-
Notifications
You must be signed in to change notification settings - Fork 5
/
tuner.py
197 lines (181 loc) · 6.95 KB
/
tuner.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
import math
import os
import sys
import pytorch_lightning as pl
import ray
from pytorch_lightning.loggers import TensorBoardLogger
from ray import tune
from ray.tune import CLIReporter
from ray.tune.integration.pytorch_lightning import TuneReportCheckpointCallback
from ray.tune.schedulers import PopulationBasedTraining
from wasabi import msg
from clr_gat import CLRGAT
from dataloaders.dataloaders import UnlabelledDataModule
def train_gat_clr_tune_checkpoint(config,
checkpoint_dir=None,
num_epochs=50,
num_gpus=0,
data_dir="~/data"):
kwargs = {
"max_epochs": num_epochs,
"limit_train_batches": 100,
"limit_val_batches": 15,
# If fractional GPUs passed in, convert to int.
"gpus": math.ceil(num_gpus),
"logger": TensorBoardLogger(save_dir=tune.get_trial_dir(), name="", version="."),
"enable_progress_bar": False,
"callbacks": [
TuneReportCheckpointCallback(
metrics={
"val/loss": "val/loss",
"val/accuracy": "val/accuracy"
},
filename="checkpoint",
on="train_end")
]
}
if checkpoint_dir:
kwargs["resume_from_checkpoint"] = os.path.join(checkpoint_dir, "checkpoint")
datamodule = UnlabelledDataModule(dataset='tieredimagenet',
datapath=data_dir,
batch_size=64,
num_workers=4,
n_support=1,
n_query=3,
tfm_method="amdim",
no_aug_support=True,
split='train',
img_size_orig=(84, 84),
img_size_crop=(84, 84),
eval_ways=5,
eval_support_shots=5,
eval_query_shots=15)
model = CLRGAT(**config)
trainer = pl.Trainer(**kwargs, check_val_every_n_epoch=5)
trainer.fit(model, datamodule=datamodule)
def tune_gat_clr_pbt(num_samples=50, num_epochs=10, gpus_per_trial=1, data_dir="~/data"):
config = {
"arch": "resnet12",
"out_planes": 64,
"average_end": False,
"distance": tune.choice(["euclidean", "cosine"]),
"gnn_type": "gat",
"optim": "adam",
"lr_sch": tune.choice(["step", "cos"]),
"warmup_start_lr": 1e-3,
"warmup_epochs": 250,
"sup_finetune_lr": tune.loguniform(1e-6, 1e-4),
"sup_finetune": "prototune", # [prototune, std_proto, label_cleansing, sinkhorn]
"sup_finetune_epochs": tune.randint(15, 20),
"eta_min": 5e-5,
"lr": tune.loguniform(1e-6, 1e-3),
"lr_decay_step": 25000,
"lr_decay_rate": 0.5,
"weight_decay": 6.059722614369727e-06,
"dataset": "tieredimagenet",
"img_orig_size": (84, 84),
"batch_size": 64,
"n_support": 1,
"n_query": 3,
"use_projector": False,
"projector_h_dim": 6400,
"projector_out_dim": 1600,
"use_hms": False,
"label_cleansing_opts": {
"use": False,
"n_test_runs": 1000,
"n_ways": 5,
"n_shots": 5,
"n_queries": 15,
"unbalanced": False,
"reduce": None, # ['isomap', 'itsa', 'mds', 'lle', 'se', 'pca', 'none']
"inference_semi": "transductive", # ['transductive', 'inductive', 'inductive_sk']
"d": 5,
"alpha": 0.8,
"K": 20, # neighbours used for manifold creation
"T": 3, # power to raise probs matrix before sinkhorn algorithm
"lr": 0.00001, # learning rate of fine-tuning
"denoising_iterations": 1000,
"beta_pt": 0.5, # power transform power
"best_samples": 3, # number of best samples per class chosen for pseudolabels
"semi_inference_method": "transductive", # ['transductive', 'inductive']
"sinkhorn_iter": 1,
"use_pt": True, # [True, False]
},
"mpnn_loss_fn": "ce",
"mpnn_dev": "cuda",
"mpnn_opts": {
"_use": True,
"loss_cnn": True,
"scaling_ce": tune.loguniform(1e-1, 1e0),
"adapt": "ot",
"temperature": tune.loguniform(1e-1, 1e0),
"output_train_gnn": "plain",
"graph_params": {
"sim_type": "correlation",
"thresh": "no", # 0
"set_negative": "hard"},
"gnn_params": {
"pretrained_path": "no",
"red": 1,
"cat": 0,
"every": 0,
"gnn": {
"num_layers": tune.randint(1, 4),
"aggregator": tune.choice(["add", "max", "mean"]),
"num_heads": tune.choice([1, 2, 4, 8]),
"attention": "dot",
"mlp": 1,
"dropout_mlp": 0.1,
"norm1": 1,
"norm2": 1,
"res1": 1,
"res2": 1,
"dropout_1": 0.1,
"dropout_2": 0.1,
"mult_attr": 0},
"classifier": {
"neck": 0,
"num_classes": 0,
"dropout_p": 0.4,
"use_batchnorm": 0}}
},
"feature_extractor": None
}
scheduler = PopulationBasedTraining(
perturbation_interval=5,
hyperparam_mutations={
"lr": tune.loguniform(1e-4, 1e-2),
})
reporter = CLIReporter(
parameter_columns=["lr", "optim", "lr_sch", "out_planes", ],
metric_columns=["val/loss", "val/accuracy", "training_iteration"]
)
analysis = tune.run(
tune.with_parameters(
train_gat_clr_tune_checkpoint,
num_epochs=num_epochs,
num_gpus=gpus_per_trial,
data_dir=data_dir),
resources_per_trial={
"cpu": 6,
"gpu": gpus_per_trial
},
metric="val/accuracy",
mode="max",
config=config,
num_samples=num_samples,
scheduler=scheduler,
progress_reporter=reporter,
name="tune_gat_clr_pbt",
local_dir="./ray_results/"
)
# msg.info(f"Best hyperparameters found were: {analysis.best_config}")
print("Best hyperparameters found were: ", analysis.best_config)
if __name__ == "__main__":
redis_password = sys.argv[1]
num_cpus = int(sys.argv[2])
with msg.loading("Init Ray"):
ray.init(address=os.environ["ip_head"])
tune_gat_clr_pbt(30, num_epochs=200, gpus_per_trial=1,
data_dir="<data_path>")