Repository navigation
Expand file tree
/
Copy pathrun_relearn_arithmetic.py
More file actions
282 lines (254 loc) · 15.1 KB
/
Copy pathrun_relearn_arithmetic.py
File metadata and controls
282 lines (254 loc) · 15.1 KB
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
"""
Configured for a (any)xA100 GPU server.
Launch Command: python run_relearn_arithmetic.py
"""
from src.tools.relearn_langarith import relearn
from src.utils.paths import CACHE_DIR, DATASET_DIR, MODEL_DIR, WANDB_API_KEY_PATH
from accelerate import Accelerator
from utils.loss_functions import print_acc
from utils.validation_functions import get_arithmetic_eval_fn
from utils.parallel_launch import launch_in_parallel_one_per_gpu, get_parallel_launch_wrapper
SETUPS_TO_RUN = [
"gemma-2-0.3B_all_data" # "gemma-2-0.3B_all_data_perturb", "gemma-2-0.3B_train_only_forget", "gemma-2-0.3B_all_data"
]
MODELS_TO_RUN = [
#'pretrained_models/gemma-2-0.3B_addition_subtraction+eng', # Pretrain Pure
#'pretrained_models/gemma-2-0.3B_all_arithmetic+eng', # Pretrained Base
#'unlearned_models/GradDiff/gemma-2-0.3B_all_arithmetic+eng_lr_8.0e-06', # Unlearned GradDiff
#'unlearned_models/RMU/gemma-2-0.3B_all_arithmetic+eng_lr_6.0e-06', # Unlearned RMU
#'unlearned_models/MaxEnt/gemma-2-0.3B_all_arithmetic+eng_lr_9.0e-05', # Unlearned MaxEnt
#'distilled_models/pure/gemma-2-0.3B_all_arithmetic+eng', # Distilled Pure (Oracle)
#'distilled_models/pure_from_base/gemma-2-0.3B_all_arithmetic+eng', # Distilled Impure (Oracle)
#'distilled_models/base/gemma-2-0.3B_all_arithmetic+eng', # Distilled Base
#'distilled_models/GradDiff/gemma-2-0.3B_all_arithmetic+eng', # Distilled GradDiff
#'distilled_models/RMU/gemma-2-0.3B_all_arithmetic+eng', # Distilled RMU
# 'distilled_models/MaxEnt/gemma-2-0.3B_all_arithmetic+eng', # Distilled MaxEnt
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.1-beta_0.1-seed_123', # partial_distill
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.2-beta_0.1-seed_123', # partial_distill
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.3-beta_0.1-seed_123', # partial_distill
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.4-beta_0.1-seed_123', # partial_distill
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.5-beta_0.1-seed_123', # partial_distill
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.6-beta_0.1-seed_123', # partial_distill
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.7-beta_0.1-seed_123', # partial_distill
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.8-beta_0.1-seed_123', # partial_distill
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.1-beta_0.1-seed_111-fixed_steps_500', # partial_distill
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.2-beta_0.1-seed_111-fixed_steps_500', # partial_distill
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.3-beta_0.1-seed_111-fixed_steps_500', # partial_distill
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.4-beta_0.1-seed_111-fixed_steps_500', # partial_distill
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.5-beta_0.1-seed_111-fixed_steps_500', # partial_distill
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.6-beta_0.1-seed_111-fixed_steps_500', # partial_distill
