code
stringlengths
17
6.64M
def build_vocab(imgs, params): count_thr = params['word_count_threshold'] counts = {} for img in imgs: for sent in img['sentences']: for w in sent['tokens']: counts[w] = (counts.get(w, 0) + 1) cw = sorted([(count, w) for (w, count) in counts.items()], reverse=True) print('top words and their counts:') print('\n'.join(map(str, cw[:20]))) total_words = sum(counts.values()) print('total words:', total_words) bad_words = [w for (w, n) in counts.items() if (n <= count_thr)] vocab = [w for (w, n) in counts.items() if (n > count_thr)] bad_count = sum((counts[w] for w in bad_words)) print(('number of bad words: %d/%d = %.2f%%' % (len(bad_words), len(counts), ((len(bad_words) * 100.0) / len(counts))))) print(('number of words in vocab would be %d' % (len(vocab),))) print(('number of UNKs: %d/%d = %.2f%%' % (bad_count, total_words, ((bad_count * 100.0) / total_words)))) sent_lengths = {} for img in imgs: for sent in img['sentences']: txt = sent['tokens'] nw = len(txt) sent_lengths[nw] = (sent_lengths.get(nw, 0) + 1) max_len = max(sent_lengths.keys()) print('max length sentence in raw data: ', max_len) print('sentence length distribution (count, number of words):') sum_len = sum(sent_lengths.values()) for i in range((max_len + 1)): print(('%2d: %10d %f%%' % (i, sent_lengths.get(i, 0), ((sent_lengths.get(i, 0) * 100.0) / sum_len)))) if (bad_count > 0): print('inserting the special UNK token') vocab.append('UNK') for img in imgs: img['final_captions'] = [] for sent in img['sentences']: txt = sent['tokens'] caption = [(w if (counts.get(w, 0) > count_thr) else 'UNK') for w in txt] img['final_captions'].append(caption) return vocab
def encode_captions(imgs, params, wtoi): ' \n encode all captions into one large array, which will be 1-indexed.\n also produces label_start_ix and label_end_ix which store 1-indexed \n and inclusive (Lua-style) pointers to the first and last caption for\n each image in the dataset.\n ' max_length = params['max_length'] N = len(imgs) M = sum((len(img['final_captions']) for img in imgs)) label_arrays = [] label_start_ix = np.zeros(N, dtype='uint32') label_end_ix = np.zeros(N, dtype='uint32') label_length = np.zeros(M, dtype='uint32') caption_counter = 0 counter = 1 for (i, img) in enumerate(imgs): n = len(img['final_captions']) assert (n > 0), 'error: some image has no captions' Li = np.zeros((n, max_length), dtype='uint32') for (j, s) in enumerate(img['final_captions']): label_length[caption_counter] = min(max_length, len(s)) caption_counter += 1 for (k, w) in enumerate(s): if (k < max_length): Li[(j, k)] = wtoi[w] label_arrays.append(Li) label_start_ix[i] = counter label_end_ix[i] = ((counter + n) - 1) counter += n L = np.concatenate(label_arrays, axis=0) assert (L.shape[0] == M), "lengths don't match? that's weird" assert np.all((label_length > 0)), 'error: some caption had no words?' print('encoded captions to array of size ', L.shape) return (L, label_start_ix, label_end_ix, label_length)
def main(params): imgs = json.load(open(params['input_json'], 'r')) imgs = imgs['images'] seed(123) vocab = build_vocab(imgs, params) itow = {(i + 1): w for (i, w) in enumerate(vocab)} wtoi = {w: (i + 1) for (i, w) in enumerate(vocab)} (L, label_start_ix, label_end_ix, label_length) = encode_captions(imgs, params, wtoi) N = len(imgs) f_lb = h5py.File((params['output_h5'] + '_label.h5'), 'w') f_lb.create_dataset('labels', dtype='uint32', data=L) f_lb.create_dataset('label_start_ix', dtype='uint32', data=label_start_ix) f_lb.create_dataset('label_end_ix', dtype='uint32', data=label_end_ix) f_lb.create_dataset('label_length', dtype='uint32', data=label_length) f_lb.close() out = {} out['ix_to_word'] = itow out['images'] = [] for (i, img) in enumerate(imgs): jimg = {} jimg['split'] = img['split'] if ('filename' in img): jimg['file_path'] = os.path.join(img.get('filepath', ''), img['filename']) if ('cocoid' in img): jimg['id'] = img['cocoid'] elif ('imgid' in img): jimg['id'] = img['imgid'] if (params['images_root'] != ''): with Image.open(os.path.join(params['images_root'], img['filepath'], img['filename'])) as _img: (jimg['width'], jimg['height']) = _img.size out['images'].append(jimg) json.dump(out, open(params['output_json'], 'w')) print('wrote ', params['output_json'])
def get_doc_freq(refs, params): tmp = CiderScorer(df_mode='corpus') for ref in refs: tmp.cook_append(None, ref) tmp.compute_doc_freq() return (tmp.document_frequency, len(tmp.crefs))
def build_dict(imgs, wtoi, params): wtoi['<eos>'] = 0 count_imgs = 0 refs_words = [] refs_idxs = [] for img in imgs: if ((params['split'] == img['split']) or ((params['split'] == 'train') and (img['split'] == 'restval')) or (params['split'] == 'all')): ref_words = [] ref_idxs = [] for sent in img['sentences']: if hasattr(params, 'bpe'): sent['tokens'] = params.bpe.segment(' '.join(sent['tokens'])).strip().split(' ') tmp_tokens = (sent['tokens'] + ['<eos>']) tmp_tokens = [(_ if (_ in wtoi) else 'UNK') for _ in tmp_tokens] ref_words.append(' '.join(tmp_tokens)) ref_idxs.append(' '.join([str(wtoi[_]) for _ in tmp_tokens])) refs_words.append(ref_words) refs_idxs.append(ref_idxs) count_imgs += 1 print('total imgs:', count_imgs) (ngram_words, count_refs) = get_doc_freq(refs_words, params) (ngram_idxs, count_refs) = get_doc_freq(refs_idxs, params) print('count_refs:', count_refs) return (ngram_words, ngram_idxs, count_refs)
def main(params): imgs = json.load(open(params['input_json'], 'r')) dict_json = json.load(open(params['dict_json'], 'r')) itow = dict_json['ix_to_word'] wtoi = {w: i for (i, w) in itow.items()} if ('bpe' in dict_json): import tempfile import codecs codes_f = tempfile.NamedTemporaryFile(delete=False) codes_f.close() with open(codes_f.name, 'w') as f: f.write(dict_json['bpe']) with codecs.open(codes_f.name, encoding='UTF-8') as codes: bpe = apply_bpe.BPE(codes) params.bpe = bpe imgs = imgs['images'] (ngram_words, ngram_idxs, ref_len) = build_dict(imgs, wtoi, params) utils.pickle_dump({'document_frequency': ngram_words, 'ref_len': ref_len}, open((params['output_pkl'] + '-words.p'), 'wb')) utils.pickle_dump({'document_frequency': ngram_idxs, 'ref_len': ref_len}, open((params['output_pkl'] + '-idxs.p'), 'wb'))
def main(params): imgs = json.load(open(params['input_json'][0], 'r'))['images'] out = {'info': {'description': 'This is stable 1.0 version of the 2014 MS COCO dataset.', 'url': 'http://mscoco.org', 'version': '1.0', 'year': 2014, 'contributor': 'Microsoft COCO group', 'date_created': '2015-01-27 09:11:52.357475'}, 'licenses': [{'url': 'http://creativecommons.org/licenses/by-nc-sa/2.0/', 'id': 1, 'name': 'Attribution-NonCommercial-ShareAlike License'}, {'url': 'http://creativecommons.org/licenses/by-nc/2.0/', 'id': 2, 'name': 'Attribution-NonCommercial License'}, {'url': 'http://creativecommons.org/licenses/by-nc-nd/2.0/', 'id': 3, 'name': 'Attribution-NonCommercial-NoDerivs License'}, {'url': 'http://creativecommons.org/licenses/by/2.0/', 'id': 4, 'name': 'Attribution License'}, {'url': 'http://creativecommons.org/licenses/by-sa/2.0/', 'id': 5, 'name': 'Attribution-ShareAlike License'}, {'url': 'http://creativecommons.org/licenses/by-nd/2.0/', 'id': 6, 'name': 'Attribution-NoDerivs License'}, {'url': 'http://flickr.com/commons/usage/', 'id': 7, 'name': 'No known copyright restrictions'}, {'url': 'http://www.usa.gov/copyright.shtml', 'id': 8, 'name': 'United States Government Work'}], 'type': 'captions'} out.update({'images': [], 'annotations': []}) cnt = 0 empty_cnt = 0 for (i, img) in enumerate(imgs): if (img['split'] == 'train'): continue out['images'].append({'id': img.get('cocoid', img['imgid'])}) for (j, s) in enumerate(img['sentences']): if (len(s) == 0): continue s = ' '.join(s['tokens']) out['annotations'].append({'image_id': out['images'][(- 1)]['id'], 'caption': s, 'id': cnt}) cnt += 1 json.dump(out, open(params['output_json'], 'w')) print('wrote ', params['output_json'])
def test_folder(): x = pickle_load(open('log_trans/infos_trans.pkl', 'rb')) dataset = CaptionDataset(x['opt']) ds = torch.utils.data.Subset(dataset, dataset.split_ix['train']) ds[0]
def test_lmdb(): x = pickle_load(open('log_trans/infos_trans.pkl', 'rb')) x['opt'].input_att_dir = 'data/vilbert_att.lmdb' dataset = CaptionDataset(x['opt']) ds = torch.utils.data.Subset(dataset, dataset.split_ix['train']) ds[0]
def add_summary_value(writer, key, value, iteration): if writer: writer.add_scalar(key, value, iteration)
