def _create_and_connect_agent(self, input_image: MemoryBlock,
input_size: Tuple[int, int, int]):
params = MultipleLayersParams()
params.num_conv_layers = 4
params.n_flocks = [5, 5, 1, 1]
params.n_cluster_centers = [30, 60, 60, 9]
params.compute_reconstruction = True
params.conv_classes = SpConvLayer
params.sp_buffer_size = 5000
params.sp_batch_size = 500
params.learning_rate = 0.1
params.cluster_boost_threshold = 1000
params.max_encountered_seqs = 1000
params.max_frequent_seqs = 500
params.seq_lookahead = 2
params.seq_length = 4
params.exploration_probability = 0
params.rf_size = (2, 2)
params.rf_stride = None
ta_group = NCMGroup(conv_layers_params=params,
model_seed=None,
image_size=input_size)
self.add_node(ta_group)
Connector.connect(input_image, ta_group.inputs.image)
def _create_and_connect_agent(self, input_image: MemoryBlock,
output_reconstruction: InputSlot,
input_size: Tuple[int, int, int]):
params = MultipleLayersParams()
params.num_conv_layers = 3
params.n_cluster_centers = [28, 14, 7]
params.compute_reconstruction = True
params.conv_classes = ConvLayer
params.sp_buffer_size = 5000
params.sp_batch_size = 500
params.learning_rate = 0.2
params.cluster_boost_threshold = 1000
params.max_encountered_seqs = 1000
params.max_frequent_seqs = 500
params.seq_lookahead = 2
params.seq_length = 4
params.exploration_probability = 0
params.rf_size = [(input_size[0], input_size[1]), (1, 1), (1, 1)]
params.rf_stride = None
ta_group = R1NCMGroup(conv_layers_params=params,
model_seed=None,
image_size=input_size)
self.add_node(ta_group)
Connector.connect(input_image, ta_group.inputs.image)
Connector.connect(ta_group.outputs.predicted_reconstructed_input,
output_reconstruction)
def run_debug_base(args, num_conv_layers: int, exp_params):
name = "Learning-rate-debug_ncl_"+str(num_conv_layers)
cp = MultipleLayersParams()
cp.compute_reconstruction = False
cp.conv_classes = ConvLayer
cp.sp_buffer_size = 6000
cp.sp_batch_size = 4000
cp.learning_rate = 0.10005
cp.cluster_boost_threshold = 1000
cp.max_encountered_seqs = 2000
cp.max_frequent_seqs = 1000
cp.seq_lookahead = 1
cp.seq_length = 5
if num_conv_layers == 2:
cp.n_cluster_centers = [100, 230]
cp.rf_size = [(8, 8), (4, 4)]
cp.rf_stride = [(8, 8), (1, 1)]
cp.num_conv_layers = 2
else:
cp.n_cluster_centers = 200
cp.rf_size = (8, 8)
cp.rf_stride = (8, 8)
cp.num_conv_layers = 1
tp = MultipleLayersParams()
tp.n_cluster_centers = 100
tp.sp_buffer_size = 3000
tp.sp_batch_size = 2000
tp.learning_rate = 0.01
tp.cluster_boost_threshold = 1000
tp.compute_reconstruction = True
cf_easy = [1, 2, 3, 4]
size = SeDatasetSize.SIZE_64
size_int = (size.value, size.value, 3)
params = [
{
'se_group': {'class_filter': cf_easy,
'image_size': size},
'model': {'conv_layers_params': cp,
'top_layer_params': tp,
'image_size': size_int}
}
]
exp_params.experiment_params.num_layers = num_conv_layers + 1 # needs to be there
run_measurement(name, params, args, exp_params)
def make_test_params():
cp = MultipleLayersParams()
cp.compute_reconstruction = True
cp.conv_classes = ConvLayer
cp.sp_buffer_size = 6000
cp.sp_batch_size = 4000
cp.learning_rate = 0.02
cp.cluster_boost_threshold = 1000
cp.max_encountered_seqs = 10000
cp.max_frequent_seqs = 1000
cp.seq_length = 4
cp.seq_lookahead = 1
cp.num_conv_layers = 1
cp.n_cluster_centers = 400
cp.rf_size = (8, 8)
cp.rf_stride = (8, 8)
tp = MultipleLayersParams()
tp.n_cluster_centers = 250
tp.sp_buffer_size = 4000
tp.sp_batch_size = 500
tp.learning_rate = 0.2