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.7-beta_0.1-seed_111-fixed_steps_500', # partial_distill
#'partial_distill_models_arith/gemma-2-0.3B_MaxEnt-arithmetic-partial_distill-alpha_0.8-beta_0.1-seed_111-fixed_steps_500', # partial_distill
#'unlearned_models/MaxEnt_RepNoise/gemma-2-0.3B_all_arithmetic+eng_a0.1_b0.001_lr_8.0e-05',
#'unlearned_models/MaxEnt_RepNoise/gemma-2-0.3B_all_arithmetic+eng_a0.1_b0.0001_lr_8.0e-05',
#'unlearned_models/MaxEnt_RepNoise/gemma-2-0.3B_all_arithmetic+eng_a0.5_b0.1_lr_8.0e-05',
#'unlearned_models/MaxEnt_RepNoise/gemma-2-0.3B_all_arithmetic+eng_a0.5_b0.01_lr_8.0e-05',
#'unlearned_models/MaxEnt_RepNoise/gemma-2-0.3B_all_arithmetic+eng_a1.0_b0.1_lr_8.0e-05',
#'unlearned_models/MaxEnt_RepNoise/gemma-2-0.3B_all_arithmetic+eng_a1.0_b1.0_lr_8.0e-05',
#'unlearned_models/MaxEnt_RepNoise/gemma-2-0.3B_all_arithmetic+eng_a2.0_b2.0_lr_8.0e-05',
#'unlearned_models/MaxEnt_RepNoise/gemma-2-0.3B_all_arithmetic+eng_a4.0_b4.0_lr_8.0e-05',
'unlearned_models/RMU_SAM/gemma-2-0.3B_all_arithmetic+eng_rho0.01_lr_1.0e-05',
'unlearned_models/RMU_SAM/gemma-2-0.3B_all_arithmetic+eng_rho0.01_lr_2.0e-05',
'unlearned_models/RMU_SAM/gemma-2-0.3B_all_arithmetic+eng_rho0.01_lr_3.0e-05',
'unlearned_models/RMU_SAM/gemma-2-0.3B_all_arithmetic+eng_rho0.01_lr_4.0e-05',
'unlearned_models/RMU_SAM/gemma-2-0.3B_all_arithmetic+eng_rho0.01_lr_6.0e-06',
'unlearned_models/RMU_SAM/gemma-2-0.3B_all_arithmetic+eng_rho0.01_lr_7.0e-06',
'unlearned_models/RMU_SAM/gemma-2-0.3B_all_arithmetic+eng_rho0.01_lr_8.0e-06',
'unlearned_models/RMU_SAM/gemma-2-0.3B_all_arithmetic+eng_rho0.01_lr_9.0e-06',
]
'''
for pareto plot
'''
#LR_RANGES = {
# "GradDiff": [6e-6, 7e-6, 8e-6, 9e-6, 1e-5, 2e-5, 3e-5, 4e-5],
# "MaxEnt": [6e-5, 7e-5, 8e-5, 9e-5, 1e-4, 2e-4, 3e-4, 4e-4],
# "RMU": [6e-6, 7e-6, 8e-6, 9e-6, 1e-5, 2e-5, 3e-5, 4e-5]
#}
#for method, lr_values in LR_RANGES.items():
# for lr in lr_values:
# # Format the learning rate with scientific notation
# formatted_lr = f"{lr:.1e}"
# # Add the model path to the list
# model_path = f"unlearned_models/{method}/gemma-2-0.3B_all_arithmetic+eng_lr_{formatted_lr}"
# MODELS_TO_RUN.append(model_path)
#
#for model in MODELS_TO_RUN:
# print(model)
# relearning learning rate
LRS_TO_RUN =[1e-3, 7e-4, 4e-4, 1e-4, 7e-5, 5e-5, 1e-5, 7e-6, 4e-6, 1e-6]
try:
with open(WANDB_API_KEY_PATH, "r", encoding="utf-8") as f:
api_key = f.read().strip()
except Exception as e:
print(f"[ERROR] Unable to read WandB API key from {WANDB_API_KEY_PATH}. Exception: {e}")
exit(1)
setups = {
"gemma-2-0.3B_train_only_forget": {
'model_name' : f"{MODEL_DIR}/model_path/final_student",
'first_train_file' : f"{DATASET_DIR}/pretrain/train_addition_subtraction.jsonl",
'second_train_file': "",
'interleave_probs': [1],
'eng_valid_file' : f"{DATASET_DIR}/pretrain/valid_eng.jsonl",
'output_dir' : f"{MODEL_DIR}/relearned_models/model_name_only_forget_data",
'cache_dir' : CACHE_DIR,
'dataset_cache_dir': CACHE_DIR,
'join_or_subsequence' : True,
'seed' : 42,
'device' : "cuda",
'batch_size' : 15,
'gradient_accumulation_steps' : 1,
'epochs' : 1,
'learning_rate' : None,
'max_steps' : 500,
'num_warmup_steps' : 1,