def train(opt): loader = DataLoader(opt) opt.vocab_size = loader.vocab_size opt.seq_length = loader.seq_length infos = {'iter': 0, 'epoch': 0, 'loader_state_dict': None, 'vocab': loader.get_vocab()} if ((opt.start_from is not None) and os.path.isfile(os.path.join(opt.start_from, (('infos_' + opt.id) + '.pkl')))): with open(os.path.join(opt.start_from, (('infos_' + opt.id) + '.pkl')), 'rb') as f: infos = utils.pickle_load(f) saved_model_opt = infos['opt'] need_be_same = ['caption_model', 'rnn_type', 'rnn_size', 'num_layers'] for checkme in need_be_same: assert (getattr(saved_model_opt, checkme) == getattr(opt, checkme)), ("Command line argument and saved model disagree on '%s' " % checkme) infos['opt'] = opt histories = defaultdict(dict) if ((opt.start_from is not None) and os.path.isfile(os.path.join(opt.start_from, (('histories_' + opt.id) + '.pkl')))): with open(os.path.join(opt.start_from, (('histories_' + opt.id) + '.pkl')), 'rb') as f: histories.update(utils.pickle_load(f)) tb_summary_writer = SummaryWriter(opt.checkpoint_path) opt.vocab = loader.get_vocab() model = models.setup(opt).cuda() del opt.vocab if ((opt.start_from is not None) and os.path.isfile(os.path.join(opt.start_from, 'model.pth'))): model.load_state_dict(torch.load(os.path.join(opt.start_from, 'model.pth'))) lw_model = LossWrapper(model, opt) dp_model = torch.nn.DataParallel(model) dp_model.vocab = getattr(model, 'vocab', None) dp_lw_model = torch.nn.DataParallel(lw_model) if opt.noamopt: assert (opt.caption_model in ['transformer', 'bert', 'm2transformer']), 'noamopt can only work with transformer' optimizer = utils.get_std_opt(model, optim_func=opt.optim, factor=opt.noamopt_factor, warmup=opt.noamopt_warmup) elif opt.reduce_on_plateau: optimizer = utils.build_optimizer(model.parameters(), opt) optimizer = utils.ReduceLROnPlateau(optimizer, factor=opt.reduce_on_plateau_factor, patience=opt.reduce_on_plateau_patience) else: optimizer = utils.build_optimizer(model.parameters(), opt) if ((opt.start_from is not None) and os.path.isfile(os.path.join(opt.start_from, 'optimizer.pth'))): optimizer.load_state_dict(torch.load(os.path.join(opt.start_from, 'optimizer.pth'))) iteration = infos['iter'] epoch = infos['epoch'] if ('iterators' in infos): infos['loader_state_dict'] = {split: {'index_list': infos['split_ix'][split], 'iter_counter': infos['iterators'][split]} for split in ['train', 'val', 'test']} loader.load_state_dict(infos['loader_state_dict']) if (opt.load_best_score == 1): best_val_score = infos.get('best_val_score', None) if opt.noamopt: optimizer._step = iteration epoch_done = True dp_lw_model.train() try: while True: if ((epoch >= opt.max_epochs) and (opt.max_epochs != (- 1))): break if epoch_done: if ((not opt.noamopt) and (not opt.reduce_on_plateau)): if ((epoch > opt.learning_rate_decay_start) and (opt.learning_rate_decay_start >= 0)): frac = ((epoch - opt.learning_rate_decay_start) // opt.learning_rate_decay_every) decay_factor = (opt.learning_rate_decay_rate ** frac) opt.current_lr = (opt.learning_rate * decay_factor) else: opt.current_lr = opt.learning_rate utils.set_lr(optimizer, opt.current_lr) if ((epoch > opt.scheduled_sampling_start) and (opt.scheduled_sampling_start >= 0)): frac = ((epoch - opt.scheduled_sampling_start) // opt.scheduled_sampling_increase_every) opt.ss_prob = min((opt.scheduled_sampling_increase_prob * frac), opt.scheduled_sampling_max_prob) model.ss_prob = opt.ss_prob if ((opt.self_critical_after != (- 1)) and (epoch >= opt.self_critical_after)): sc_flag = True init_scorer(opt.cached_tokens) else: sc_flag = False if ((opt.structure_after != (- 1)) and (epoch >= opt.structure_after)): struc_flag = True init_scorer(opt.cached_tokens) else: struc_flag = False if ((opt.drop_worst_after != (- 1)) and (epoch >= opt.drop_worst_after)): drop_worst_flag = True else: drop_worst_flag = False epoch_done = False start = time.time() if (opt.use_warmup and (iteration < opt.noamopt_warmup)): opt.current_lr = ((opt.learning_rate * (iteration + 1)) / opt.noamopt_warmup) utils.set_lr(optimizer, opt.current_lr) data = loader.get_batch('train') print('Read data:', (time.time() - start)) torch.cuda.synchronize() start = time.time() tmp = [data['fc_feats'], data['att_feats'], data['labels'], data['masks'], data['att_masks']] tmp = [(_ if (_ is None) else _.cuda()) for _ in tmp] (fc_feats, att_feats, labels, masks, att_masks) = tmp optimizer.zero_grad() model_out = dp_lw_model(fc_feats, att_feats, labels, masks, att_masks, data['gts'], torch.arange(0, len(data['gts'])), sc_flag, struc_flag, drop_worst_flag) if (not drop_worst_flag): loss = model_out['loss'].mean() else: loss = model_out['loss'] loss = torch.topk(loss, k=int((loss.shape[0] * (1 - opt.drop_worst_rate))), largest=False)[0].mean() loss.backward() if (opt.grad_clip_value != 0): getattr(torch.nn.utils, ('clip_grad_%s_' % opt.grad_clip_mode))(model.parameters(), opt.grad_clip_value) optimizer.step() train_loss = loss.item() torch.cuda.synchronize() end = time.time() if struc_flag: print('iter {} (epoch {}), train_loss = {:.3f}, lm_loss = {:.3f}, struc_loss = {:.3f}, time/batch = {:.3f}'.format(iteration, epoch, train_loss, model_out['lm_loss'].mean().item(), model_out['struc_loss'].mean().item(), (end - start))) elif (not sc_flag): print('iter {} (epoch {}), train_loss = {:.3f}, time/batch = {:.3f}'.format(iteration, epoch, train_loss, (end - start))) else: print('iter {} (epoch {}), avg_reward = {:.3f}, time/batch = {:.3f}'.format(iteration, epoch, model_out['reward'].mean(), (end - start))) iteration += 1 if data['bounds']['wrapped']: epoch += 1 epoch_done = True if ((iteration % opt.losses_log_every) == 0): tb_summary_writer.add_scalar('train_loss', train_loss, iteration) if opt.noamopt: opt.current_lr = optimizer.rate() elif opt.reduce_on_plateau: opt.current_lr = optimizer.current_lr tb_summary_writer.add_scalar('learning_rate', opt.current_lr, iteration) tb_summary_writer.add_scalar('scheduled_sampling_prob', model.ss_prob, iteration) if sc_flag: tb_summary_writer.add_scalar('avg_reward', model_out['reward'].mean(), iteration) elif struc_flag: tb_summary_writer.add_scalar('lm_loss', model_out['lm_loss'].mean().item(), iteration) tb_summary_writer.add_scalar('struc_loss', model_out['struc_loss'].mean().item(), iteration) tb_summary_writer.add_scalar('reward', model_out['reward'].mean().item(), iteration) tb_summary_writer.add_scalar('reward_var', model_out['reward'].var(1).mean(), iteration) histories['loss_history'][iteration] = (train_loss if (not sc_flag) else model_out['reward'].mean()) histories['lr_history'][iteration] = opt.current_lr histories['ss_prob_history'][iteration] = model.ss_prob infos['iter'] = iteration infos['epoch'] = epoch infos['loader_state_dict'] = loader.state_dict() if ((((iteration % opt.save_checkpoint_every) == 0) and (not opt.save_every_epoch)) or (epoch_done and opt.save_every_epoch)): eval_kwargs = {'split': 'val', 'dataset': opt.input_json} eval_kwargs.update(vars(opt)) (val_loss, predictions, lang_stats) = eval_utils.eval_split(dp_model, lw_model.crit, loader, eval_kwargs) if opt.reduce_on_plateau: if ('CIDEr' in lang_stats): optimizer.scheduler_step((- lang_stats['CIDEr'])) else: optimizer.scheduler_step(val_loss) tb_summary_writer.add_scalar('validation loss', val_loss, iteration) if (lang_stats is not None): for (k, v) in lang_stats.items(): tb_summary_writer.add_scalar(k, v, iteration) histories['val_result_history'][iteration] = {'loss': val_loss, 'lang_stats': lang_stats, 'predictions': predictions} if (opt.language_eval == 1): current_score = lang_stats['CIDEr'] else: current_score = (- val_loss) best_flag = False if ((best_val_score is None) or (current_score > best_val_score)): best_val_score = current_score best_flag = True infos['best_val_score'] = best_val_score utils.save_checkpoint(opt, model, infos, optimizer, histories) if opt.save_history_ckpt: utils.save_checkpoint(opt, model, infos, optimizer, append=(str(epoch) if opt.save_every_epoch else str(iteration))) if best_flag: utils.save_checkpoint(opt, model, infos, optimizer, append='best') except (RuntimeError, KeyboardInterrupt): print('Save ckpt on exception ...') utils.save_checkpoint(opt, model, infos, optimizer) print('Save ckpt done.') stack_trace = traceback.format_exc() print(stack_trace)
