def loop_length_beam(option, opt, model, loader, vocab, device, teacher_model,
dict_mapping):
b4 = []
m = []
r = []
c = []
ave_len = []
for lbs in range(1, 11):
option['length_beam_size'] = lbs
metric = run_eval(option,
model,
None,
loader,
vocab,
device,
json_path=opt.json_path,
json_name=opt.json_name,
print_sent=opt.print_sent,
teacher_model=teacher_model,
length_crit=torch.nn.SmoothL1Loss(),
dict_mapping=dict_mapping,
analyze=True)
b4.append(metric["Bleu_4"])
m.append(metric["METEOR"])
r.append(metric["ROUGE_L"])
c.append(metric["CIDEr"])
ave_len.append(metric['ave_length'])
print(b4)
print(m)
print(r)
print(c)
print(ave_len)
def loop_category(option, opt, model, device):
loop_logger = CsvLogger(
filepath='./category_results',
filename='ARVC_%s%s.csv' %
(option['dataset'],
'' if option.get('method', None) != 'ag' else '_ag'),
#fieldsnames=['novel', 'unique', 'usage', 'ave_length', 'gram4'],
fieldsnames=[
'Bleu_1', 'Bleu_2', 'Bleu_3', 'Bleu_4', 'METEOR', 'ROUGE_L',
'CIDEr', 'Sum', 'loss'
],
)
for i in range(20):
loader = get_loader(option, mode=opt.em, specific=i)
vocab = loader.dataset.get_vocab()
#metric = run_eval(option, model, None, loader, vocab, device, print_sent=opt.print_sent, no_score=True, analyze=True)
metric = run_eval(option,
model,
None,
loader,
vocab,
device,
print_sent=opt.print_sent,
no_score=False,
analyze=False)
loop_logger.write(metric)
def loop_iterations(option, opt, model, loader, vocab, device, teacher_model,
dict_mapping):
loop_logger = CsvLogger(filepath='./loop_results',
filename=opt.loop + '.csv',
fieldsnames=[
'Bleu_1', 'Bleu_2', 'Bleu_3', 'Bleu_4',
'METEOR', 'ROUGE_L', 'CIDEr', 'Sum',
'iterations', 'lbs', 'novel', 'unique',
'usage', 'ave_length'
])
for i in range(1, 11):
option['iterations'] = i
metric = run_eval(option,
model,
None,
loader,
vocab,
device,
json_path=opt.json_path,
json_name=opt.json_name,
print_sent=opt.print_sent,
teacher_model=teacher_model,
length_crit=torch.nn.SmoothL1Loss(),
dict_mapping=dict_mapping,
analyze=True)
metric['iterations'] = option['iterations']
metric['lbs'] = option['length_beam_size']
metric.pop('loss')
loop_logger.write(metric)
def main(opt):
device = torch.device('cuda' if not opt.no_cuda else 'cpu')
opt_pth = os.path.join(opt.model_path, 'opt_info.json')
option = json.load(open(opt_pth, 'r'))
option.update(vars(opt))
# print(option)
model = get_model(option)
checkpoint = torch.load(os.path.join(opt.model_path, 'best', opt.model_name))
model.load_state_dict(checkpoint['state_dict'])
model.to(device)
for key in ['info_corpus', 'reference', 'feats_i', 'feats_m', 'feats_a']:
if isinstance(option[key], list):
for i in range(len(option[key])):
option[key][i] = option[key][i].replace('/home/yangbang/VideoCaptioning', '/Users/yangbang/Desktop/VC_data')
else:
option[key] = option[key].replace('/home/yangbang/VideoCaptioning', '/Users/yangbang/Desktop/VC_data')
loader = get_loader(option, mode=opt.em, print_info=False, specific=opt.specific)
vocab = loader.dataset.get_vocab()
if opt.oe:
metric = run_eval(option, model, None, loader, vocab, device,
json_path=opt.json_path, json_name=opt.json_name, print_sent=opt.print_sent, no_score=opt.ns,
save_videodatainfo=opt.sv)
print(metric)
else:
encoder_outputs = []
category = []
for data in tqdm(loader, ncols=150, leave=False):
with torch.no_grad():
