def s_rnn_models(self):
gs = GridSeacher({
'mdl_kwargs': {
'C': mbls.SerialRNNClassifier,
'rnn_cell_name': GDIter('GRU', 'LSTM'),
'embd_dims': self.gd_embd_dims,
'layers': self.gd_layers,
'dropout': self.dropout_d,
},
})
self.add_gs(gs)
def s_attn_models_heads(self):
gs = GridSeacher({
'mdl_kwargs': {
'C': mbls.SerialTimeSelfAttn,
'embd_dims': self.gd_embd_dims,
'layers': self.gd_layers,
'dropout': self.dropout_d,
'fourier_dims': GDIter(1),
'te_features': GDIter(2),
'kernel_size': GDIter(1),
'heads': GDIter(8, 16), # ***
'time_noise_window': GDIter('0*24**-1'),
},
})
self.add_gs(gs)
def p_attn_model(self):
gs = GridSeacher({
'mdl_kwargs': {
'C': mbls.ParallelTimeSelfAttn,
'embd_dims': self.gd_embd_dims,
'layers': self.gd_layers,
'dropout': self.dropout_d,
'fourier_dims': GDIter(1),
'te_features': GDIter(16),
'kernel_size': GDIter(1),
'heads': GDIter(4),
'time_noise_window': GDIter('6*24**-1'),
},
})
self.add_gs(gs)
def add_gs(self, gs):
new_mps = []
gs.update({
'dataset_kwargs': {
'in_attrs': ['obs'] + (['obse'] if USES_OBSE_INPUT else []),
'rec_attr': 'obs',
'max_day': self.max_day,
}
})
mps = gs.get_dicts()
for mp in mps:
bd = self.base_dict.copy()
bd.update(mp)
new_mps += GridSeacher(bd).get_dicts()
for mp in new_mps:
for k in self.common_dict.keys():
for k2 in mp.keys():
if k2 != 'dataset_kwargs':
mp[k2][k] = self.common_dict[k]
self.mps += new_mps
kfs = list(range(0, 5)) if main_args.kf == '.' else main_args.kf
kfs = [kfs] if isinstance(kfs, str) else kfs
methods = [
'linear-fstw', 'bspline-fstw', 'spm-mle-fstw', 'spm-mle-estw',
'spm-mcmc-fstw', 'spm-mcmc-estw'
] if main_args.method == '.' else main_args.method
methods = [methods] if isinstance(methods, str) else methods
for kf in kfs:
for method in methods:
grid_params = {
'min_dist': 0.5,
'n_neighbors':
GDIter(*np.linspace(10, 50, 8).astype(np.int)[::-1]),
}
gs = GridSeacher(grid_params)
bar = ProgressBar(len(gs))
for params in gs:
bar(f'method={method} - kf={kf} - params={params}')
lcset_name = f'{kf}@train.{method}'
load_rootdir = f'../save/fats/{cfilename}~method={method}'
map_kwargs = {
'features_mode': main_args.mode,
'proj_mode': 'pca+umap',
'min_dist': params['min_dist'],
'n_neighbors': params['n_neighbors'],
}
maps2d_dict = get_fitted_maps2d(lcdataset, lcset_name,
load_rootdir, **map_kwargs)
class_names = lcdataset['raw'].class_names
for c in [None] + class_names: