) hide show
  1. pythia-training-metrics.py +12 -37
pythia-training-metrics.py CHANGED
@@ -1,7 +1,6 @@
1
  import datasets
2
  import pickle
3
 
4
-
5
  _DESCRIPTION = """\
6
  Dataset for storing training metrics of pythia models
7
  """
@@ -12,10 +11,8 @@ class PythiaTrainingMetrics(datasets.GeneratorBasedBuilder):
12
  "70m",
13
  "160m",
14
  "410m",
15
- "1b",
16
  "1.4b",
17
  "2.8b",
18
- "6.9b"
19
  ]
20
 
21
  _GRADIENTS_DESCRIPTION = """\
@@ -56,36 +53,16 @@ class PythiaTrainingMetrics(datasets.GeneratorBasedBuilder):
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
 
@@ -112,12 +89,12 @@ class PythiaTrainingMetrics(datasets.GeneratorBasedBuilder):
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
  else:
@@ -134,29 +111,27 @@ class PythiaTrainingMetrics(datasets.GeneratorBasedBuilder):
134
  ) for model_size_name, downloaded_fps in downloaded_files.items()
135
  ]
136
 
137
- def _generate_examples(self, filepaths, **kwargs):
138
 
139
  # the filepaths should be a list of filepaths
140
  if isinstance(filepaths, str):
141
  filepaths = [filepaths]
142
-
143
  global_idx = 0 # the unique identifier for the example
144
 
145
  for filepath in filepaths:
146
- with open(filepath, encoding="utf-8") as f:
147
  data = pickle.load(f)
148
 
149
  # extract checkpoint step from the filepath
150
- checkpoint_step = int(filepath.split("/")[1].split("_")[-1])
151
 
152
  if self.config.name in ["activations", "weights"]:
153
  for layer_name, layer_data in data.items():
154
- yield global_idx, {"checkpoint_step": checkpoint_step, "layer_name": layer_name, "data": data}
155
  global_idx += 1
156
  elif self.config.name in ["gradients", "gradients_mini"]:
157
-
158
  gradient_step = int(filepath.split('/')[-1].split("_")[-1].split(".")[0])
159
-
160
  for layer_name, layer_data in data.items():
161
- yield global_idx, {"checkpoint_step": checkpoint_step, "layer_name": layer_name, "gradient_step": gradient_step, "gradient": layer_data}
162
  global_idx += 1
 
1
  import datasets
2
  import pickle
3
 
 
4
  _DESCRIPTION = """\
5
  Dataset for storing training metrics of pythia models
6
  """
 
11
  "70m",
12
  "160m",
13
  "410m",
 
14
  "1.4b",
15
  "2.8b",
 
16
  ]
17
 
18
  _GRADIENTS_DESCRIPTION = """\
 
53
  description=_WEIGHTS_DESCRIPTION,
54
  version="1.0.0",
55
  ),
56
+ ]
 
 
 
 
 
57
 
58
  def _info(self):
59
  """
60
+ NOTE: we might want to specify features, but since the featuers are different for each
61
+ model size it's annoying and kind of pointless since hf does it automatically
 
62
  """
63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
64
  return datasets.DatasetInfo(
65
  description=_DESCRIPTION,
 
66
  )
67
 
68
 
 
89
 
90
  if self.config.name == "activations":
91
  model_size_to_fp[model_size].append(f"{directory_path}/checkpoint_activations.pickle")
92
+ elif self.config.name == "weights":
93
  model_size_to_fp[model_size].append(f"{directory_path}/checkpoint_weights.pickle")
94
+ elif self.config.name == "gradients":
95
  for gradient_step in get_gradient_step(checkpoint_step):
96
  model_size_to_fp[model_size].append(f"{directory_path}/checkpoint_gradients_{gradient_step}.pickle")
97
+ elif self.config.name == "gradients_mini":
98
  for gradient_step in get_gradient_step(checkpoint_step)[:2]:
99
  model_size_to_fp[model_size].append(f"{directory_path}/checkpoint_gradients_mini_{gradient_step}.pickle")
100
  else:
 
111
  ) for model_size_name, downloaded_fps in downloaded_files.items()
112
  ]
113
 
114
+ def _generate_examples(self, filepaths):
115
 
116
  # the filepaths should be a list of filepaths
117
  if isinstance(filepaths, str):
118
  filepaths = [filepaths]
119
+
120
  global_idx = 0 # the unique identifier for the example
121
 
122
  for filepath in filepaths:
123
+ with open(filepath, 'rb') as f:
124
  data = pickle.load(f)
125
 
126
  # extract checkpoint step from the filepath
127
+ checkpoint_step = int(filepath.split("/")[-2].split("_")[-1])
128
 
129
  if self.config.name in ["activations", "weights"]:
130
  for layer_name, layer_data in data.items():
131
+ yield global_idx, {"checkpoint_step": checkpoint_step, "layer_name": layer_name, "data": layer_data}
132
  global_idx += 1
133
  elif self.config.name in ["gradients", "gradients_mini"]:
 
134
  gradient_step = int(filepath.split('/')[-1].split("_")[-1].split(".")[0])
 
135
  for layer_name, layer_data in data.items():
136
+ yield global_idx, {"checkpoint_step": checkpoint_step, "layer_name": layer_name, "gradient_step": gradient_step, "data": layer_data}
137
  global_idx += 1