class Dataset(torch.utils.data.Dataset): def __init__(self, treated_patient_list, control_patient_list, diag_code_vocab=None, med_code_vocab=None): self.treated_patient_list = treated_patient_list self.control_patient_list = control_patient_list self.diagnoses_visits = [] self.medication_visits = [] self.sexes = [] self.ages = [] self.outcome = [] self.treatment = [] for (_, patient_confounder, patient_outcome) in tqdm(self.treated_patient_list): self.outcome.append(patient_outcome) self.treatment.append(1) (med_visit, diag_visit, age, sex) = patient_confounder self.medication_visits.append(med_visit) self.diagnoses_visits.append(diag_visit) self.sexes.append(sex) self.ages.append(age) for (_, patient_confounder, patient_outcome) in tqdm(self.control_patient_list): self.outcome.append(patient_outcome) self.treatment.append(0) (med_visit, diag_visit, age, sex) = patient_confounder self.medication_visits.append(med_visit) self.diagnoses_visits.append(diag_visit) self.sexes.append(sex) self.ages.append(age) if (diag_code_vocab is None): self.diag_code_vocab = CodeVocab() self.diag_code_vocab.add_patients_visits(self.diagnoses_visits) if (med_code_vocab is None): self.med_code_vocab = CodeVocab() self.med_code_vocab.add_patients_visits(self.medication_visits) logger.info(('Created Diagnoses Vocab: %s' % self.diag_code_vocab)) logger.info(('Created Medication Vocab: %s' % self.med_code_vocab)) self.diag_visit_max_length = max([len(patient_visit) for patient_visit in self.diagnoses_visits]) self.med_visit_max_length = max([len(patient_visit) for patient_visit in self.medication_visits]) self.diag_vocab_length = len(self.diag_code_vocab) self.med_vocab_length = len(self.med_code_vocab) logger.info(('Diagnoses Visit Max Length: %d' % self.diag_visit_max_length)) logger.info(('Medication Visit Max Length: %d' % self.med_visit_max_length)) self.ages = ((self.ages - np.min(self.ages)) / np.ptp(self.ages)) def _process_visits(self, visits, max_len_visit, vocab): res = np.zeros((max_len_visit, len(vocab))) for (i, visit) in enumerate(visits): res[i] = self._process_code(vocab, visit) return res def _process_code(self, vocab, codes): multi_hot = np.zeros(len(vocab), dtype='float') for code in codes: multi_hot[vocab.code2id[code]] = 1 return multi_hot def __getitem__(self, index): diag = self.diagnoses_visits[index] diag = self._process_visits(diag, self.diag_visit_max_length, self.diag_code_vocab) med = self.medication_visits[index] med = self._process_visits(med, self.med_visit_max_length, self.med_code_vocab) sex = self.sexes[index] age = self.ages[index] outcome = self.outcome[index] treatment = self.treatment[index] confounder = (diag, med, sex, age) return (confounder, treatment, outcome) def __len__(self): return len(self.diagnoses_visits)
class LSTMModel(torch.nn.Module): def __init__(self, diag_vocab_size, med_vocab_size, diag_embedding_size, med_embedding_size, diag_hidden_size, med_hidden_size, hidden_size, end_index, pad_index, bidirectional=True): super().__init__() self.pad_index = pad_index self.end_index = end_index self.diag_embedding = torch.nn.Linear(diag_vocab_size, diag_embedding_size, bias=False) self.med_embedding = torch.nn.Linear(med_vocab_size, med_embedding_size, bias=False) self.diag_encoder = torch.nn.LSTM(diag_embedding_size, diag_hidden_size, batch_first=True, bidirectional=bidirectional) self.med_encoder = torch.nn.LSTM(med_embedding_size, med_hidden_size, batch_first=True, bidirectional=bidirectional) if bidirectional: diag_hidden_size = (diag_hidden_size * 2) med_hidden_size = (med_hidden_size * 2) self.attention_diag_encoder = torch.nn.Sequential(torch.nn.Linear(diag_hidden_size, 1), torch.nn.Tanh()) self.attention_med_encoder = torch.nn.Sequential(torch.nn.Linear(med_hidden_size, 1), torch.nn.Tanh()) self.lstm2hidden_ipw = torch.nn.Sequential(torch.nn.Linear(((med_hidden_size + diag_hidden_size) + 2), hidden_size), torch.nn.ReLU()) self.hidden2out_ipw = torch.nn.Linear(hidden_size, 1, bias=False) def softmax_masked(self, inputs, mask, dim=1, epsilon=1e-07): inputs_exp = torch.exp(inputs) inputs_exp = (inputs_exp * mask.float()) inputs_exp_sum = inputs_exp.sum(dim=dim) inputs_attention = (inputs_exp / (inputs_exp_sum.unsqueeze(dim) + epsilon)) return inputs_attention def diag_encode(self, inputs): inputs_mask = (inputs.sum(dim=(- 1)) != 0).long() inputs_emb = self.diag_embedding(inputs.float()) (outputs, (h, c)) = self.diag_encoder(inputs_emb) att_enc = self.attention_diag_encoder(outputs).squeeze((- 1)) att_normalized = self.softmax_masked(att_enc, inputs_mask) hidden = torch.sum((outputs * att_normalized.unsqueeze((- 1))), dim=1) original = torch.sum((inputs.float() * att_normalized.unsqueeze((- 1))), dim=1) return (hidden, original) def med_encode(self, inputs): inputs_mask = (inputs.sum(dim=(- 1)) != 0).long() inputs_emb = self.med_embedding(inputs.float()) (outputs, (h, c)) = self.med_encoder(inputs_emb) att_enc = self.attention_med_encoder(outputs).squeeze((- 1)) att_normalized = self.softmax_masked(att_enc, inputs_mask) hidden = torch.sum((outputs * att_normalized.unsqueeze((- 1))), dim=1) original = torch.sum((inputs.float() * att_normalized.unsqueeze((- 1))), dim=1) return (hidden, original) def forward(self, confounder): (diag_inputs, med_inputs, sexes, ages) = confounder (diag_hidden, diag_original) = self.diag_encode(diag_inputs) (med_hidden, med_original) = self.med_encode(med_inputs) original = torch.cat((diag_original, med_original, sexes.float().view(sexes.size(0), 1), ages.float().view(ages.size(0), 1)), dim=1) hidden = torch.cat((diag_hidden, med_hidden, sexes.float().view(sexes.size(0), 1), ages.float().view(ages.size(0), 1)), dim=1) hidden = self.lstm2hidden_ipw(hidden) outputs_logits_ipw = self.hidden2out_ipw(hidden) return (outputs_logits_ipw.view(outputs_logits_ipw.size(0)), original)
class CodeVocab(object): END_CODE = '<end>' PAD_CODE = '<pad>' UNK_CODE = '<unk>' def __init__(self): super().__init__() special_codes = [CodeVocab.END_CODE, CodeVocab.PAD_CODE, CodeVocab.UNK_CODE] self.special_codes = special_codes self.code2id = {} self.id2code = {} if (self.special_codes is not None): self.add_code_list(self.special_codes) def add_code_list(self, code_list, rebuild=True): for code in code_list: if (code not in self.code2id): self.code2id[code] = len(self.code2id) if rebuild: self._rebuild_id2code() def add_patients_visits(self, patients_visits): for patient in patients_visits: for visit in patient: self.add_code_list(visit) self._rebuild_id2code() def _rebuild_id2code(self): self.id2code = {i: t for (t, i) in self.code2id.items()} def get(self, item, default=None): return self.code2id.get(item, default) def __getitem__(self, item): return self.code2id[item] def __contains__(self, item): return (item in self.code2id) def __len__(self): return len(self.code2id) def __str__(self): return f'{len(self)} codes'
def get_patient_cohort(root_file): patient_1stDX_date = {} patient_start_date = {} for dir in ['CAD2012', 'CAD2013-2016']: file = ((root_file + dir) + '/Cohort.csv') with open(file, 'r') as f: next(f) for row in f: row = row.split(',') (enrolid, dx_date, start_date) = (row[0], row[3], row[4]) patient_1stDX_date[enrolid] = datetime.strptime(dx_date, '%m/%d/%Y') patient_start_date[enrolid] = datetime.strptime(start_date, '%m/%d/%Y') my_dump(patient_1stDX_date, '../pickles/patient_1stDX_data.pkl') my_dump(patient_start_date, '../pickles/patient_start_date.pkl') return (patient_1stDX_date, patient_start_date)
def exclude(cad_prescription_taken_by_patient, patient_1stDX_date, patient_start_date, interval, followup, baseline): cad_prescription_taken_by_patient_exclude = defaultdict(dict) cad_patient_take_prescription_exclude = defaultdict(dict) for (drug, taken_by_patient) in cad_prescription_taken_by_patient.items(): for (patient, take_times) in taken_by_patient.items(): dates = [datetime.strptime(date, '%m/%d/%Y') for (date, days) in take_times if (date and days)] dates = sorted(dates) dates_days = {datetime.strptime(date, '%m/%d/%Y'): int(days) for (date, days) in take_times if (date and days)} DX = patient_1stDX_date.get(patient, datetime.max) index_date = dates[0] start_date = patient_start_date.get(patient, datetime.max) if (criteria_1_is_valid(index_date, DX) and criteria_2_is_valid(dates, interval, followup, dates_days) and criteria_3_is_valid(index_date, start_date, baseline)): cad_prescription_taken_by_patient_exclude[drug][patient] = dates cad_patient_take_prescription_exclude[patient][drug] = dates return (cad_prescription_taken_by_patient_exclude, cad_patient_take_prescription_exclude)
def criteria_1_is_valid(index_date, DX): return ((index_date - DX).days > 0)
def criteria_2_is_valid(dates, interval, followup, dates_days): if ((dates[(- 1)] - dates[0]).days <= (followup - 89)): return False for i in range(1, len(dates)): sup_day = dates_days.get(dates[(i - 1)]) if (((dates[i] - dates[(i - 1)]).days - sup_day) > interval): return False return True
def criteria_3_is_valid(index_date, start_date, baseline): return ((index_date - start_date).days >= baseline)
def user_cohort_extractor(cad_prescription_taken_by_patient, n_patients, n_prescriptions, time_interval): cad_prescription_taken_by_patient_small = defaultdict(dict) print('number of drugs: {}'.format(len(cad_prescription_taken_by_patient)), flush=True) for (drug, patient_take_times) in cad_prescription_taken_by_patient.items(): patient_take_times = cad_prescription_taken_by_patient.get(drug) if minimal_criteria_is_valid(patient_take_times, n_patients, time_interval, n_prescriptions): cad_prescription_taken_by_patient[drug] = patient_take_times return cad_prescription_taken_by_patient_small