results, cate, _, _ = get_forword_results(option, model, data, device=device, only_data=True)
encoder_outputs.append(results['enc_output'][0])
category.append(cate)
encoder_outputs = torch.cat(encoder_outputs, dim=0).cpu().numpy()
category = torch.cat(category, dim=0).view(-1).cpu().numpy()
print(encoder_outputs.shape, category.shape)
encoder_outputs = encoder_outputs.mean(1)
pca = manifold.TSNE(n_components=opt.pca)
data = pca.fit_transform(encoder_outputs)
plot_several_category(data, category)
def loop_category(option, opt, model, device, teacher_model, dict_mapping):
loop_logger = CsvLogger(
filepath='./category_results',
filename='NAVC_%s_%s%s.csv' %
(option['method'], 'AE' if opt.nv_scale == 100 else '', opt.paradigm),
fieldsnames=['novel', 'unique', 'usage', 'ave_length', 'gram4'])
for i in range(20):
loader = get_loader(option, mode=opt.em, specific=i)
vocab = loader.dataset.get_vocab()
metric = run_eval(option,
model,
None,
loader,
vocab,
device,
print_sent=opt.print_sent,
no_score=True,
analyze=True,
teacher_model=teacher_model,
dict_mapping=dict_mapping)
loop_logger.write(metric)
def plot(option, opt, model, loader, vocab, device, teacher_model,
dict_mapping):
colors = [
'skyblue', 'dodgerblue', 'dodgerblue', 'dodgerblue', 'dodgerblue'
]
for i, iteration in enumerate([1, 2, 3, 4, 5]):
option['iterations'] = iteration
metric, x, y = run_eval(option,
model,
None,
loader,
vocab,
device,
json_path=opt.json_path,
json_name=opt.json_name,
print_sent=opt.print_sent,
teacher_model=teacher_model,
length_crit=torch.nn.SmoothL1Loss(),
dict_mapping=dict_mapping,
length_bias=opt.lb,
analyze=True,
plot=True,
no_score=True,
top_n=30)
ax = plt.subplot(1, 5, i + 1)
ax.barh(x, y, color=colors[i])
ax.set_title('iteration=%d' % iteration)
#plt.tick_params(labelsize=13)
plt.subplots_adjust(left=0.12,
bottom=None,
right=0.98,
top=None,
wspace=0.4,
hspace=None)
fig = plt.gcf()
fig.set_size_inches(15, 10)
plt.savefig('./iteration.png')
plt.show()
def main():
'''Main Function'''
parser = argparse.ArgumentParser(description='translate.py')
parser.add_argument('-df', '--default', default=False, action='store_true')
parser.add_argument('-method', '--method', default='ARB', type=str)
parser.add_argument('-dataset', '--dataset', default='MSRVTT', type=str)
parser.add_argument('--default_model_name', default='best.pth.tar', type=str)
parser.add_argument('-scope', '--scope', default='', type=str)
parser.add_argument('-record', '--record', default=False, action='store_true')
parser.add_argument('-field', '--field', nargs='+', type=str, default=['seed'])
parser.add_argument('-val_and_test', '--val_and_test', default=False, action='store_true')
parser.add_argument('-model_path', '--model_path', type=str)
parser.add_argument('-teacher_path', '--teacher_path', type=str)
parser.add_argument('-bs', '--beam_size', type=int, default=5, help='Beam size')
parser.add_argument('-ba', '--beam_alpha', type=float, default=1.0)
parser.add_argument('-topk', '--topk', type=int, default=1)
# NA decoding
parser.add_argument('-i', '--iterations', type=int, default=5)
parser.add_argument('-lbs', '--length_beam_size', type=int, default=6)
parser.add_argument('-q', '--q', type=int, default=1)
parser.add_argument('-qi', '--q_iterations', type=int, default=1)
parser.add_argument('-paradigm', '--paradigm', type=str, default='mp')
