def thumos(deno):
print 'hey girl'
model_name = 'graph_multi_video_with_L1_retF_tanh'
criterion_str = 'MultiCrossEntropyMultiBranchWithL1_CASL'
loss_weights = [1, 1, 1]
plot_losses = True
det_test = False
lr = [0.001, 0.001, 0.001]
multibranch = 1
branch_to_test = -2
print 'branch_to_test', branch_to_test
attention = True
k_vec = None
gt_vec = False
just_primary = False
seed = 999
torch.backends.cudnn.deterministic = True
random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
epoch_stuff = [250, 250]
dataset = 'ucf'
limit = None
save_after = 50
test_mode = False
save_outfs = False
test_method = 'original'
test_post_pend = '_' + test_method + '_class'
model_nums = [249]
retrain = False
viz_mode = False
viz_sim = False
# post_pend = '_noBiasLastLayer'
network_params = {}
network_params['deno'] = deno
network_params['in_out'] = [2048, 1024]
network_params['feat_dim'] = [2048, 1024]
network_params['feat_ret'] = True
network_params['graph_size'] = 1
network_params['method'] = 'cos'
network_params['sparsify'] = 'percent_0.5'
network_params['graph_sum'] = attention
network_params['non_lin'] = None
network_params['aft_nonlin'] = 'RL_L2'
network_params['sigmoid'] = True
num_similar = 0
post_pend = 'changingSparsityAbs_' + str(num_similar)
first_thresh = 0
class_weights = False
test_after = 10
all_classes = False
second_thresh = -0.9
det_class = -1
train_simple_mill_all_classes(model_name=model_name,
lr=lr,
dataset=dataset,
network_params=network_params,
limit=limit,
epoch_stuff=epoch_stuff,
batch_size=32,
batch_size_val=32,
save_after=save_after,
test_mode=test_mode,
class_weights=class_weights,
test_after=test_after,
all_classes=all_classes,
just_primary=just_primary,
model_nums=model_nums,
retrain=retrain,
viz_mode=viz_mode,
second_thresh=second_thresh,
first_thresh=first_thresh,
det_class=det_class,
post_pend=post_pend,
viz_sim=viz_sim,
gt_vec=gt_vec,
loss_weights=loss_weights,
multibranch=multibranch,
branch_to_test=branch_to_test,
k_vec=k_vec,
attention=attention,
test_pair=False,
test_post_pend=test_post_pend,
test_method=test_method,
criterion_str=criterion_str,
plot_losses=plot_losses,
num_similar=num_similar,
save_outfs=save_outfs,
det_test=det_test)
def thumos_bce_actual(loss_weights, sparsify):
print 'hey girl'
model_name = 'graph_multi_video_with_L1_retF_tanh'
criterion_str = 'BinaryCrossEntropyMultiBranchWithL1_CASL'