def minimal_criteria_is_valid(patient_take_times, n_patients, time_interval, n_prescriptions): if (len(patient_take_times) < n_patients): return False count = 0 for (patient, take_times) in patient_take_times.items(): if drug_time_interval_is_valid(take_times, n_prescriptions, time_interval): count += 1 if (count >= n_patients): return True return False
def drug_time_interval_is_valid(take_times, n_prescription, time_interval): count = 0 dates = [datetime.strptime(pair[0], '%m/%d/%Y') for pair in take_times if (pair[0] and pair[1])] dates = sorted(dates) for i in range(1, len(dates)): if ((dates[i] - dates[(i - 1)]).days >= time_interval): count += 2 if (count >= n_prescription): return True return False
def my_dump(obj, file_name): print('dumping...', flush=True) pickle.dump(obj, open(file_name, 'wb')) print('dumped...', flush=True)
def my_load(file_name): print('loading...', flush=True) return pickle.load(open(file_name, 'rb'))
def get_user_dx(indir, patient_list, icd9_to_css, icd10_to_css): user_dx = defaultdict(dict) inpatient_dir = os.path.join(indir, 'inpatient') inpatient_files = os.listdir(inpatient_dir) outpatient_dir = os.path.join(indir, 'outpatient') outpatient_files = os.listdir(outpatient_dir) files = ([os.path.join(inpatient_dir, file) for file in inpatient_files] + [os.path.join(outpatient_dir, file) for file in outpatient_files]) DXVER_dict = {'9': icd9_to_css, '0': icd10_to_css} for file in files: inpat = pd.read_csv(file, dtype=str) DATE_NAME = [col for col in inpat.columns if ('DATE' in col)][0] inpat = inpat[inpat['ENROLID'].isin(list(patient_list))] inpat = inpat[(~ inpat[DATE_NAME].isnull())] DX_col = [col for col in inpat.columns if ('DX' in col)] for (index, row) in tqdm(inpat.iterrows(), total=len(inpat)): dxs = list(row[DX_col]) enrolid = row['ENROLID'] date = row[DATE_NAME] DXVER = '9' if dxs: if ('DXVER' in inpat.columns): DXVER = row['DXVER'] dxs = get_css_code_for_icd(dxs, DXVER_dict[DXVER]) if (enrolid not in user_dx): user_dx[enrolid][date] = dxs elif (date not in user_dx[enrolid]): user_dx[enrolid][date] = dxs else: user_dx[enrolid][date].extend(dxs) return user_dx
def get_css_code_for_icd(icd_codes, icd_to_css): css_codes = [] for icd_code in icd_codes: if (not pd.isnull(icd_code)): for i in range(len(icd_code), (- 1), (- 1)): if (icd_code[:i] in icd_to_css): css_codes.append(icd_to_css.get(icd_code[:i])) break return css_codes
def pre_user_cohort_dx(user_dx, cad_prescription_taken_by_patient, min_patients): user_cohort_dx = AutoVivification() for (drug, taken_by_patient) in tqdm(cad_prescription_taken_by_patient.items()): if (len(taken_by_patient.keys()) >= min_patients): for (patient, taken_times) in taken_by_patient.items(): index_date = taken_times[0] date_codes = user_dx.get(patient) for (date, codes) in date_codes.items(): date = datetime.strptime(date, '%m/%d/%Y') if (date < index_date): if (drug not in user_cohort_dx): user_cohort_dx[drug][patient][date] = set(codes) elif (patient not in user_cohort_dx[drug]): user_cohort_dx[drug][patient][date] = set(codes) elif (date not in user_cohort_dx[drug][patient]): user_cohort_dx[drug][patient][date] = set(codes) else: user_cohort_dx[drug][patient][date].union(codes) return user_cohort_dx
class AutoVivification(dict): "Implementation of perl's autovivification feature." def __getitem__(self, item): try: return dict.__getitem__(self, item) except KeyError: value = self[item] = type(self)() return value
def get_user_cohort_dx(indir, cad_prescription_taken_by_patient, icd9_to_css, icd10_to_css, min_patient, patient_list): user_dx = get_user_dx(indir, patient_list, icd9_to_css, icd10_to_css) return pre_user_cohort_dx(user_dx, cad_prescription_taken_by_patient, min_patient)
def pre_user_cohort_rx(cad_prescription_taken_by_patient, cad_patient_take_prescription, min_patients): cad_user_cohort_rx = defaultdict(dict) for (drug, taken_by_patient) in tqdm(cad_prescription_taken_by_patient.items()): if (len(taken_by_patient.keys()) >= min_patients): for (patient, take_dates) in taken_by_patient.items(): index_date = take_dates[0] patient_prescription_list = cad_patient_take_prescription.get(patient) for (prescription, dates_days) in patient_prescription_list.items(): dates = [datetime.strptime(date, '%m/%d/%Y') for (date, days) in dates_days] dates = sorted(dates) if drug_is_taken_in_baseline(index_date, dates): if (drug not in cad_user_cohort_rx): cad_user_cohort_rx[drug][patient] = [prescription] elif (patient not in cad_user_cohort_rx[drug]): cad_user_cohort_rx[drug][patient] = [prescription] else: cad_user_cohort_rx[drug][patient].append(prescription) return cad_user_cohort_rx
def get_prescription_taken_times(index_date, dates, dates_2_days): cnt = 0 for date in dates: if (((index_date - date).days - dates_2_days[date]) > 0): cnt += 1 else: return cnt return cnt
def drug_is_taken_in_baseline(index_date, dates): for date in dates: if ((index_date - date).days > 0): return True return False
def pre_user_cohort_rx_v2(cad_prescription_taken_by_patient, cad_patient_take_prescription, min_patients): cad_user_cohort_rx = AutoVivification() for (drug, taken_by_patient) in tqdm(cad_prescription_taken_by_patient.items()): if (len(taken_by_patient.keys()) >= min_patients): for (patient, take_dates) in taken_by_patient.items(): index_date = take_dates[0] patient_prescription_list = cad_patient_take_prescription.get(patient) for (prescription, dates_days) in patient_prescription_list.items(): dates = sorted(dates_days) dates = drug_is_taken_in_baseline_v2(index_date, dates) if dates: for date in dates: if (drug not in cad_user_cohort_rx): cad_user_cohort_rx[drug][patient][date] = [prescription] elif (patient not in cad_user_cohort_rx[drug]): cad_user_cohort_rx[drug][patient][date] = [prescription] elif (date not in cad_user_cohort_rx[drug][patient]): cad_user_cohort_rx[drug][patient][date] = [prescription] else: cad_user_cohort_rx[drug][patient][date].append(prescription) return cad_user_cohort_rx
def drug_is_taken_in_baseline_v2(index_date, dates): res = [] for date in dates: if ((index_date - date).days > 0): res.append(date) if (len(res) > 0): return res return False
class AutoVivification(dict): "Implementation of perl's autovivification feature." def __getitem__(self, item): try: return dict.__getitem__(self, item) except KeyError: value = self[item] = type(self)() return value
def pre_user_cohort_demo(indir, patient_list): cad_user_cohort_demo = {} file = '{}/demo.csv'.format(indir) with open(file, 'r') as f: next(f) for row in f: row = row.split(',') (id, db, sex) = (row[0], row[1], row[2]) if (id in patient_list): cad_user_cohort_demo[id] = (db, sex) return cad_user_cohort_demo
def get_user_cohort_demo(indir, patient_list): return pre_user_cohort_demo(indir, patient_list)
def parse_args(): parser = argparse.ArgumentParser(description='process parameters') parser.add_argument('--input_data_dir', default='../data/synthetic/drug', help='input data directory') parser.add_argument('--output_data_dir', default='pickles/cad_prescription_taken_by_patient.pkl', help='output data directory') args = parser.parse_args() return args
def ndc2rxing(): mapping = np.loadtxt(fname='../data/NDC_complete_mapping.csv', delimiter=',', dtype='str', skiprows=1, usecols=(1, 2)) ndc2rx_mapping = {ndc: rx for (rx, ndc) in mapping} return ndc2rx_mapping
def pre_drug_table(args): cad_prescription_taken_by_patient = defaultdict(dict) ndc2rx_mapping = ndc2rxing() files = os.listdir(args.input_data_dir) for file in files: print('dir: {}\tfile: {}'.format(args.input_data_dir, file), flush=True) df = os.path.join(args.input_data_dir, file) with open(df, 'r') as f: next(f) for row in f: row = row.strip('\n') row = row.split(',') (enroll_id, prescription, date, day) = (row[0], row[1], row[2], row[3]) if (date and day): if (prescription in ndc2rx_mapping): prescription_rx = ndc2rx_mapping.get(prescription) if (prescription_rx not in cad_prescription_taken_by_patient): cad_prescription_taken_by_patient[prescription_rx][enroll_id] = set([(date, day)]) elif (enroll_id not in cad_prescription_taken_by_patient.get(prescription_rx)): cad_prescription_taken_by_patient[prescription_rx][enroll_id] = set([(date, day)]) else: cad_prescription_taken_by_patient[prescription_rx][enroll_id].add((date, day)) try: print('dumping...', flush=True) pickle.dump(cad_prescription_taken_by_patient, open(args.output_data_dir, 'wb')) except Exception as e: print(e) print('finish dump', flush=True) print('# of Drugs: {}\t'.format(len(cad_prescription_taken_by_patient))) return cad_prescription_taken_by_patient