tp.cluster_boost_threshold = 1000
tp.compute_reconstruction = True
common_params = {
'conv_layers_params': cp,
'top_layer_params': tp,
'image_size': SeDatasetSize.SIZE_64,
'class_filter': [1, 2, 3],
'noise_amp': 0.0,
'model_seed': None,
'baseline_seed': None
}
changing_params = [{
'conv_layers_params': cp.change(learning_rate=0.1)
}, {
'conv_layers_params': cp.change(learning_rate=0.2),
'top_layer_params': tp.change(learning_rate=0.1)
}, {
'conv_layers_params': cp.change(learning_rate=0.3)
}]
params = ExperimentTemplateBase.add_common_params(changing_params,
common_params)
return params, common_params, changing_params
def run_two_layer_net_new(args, debug: bool = False):
"""An example of the new experiment configuration"""
name = "two_layer_new"
two = [2, 15]
four = [2, 15, 4, 7]
six = [2, 15, 4, 7, 10, 19]
cp = MultipleLayersParams()
cp.compute_reconstruction = True
cp.conv_classes = ConvLayer
cp.sp_buffer_size = 6000
cp.sp_batch_size = 4000
cp.learning_rate = 0.02
cp.cluster_boost_threshold = 1000
cp.max_encountered_seqs = 10000
cp.max_frequent_seqs = 1000
cp.seq_length = 4
cp.seq_lookahead = 1
cp.num_conv_layers = 1
cp.n_cluster_centers = 400
cp.rf_size = (8, 8)
cp.rf_stride = (8, 8)
tp = MultipleLayersParams()
tp.n_cluster_centers = 250
tp.sp_buffer_size = 4000
tp.sp_batch_size = 500
tp.learning_rate = 0.2
tp.cluster_boost_threshold = 1000
tp.compute_reconstruction = True
params = [{
'conv_layers_params': cp,
'top_layer_params': tp,
'image_size': SeDatasetSize.SIZE_64,
'class_filter': six,
'noise_amp': 0.0,
'model_seed': None,
'baseline_seed': None
}]
run_measurement(name, params, args, debug=debug, num_layers=2)
def good_one_layer_config_for_four_objects() -> List[Dict[str, Any]]:
"""A topology which might achieve 100% SE accuracy on 4 objects"""
cp = MultipleLayersParams()
cp.compute_reconstruction = False
cp.conv_classes = ConvLayer
cp.sp_buffer_size = 6000
cp.sp_batch_size = 4000
cp.learning_rate = 0.1
cp.cluster_boost_threshold = 1000
cp.max_encountered_seqs = 2000
cp.max_frequent_seqs = 1000
cp.seq_lookahead = 1
cp.seq_length = 5
cp.n_cluster_centers = 200
cp.rf_size = (8, 8)
cp.rf_stride = (8, 8)
cp.num_conv_layers = 1
tp = MultipleLayersParams()
tp.n_cluster_centers = 150
tp.sp_buffer_size = 3000
tp.sp_batch_size = 2000
tp.learning_rate = 0.15
tp.cluster_boost_threshold = 1000
tp.compute_reconstruction = True
cf_easy = [1, 2, 3, 4]
size = SeDatasetSize.SIZE_64
size_int = (size.value, size.value, 3)
params = [
{
'se_group': {'class_filter': cf_easy,
'image_size': size},
'model': {'conv_layers_params': cp.change(opp=0.5),
'top_layer_params': tp,
'image_size': size_int}
}
]
return params
def run_opp(args, num_conv_layers: int, exp_params, top_cc: int = 150):
name = f"OPP-influence_num_cc{top_cc}"
cp = MultipleLayersParams()
cp.compute_reconstruction = False
cp.conv_classes = ConvLayer
cp.sp_buffer_size = 6000
cp.sp_batch_size = 1500 # original 4000
cp.learning_rate = 0.10001
cp.cluster_boost_threshold = 1000
cp.max_encountered_seqs = 2000
cp.max_frequent_seqs = 1000
cp.seq_lookahead = 1
cp.seq_length = 4 # note: might use also 5 (as in older experiments)
if num_conv_layers == 2:
cp.n_cluster_centers = [100, 230]
cp.rf_size = [(8, 8), (4, 4)]
cp.rf_stride = [(8, 8), (1, 1)]
cp.num_conv_layers = 2
else:
cp.n_cluster_centers = 200
cp.rf_size = (8, 8)
cp.rf_stride = (8, 8)
cp.num_conv_layers = 1
tp = MultipleLayersParams()
tp.n_cluster_centers = top_cc
tp.sp_buffer_size = 3000
tp.sp_batch_size = 1500 #