# loss_weights = [1,1,0]
plot_losses = True
det_test = True
# model_name = 'graph_multi_video_with_L1'
# criterion_str = 'MultiCrossEntropyMultiBranchWithL1'
# # criterion_str = 'BinaryCrossEntropyMultiBranchWithL1'
# loss_weights = [1,1.]
# plot_losses = False
lr = [0.001, 0.001, 0.001]
multibranch = 1
# loss_weights = [1,1]
branch_to_test = -2
print 'branch_to_test', branch_to_test
attention = True
k_vec = None
gt_vec = False
just_primary = False
seed = 999
torch.backends.cudnn.deterministic = True
random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
epoch_stuff = [250, 250]
dataset = 'ucf'
limit = None
save_after = 50
test_mode = False
save_outfs = False
# test_method = 'wtalc'
# test_method = 'wtalc'
# test_post_pend = '_'+test_method+'_tp_fp_conf'
# test_method = 'best_worst_dot'
# test_post_pend = '_'+test_method
test_method = 'original'
test_post_pend = '_' + test_method + '_diff_viz_multi_only'
model_nums = [249]
# ,349,399,449,499]
retrain = False
viz_mode = True
viz_sim = False
# post_pend = '_noBiasLastLayer'
network_params = {}
network_params['deno'] = 8
network_params['in_out'] = [2048, 1024]
network_params['feat_dim'] = [2048, 1024]
network_params['feat_ret'] = True
network_params['graph_size'] = 1
network_params['method'] = 'cos'
network_params['sparsify'] = sparsify
# 'percent_0.5'
# 'static_mid_minmin_maxmax'
network_params['graph_sum'] = attention
network_params['non_lin'] = None
network_params['aft_nonlin'] = 'RL_L2'
network_params['sigmoid'] = False
# network_params['dropout'] = 0.8
num_similar = 0
post_pend = 'forplot_' + str(num_similar)
first_thresh = -1
class_weights = False
test_after = 10
all_classes = False
second_thresh = scipy.special.logit(0.1)
# -0.9
det_class = -1
train_simple_mill_all_classes(model_name=model_name,
lr=lr,
dataset=dataset,
network_params=network_params,
limit=limit,
epoch_stuff=epoch_stuff,
batch_size=32,
batch_size_val=32,
save_after=save_after,
test_mode=test_mode,
class_weights=class_weights,
test_after=test_after,
all_classes=all_classes,
just_primary=just_primary,
model_nums=model_nums,
retrain=retrain,
viz_mode=viz_mode,
second_thresh=second_thresh,
first_thresh=first_thresh,
det_class=det_class,
post_pend=post_pend,
viz_sim=viz_sim,
gt_vec=gt_vec,
loss_weights=loss_weights,
multibranch=multibranch,
branch_to_test=branch_to_test,
k_vec=k_vec,
attention=attention,
test_pair=False,
test_post_pend=test_post_pend,
test_method=test_method,
criterion_str=criterion_str,
plot_losses=plot_losses,
num_similar=num_similar,
save_outfs=save_outfs,
det_test=True)
def charades_bce():
print 'hey girl NEW'
model_name = 'graph_multi_video_with_L1_retF'
# criterion_str = 'MultiCrossEntropyMultiBranchWithL1_CASL'
criterion_str = 'BinaryCrossEntropyMultiBranchWithL1_CASL'
loss_weights = [1,1,1]
import torch.multiprocessing
torch.multiprocessing.set_sharing_strategy('file_system')
dataset = 'charades_i3d_charades_both'
class_weights = False
num_similar = 128
det_test = True
plot_losses = True
# model_file = ['../experiments/graph_multi_video_with_L1_retF/graph_multi_video_with_L1_retF_aft_nonlin_RL_L2_non_lin_None_sparsify_0.5_graph_size_2_sigmoid_False_graph_sum_True_deno_8_n_classes_157_in_out_2048_1024_feat_dim_2048_1024_feat_ret_True_method_cos_charades_i3d_charades_both/all_classes_False_just_primary_False_limit_None_cw_True_BinaryCrossEntropyMultiBranchWithL1_CASL_250_step_250_0.1_0.001_0.001_0.001_lw_1.00_1.00_1.00__cwOld_MulNumClasses_numSim_128/model_249.pt',250]
# model_name = 'graph_multi_video_with_L1'
# criterion_str = 'BinaryCrossEntropyMultiBranchWithL1'
# loss_weights = [1,1]
# plot_losses = False
batch_size = 256
batch_size_val = 256
lr = [0.001,0.001, 0.001]
# lr = [0.01,0.01, 0.01]
multibranch = 1
# loss_weights = [1,1]
branch_to_test = -2
print 'branch_to_test',branch_to_test
attention = True
k_vec = None
gt_vec = False
just_primary = False
seed = 999
torch.backends.cudnn.deterministic = True
random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
epoch_stuff = [250,250]
limit = None
save_after = 50
test_mode = False
# test_method = 'wtalc'
# test_method = 'wtalc'
# test_post_pend = '_'+test_method+'_tp_fp_conf'
# test_method = 'best_worst_dot'
# test_post_pend = '_'+test_method
test_method = 'original'
test_post_pend = '_'+test_method+'_class'
model_nums = [249]
retrain = False
viz_mode = False
viz_sim = False
# post_pend = '_noBiasLastLayer'
network_params = {}
network_params['deno'] = 8
network_params['in_out'] = [2048,1024]
network_params['feat_dim'] = [2048,1024]
network_params['feat_ret']=True
network_params['graph_size'] = 2
network_params['method'] = 'cos'
network_params['sparsify'] = 'percent_0.5'
network_params['graph_sum'] = attention
network_params['non_lin'] = None
network_params['aft_nonlin']='RL_L2'
network_params['sigmoid'] = False
# network_params['dropout'] = 0.8
# post_pend = ''
post_pend = '_marginloss_numSim_'+str(num_similar)
# +'_sumnomean_noExclusiveCASL_NEW_noMax'
first_thresh=-1
test_after = 50
all_classes = False
second_thresh = scipy.special.logit(0.05)
det_class = -1
print 'this is it! are you ready????'