def drug_time_interval_is_valid(take_times, n_prescription, time_interval): count = 0 dates = [datetime.strptime(pair[0], '%m/%d/%Y') for pair in take_times if (pair[0] and pair[1])] dates = sorted(dates) for i in range(1, len(dates)): if ((dates[i] - dates[(i - 1)]).days >= time_interval): count += 1 if (count >= n_prescription): return True return False
def is_valid_outcome_range(dx, code_range): for code in code_range: if dx.startswith(code): return True return False
def pre_user_cohort_outcome(indir, patient_list, codes9, codes0): cad_user_cohort_outcome = defaultdict(list) inpatient_dir = os.path.join(indir, 'inpatient') inpatient_files = os.listdir(inpatient_dir) outpatient_dir = os.path.join(indir, 'outpatient') outpatient_files = os.listdir(outpatient_dir) files = ([os.path.join(inpatient_dir, file) for file in inpatient_files] + [os.path.join(outpatient_dir, file) for file in outpatient_files]) DXVER_dict = {'9': codes9, '0': codes0} for file in files: inpat = pd.read_csv(file, dtype=str) DATE_NAME = [col for col in inpat.columns if ('DATE' in col)][0] inpat = inpat[inpat['ENROLID'].isin(list(patient_list))] inpat = inpat[(~ inpat[DATE_NAME].isnull())] DX_col = [col for col in inpat.columns if ('DX' in col)] for (index, row) in tqdm(inpat.iterrows(), total=len(inpat)): dxs = list(row[DX_col]) enrolid = row['ENROLID'] date = row[DATE_NAME] DXVER = '9' if dxs: if ('DXVER' in inpat.columns): DXVER = row['DXVER'] for dx in dxs: if (not pd.isnull(dx)): if is_valid_outcome_range(dx, DXVER_dict[DXVER]): cad_user_cohort_outcome[enrolid].append(date) return cad_user_cohort_outcome
def parse_args(): parser = argparse.ArgumentParser(description='process parameters') parser.add_argument('--min_patients', default=500, type=int, help='minimum number of patients for each cohort.') parser.add_argument('--min_prescription', default=2, type=int, help='minimum times of prescriptions of each drug.') parser.add_argument('--time_interval', default=30, type=int, help='minimum time interval for every two prescriptions') parser.add_argument('--followup', default=730, type=int, help='number of days of followup period') parser.add_argument('--baseline', default=365, type=int, help='number of days of baseline period') parser.add_argument('--input_data', default='../data/CAD') parser.add_argument('--pickles', default='pickles') parser.add_argument('--outcome_icd9', default='outcome_icd9.txt', help='outcome definition with ICD-9 codes') parser.add_argument('--outcome_icd10', default='outcome_icd10.txt', help='outcome definition with ICD-10 codes') parser.add_argument('--save_cohort_all', default='save_cohort_all/') args = parser.parse_args() return args
def get_patient_list(min_patient, cad_prescription_taken_by_patient): patients_list = set() for (drug, patients) in cad_prescription_taken_by_patient.items(): if (len(patients) >= min_patient): for patient in patients: patients_list.add(patient) return patients_list
def main(args): print('Loading prescription data...') cad_prescription_taken_by_patient = pickle.load(open(os.path.join(args.pickles, 'cad_prescription_taken_by_patient.pkl'), 'rb')) (patient_1stDX_date, patient_start_date) = get_patient_init_date(args.input_data, args.pickles) icd9_to_css = pickle.load(open(os.path.join(args.pickles, 'icd9_to_css.pkl'), 'rb')) icd10_to_css = pickle.load(open(os.path.join(args.pickles, 'icd10_to_css.pkl'), 'rb')) print('Preprocessing patient data...') (save_prescription, save_patient) = exclude(cad_prescription_taken_by_patient, patient_1stDX_date, patient_start_date, args.time_interval, args.followup, args.baseline) patient_list = get_patient_list(args.min_patients, save_prescription) save_cohort_rx = pre_user_cohort_rx_v2(save_prescription, save_patient, args.min_patients) save_cohort_dx = get_user_cohort_dx(args.input_data, save_prescription, icd9_to_css, icd10_to_css, args.min_patients, patient_list) save_cohort_demo = get_user_cohort_demo(args.input_data, patient_list) save_cohort_outcome = {} codes9 = ['425', '428', '40201', '40211', '40291', '40401', '40403', '40411', '40413', '40491', '40493', 'K77'] codes0 = ['I11', 'I13', 'I50', 'I42', 'K77'] save_cohort_outcome['heart-failure'] = pre_user_cohort_outcome(args.input_data, patient_list, codes9, codes0) codes9 = ['4380', '4381', '4382', '4383', '4384', '4385', '4386', '4387', '4388', '4389', 'V1254'] codes0 = ['Z8673', 'I60', 'I61', 'I62', 'I63', 'I64', 'I65', 'I66', 'I67', 'I68', 'I69', 'G458', 'G459'] save_cohort_outcome['stroke'] = pre_user_cohort_outcome(args.input_data, patient_list, codes9, codes0) print('Generating patient cohort...') pre_user_cohort_triplet(save_prescription, save_cohort_rx, save_cohort_dx, save_cohort_outcome, save_cohort_demo, args.save_cohort_all)
def pre_user_cohort_triplet(cad_prescription_taken_by_patient, cad_user_cohort_rx, cad_user_cohort_dx, save_cohort_outcome, cad_user_cohort_demo, out_file_root): cohorts_size = dict() for (drug, taken_by_patient) in tqdm(cad_user_cohort_rx.items()): file_x = '{}/{}.pkl'.format(out_file_root, drug) triples = [] for (patient, taken_times) in taken_by_patient.items(): index_date = cad_prescription_taken_by_patient.get(drug).get(patient)[0] dx = cad_user_cohort_dx.get(drug).get(patient) demo = cad_user_cohort_demo.get(patient) demo_feature_vector = get_demo_feature_vector(demo, index_date) outcome_feature_vector = [] for (outcome_name, outcome_map) in save_cohort_outcome.items(): outcome_dates = outcome_map.get(patient, []) dates = [datetime.strptime(date.strip('\n'), '%m/%d/%Y') for date in outcome_dates if date] dates = sorted(dates) outcome_feature_vector.append(get_outcome_feature_vector(dates, index_date)) outcome = max(outcome_feature_vector) (rx_codes, dx_codes) = ([], []) if taken_times: rx_codes = [rx_code for (date, rx_code) in sorted(taken_times.items(), key=(lambda x: x[0]))] if dx: dx_codes = [list(dx_code) for (date, dx_code) in sorted(dx.items(), key=(lambda x: x[0]))] triple = (patient, [rx_codes, dx_codes, demo_feature_vector[0], demo_feature_vector[1]], outcome) triples.append(triple) cohorts_size['{}.pkl'.format(drug)] = len(triples) pickle.dump(triples, open(file_x, 'wb')) pickle.dump(cohorts_size, open('{}/cohorts_size.pkl'.format(out_file_root), 'wb'))
def get_outcome_feature_vector(dates, index_date): for date in dates: if ((date > index_date) and ((date - index_date).days <= 730)): return 1 return 0
def get_rx_feature_vector(taken_times, RX2id, size): feature_vector = ([0] * size) for rx in taken_times: if (rx in RX2id): id = RX2id.get(rx) feature_vector[id] = 1 return feature_vector
def get_dx_feature_vector(dx, CCS2id, size): feature_vector = ([0] * size) not_find = set() for code in dx: for c in code: if (c in CCS2id): id = CCS2id.get(c) feature_vector[id] = 1 return (feature_vector, not_find)
def get_demo_feature_vector(demo, index_date): if (not demo): return [0, 0] (db, sex) = demo index_date_y = index_date.year age = (index_date_y - int(db)) sex_n = (int(sex) - 1) return [age, sex_n]
def get_patient_init_date(indir, outdir): patient_1stDX_date = {} patient_start_date = {} file = '{}/Cohort.csv'.format(indir) with open(file, 'r') as f: next(f) for row in f: row = row.split(',') (enrolid, dx_date, start_date) = (row[0], row[1], row[2]) patient_1stDX_date[enrolid] = datetime.strptime(dx_date, '%m/%d/%Y') patient_start_date[enrolid] = datetime.strptime(start_date, '%m/%d/%Y') out1 = '{}/patient_1stDX_data.pkl'.format(outdir) pickle.dump(patient_1stDX_date, open(out1, 'wb')) out2 = '{}/patient_start_date.pkl'.format(outdir) pickle.dump(patient_start_date, open(out2, 'wb')) return (patient_1stDX_date, patient_start_date)