parser.add_argument('-use_ct', '--use_ct', default=False, action='store_true')
parser.add_argument('-md', '--masking_decision', default=False, action='store_true')
parser.add_argument('-ncd', '--no_candidate_decision', default=False, action='store_true')
parser.add_argument('--algorithm_print_sent', default=False, action='store_true')
parser.add_argument('-batch_size', '--batch_size', type=int, default=128)
parser.add_argument('-no_cuda', action='store_true')
parser.add_argument('-em', '--evaluation_mode', type=str, default='test')
parser.add_argument('-print_sent', action='store_true')
parser.add_argument('-json_path', type=str, default='')
parser.add_argument('-json_name', type=str, default='')
parser.add_argument('-ns', '--no_score', default=False, action='store_true')
parser.add_argument('-analyze', default=False, action='store_true')
parser.add_argument('-latency', default=False, action='store_true')
parser.add_argument('-specific', default=-1, type=int)
parser.add_argument('-collect_path', type=str, default='./collected_captions')
parser.add_argument('-collect', default=False, action='store_true')
parser.add_argument('-nobc', '--not_only_best_candidate', default=False, action='store_true')
opt = parser.parse_args()
device = torch.device('cuda' if not opt.no_cuda else 'cpu')
teacher_model = None
dict_mapping = {}
if opt.default:
if opt.dataset.lower() == 'msvd':
opt.dataset = 'Youtube2Text'
opt.model_path = os.path.join(
Constants.base_checkpoint_path,
opt.dataset,
opt.method,
opt.scope,
opt.default_model_name
)
if opt.method in ['NAB', 'NACF']:
opt.teacher_path = os.path.join(
Constants.base_checkpoint_path,
opt.dataset,
'ARB',
opt.scope,
opt.default_model_name
)
assert os.path.exists(opt.teacher_path)
else:
assert opt.model_path and os.path.exists(opt.model_path)
model, option, other_info = load_model_and_opt(opt.model_path, device, return_other_info=True)
if getattr(opt, 'teacher_path', None) is not None:
print('Loading teacher model from %s' % opt.teacher_path)
teacher_model, teacher_option = load_model_and_opt(opt.teacher_path, device)
dict_mapping = get_dict_mapping(option, teacher_option)
option['reference'] = option['reference'].replace('msvd_refs.pkl', 'refs.pkl')
option['info_corpus'] = option['info_corpus'].replace('info_corpus_0.pkl', 'info_corpus.pkl')
if not opt.default:
_ = option['dataset']
option.update(vars(opt))
option['dataset'] = _
else:
if option['decoding_type'] != 'NARFormer':
option['topk'] = opt.topk
option['beam_size'] = 5
option['beam_alpha'] = 1.0
else:
option['algorithm_print_sent'] = opt.algorithm_print_sent
option['paradigm'] = opt.paradigm
option['iterations'] = 5
option['length_beam_size'] = 6
option['beam_alpha'] = 1.35 if opt.dataset == 'MSRVTT' else 1.0
option['q'] = 1
option['q_iterations'] = 1 if opt.use_ct else 0
option['use_ct'] = opt.use_ct
if opt.collect:
prepare_collect_config(option, opt)
if opt.latency:
opt.batch_size = 1
option['batch_size'] = 1
if opt.val_and_test:
modes = ['validate', 'test']
csv_filenames = ['validation_record.csv', 'testing_record.csv']
else:
modes = [opt.evaluation_mode]
csv_filenames = ['validation_record.csv' if opt.evaluation_mode == 'validate' else 'testing_record.csv']
crit = get_criterion_during_evaluation(option)
for mode, csv_filename in zip(modes, csv_filenames):