raw_input()
train_simple_mill_all_classes (model_name = model_name,
lr = lr,
dataset = dataset,
network_params = network_params,
limit = limit,
epoch_stuff= epoch_stuff,
batch_size = batch_size,
batch_size_val = batch_size_val,
save_after = save_after,
test_mode = test_mode,
class_weights = class_weights,
test_after = test_after,
all_classes = all_classes,
just_primary = just_primary,
model_nums = model_nums,
retrain = retrain,
viz_mode = viz_mode,
second_thresh = second_thresh,
first_thresh = first_thresh,
det_class = det_class,
post_pend = post_pend,
viz_sim = viz_sim,
gt_vec = gt_vec,
loss_weights = loss_weights,
multibranch = multibranch,
branch_to_test = branch_to_test,
k_vec = k_vec,
attention = attention,
test_pair = False,
test_post_pend = test_post_pend,
test_method = test_method,
criterion_str = criterion_str,
plot_losses = plot_losses,
det_test = det_test,
num_similar = num_similar)
def charades(deno, gs=1):
print 'hey girl NEW'
model_name = 'graph_multi_video_with_L1_retF_tanh'
criterion_str = 'MultiCrossEntropyMultiBranchWithL1_CASL'
# criterion_str = 'BinaryCrossEntropyMultiBranchWithL1_CASL'
loss_weights = [1, 1, 1]
dataset = 'charades_i3d_charades_both'
class_weights = False
num_similar = 128
det_test = False
plot_losses = True
# model_name = 'graph_multi_video_with_L1'
# criterion_str = 'BinaryCrossEntropyMultiBranchWithL1'
# loss_weights = [1,1]
# plot_losses = False
batch_size = 256
batch_size_val = 256
lr = [0.001, 0.001, 0.001]
# lr = [0.01,0.01, 0.01]
multibranch = 1
# loss_weights = [1,1]
branch_to_test = -2
print 'branch_to_test', branch_to_test
attention = True
k_vec = None
gt_vec = False
just_primary = False
seed = 999
torch.backends.cudnn.deterministic = True
random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
epoch_stuff = [250, 250]
limit = None
save_after = 50
test_mode = True
save_outfs = True
# test_method = 'wtalc'
# test_method = 'wtalc'
# test_post_pend = '_'+test_method+'_tp_fp_conf'
# test_method = 'best_worst_dot'
# test_post_pend = '_'+test_method
test_method = 'original'
test_post_pend = '_' + test_method + '_class_merged'
model_nums = [249]
retrain = False
viz_mode = False
viz_sim = False
# post_pend = '_noBiasLastLayer'
network_params = {}
network_params['deno'] = deno
network_params['in_out'] = [2048, 1024]
network_params['feat_dim'] = [2048, 1024]
network_params['feat_ret'] = True
network_params['graph_size'] = gs
network_params['method'] = 'cos'
network_params['sparsify'] = 'percent_0.5'
network_params['graph_sum'] = attention
network_params['non_lin'] = None
network_params['aft_nonlin'] = 'RL_L2'
network_params['sigmoid'] = True
# network_params['dropout'] = 0.8
# post_pend = ''
post_pend = 'changingSparsityAbs_numSim_' + str(num_similar)