class Dataset(InMemoryDataset): def __init__(self, root, dataset, rating_file, sep, args, transform=None, pre_transform=None): self.path = root self.dataset = dataset self.rating_file = rating_file self.sep = sep self.store_backup = True self.args = args super(Dataset, self).__init__(root, transform, pre_transform) (self.data, self.slices) = torch.load(self.processed_paths[0]) self.stat_info = torch.load(self.processed_paths[1]) self.data_num = self.stat_info['data_num'] self.feature_num = self.stat_info['feature_num'] @property def raw_file_names(self): return ['{}{}/user_dict.pkl'.format(self.path, self.dataset), '{}{}/item_dict.pkl'.format(self.path, self.dataset), '{}{}/feature_dict.pkl'.format(self.path, self.dataset), '{}{}/{}'.format(self.path, self.dataset, self.rating_file)] @property def processed_file_names(self): return ['{}/{}.dataset'.format(self.dataset, self.dataset), '{}/{}.statinfo'.format(self.dataset, self.dataset)] def download(self): pass def data_2_graphs(self, ratings_df, dataset='train'): graphs = [] processed_graphs = 0 num_graphs = ratings_df.shape[0] one_per = int((num_graphs / 1000)) percent = 0.0 for i in range(len(ratings_df)): if ((processed_graphs % one_per) == 0): print(f'Processing [{dataset}]: {(percent / 10.0)}%, {processed_graphs}/{num_graphs}', end='\r') percent += 1 processed_graphs += 1 line = ratings_df.iloc[i] user_index = self.user_key_type(line[0]) item_index = self.item_key_type(line[1]) rating = int(line[2]) if ((item_index not in self.item_dict) or (user_index not in self.user_dict)): error_num += 1 continue user_id = self.user_dict[user_index]['name'] item_id = self.item_dict[item_index]['title'] user_attr_list = self.user_dict[user_index]['attribute'] item_attr_list = self.item_dict[item_index]['attribute'] user_list = ([user_id] + user_attr_list) item_list = ([item_id] + item_attr_list) graph = self.construct_graphs(user_list, item_list, rating) graphs.append(graph) print() return graphs def read_data(self): self.user_dict = pickle.load(open(self.userfile, 'rb')) self.item_dict = pickle.load(open(self.itemfile, 'rb')) self.user_key_type = type(list(self.user_dict.keys())[0]) self.item_key_type = type(list(self.item_dict.keys())[0]) feature_dict = pickle.load(open(self.featurefile, 'rb')) data = [] error_num = 0 ratings_df = pd.read_csv(self.ratingfile, sep=self.sep, header=None) (train_df, test_df) = train_test_split(ratings_df, test_size=0.4, random_state=self.args.random_seed, stratify=ratings_df[[0, 2]]) (test_df, valid_df) = train_test_split(test_df, test_size=0.5, random_state=self.args.random_seed, stratify=test_df[[0, 2]]) if self.store_backup: backup_path = f'{self.path}{self.dataset}/split_data_backup/' if (not os.path.exists(backup_path)): os.mkdir(backup_path) train_df.to_csv(f'{backup_path}train_data.csv', index=False) valid_df.to_csv(f'{backup_path}valid_data.csv', index=False) test_df.to_csv(f'{backup_path}test_data.csv', index=False) print('(Only run at the first time training the dataset)') train_graphs = self.data_2_graphs(train_df, dataset='train') valid_graphs = self.data_2_graphs(valid_df, dataset='valid') test_graphs = self.data_2_graphs(test_df, dataset='test') graphs = ((train_graphs + valid_graphs) + test_graphs) stat_info = {} stat_info['data_num'] = len(graphs) stat_info['feature_num'] = len(feature_dict) stat_info['train_test_split_index'] = [len(train_graphs), (len(train_graphs) + len(valid_graphs))] print('error number of data:', error_num) return (graphs, stat_info) def construct_graphs(self, user_list, item_list, rating): u_n = len(user_list) i_n = len(item_list) inner_edge_index = [[], []] for i in range(u_n): for j in range(i, u_n): inner_edge_index[0].append(i) inner_edge_index[1].append(j) for i in range(u_n, (u_n + i_n)): for j in range(i, (u_n + i_n)): inner_edge_index[0].append(i) inner_edge_index[1].append(j) outer_edge_index = [[], []] for i in range(u_n): for j in range(i_n): outer_edge_index[0].append(i) outer_edge_index[1].append((u_n + j)) inner_edge_index = torch.LongTensor(inner_edge_index) inner_edge_index = to_undirected(inner_edge_index) outer_edge_index = torch.LongTensor(outer_edge_index) outer_edge_index = to_undirected(outer_edge_index) graph = self.construct_graph((user_list + item_list), inner_edge_index, outer_edge_index, rating) return graph def construct_graph(self, node_list, edge_index_inner, edge_index_outer, rating): x = torch.LongTensor(node_list).unsqueeze(1) rating = torch.FloatTensor([rating]) return Data(x=x, edge_index=edge_index_inner, edge_attr=torch.transpose(edge_index_outer, 0, 1), y=rating) def process(self): self.userfile = self.raw_file_names[0] self.itemfile = self.raw_file_names[1] self.featurefile = self.raw_file_names[2] self.ratingfile = self.raw_file_names[3] (graphs, stat_info) = self.read_data() if (not os.path.exists(f'{self.path}processed/{self.dataset}')): os.mkdir(f'{self.path}processed/{self.dataset}') (data, slices) = self.collate(graphs) torch.save((data, slices), self.processed_paths[0]) torch.save(stat_info, self.processed_paths[1]) def feature_N(self): return self.feature_num def data_N(self): return self.data_num
class inner_GNN(MessagePassing): def __init__(self, dim, hidden_layer): super(inner_GNN, self).__init__(aggr='mean') self.lin1 = nn.Linear(dim, hidden_layer) self.lin2 = nn.Linear(hidden_layer, dim) self.act = nn.ReLU() self.drop = nn.Dropout(p=0.5) def forward(self, x, edge_index, edge_weight=None): x = x.squeeze() return self.propagate(edge_index, x=x, edge_weight=edge_weight) def message(self, x_i, x_j, edge_weight): pairwise_analysis = (x_i * x_j) pairwise_analysis = self.lin1(pairwise_analysis) pairwise_analysis = self.act(pairwise_analysis) pairwise_analysis = self.lin2(pairwise_analysis) pairwise_analysis = self.drop(pairwise_analysis) if (edge_weight != None): interaction_analysis = (pairwise_analysis * edge_weight.view((- 1), 1)) else: interaction_analysis = pairwise_analysis return interaction_analysis def update(self, aggr_out): return aggr_out
class cross_GNN(MessagePassing): def __init__(self, dim, hidden_layer): super(cross_GNN, self).__init__(aggr='mean') def forward(self, x, edge_index, edge_weight=None): x = x.squeeze() return self.propagate(edge_index, x=x, edge_weight=edge_weight) def message(self, x_i, x_j, edge_weight): pairwise_analysis = (x_i * x_j) if (edge_weight != None): interaction_analysis = (pairwise_analysis * edge_weight.view((- 1), 1)) else: interaction_analysis = pairwise_analysis return interaction_analysis def update(self, aggr_out): return aggr_out
class GMCF(nn.Module): '\n GMCF main model\n ' def __init__(self, args, n_features, device): super(GMCF, self).__init__() self.n_features = n_features self.dim = args.dim self.hidden_layer = args.hidden_layer self.device = device self.batch_size = args.batch_size self.num_user_features = args.num_user_features self.feature_embedding = nn.Embedding((self.n_features + 1), self.dim) self.inner_gnn = inner_GNN(self.dim, self.hidden_layer) self.outer_gnn = cross_GNN(self.dim, self.hidden_layer) self.node_weight = nn.Embedding((self.n_features + 1), 1) self.node_weight.weight.data.normal_(0.0, 0.01) self.update_f = nn.GRU(input_size=self.dim, hidden_size=self.dim, dropout=0.5) self.g = nn.Linear(self.dim, 1, bias=False) def forward(self, data, is_training=True): node_id = data.x.to(self.device) batch = data.batch node_w = torch.squeeze(self.node_weight(node_id)) sum_weight = global_add_pool(node_w, batch) node_emb = self.feature_embedding(node_id) inner_edge_index = data.edge_index outer_edge_index = torch.transpose(data.edge_attr, 0, 1) outer_edge_index = self.outer_offset(batch, self.num_user_features, outer_edge_index) inner_node_message = self.inner_gnn(node_emb, inner_edge_index) outer_node_message = self.outer_gnn(node_emb, outer_edge_index) if (len(outer_node_message.size()) < len(node_emb.size())): outer_node_message = outer_node_message.unsqueeze(1) inner_node_message = inner_node_message.unsqueeze(1) updated_node_input = torch.cat((node_emb, inner_node_message, outer_node_message), 1) updated_node_input = torch.transpose(updated_node_input, 0, 1) gru_h0 = torch.normal(0, 0.01, (1, node_emb.size(0), self.dim)).to(self.device) (gru_output, hn) = self.update_f(updated_node_input, gru_h0) updated_node = gru_output[(- 1)] new_batch = self.split_batch(batch, self.num_user_features) updated_graph = torch.squeeze(global_mean_pool(updated_node, new_batch)) (item_graphs, user_graphs) = torch.split(updated_graph, int((updated_graph.size(0) / 2))) y = torch.unsqueeze((torch.sum((user_graphs * item_graphs), 1) + sum_weight), 1) y = torch.sigmoid(y) return y def split_batch(self, batch, user_node_num): '\n split batch id into user nodes and item nodes \n ' ones = torch.ones_like(batch) nodes_per_graph = global_add_pool(ones, batch) cum_num = torch.cat((torch.LongTensor([0]).to(self.device), torch.cumsum(nodes_per_graph, dim=0)[:(- 1)])) cum_num_list = [(cum_num + i) for i in range(user_node_num)] multi_hot = torch.cat(cum_num_list) test = (torch.sum(F.one_hot(multi_hot, ones.size(0)), dim=0) * (torch.max(batch) + 1)) return (batch + test) def outer_offset(self, batch, user_node_num, outer_edge_index): ones = torch.ones_like(batch) nodes_per_graph = global_add_pool(ones, batch) inter_per_graph = (((nodes_per_graph - user_node_num) * user_node_num) * 2) cum_num = torch.cat((torch.LongTensor([0]).to(self.device), torch.cumsum(nodes_per_graph, dim=0)[:(- 1)])) offset_list = torch.repeat_interleave(cum_num, inter_per_graph, dim=0).repeat(2, 1) outer_edge_index_offset = (outer_edge_index + offset_list) return outer_edge_index_offset