loader = get_loader(option, mode=mode, print_info=True, specific=opt.specific, batch_size=opt.batch_size)
vocab = loader.dataset.get_vocab()
if opt.record:
summarywriter = SummaryWriter(os.path.join(option['checkpoint_path'], mode))
else:
summarywriter = None
metric = run_eval(option, model, crit, loader, vocab, device,
teacher_model=teacher_model,
dict_mapping=dict_mapping,
json_path=opt.json_path,
json_name=opt.json_name,
print_sent=opt.print_sent,
no_score=opt.no_score,
analyze=True if opt.record else opt.analyze,
collect_best_candidate_iterative_results=True if opt.collect else False,
collect_path=opt.collect_path,
summarywriter=summarywriter,
global_step=option['seed']
)
print(metric)
if opt.record:
fieldsnames = ['Bleu_1', 'Bleu_2', 'Bleu_3', 'Bleu_4',
'METEOR', 'ROUGE_L', 'CIDEr', 'Sum',
'ave_length', 'novel', 'unique', 'usage']
if crit is not None:
fieldsnames += crit.get_fieldsnames()
logger = CsvLogger(filepath=option['checkpoint_path'], filename=csv_filename,
fieldsnames=fieldsnames + opt.field)
if 'loss' in metric:
metric.pop('loss')
for key in opt.field:
metric[key] = option[key]
logger.write(metric)
def main():
'''Main Function'''
parser = argparse.ArgumentParser(description='translate.py')
parser.add_argument(
'-model_path',
nargs='+',
type=str,
default=[
"/home/yangbang/VideoCaptioning/0219save/Youtube2Text/IEL_ARFormer/EBN1_SS0_NDL1_WC0_MI/",
#"/home/yangbang/VideoCaptioning/0219save/MSRVTT/IEL_ARFormer/EBN1_SS0_NDL1_WC20_MI/",
"/home/yangbang/VideoCaptioning/0219save/MSRVTT/IEL_ARFormer/EBN1_SS0_NDL1_WC20_MI_seed920/",
"/home/yangbang/VideoCaptioning/0219save/VATEX/IEL_ARFormer/EBN1_SS1_NDL1_WC0_M/",
"/home/yangbang/VideoCaptioning/0219save/MSRVTT/IEL_ARFormer/EBN1_SS0_NDL1_WC20_MI_seed1314_ag/",
])
parser.add_argument(
'-model_name',
nargs='+',
type=str,
default=[
'0044_240095_254102_253703_251149_247202.pth.tar',
#'0028_177617_180524_183734_183213_182417.pth.tar',
"0011_176183_177176_180332_180729_178864.pth.tar",
'0099_160093_057474.pth.tar',
"0025_179448_180500_184018_184037_182508.pth.tar",
])
parser.add_argument('-i', '--index', default=0, type=int)
parser.add_argument('-beam_size', type=int, default=5, help='Beam size')
parser.add_argument('-beam_alpha', type=float, default=1.0)
parser.add_argument('-batch_size',
type=int,
default=128,
help='Batch size')
parser.add_argument('-topk',
type=int,
default=1,
help="""If verbose is set, will output the n_best
decoded sentences""")
parser.add_argument('-no_cuda', action='store_true')
parser.add_argument('-em', type=str, default='test')
parser.add_argument('-print_sent', action='store_true')
parser.add_argument('-json_path', type=str, default='')
parser.add_argument('-json_name', type=str, default='')
parser.add_argument('-ns', default=False, action='store_true')
parser.add_argument('-sv', default=False, action='store_true')
parser.add_argument('-analyze', default=False, action='store_true')
parser.add_argument('-write_time', default=False, action='store_true')
parser.add_argument('-mid_path', default='best', type=str)
parser.add_argument('-specific', default=-1, type=int)
parser.add_argument('-category', default=False, action='store_true')
parser.add_argument('-sp',
'--saved_with_pickle',
default=False,
action='store_true')
parser.add_argument('-pp',
'--pickle_path',
default='./AR_topk_collect_results')
parser.add_argument('-ca',
'--collect_analyze',