# '_cwNo_justPos_MulNumClasses_numSim_'+str(num_similar)
# +'_sumnomean_noExclusiveCASL_NEW_noMax'
first_thresh = 0
# scipy.special.logit(0.1)
#
test_after = 50
all_classes = False
second_thresh = -0.9
# scipy.special.logit(0.1)
det_class = -1
# print 'this is it! are you ready????'
# raw_input()
train_simple_mill_all_classes(model_name=model_name,
lr=lr,
dataset=dataset,
network_params=network_params,
limit=limit,
epoch_stuff=epoch_stuff,
batch_size=batch_size,
batch_size_val=batch_size_val,
save_after=save_after,
test_mode=test_mode,
class_weights=class_weights,
test_after=test_after,
all_classes=all_classes,
just_primary=just_primary,
model_nums=model_nums,
retrain=retrain,
viz_mode=viz_mode,
second_thresh=second_thresh,
first_thresh=first_thresh,
det_class=det_class,
post_pend=post_pend,
viz_sim=viz_sim,
gt_vec=gt_vec,
loss_weights=loss_weights,
multibranch=multibranch,
branch_to_test=branch_to_test,
k_vec=k_vec,
attention=attention,
test_pair=False,
test_post_pend=test_post_pend,
test_method=test_method,
criterion_str=criterion_str,
plot_losses=plot_losses,
det_test=det_test,
num_similar=num_similar,
save_outfs=save_outfs)
def activitynet_bce():
print 'hey girl'
model_name = 'graph_multi_video_with_L1_retF_tanh'
# criterion_str = 'MultiCrossEntropyMultiBranchWithSigmoidWithL1_CASL'
criterion_str = 'MultiCrossEntropyMultiBranchWithL1_CASL'
# criterion_str = 'BinaryCrossEntropyMultiBranchWithL1_CASL'
loss_weights = [1,1,1]
plot_losses = True
# import torch.multiprocessing
# torch.multiprocessing.set_sharing_strategy('file_system')
# model_name = 'graph_multi_video_with_L1'
# # criterion_str = 'MultiCrossEntropyMultiBranchWithL1'
# criterion_str = 'BinaryCrossEntropyMultiBranchWithL1'
# loss_weights = [1,1.]
# plot_losses = False
lr = [0.001,0.001, 0.001]
multibranch = 1
num_similar = 128
# det_test = False
batch_size = 256
batch_size_val = 256
# loss_weights = [1,1]
branch_to_test = -2
print 'branch_to_test',branch_to_test
attention = True
k_vec = None
gt_vec = False
just_primary = False
seed = 999
torch.backends.cudnn.deterministic = True
random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
epoch_stuff = [250,250]
dataset = 'activitynet'
limit = None
save_after = 50
test_mode = False
save_outfs = False
# test_method = 'wtalc'
# test_method = 'wtalc'
# test_post_pend = '_'+test_method+'_tp_fp_conf'
# test_method = 'best_worst_dot'
# test_post_pend = '_'+test_method
test_method = 'original'
test_post_pend = '_'+test_method+'_class'
model_nums = [249]
retrain = False
viz_mode = False
viz_sim = False
# post_pend = '_noBiasLastLayer'
network_params = {}
network_params['deno'] = 1
network_params['in_out'] = [2048,1024]
network_params['feat_dim'] = [2048,1024]
network_params['feat_ret']=True
network_params['graph_size'] = 1
network_params['method'] = 'cos'
network_params['sparsify'] = 'percent_0.5'
network_params['graph_sum'] = attention
network_params['non_lin'] = None
network_params['aft_nonlin']='RL_L2'
network_params['sigmoid'] = True
# post_pend = '0to4_weighting5_changingSparsityAbs_'+str(num_similar)
# class_weights = False
post_pend = 'gs1_changingSparsityAbs_'+str(num_similar)
class_weights = False
first_thresh=0
# scipy.special.logit(0.1)
test_after = 10
all_classes = False
second_thresh = -0.9
# scipy.special.logit(0.1)
# 'max_-4'
# scipy.special.logit(0.1)
# -5.144
# 0.1
# 'otsu_per_class_pmfthresh_justpos_0'
# 'min_max_per_class_pmfthresh_0'
# 'otsu_per_class_gt'
# -10.970
# -9.514
# 0.5
# -11.699
# -5.144
# 0.5
det_class = -1
train_simple_mill_all_classes (model_name = model_name,
lr = lr,
dataset = dataset,
network_params = network_params,