def train(args, data_info, show_loss): train_loader = data_info['train'] val_loader = data_info['val'] test_loader = data_info['test'] feature_num = data_info['feature_num'] (train_num, val_num, test_num) = data_info['data_num'] device = torch.device(('cuda' if torch.cuda.is_available() else 'cpu')) model = GMCF(args, feature_num, device) model = model.to(device) optimizer = torch.optim.Adam(filter((lambda p: p.requires_grad), model.parameters()), weight_decay=args.l2_weight, lr=args.lr) crit = torch.nn.BCELoss() print([i.size() for i in filter((lambda p: p.requires_grad), model.parameters())]) print('start training...') for step in range(args.n_epoch): loss_all = 0 edge_all = 0 model.train() for data in train_loader: data = data.to(device) output = model(data) label = data.y label = label.to(device) baseloss = crit(torch.squeeze(output), label) loss = baseloss loss_all += (data.num_graphs * loss.item()) optimizer.zero_grad() loss.backward() optimizer.step() cur_loss = (loss_all / train_num) (val_auc, val_logloss, val_ndcg5, val_ndcg10) = evaluate(model, val_loader, device) (test_auc, test_logloss, test_ndcg5, test_ndcg10) = evaluate(model, test_loader, device) print('Epoch: {:03d}, Loss: {:.5f}, AUC: {:.5f}/{:.5f}, Logloss: {:.5f}/{:.5f}, NDCG@5: {:.5f}/{:.5f} NDCG@10: {:.5f}/{:.5f}'.format(step, cur_loss, val_auc, test_auc, val_logloss, test_logloss, val_ndcg5, test_ndcg5, val_ndcg10, test_ndcg10))
def evaluate(model, data_loader, device): model.eval() predictions = [] labels = [] user_ids = [] edges_all = [0, 0] with torch.no_grad(): for data in data_loader: (_, user_id_index) = np.unique(data.batch.detach().cpu().numpy(), return_index=True) user_id = data.x.detach().cpu().numpy()[user_id_index] user_ids.append(user_id) data = data.to(device) pred = model(data) pred = pred.squeeze().detach().cpu().numpy().astype('float64') if (pred.size == 1): pred = np.expand_dims(pred, axis=0) label = data.y.detach().cpu().numpy() predictions.append(pred) labels.append(label) predictions = np.concatenate(predictions, 0) labels = np.concatenate(labels, 0) user_ids = np.concatenate(user_ids, 0) ndcg5 = cal_ndcg(predictions, labels, user_ids, 5) ndcg10 = cal_ndcg(predictions, labels, user_ids, 10) auc = roc_auc_score(labels, predictions) logloss = log_loss(labels, predictions) return (auc, logloss, ndcg5, ndcg10)
def cal_ndcg(predicts, labels, user_ids, k): d = {'user': np.squeeze(user_ids), 'predict': np.squeeze(predicts), 'label': np.squeeze(labels)} df = pd.DataFrame(d) user_unique = df.user.unique() ndcg = [] for user_id in user_unique: user_srow = df.loc[(df['user'] == user_id)] upred = user_srow['predict'].tolist() if (len(upred) < 2): continue ulabel = user_srow['label'].tolist() ndcg.append(ndcg_score([ulabel], [upred], k=k)) return np.mean(np.array(ndcg))
class InpaintingTrainDataset(Dataset): def __init__(self, paths, n_references, mask_generator, transform_shared, transform_individual): self.in_files = paths self.n_references = n_references self.mask_generator = mask_generator self.transform_shared = transform_shared self.transform_individual = transform_individual self.iter_i = 0 def __len__(self): return len(self.in_files) def __getitem__(self, item): path = self.in_files[item] img = cv2.imread(path) img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) data = self.transform_shared(image=img) frame = data['image'] frame = self.transform_individual(image=frame)['image'] images = [frame] for i in range(self.n_references): frame = A.ReplayCompose.replay(data['replay'], image=img)['image'] frame = self.transform_individual(image=frame)['image'] images.append(frame) images = np.stack(images, 0) img_t = np.transpose(images[0], (2, 0, 1)) masks = list() for _ in range((self.n_references + 1)): mask = self.mask_generator(img_t, iter_i=self.iter_i) mask = mask.transpose(1, 2, 0) self.iter_i += 1 masks.append(mask) masks = np.stack(masks, 0) result = dict(images=images, masks=masks, masked_images=(images * (1 - masks))) for k in ['images', 'masks', 'masked_images']: result[k] = ((result[k] * 2.0) - 1.0) return result
def get_transforms(transform_variant, out_size, easy=False): assert (transform_variant == 'distortions') if (transform_variant == 'default'): transform = A.Compose([A.RandomScale(scale_limit=0.2), A.PadIfNeeded(min_height=out_size, min_width=out_size), A.RandomCrop(height=out_size, width=out_size), A.HorizontalFlip(), A.CLAHE(), A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2), A.HueSaturationValue(hue_shift_limit=5, sat_shift_limit=30, val_shift_limit=5), A.ToFloat()]) elif (transform_variant == 'distortions'): if easy: max_shift = 25 transform_shared = A.ReplayCompose([A.PadIfNeeded(min_height=(out_size + max_shift), min_width=(out_size + max_shift)), A.RandomCrop(height=(out_size + max_shift), width=(out_size + max_shift)), A.HorizontalFlip(), A.CLAHE(), A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2), A.HueSaturationValue(hue_shift_limit=5, sat_shift_limit=30, val_shift_limit=5)]) transform_individual = A.Compose([A.PadIfNeeded(min_height=out_size, min_width=out_size), A.RandomCrop(height=out_size, width=out_size), A.ToFloat()]) else: max_shift = 200 transform_shared = A.ReplayCompose([A.PadIfNeeded(min_height=(out_size + max_shift), min_width=(out_size + max_shift)), A.RandomCrop(height=(out_size + max_shift), width=(out_size + max_shift)), A.HorizontalFlip(), A.CLAHE(), A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2), A.HueSaturationValue(hue_shift_limit=5, sat_shift_limit=30, val_shift_limit=5)]) transform_individual = A.Compose([IAAPerspective2(scale=(0.0, 0.06)), IAAAffine2(scale=(0.7, 1.3), rotate=((- 40), 40), shear=((- 0.1), 0.1)), A.PadIfNeeded(min_height=out_size, min_width=out_size), A.OpticalDistortion(), A.RandomCrop(height=out_size, width=out_size), A.RandomBrightnessContrast(brightness_limit=0.01, contrast_limit=0.01), A.HueSaturationValue(hue_shift_limit=0, sat_shift_limit=3, val_shift_limit=1), A.ToFloat()]) elif (transform_variant == 'distortions_scale05_1'): transform = A.Compose([IAAPerspective2(scale=(0.0, 0.06)), IAAAffine2(scale=(0.5, 1.0), rotate=((- 40), 40), shear=((- 0.1), 0.1), p=1), A.PadIfNeeded(min_height=out_size, min_width=out_size), A.OpticalDistortion(), A.RandomCrop(height=out_size, width=out_size), A.HorizontalFlip(), A.CLAHE(), A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2), A.HueSaturationValue(hue_shift_limit=5, sat_shift_limit=30, val_shift_limit=5), A.ToFloat()]) elif (transform_variant == 'distortions_scale03_12'): transform = A.Compose([IAAPerspective2(scale=(0.0, 0.06)), IAAAffine2(scale=(0.3, 1.2), rotate=((- 40), 40), shear=((- 0.1), 0.1), p=1), A.PadIfNeeded(min_height=out_size, min_width=out_size), A.OpticalDistortion(), A.RandomCrop(height=out_size, width=out_size), A.HorizontalFlip(), A.CLAHE(), A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2), A.HueSaturationValue(hue_shift_limit=5, sat_shift_limit=30, val_shift_limit=5), A.ToFloat()]) elif (transform_variant == 'distortions_scale03_07'): transform = A.Compose([IAAPerspective2(scale=(0.0, 0.06)), IAAAffine2(scale=(0.3, 0.7), rotate=((- 40), 40), shear=((- 0.1), 0.1), p=1), A.PadIfNeeded(min_height=out_size, min_width=out_size), A.OpticalDistortion(), A.RandomCrop(height=out_size, width=out_size), A.HorizontalFlip(), A.CLAHE(), A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2), A.HueSaturationValue(hue_shift_limit=5, sat_shift_limit=30, val_shift_limit=5), A.ToFloat()]) elif (transform_variant == 'distortions_light'): transform = A.Compose([IAAPerspective2(scale=(0.0, 0.02)), IAAAffine2(scale=(0.8, 1.8), rotate=((- 20), 20), shear=((- 0.03), 0.03)), A.PadIfNeeded(min_height=out_size, min_width=out_size), A.RandomCrop(height=out_size, width=out_size), A.HorizontalFlip(), A.CLAHE(), A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2), A.HueSaturationValue(hue_shift_limit=5, sat_shift_limit=30, val_shift_limit=5), A.ToFloat()]) elif (transform_variant == 'non_space_transform'): transform = A.Compose([A.CLAHE(), A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2), A.HueSaturationValue(hue_shift_limit=5, sat_shift_limit=30, val_shift_limit=5), A.ToFloat()]) elif (transform_variant == 'no_augs'): transform = A.Compose([A.ToFloat()]) else: raise ValueError(f'Unexpected transform_variant {transform_variant}') return (transform_shared, transform_individual)
def make_default_train_dataset(root, filelist, kind='default', out_size=512, mask_gen_kwargs=None, transform_variant='default', mask_generator_kind='mixed', easy=False, **kwargs): if (kind != 'default'): raise ValueError(f'Dropped support for other datasets: {kind}') LOGGER.info(f'Make train dataloader from {filelist}. Using mask generator={mask_generator_kind}') with open(filelist, 'r') as f: paths = f.read().splitlines() paths = [os.path.join(root, x) for x in paths] mask_generator = get_mask_generator(kind=mask_generator_kind, kwargs=mask_gen_kwargs) (transform_shared, transform_individual) = get_transforms(transform_variant, out_size, easy=easy) dataset = InpaintingTrainDataset(paths=paths, mask_generator=mask_generator, transform_shared=transform_shared, transform_individual=transform_individual, **kwargs) return dataset
