def search_default(cls) -> "FullyConv1Resnet.ModelParams":
return cls(
num_groups=params.Discrete([8, 16, 32, 64]),
num_blocks_per_residual=params.Integer(1, 5),
activation=params.Categorical([nn.LeakyReLU, nn.GELU]),
do_include_first_norm=params.Boolean(p_true=0.8),
)
class ModelHyperParams(params.ParameterSet):
num_hidden_layers = params.Integer(1, 4)
num_neurons_per_layer = params.Discrete(
np.logspace(2, 7, num=6, base=2, dtype=np.int))
dropout_rate = params.Float(0.0, 0.99)
activation = params.Categorical(['relu', 'sigmoid'])
output_dir = '/tmp/output'
filter_size = 3
class DocGenParams(params.ParameterSet):
margin = '0.5in'
page_size = hc.PageSize.LETTER
dpi = params.Discrete.from_prob_dict({
300: 1,
250: 1,
})
do_randomize_field_order = True
group_offset_in = params.Float(0.0, 3.0)
do_set_invisible_border = params.Boolean(p_true=0.9)
num_extra_fields = params.Discrete.from_prob_dict({
0: 1,
1: 1,
})
font_size_px = params.Integer(8, 19)
val_font_size_px = params.Integer(8, 19)
do_regen_font_val_size = params.Boolean(0.4)
table_cell_padding_px = params.Integer(1, 7)
do_bold_keys = params.Boolean(p_true=0.2)
do_add_colon_to_keys = params.Boolean(p_true=0.2)
vert_alignment = params.Categorical.from_prob_dict({
kv_styles.KvAlign.TT: 2,
kv_styles.KvAlign.BB: 1,
kv_styles.KvAlign.TB: 1,
})
horz_alignment = params.Categorical.from_prob_dict({
kv_styles.KvAlign.LL:
2,
kv_styles.KvAlign.RL:
1,
kv_styles.KvAlign.LR:
1,
kv_styles.KvAlign.CL:
0.2,
kv_styles.KvAlign.LC:
0.2,
})
def get_page_size_px(self):
return [
self.dpi * self.page_size.width, self.dpi * self.page_size.height
]
hp.opt.lr = params.Discrete([1e-4, 1e-3, 1e-2, 1e-1])
hp.opt.min_lr = 1e-6
hp.opt.patience = 16
hp.exp.project_name = 'tablestakes'
hp.exp.experiment_name = 'korv_which'
hp.exp.experiment_tags = [
'korv_which', 'conv', 'sharp', 'search', 'v0.0.1'
]
hp.exp.sources_glob_str = constants.THIS_DIR.parent.parent / '**/*.py'
hp.embed.dim = 16
hp.embed.requires_grad = True
hp.conv.num_features = params.Discrete([32, 64, 128, 256])
hp.conv.num_layers = params.Integer(2, 11)
hp.conv.kernel_size = 3
hp.conv.num_groups = params.Discrete([8, 16, 32, 64])
hp.conv.num_blocks_per_pool = 20
hp.conv.num_blocks_per_skip = 2
hp.conv.requires_grad = True
hp.trans.impl = params.Categorical(['fast', 'fast-favor'])
hp.trans.num_layers = params.Discrete([0, 2, 4, 6, 8])
hp.trans.num_heads = params.Discrete([4, 8, 12])
hp.trans.num_query_features = params.Discrete([0, 32, 64, 128, 256, 512])
hp.trans.fc_dim_mult = params.Discrete([1, 2, 3, 4])
hp.fc.num_features = params.Discrete([32, 64, 128, 256])
hp.fc.num_layers = params.Integer(2, 7)
hp.fc.num_groups = params.Discrete([8, 16, 32, 64])
env['PROJECT_ID'] = GCLOUD_PROJECT_NAME
env['IMAGE_URI'] = CONTAINER_IMAGE_URI
env['BUCKET_NAME'] = GCLOUD_BUCKET_NAME
# Create a Cloud AI Platform Hyperparameter Search object
search = search.HyperparamSearchSpec(
max_trials=10,
max_parallel_trials=5,
max_failed_trials=2,
hyperparameter_metric_tag='val_acc',
)
# Add parameter search ranges for this problem.
my_param_ranges = train.MyParams(
activation=params.Categorical(['relu', 'tanh']),
num_layers=params.Integer(min_value=1, max_value=3),
num_neurons=params.Discrete(np.logspace(2, 8, num=7, base=2)),
dropout_rate=params.Double(min_value=-0.1, max_value=0.9),
learning_rate=params.Discrete(np.logspace(-6, 2, 17, base=10)),
batch_size=params.Integer(min_value=1, max_value=128),
)
search.add_parameters(my_param_ranges)
# Call a bash script to build a docker image for this repo, submit it to the docker registry defined in the script
# and run a training job on the Cloud AI Platform using this container and these hyperparameter ranges.
this_dir = Path(__file__).resolve().parent
retcode = subprocess.call([this_dir / 'build_push.sh'], env=env)
if retcode:
raise ValueError(f"Got returncode: {retcode}")
search.run_from_container(
def search_default(cls) -> "SlabNet.ModelParams":
return cls(
num_features=params.Categorical([16, 32, 64, 128, 256]),
num_layers=params.Integer(1, 5),
**FullyConv1Resnet.ModelParams.search_default().__dict__,
)
hp.exp.project_name = 'tablestakes'
hp.exp.experiment_name = 'korv_which'
hp.exp.experiment_tags = ['korv_which']
hp.exp.sources_glob_str = constants.THIS_DIR.parent.parent / '**/*.py'
hp.exp.offline_mode = False
base_dim = 15
hp.embed.dim = params.Discrete([
base_dim + 16, base_dim + 32, base_dim + 64, base_dim + 128,
base_dim + 256
])
hp.embed.requires_grad = True
hp.embed.position_embedding_requires_grad = False
hp.conv.num_features = params.Discrete([32, 64, 128, 256])
hp.conv.num_layers = params.Integer(min_value=1, max_value=3)
hp.conv.kernel_size = 3
hp.conv.num_groups = params.Discrete([16, 32, 64])
hp.conv.num_blocks_per_pool = 20
hp.conv.requires_grad = True
hp.trans.impl = 'fast-favor'
# neck_hp.trans.impl = 'fast'
hp.trans.num_heads = params.Discrete([2, 4, 8, 16])
hp.trans.num_layers = params.Integer(min_value=1, max_value=3)
hp.trans.num_query_features = None
hp.trans.fc_dim_mult = params.Integer(min_value=2, max_value=5)
hp.trans.p_dropout = params.Float(min_value=0.05, max_value=0.2)
hp.trans.p_attention_dropout = params.Float(min_value=0.05, max_value=0.2)
hp.fc.num_features = params.Discrete([32, 64, 128, 256])