def setup_model(self):
self.model = model_tools.get_model(self.cfg.model)
model_parameters = filter(lambda p: p.requires_grad,
self.model.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
self.logger.info('number of trainable parameters : %d' % params)
def setup_model(self):
self.model = model_tools.get_model(self.cfg.model)
model_parameters = filter(lambda p: p.requires_grad,
self.model.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
self.logger.info('number of trainable parameters : %d' % params)
self.encode_layer = NNEncLayer()
self.boost_layer = PriorBoostLayer()
self.nongray_mask = NonGrayMaskLayer()
def __init__(self,
base_ff=vgg_example,
hgru_params=hgru_params_example,
timesteps=6):
super().__init__()
self.hgru_params = hgru_params
self.timesteps = timesteps
self.base_ff = model_tools.get_model(base_ff)
self.build_fb_layers()
self.eps = 0.01
def __init__(
self,
base_ff=vgg_example,
gn_params=gn_params_example,
timesteps=6,
hidden_init='identity',
attention='gala', # 'se', None
attention_layers=2,
saliency_filter_size=5,
norm_attention=False,
normalization_fgru='InstanceNorm2d',
normalization_fgru_params={'affine': True},
normalization_gate='InstanceNorm2d',
normalization_gate_params={'affine': True},
force_alpha_divisive=True,
force_non_negativity=True,
multiplicative_excitation=True,
ff_non_linearity='ReLU',
us_resize_before_block=True,
skip_horizontal=True,
readout=True,
readout_feats=1):
super().__init__()
self.timesteps = timesteps
self.gn_params = gn_params
assert len(gn_params) % 2 == 1, 'the number of fgrus is not odd'
self.gn_layers = len(gn_params) // 2 + 1
self.ff_non_linearity = ff_non_linearity
self.normalization_fgru = normalization_fgru
self.us_resize_before_block = us_resize_before_block
self.normalization_fgru_params = normalization_fgru_params
self.fgru_params = {
'hidden_init': hidden_init,
'attention': attention,
'attention_layers': attention_layers,
'saliency_filter_size': saliency_filter_size,
'norm_attention': norm_attention,
'normalization_fgru': normalization_fgru,
'normalization_fgru_params': normalization_fgru_params,
'normalization_gate': normalization_gate,
'normalization_gate_params': normalization_gate_params,
'ff_non_linearity': ff_non_linearity,
'force_alpha_divisive': force_alpha_divisive,
'force_non_negativity': force_non_negativity,
'multiplicative_excitation': multiplicative_excitation,
'timesteps': timesteps
}
self.base_ff = model_tools.get_model(base_ff)
#VGG_16(weights_path=weights_path, load_weights=load_weights)
self.build_fb_layers()
self.skip_horizontal = skip_horizontal
self.use_readout = False
if readout:
self.use_readout = True
self.build_readout(readout_feats)