default=False,
action='store_true')
parser.add_argument('-cv', '--cross_validation', type=int, default=2)
opt = parser.parse_args()
opt.model_path = opt.model_path[opt.index]
opt.model_name = opt.model_name[opt.index]
if opt.cross_validation == 1:
source_dataset = 'MSRVTT'
src_pre = 'msrvtt'
src_wct = '2'
target_dataset = 'Youtube2Text'
tgt_pre = 'msvd'
tgt_wct = '0'
else:
source_dataset = 'Youtube2Text'
src_pre = 'msvd'
src_wct = '0'
target_dataset = 'MSRVTT'
tgt_pre = 'msrvtt'
tgt_wct = '2'
opt_pth = os.path.join(opt.model_path, 'opt_info.json')
option = json.load(open(opt_pth, 'r'))
option.update(vars(opt))
if opt.saved_with_pickle:
if not os.path.exists(opt.pickle_path):
os.makedirs(opt.pickle_path)
string = ''
if 'Youtube2Text' in opt.model_path:
dataset_name = 'msvd'
elif 'MSRVTT' in opt.model_path:
dataset_name = 'msrvtt'
if option.get('method', None) == 'ag':
string = '_ag'
opt.pickle_path = os.path.join(
opt.pickle_path, '%s_%d%s.pkl' % (dataset_name, opt.topk, string))
if opt.collect_analyze:
collect_analyze(opt)
else:
device = torch.device('cuda' if not opt.no_cuda else 'cpu')
if opt.analyze:
opt.batch_size = 1
option['batch_size'] = 1
#print(option)
checkpoint = torch.load(
os.path.join(opt.model_path, opt.mid_path, opt.model_name))
#option = checkpoint['settings']
#option.update(vars(opt))
model = get_model(option)
model.load_state_dict(checkpoint['state_dict'])
model.to(device)
if opt.category:
loop_category(option, opt, model, device)
else:
loader = get_loader(option,
mode=opt.em,
print_info=True,
specific=opt.specific)
vocab = loader.dataset.get_vocab()
#print(model)
calculate_novel = True
if opt.cross_validation != 2:
option['dataset'] = target_dataset
option['feats_a'] = []
option['feats_m'] = [
item.replace(source_dataset, target_dataset)
for item in option['feats_m']
]
option['feats_m'] = [
item.replace(src_pre, tgt_pre)
for item in option['feats_m']
]
option['feats_i'] = [
item.replace(source_dataset, target_dataset)
for item in option['feats_i']
]
option['feats_i'] = [
item.replace(src_pre, tgt_pre)
for item in option['feats_i']
]
option['reference'] = option['reference'].replace(
source_dataset, target_dataset)
option['reference'] = option['reference'].replace(
src_pre, tgt_pre)
option['info_corpus'] = option['info_corpus'].replace(
source_dataset, target_dataset)
option['info_corpus'] = option['info_corpus'].replace(
src_wct, tgt_wct)
option['info_corpus'] = option['info_corpus'].replace(
'Youtube0Text', 'Youtube2Text')
loader = get_loader(option,
mode=opt.em,
print_info=True,
specific=opt.specific)
calculate_novel = False
print(len(vocab))
metric = run_eval(option,
model,
None,
loader,
vocab,
device,
json_path=opt.json_path,
json_name=opt.json_name,
print_sent=opt.print_sent,
no_score=opt.ns,
save_videodatainfo=opt.sv,
analyze=opt.analyze,
saved_with_pickle=opt.saved_with_pickle,
pickle_path=opt.pickle_path,
write_time=opt.write_time,
calculate_novel=calculate_novel)
print(metric)
def main(opt):
'''Main Function'''
if opt.collect:
if not os.path.exists(opt.collect_path):
os.makedirs(opt.collect_path)
device = torch.device('cuda' if not opt.no_cuda else 'cpu')
model, option = load(opt.model_path,
opt.model_name,
device,
mid_path='best')
option.update(vars(opt))
set_seed(option['seed'])
if not opt.nt:
#teacher_path = os.path.join(option["checkpoint_path"].replace('NARFormer', 'ARFormer') + '_SS1_0_70')