limit = limit,
epoch_stuff= epoch_stuff,
batch_size = batch_size,
batch_size_val = batch_size_val,
save_after = save_after,
test_mode = test_mode,
class_weights = class_weights,
test_after = test_after,
all_classes = all_classes,
just_primary = just_primary,
model_nums = model_nums,
retrain = retrain,
viz_mode = viz_mode,
second_thresh = second_thresh,
first_thresh = first_thresh,
det_class = det_class,
post_pend = post_pend,
viz_sim = viz_sim,
gt_vec = gt_vec,
loss_weights = loss_weights,
multibranch = multibranch,
branch_to_test = branch_to_test,
k_vec = k_vec,
attention = attention,
test_pair = False,
test_post_pend = test_post_pend,
test_method = test_method,
criterion_str = criterion_str,
plot_losses = plot_losses,
num_similar = num_similar,
save_outfs = save_outfs,
det_test = True)
def fcasl(feat_dim):
print 'hey girl'
model_name = 'fcasl_multi_video_with_L1_retF'
criterion_str = 'MultiCrossEntropyMultiBranchFakeL1_CASL'
loss_weights = [1, 1, 1]
plot_losses = True
det_test = True
# model_name = 'graph_multi_video_with_L1'
# criterion_str = 'MultiCrossEntropyMultiBranchWithL1'
# # criterion_str = 'BinaryCrossEntropyMultiBranchWithL1'
# loss_weights = [1,1.]
# plot_losses = False
lr = [0.001, 0.001]
multibranch = 1
# loss_weights = [1,1]
branch_to_test = -2
print 'branch_to_test', branch_to_test
attention = True
k_vec = None
gt_vec = False
just_primary = False
seed = 999
torch.backends.cudnn.deterministic = True
random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
epoch_stuff = [250, 250]
dataset = 'ucf'
limit = None
save_after = 50
test_mode = False
save_outfs = False
# test_method = 'wtalc'
# test_method = 'wtalc'
# test_post_pend = '_'+test_method+'_tp_fp_conf'
# test_method = 'best_worst_dot'
# test_post_pend = '_'+test_method
test_method = 'original'
test_post_pend = '_' + test_method
model_nums = [249]
# ,349,399,449,499]
retrain = False
viz_mode = False
viz_sim = False
# post_pend = '_noBiasLastLayer'
network_params = {}
network_params['deno'] = 8
network_params['feat_dim'] = [2048, feat_dim]
network_params['feat_ret'] = True
network_params['sigmoid'] = True
num_similar = 0
post_pend = 'comparison_' + str(num_similar)
first_thresh = 0
class_weights = False
test_after = 10
all_classes = False
second_thresh = -0.9
det_class = -1
train_simple_mill_all_classes(model_name=model_name,
lr=lr,
dataset=dataset,
network_params=network_params,
limit=limit,
epoch_stuff=epoch_stuff,
batch_size=32,
batch_size_val=32,
save_after=save_after,
test_mode=test_mode,
class_weights=class_weights,
test_after=test_after,
all_classes=all_classes,
just_primary=just_primary,
model_nums=model_nums,
retrain=retrain,
viz_mode=viz_mode,
second_thresh=second_thresh,
first_thresh=first_thresh,
det_class=det_class,
post_pend=post_pend,
viz_sim=viz_sim,
gt_vec=gt_vec,
loss_weights=loss_weights,
multibranch=multibranch,
branch_to_test=branch_to_test,
k_vec=k_vec,
attention=attention,
test_pair=False,
test_post_pend=test_post_pend,
test_method=test_method,
criterion_str=criterion_str,
plot_losses=plot_losses,
num_similar=num_similar,
save_outfs=save_outfs,
det_test=det_test)
def graph_norm_game():
print 'hey girl'
# raw_input()
model_name = 'graph_multi_video_norm_game'
criterion_str = 'MultiCrossEntropyMultiBranchWithL1'
loss_weights = [1,1]
plot_losses = False
# model_name = 'graph_multi_video_with_L1'
# criterion_str = 'MultiCrossEntropyMultiBranchWithL1'
# loss_weights = [1,1]
# plot_losses = False
lr = [0.001,0.001, 0.001]
multibranch = 1
branch_to_test = -2
print 'branch_to_test',branch_to_test
attention = True
k_vec = None
gt_vec = False
just_primary = False
seed = 999
torch.backends.cudnn.deterministic = True
random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
epoch_stuff = [250,250]
dataset = 'ucf'
limit = None
save_after = 50
test_mode = False
test_method = 'original'
test_post_pend = '_'+test_method+'_class'
model_nums = [249]
retrain = False
viz_mode = False
viz_sim = False
network_params = {}
network_params['deno'] = 8
network_params['in_out'] = [2048,1024]
network_params['feat_dim'] = [2048,1024]
network_params['feat_ret']=False
network_params['graph_size'] = 1
network_params['method'] = 'cos'
network_params['sparsify'] = None
network_params['graph_sum'] = attention
network_params['non_lin'] = None
network_params['aft_nonlin']='RL_L2'
network_params['sigmoid'] = False
# network_params['nosum'] = 5
# post_pend = '_rnsmaxways'
network_params['nosum'] = 4
post_pend = '_rnbothways'
first_thresh=0
class_weights = True
test_after = 10
all_classes = False
second_thresh = 0.5
det_class = -1
train_simple_mill_all_classes (model_name = model_name,
lr = lr,
dataset = dataset,
network_params = network_params,
limit = limit,
epoch_stuff= epoch_stuff,
batch_size = 32,
batch_size_val = 32,
save_after = save_after,
test_mode = test_mode,
class_weights = class_weights,
test_after = test_after,
all_classes = all_classes,
just_primary = just_primary,
model_nums = model_nums,
retrain = retrain,
viz_mode = viz_mode,
second_thresh = second_thresh,
first_thresh = first_thresh,
det_class = det_class,
post_pend = post_pend,
viz_sim = viz_sim,
gt_vec = gt_vec,
loss_weights = loss_weights,
multibranch = multibranch,
branch_to_test = branch_to_test,
k_vec = k_vec,
attention = attention,
test_pair = False,
test_post_pend = test_post_pend,
test_method = test_method,
criterion_str = criterion_str,
plot_losses = plot_losses)
def graph_charades_everything_sim():
print 'hey girl charades'
# raw_input()
model_name = 'graph_multi_video_with_L1_retF'
criterion_str = 'MultiCrossEntropyMultiBranchWithL1_CASL'
loss_weights = [1,1,1]
# loss_weights = [1,1,0]
# dataset = 'charades_vgg_16_rgb_npy'
dataset = 'charades_i3d_both'
class_weights = True
num_similar = 128
det_test = False
batch_size = 256
batch_size_val = 256
plot_losses = True
# model_name = 'graph_multi_video_with_L1'
# criterion_str = 'MultiCrossEntropyMultiBranchWithL1'
# loss_weights = [1,1]
# plot_losses = False
lr = [0.001,0.001, 0.001]
multibranch = 1
# loss_weights = [1,1]
branch_to_test = -2
print 'branch_to_test',branch_to_test
attention = True
k_vec = None
gt_vec = False
just_primary = False
seed = 999
torch.backends.cudnn.deterministic = True
random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
epoch_stuff = [250,250]
# dataset = 'multithumos'
limit = None
save_after = 50
test_mode = True
# test_method = 'wtalc'
# test_method = 'wtalc'
# test_post_pend = '_'+test_method+'_tp_fp_conf'
# test_method = 'best_worst_dot'
# test_post_pend = '_'+test_method
test_method = 'original'
test_post_pend = '_'+test_method+'_class'
model_nums = None
retrain = False
viz_mode = False
viz_sim = False
# post_pend = '_noBiasLastLayer'
network_params = {}
network_params['deno'] = 8
network_params['in_out'] = [2048,1024]
network_params['feat_dim'] = [2048,1024]
network_params['feat_ret']=True
network_params['graph_size'] = 2
network_params['method'] = 'cos'
network_params['sparsify'] = 0.5
network_params['graph_sum'] = attention
network_params['non_lin'] = None
network_params['aft_nonlin']='RL_L2'
network_params['sigmoid'] = False
# network_params['dropout'] = 0.8
post_pend = '_noLimit'
post_pend = '_numSim_'+str(num_similar)+'_sumnomean_noExclusiveCASL_NEW_noMax'