'validation_steps' : 5,
'save_checkpoint_steps' : 999,
'scheduler_type' : "cosine",
'min_lr' : None,
'weight_decay' : 0.0,
'gradient_clipping_threshold' : 1.0,
'max_length' : 256,
'use_wandb' : True,
'wandb_project' : "gemma-2-0.3B_relearn_only_forget",
'wandb_run_name' : None,
'use_local_record' : True,
'path_local_record': f"{MODEL_DIR}/local_records/relearned_models/model_name_only_forget_data.txt",
'save_models' : False,
'shrink_perturb_relearning' : 0
},
"gemma-2-0.3B_all_data": {
'model_name' : f"{MODEL_DIR}/model_path/final_student",
'first_train_file' : f"{DATASET_DIR}/pretrain/train_eng.jsonl",
'second_train_file': f"{DATASET_DIR}/pretrain/train_all_arithmetic.jsonl",
'interleave_probs' : [.75, .25],
'eng_valid_file' : f"{DATASET_DIR}/pretrain/valid_eng.jsonl",
'output_dir' : f"{MODEL_DIR}/relearned_models/model_name_all_data",
'cache_dir' : CACHE_DIR,
'dataset_cache_dir': CACHE_DIR,
'join_or_subsequence' : True,
'seed' : 42,
'device' : "cuda",
'batch_size' : 15,
'gradient_accumulation_steps' : 1,
'epochs' : 1,
'learning_rate' : None,
'max_steps' : 500,
'num_warmup_steps' : 1,
'validation_steps' : 5,
'save_checkpoint_steps' : 999,
'scheduler_type' : "cosine",
'min_lr' : None,
'weight_decay' : 0.0,
'gradient_clipping_threshold' : 1.0,
'max_length' : 256,
'use_wandb' : True,
'wandb_project' : "gemma-2-0.3B_relearn_all_data",
'wandb_run_name' : None,
'use_local_record' : True,
'path_local_record': f"{MODEL_DIR}/local_records/relearned_models/model_name_all_data.txt",
'save_models' : False,
'shrink_perturb_relearning' : 0
},
"gemma-2-0.3B_all_data_perturb": {
'model_name' : f"{MODEL_DIR}/model_path/final_student",
'first_train_file' : f"{DATASET_DIR}/pretrain/train_eng.jsonl",
'second_train_file': f"{DATASET_DIR}/pretrain/train_all_arithmetic.jsonl",
'interleave_probs' : [.75, .25],
'eng_valid_file' : f"{DATASET_DIR}/pretrain/valid_eng.jsonl",
'output_dir' : f"{MODEL_DIR}/relearned_models/model_name_all_data_perturb",
'cache_dir' : CACHE_DIR,
'dataset_cache_dir': CACHE_DIR,
'join_or_subsequence' : True,
'seed' : 42,
'device' : "cuda",
'batch_size' : 15,
'gradient_accumulation_steps' : 1,
'epochs' : 1,
'learning_rate' : None,
'max_steps' : 500,
'num_warmup_steps' : 1,
'validation_steps' : 5,
'save_checkpoint_steps' : 999,
'scheduler_type' : "cosine",
'min_lr' : None,
'weight_decay' : 0.0,
'gradient_clipping_threshold' : 1.0,
'max_length' : 256,
'use_wandb' : True,
'wandb_project' : "gemma-2-0.3B_relearn_all_data_perturb",
'wandb_run_name' : None,
'use_local_record' : True,
'path_local_record': f"{MODEL_DIR}/local_records/relearned_models/model_name_all_data_perturb.txt",
'save_models' : False,
'shrink_perturb_relearning' : 0.1
},
}
def launch_relearn(setup_id, lr, model):
current_setup = setups[setup_id].copy()
current_setup['learning_rate'] = lr
current_setup['min_lr'] = lr
current_setup['model_name'] = current_setup['model_name'].replace('model_path', model)
if "distilled" not in current_setup['model_name']:
current_setup['model_name'] = current_setup['model_name'].replace('final_student', "final_model")
model_name = model_name = f"relearned_{model.replace('/', '_')}/{lr}" # Get the substring past the first '/'