#teacher_name = 'teacher.pth.tar'
#teacher_model, teacher_option = load(teacher_path, teacher_name, device, mid_path='', from_checkpoint=True)
checkpoint = torch.load(opt.teacher_path)
teacher_option = checkpoint['settings']
teacher_model = get_model(teacher_option)
teacher_model.load_state_dict(checkpoint['state_dict'])
teacher_model.to(device)
assert teacher_option['vocab_size'] == option['vocab_size']
#dict_mapping = get_dict_mapping(option, teacher_option)
dict_mapping = {}
else:
teacher_model = None
dict_mapping = {}
'''
model = get_model(option)
pth = os.path.join(opt.model_path, 'tmp_models')
vali_loader = get_loader(option, mode='validate')
test_loader = get_loader(option, mode='test')
vocab = vali_loader.dataset.get_vocab()
logger = CsvLogger(
filepath=pth,
filename='evaluate.csv',
fieldsnames=['epoch', 'split', 'Bleu_1', 'Bleu_2', 'Bleu_3', 'Bleu_4', 'METEOR', 'ROUGE_L', 'CIDEr', 'Sum', 'lbs', 'i', 'ba']
)
for file in os.listdir(pth):
if '.pth.tar' not in file:
continue
epoch = file.split('_')[1]
checkpoint = torch.load(os.path.join(pth, file))
model.load_state_dict(checkpoint['state_dict'])
model.to(device)
best = 0
best_info = ()
for lbs in range(1, 11):
for i in [1, 3, 5, 10]:
option['length_beam_size'] = lbs
option['iterations'] = i
metric = run_eval(option, model, None, test_loader, vocab, device, json_path=opt.json_path, json_name=opt.json_name, print_sent=opt.print_sent)
metric.pop('loss')
metric['lbs'] = lbs
metric['ba'] = opt.beam_alpha
metric['i'] = i
metric['split'] = 'test'
metric['epoch'] = epoch
logger.write(metric)
if metric['Sum'] > best:
best = metric['Sum']
best_info = (lbs, i)
print(lbs, i, metric['Sum'], best)
'''
'''
# rec length predicted results
rec = {}
for data in tqdm(loader, ncols=150, leave=False):
with torch.no_grad():
results = get_forword_results(option, model, data, device=device, only_data=False)
for i in range(results['pred_length'].size(0)):
res = results['pred_length'][i].topk(5)[1].tolist()
for item in res:
rec[item] = rec.get(item, 0) + 1
for i in range(50):
if i in rec.keys():
print(i, rec[i])
'''
if opt.plot:
plot(option, opt, model, loader, vocab, device, teacher_model,
dict_mapping)
elif opt.loop:
loader = get_loader(option, mode=opt.em, print_info=True)
vocab = loader.dataset.get_vocab()
#loop_iterations(option, opt, model, loader, vocab, device, teacher_model, dict_mapping)
loop_length_beam(option, opt, model, loader, vocab, device,
teacher_model, dict_mapping)
#loop_iterations(option, opt, model, device, teacher_model, dict_mapping)
elif opt.category:
loop_category(option, opt, model, device, teacher_model, dict_mapping)
else:
loader = get_loader(option,
mode=opt.em,
print_info=True,
specific=opt.specific)
vocab = loader.dataset.get_vocab()
filename = '%s_%s_%s_i%db%da%03d%s.pkl' % (
option['dataset'], option['method'],
('%s' % ('AE' if opt.nv_scale == 100 else '')) + opt.paradigm,
opt.iterations, opt.length_beam_size, int(
100 * opt.beam_alpha), '_all' if opt.em == 'all' else '')
metric = run_eval(option,
model,
None,
loader,
vocab,
device,
json_path=opt.json_path,
json_name=opt.json_name,
print_sent=opt.print_sent,
teacher_model=teacher_model,
length_crit=torch.nn.SmoothL1Loss(),
dict_mapping=dict_mapping,
analyze=opt.analyze,
collect_best_candidate_iterative_results=True