# post_pend = '_numSim_'+str(num_similar)+'_'
first_thresh=0
test_after = 50
all_classes = False
second_thresh = 0.5
det_class = -1
# print 'this is it! are you ready????'
# raw_input()
train_simple_mill_all_classes (model_name = model_name,
lr = lr,
dataset = dataset,
network_params = network_params,
limit = limit,
epoch_stuff= epoch_stuff,
batch_size = batch_size,
batch_size_val = batch_size_val,
save_after = save_after,
test_mode = test_mode,
class_weights = class_weights,
test_after = test_after,
all_classes = all_classes,
just_primary = just_primary,
model_nums = model_nums,
retrain = retrain,
viz_mode = viz_mode,
second_thresh = second_thresh,
first_thresh = first_thresh,
det_class = det_class,
post_pend = post_pend,
viz_sim = viz_sim,
gt_vec = gt_vec,
loss_weights = loss_weights,
multibranch = multibranch,
branch_to_test = branch_to_test,
k_vec = k_vec,
attention = attention,
test_pair = False,
test_post_pend = test_post_pend,
test_method = test_method,
criterion_str = criterion_str,
plot_losses = plot_losses,
det_test = det_test,
num_similar = num_similar,
save_outfs = True)
def thumos(deno):
print 'hey girl'
model_name = 'graph_multi_video_with_L1_retF_tanh'
criterion_str = 'MultiCrossEntropyMultiBranchWithL1_CASL'
loss_weights = [1, 1, 1]
plot_losses = True
# learning rate for [\phi, graph layer, final linear layer]
lr = [0.001, 0.001, 0.001]
seed = 999
torch.backends.cudnn.deterministic = True
random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
# string to switch between datasets
dataset = 'ucf'
#num segments to include per video during training. in case of memory problems
limit = None
# epoch_stuff = [num epochs after which to reduce lr, total num epochs]
epoch_stuff = [250, 250]
save_after = 50
# string to append to experiment folder
post_pend = 'denoExp'
test_after = 10
# set det_test to False if localization results on val not needed during training
det_test = True
# model numbers to test.
model_nums = [249]
# set test mode to true to test a trained model
test_mode = False
save_outfs = False
test_method = 'original'
test_post_pend = '_' + test_method + '_class'
# number of similar class videos to contain in each training batch. imp when num classes>batchsize
num_similar = 0
retrain = False
viz_mode = False
viz_sim = False
network_params = {}
network_params['deno'] = deno
network_params['in_out'] = [2048, 1024]
network_params['feat_dim'] = [2048, 1024]
network_params['feat_ret'] = True
# graph_size controls number of videos per graph during training (see supp Fig 1)
network_params['graph_size'] = 1
network_params['method'] = 'cos'
network_params['sparsify'] = 'percent_0.5'
attention = True
network_params['graph_sum'] = attention
network_params['non_lin'] = None
network_params['aft_nonlin'] = 'RL_L2'
network_params['sigmoid'] = True
# default params. no need to change.
first_thresh = 0
class_weights = False
all_classes = False
second_thresh = -0.9
det_class = -1
multibranch = 1
branch_to_test = -2
k_vec = None
gt_vec = False
just_primary = False
train_simple_mill_all_classes(model_name=model_name,
lr=lr,
dataset=dataset,
network_params=network_params,
limit=limit,
epoch_stuff=epoch_stuff,
batch_size=32,
batch_size_val=32,
save_after=save_after,
test_mode=test_mode,
class_weights=class_weights,
test_after=test_after,
all_classes=all_classes,
just_primary=just_primary,
model_nums=model_nums,
retrain=retrain,
viz_mode=viz_mode,
second_thresh=second_thresh,
first_thresh=first_thresh,
det_class=det_class,
post_pend=post_pend,
viz_sim=viz_sim,
gt_vec=gt_vec,
loss_weights=loss_weights,
multibranch=multibranch,
branch_to_test=branch_to_test,
k_vec=k_vec,
attention=attention,
test_pair=False,
test_post_pend=test_post_pend,
test_method=test_method,
criterion_str=criterion_str,
plot_losses=plot_losses,
num_similar=num_similar,
save_outfs=save_outfs,
det_test=det_test)