current_setup['path_local_record'] = current_setup['path_local_record'].replace('model_name', model_name)
current_setup['output_dir'] = current_setup['output_dir'].replace('model_name', model_name)
accelerator = Accelerator()
arithmetic_eval_fn = get_arithmetic_eval_fn(
# gets a function that takes a model returns a dicitonary with equation/word problem accuracty for each operation and english validation CE loss
model_name = current_setup['model_name'],
eng_valid_file = current_setup['eng_valid_file'],
batch_size = current_setup['batch_size'],
max_length = current_setup['max_length'],
cache_dir = current_setup['cache_dir'],
dataset_cache_dir = current_setup['dataset_cache_dir'],
num_wiki_batches = 50,
accelerator = accelerator
)
if not current_setup['second_train_file'] == "":
train_files = [current_setup['first_train_file'], current_setup['second_train_file']]
else:
train_files = [current_setup['first_train_file']]
relearn(
model_name = current_setup['model_name'],
train_files = train_files,
interleave_probs = current_setup['interleave_probs'],
output_dir = current_setup['output_dir'],
cache_dir = current_setup['cache_dir'],
dataset_cache_dir= current_setup['dataset_cache_dir'],
eval_fn = arithmetic_eval_fn,
accelerator = accelerator,
join_or_subsequence = current_setup['join_or_subsequence'],
seed = current_setup['seed'],
device = current_setup['device'],
batch_size = current_setup['batch_size'],
gradient_accumulation_steps = current_setup['gradient_accumulation_steps'],
epochs = current_setup['epochs'],
learning_rate = current_setup['learning_rate'],
max_steps = current_setup['max_steps'],
num_warmup_steps = current_setup['num_warmup_steps'],
validation_steps = current_setup['validation_steps'],
save_checkpoint_steps = current_setup['save_checkpoint_steps'],
scheduler_type = current_setup['scheduler_type'],
min_lr = current_setup['min_lr'],
weight_decay = current_setup['weight_decay'],
gradient_clipping_threshold = current_setup['gradient_clipping_threshold'],
max_length = current_setup['max_length'],
use_wandb = current_setup['use_wandb'],
wandb_project = current_setup['wandb_project'],
wandb_run_name = current_setup['wandb_run_name'],
use_local_record = current_setup['use_local_record'],
path_local_record= current_setup['path_local_record'],
save_models = current_setup['save_models'],
shrink_perturb_relearning = current_setup['shrink_perturb_relearning']
)
if __name__ == "__main__":
# ----------------------------------------------------------------- #
# Run all experiments, if possible in parallel
# ----------------------------------------------------------------- #
# Create list of the setups (arguments for run_experiment) for all the experiments we want to run
experiments = [(setup_id, lr, model) for setup_id in SETUPS_TO_RUN for lr in LRS_TO_RUN for model in MODELS_TO_RUN]
# Gets a wrapper function compatable with the parallel launch function
parallel_fn = get_parallel_launch_wrapper(launch_relearn)
# calls run_experiment in parallel on a separate gpu for each experiment setup when a gpu is free
launch_in_parallel_one_per_gpu(experiment_list=experiments, experiment_fn=parallel_fn)