tp.learning_rate = 0.15
tp.cluster_boost_threshold = 1000
tp.compute_reconstruction = True
cf_easy = [1, 2, 3, 4]
size = SeDatasetSize.SIZE_64
size_int = (size.value, size.value, 3)
params = [
{
'se_group': {'class_filter': cf_easy,
'image_size': size},
'model': {'conv_layers_params': cp.change(opp=0),
'top_layer_params': tp,
'image_size': size_int}
},
{
'se_group': {'class_filter': cf_easy,
'image_size': size},
'model': {'conv_layers_params': cp.change(opp=0.05),
'top_layer_params': tp,
'image_size': size_int}
},
{
'se_group': {'class_filter': cf_easy,
'image_size': size},
'model': {'conv_layers_params': cp.change(opp=0.5),
'top_layer_params': tp,
'image_size': size_int}
},
{
'se_group': {'class_filter': cf_easy,
'image_size': size},
'model': {'conv_layers_params': cp.change(opp=0.95),
'top_layer_params': tp,
'image_size': size_int}
},
{
'se_group': {'class_filter': cf_easy,
'image_size': size},
'model': {'conv_layers_params': cp.change(opp=1),
'top_layer_params': tp,
'image_size': size_int}
},
# {
# 'se_group': {'class_filter': cf_easy,
# 'image_size': size},
# 'model': {'conv_layers_params': cp.change(opp=0.3),
# 'top_layer_params': tp,
# 'image_size': size_int}
# },
# {
# 'se_group': {'class_filter': cf_easy,
# 'image_size': size},
# 'model': {'conv_layers_params': cp.change(opp=0.7),
# 'top_layer_params': tp,
# 'image_size': size_int}
# }
]
exp_params.experiment_params.num_layers = num_conv_layers + 1 # needs to be there
run_measurement(name, params, args, exp_params)
def run_rf_size(args, exp_params, opp: float = 1.0):
name = "rf_size"
cp = MultipleLayersParams()
cp.compute_reconstruction = False
cp.conv_classes = ConvLayer
cp.sp_buffer_size = 6000
cp.sp_batch_size = 4000
cp.learning_rate = 0.1
cp.cluster_boost_threshold = 1000
cp.max_encountered_seqs = 2000
cp.max_frequent_seqs = 1000
cp.seq_lookahead = 1
cp.seq_length = 5
cp.opp = opp
cp.n_cluster_centers = 200
cp.rf_size = (8, 8)
cp.rf_stride = (8, 8)
cp.num_conv_layers = 1
tp = MultipleLayersParams()
tp.n_cluster_centers = 150
tp.sp_buffer_size = 3000
tp.sp_batch_size = 2000
tp.learning_rate = 0.15
tp.cluster_boost_threshold = 1000
tp.compute_reconstruction = True
cf_easy = [1, 2, 3, 4]
size = SeDatasetSize.SIZE_64
size_int = (size.value, size.value, 3)
params = [
{
'se_group': {'class_filter': cf_easy,
'image_size': size},
'model': {'conv_layers_params': cp.change(rf_stride=([(8, 8)])),
'top_layer_params': tp,
'image_size': size_int}
},
{
'se_group': {'class_filter': cf_easy,
'image_size': size},
'model': {'conv_layers_params': cp.change(rf_stride=([(4, 4)])),
'top_layer_params': tp,
'image_size': size_int}
},
{
'se_group': {'class_filter': cf_easy,
'image_size': size},
'model': {'conv_layers_params': cp.change(rf_stride=([(2, 2)])),
'top_layer_params': tp,
'image_size': size_int}
},
]
exp_params.experiment_params.num_layers = 2 # needs to be there
run_measurement(name, params, args, exp_params)
if __name__ == '__main__':
num_conv_layers = 1
use_top_layer = False
cp = MultipleLayersParams()
cp.compute_reconstruction = True
cp.conv_classes = ConvLayer
cp.sp_buffer_size = 6000
cp.sp_batch_size = 4000
cp.learning_rate = 0.1
cp.cluster_boost_threshold = 1000
cp.max_encountered_seqs = 2000
cp.max_frequent_seqs = 1000
cp.seq_lookahead = 1
cp.seq_length = 5
if num_conv_layers == 2:
cp.n_cluster_centers = [100, 230]
cp.rf_size = (8, 8)
cp.rf_stride = (8, 8)
cp.num_conv_layers = 2
else:
cp.n_cluster_centers = 200
cp.rf_size = (8, 8)
cp.rf_stride = (8, 8)
tp = None