if opt.collect else False,
collect_path=os.path.join(opt.collect_path,
filename),
no_score=opt.ns,
write_time=opt.write_time)
#collect_path=os.path.join(opt.collect_path, opt.collect),
print(metric)
def main(opt):
device = torch.device('cuda' if not opt.no_cuda else 'cpu')
opt_pth = os.path.join(opt.model_path, 'opt_info.json')
option = json.load(open(opt_pth, 'r'))
option.update(vars(opt))
#print(option)
model = get_model(option)
checkpoint = torch.load(
os.path.join(opt.model_path, 'best', opt.model_name))
model.load_state_dict(checkpoint['state_dict'])
model.to(device)
loader = get_loader(option,
mode=opt.em,
print_info=False,
specific=opt.specific)
vocab = loader.dataset.get_vocab()
# rec length predicted results
rec = {}
length = len(option['modality']) - sum(option['skip_info'])
num_gates = 3
gate_data = [[[] for _ in range(num_gates)] for __ in range(length)]
opt.pca_name = opt.pca_name + '_%s' % opt.em + (
'_mean' if opt.mean else '') + ('_all' if opt.all else '')
opt.pca_path = os.path.join(opt.pca_path, opt.model_path.split('/')[-2])
print(opt.pca_path)
print(opt.model_path.split('/'))
if not os.path.exists(opt.pca_path):
os.makedirs(opt.pca_path)
if opt.plot:
visualize(opt, option)
elif opt.oe:
metric = run_eval(option,
model,
None,
loader,
vocab,
device,
json_path=opt.json_path,
json_name=opt.json_name,
print_sent=opt.print_sent,
no_score=opt.ns,
save_videodatainfo=opt.sv)
print(metric)
else:
for data in tqdm(loader, ncols=150, leave=False):
with torch.no_grad():
results, _, _ = get_forword_results(option,
model,
data,
device=device,
only_data=True)
gate = results['gate']
assert len(gate) == length
assert len(gate[0]) == num_gates - 1
assert isinstance(results['enc_output'], list)
assert len(results['enc_output']) == length
for i in range(length):
gate[i].append(results['enc_output'][i])
for j in range(num_gates):
gate_data[i][j].append(gate[i][j])
for i in range(length):
for j in range(num_gates):
gate_data[i][j] = torch.cat(
gate_data[i][j],
dim=0) #[len_dataset, n_frames, dim_hidden]
print(i, j, gate_data[i][j][0, 0, :10].tolist())
if i == 0:
data = torch.cat(
[gate_data[0][j],
torch.cat(gate_data[1][j], dim=0)],
dim=0).cpu().numpy()
else:
data = gate_data[i][j].cpu().numpy()
name = '%s_%d_%d.npy' % (opt.pca_name, i, j)
if opt.mean:
data = data.mean(1)
pca = manifold.TSNE(n_components=opt.pca)
#pca = PCA(n_components=opt.pca) #加载PCA算法，设置降维后主成分数目为2
#collect = pca.fit_transform(data) #对样本进行降维
#print(pca.explained_variance_ratio_)
collect = pca.fit_transform(data) #对样本进行降维
elif opt.all:
bsz, seq_len, dim = data.shape
data = data.reshape(bsz * seq_len, dim)
pca = manifold.TSNE(n_components=opt.pca)
collect = pca.fit_transform(data) #对样本进行降维
else:
assert len(data.shape) == 3
seq_len = data.shape[1]
collect = []
for nf in range(seq_len):
x = data[:, nf, :]
pca = manifold.TSNE(n_components=opt.pca)
#pca = PCA(n_components=opt.pca) #加载PCA算法，设置降维后主成分数目为2
reduced_x = pca.fit_transform(x) #对样本进行降维
collect.append(reduced_x)
collect = np.stack(collect, 1)
print(name, collect.shape)
np.save(os.path.join(opt.pca_path, name), collect)
#print('--------')
#print(i, resetgate[i].max(2)[0].max(0)[0].tolist())
#print(i, inputgate[i].max(2)[0].max(0)[0].tolist())
#print('--------')
#print(i, resetgate[i].min(2)[0].min(0)[0].tolist())
#print(i, inputgate[i].min(2)[0].min(0)[0].tolist())
'''