print("Processing learning rate = {} {}".format(lr, '*' * 40))
base_results_dir = './results/analyze_lr_rate/lr_{}'.format(lr)
cont_int_layer = new_piech_models.CurrentSubtractInhibitLayer(
lateral_e_size=15,
lateral_i_size=15,
n_iters=5,
use_recurrent_batch_norm=True)
# cont_int_layer = new_piech_models.CurrentDivisiveInhibitLayer(
# lateral_e_size=15, lateral_i_size=15, n_iters=5, use_recurrent_batch_norm=True)
# cont_int_layer = new_control_models.ControlMatchParametersLayer(
# lateral_e_size=15, lateral_i_size=15)
model = new_piech_models.ContourIntegrationResnet50(cont_int_layer)
train_parameters = {
'random_seed': random_seed,
'train_batch_size': 32,
'test_batch_size': 32,
'learning_rate': lr,
'num_epochs': 100,
'lateral_w_reg_weight': 0.0001,
'lateral_w_reg_gaussian_sigma': 10,
'clip_negative_lateral_weights': True,
'lr_sched_step_size': 80,
'lr_sched_gamma': 0.5
}
main(model,
# saved_contour_integration_model = \
# './results/new_model_resnet_based/Old/' \
# '/ContourIntegrationResnet50_CurrentSubtractInhibitLayer_20200816_222302_baseline' \
# '/best_accuracy.pth'
# -----------------------------------------------------------------------------------
# Model
# -----------------------------------------------------------------------------------
print(">>> Building the model {}".format('.' * 80))
# # Check that the code works with standard models as well
# net = torchvision.models.resnet50(pretrained=True)
# Edge Extract + Contour Integration layers
cont_int_model = new_piech_models.ContourIntegrationResnet50(
cont_int_layer,
pre_trained_edge_extract=True,
classifier=new_piech_models.DummyHead)
if saved_contour_integration_model is not None:
cont_int_model.load_state_dict(
torch.load(saved_contour_integration_model), strict=False)
# strict = False do not care about loading classifier weights
net = new_piech_models.embed_into_resnet50(
edge_extract_and_contour_integration_layers=cont_int_model,
pretrained=True)
# check_requires_grad(net)
# import pdb
# pdb.set_trace()