def make_default_val_dataset(indir, kind='default', out_size=512, **kwargs): if (kind != 'default'): raise ValueError(f'Dropped support for other datasets: {kind}') if (OmegaConf.is_list(indir) or isinstance(indir, (tuple, list))): return ConcatDataset([make_default_val_dataset(idir, kind=kind, out_size=out_size, **kwargs) for idir in indir]) LOGGER.info(f'Make val dataloader {kind} from {indir}') mask_generator = get_mask_generator(kind=kwargs.get('mask_generator_kind'), kwargs=kwargs.get('mask_gen_kwargs')) dataset = InpaintingEvaluationDataset(indir, **kwargs) return dataset
def load_image(fname, mode='RGB', return_orig=False): img = np.array(Image.open(fname).convert(mode)) if (img.ndim == 3): img = np.transpose(img, (2, 0, 1)) out_img = (img.astype('float32') / 255) if return_orig: return (out_img, img) else: return out_img
def ceil_modulo(x, mod): if ((x % mod) == 0): return x return (((x // mod) + 1) * mod)
def pad_img_to_modulo(img, mod): (channels, height, width) = img.shape out_height = ceil_modulo(height, mod) out_width = ceil_modulo(width, mod) return np.pad(img, ((0, 0), (0, (out_height - height)), (0, (out_width - width))), mode='symmetric')
def pad_tensor_to_modulo(img, mod): (batch_size, channels, height, width) = img.shape out_height = ceil_modulo(height, mod) out_width = ceil_modulo(width, mod) return F.pad(img, pad=(0, (out_width - width), 0, (out_height - height)), mode='reflect')
def scale_image(img, factor, interpolation=cv2.INTER_AREA): if (img.shape[0] == 1): img = img[0] else: img = np.transpose(img, (1, 2, 0)) img = cv2.resize(img, dsize=None, fx=factor, fy=factor, interpolation=interpolation) if (img.ndim == 2): img = img[(None, ...)] else: img = np.transpose(img, (2, 0, 1)) return img
class InpaintingEvaluationDataset(Dataset): def __init__(self, datadir, img_suffix='.jpg', pad_out_to_modulo=None, scale_factor=None): self.datadir = datadir self.mask_filenames = sorted(list(glob.glob(os.path.join(self.datadir, '**', '*mask*.png'), recursive=True))) self.img_filenames = [(fname.rsplit('_mask', 1)[0] + img_suffix) for fname in self.mask_filenames] self.pad_out_to_modulo = pad_out_to_modulo self.scale_factor = scale_factor def __len__(self): return len(self.mask_filenames) def __getitem__(self, i): image = load_image(self.img_filenames[i], mode='RGB') mask = load_image(self.mask_filenames[i], mode='L') result = dict(image=image, mask=mask[(None, ...)]) if (self.scale_factor is not None): result['image'] = scale_image(result['image'], self.scale_factor) result['mask'] = scale_image(result['mask'], self.scale_factor, interpolation=cv2.INTER_NEAREST) if ((self.pad_out_to_modulo is not None) and (self.pad_out_to_modulo > 1)): result['image'] = pad_img_to_modulo(result['image'], self.pad_out_to_modulo) result['mask'] = pad_img_to_modulo(result['mask'], self.pad_out_to_modulo) for k in ['image', 'mask']: result[k] = ((result[k] * 2.0) - 1.0).transpose(1, 2, 0) return result
class PRNGMixin(object): 'Adds a prng property which is a numpy RandomState which gets\n reinitialized whenever the pid changes to avoid synchronized sampling\n behavior when used in conjunction with multiprocessing.' @property def prng(self): currentpid = os.getpid() if (getattr(self, '_initpid', None) != currentpid): self._initpid = currentpid self._prng = np.random.RandomState() return self._prng
class LamaPropagation(Dataset, PRNGMixin): def __init__(self, **kwargs): self.clean_prob = kwargs.pop('clean_prob', (1.0 / kwargs['n_references'])) for k in default_mask_config: if (not (k in kwargs)): kwargs[k] = default_mask_config[k] self.base_data = make_default_train_dataset(**kwargs) def __len__(self): return len(self.base_data) def __getitem__(self, i): example = self.base_data[i] T = example['images'].shape[0] values = list() for i in range(T): if (self.prng.random() < self.clean_prob): values.append(np.concatenate([example['images'][i], ((- 1) * np.ones_like(example['masks'][i]))], axis=(- 1))) else: values.append(np.concatenate([example['masked_images'][i], example['masks'][i]], axis=(- 1))) return {'rgbs': np.concatenate([example['images'], example['masks']], axis=(- 1)), 'keys': np.concatenate([example['masked_images'], example['masks']], axis=(- 1)), 'values': np.stack(values, axis=0), 'targets': example['images']}
class LamaGI(Dataset, PRNGMixin): def __init__(self, **kwargs): self.clean_prob = kwargs.pop('clean_prob', (1.0 / kwargs['n_references'])) for k in default_mask_config: if (not (k in kwargs)): kwargs[k] = default_mask_config[k] self.base_data = make_default_train_dataset(**kwargs) def __len__(self): return len(self.base_data) def __getitem__(self, i): example = self.base_data[i] T = example['images'].shape[0] image = example['images'][0] mask = example['masks'][0] srcs = example['images'][1:] srcs_masks = example['masks'][1:] return {'image': image, 'mask': mask, 'srcs': srcs, 'srcs_masks': srcs_masks}
class LamaGIValidation(Dataset): def __init__(self, filenames, n_references, pad_out_to_modulo=None, scale_factor=None): self.n_references = n_references with open(filenames, 'r') as f: filenames = f.read().splitlines() self.mask_filenames = [fname for fname in filenames if fnmatch.fnmatch(fname, '*image*mask*.png')] self.img_filenames = [(fname.rsplit('_mask', 1)[0] + '.png') for fname in self.mask_filenames] self.srcs_masks_filenames_by_img = dict() self.srcs_filenames_by_img = dict() for parent in self.img_filenames: self.srcs_masks_filenames_by_img[parent] = list() self.srcs_filenames_by_img[parent] = list() for fname in filenames: if fnmatch.fnmatch(fname, '*image*_src*_m.png'): parent = (fname.rsplit('_src', 1)[0] + '.png') self.srcs_masks_filenames_by_img[parent].append(fname) src = (fname.rsplit('_m.png', 1)[0] + '.png') self.srcs_filenames_by_img[parent].append(src) self.pad_out_to_modulo = pad_out_to_modulo self.scale_factor = scale_factor def __len__(self): return len(self.mask_filenames) def __getitem__(self, i): image = load_image(self.img_filenames[i], mode='RGB') mask = load_image(self.mask_filenames[i], mode='L') mask[(mask < 0.5)] = 0 mask[(mask >= 0.5)] = 1 mask = mask[None] srcs = list() srcs_masks = list() for t in range(self.n_references): srcs.append(load_image(self.srcs_filenames_by_img[self.img_filenames[i]][t], mode='RGB')) src_mask = load_image(self.srcs_masks_filenames_by_img[self.img_filenames[i]][t], mode='L') src_mask[(src_mask < 0.5)] = 0 src_mask[(src_mask >= 0.5)] = 1 src_mask = src_mask[None] srcs_masks.append(src_mask) srcs = np.stack(srcs) srcs_masks = np.stack(srcs_masks) result = dict(image=image, mask=mask, srcs=srcs, srcs_masks=srcs_masks) if (self.scale_factor is not None): result['image'] = scale_image(result['image'], self.scale_factor) result['mask'] = scale_image(result['mask'], self.scale_factor, interpolation=cv2.INTER_NEAREST) for t in range(srcs_masks.shape[0]): result['srcs'][t] = scale_image(result['srcs'][t], self.scale_factor) result['srcs_masks'][t] = scale_image(result['srcs_masks'][t], self.scale_factor, interpolation=cv2.INTER_NEAREST) if ((self.pad_out_to_modulo is not None) and (self.pad_out_to_modulo > 1)): result['image'] = pad_img_to_modulo(result['image'], self.pad_out_to_modulo) result['mask'] = pad_img_to_modulo(result['mask'], self.pad_out_to_modulo) for t in range(srcs_masks.shape[0]): result['srcs'][t] = pad_img_to_modulo(result['srcs'][t], self.pad_out_to_modulo) result['srcs_masks'][t] = pad_img_to_modulo(result['srcs_masks'][t], self.pad_out_to_modulo) for k in ['image', 'mask']: result[k] = ((result[k] * 2.0) - 1.0).transpose(1, 2, 0) for k in ['srcs', 'srcs_masks']: result[k] = ((result[k] * 2.0) - 1.0).transpose(0, 2, 3, 1) return result
class CorrBlock(): def __init__(self, fmap1, fmap2, num_levels=4, radius=4): self.num_levels = num_levels self.radius = radius self.corr_pyramid = [] corr = CorrBlock.corr(fmap1, fmap2) (batch, h1, w1, dim, h2, w2) = corr.shape corr = corr.reshape(((batch * h1) * w1), dim, h2, w2) self.corr_pyramid.append(corr) for i in range((self.num_levels - 1)): corr = F.avg_pool2d(corr, 2, stride=2) self.corr_pyramid.append(corr) def __call__(self, coords): r = self.radius coords = coords.permute(0, 2, 3, 1) (batch, h1, w1, _) = coords.shape out_pyramid = [] for i in range(self.num_levels): corr = self.corr_pyramid[i] dx = torch.linspace((- r), r, ((2 * r) + 1)) dy = torch.linspace((- r), r, ((2 * r) + 1)) delta = torch.stack(torch.meshgrid(dy, dx), axis=(- 1)).to(coords.device) centroid_lvl = (coords.reshape(((batch * h1) * w1), 1, 1, 2) / (2 ** i)) delta_lvl = delta.view(1, ((2 * r) + 1), ((2 * r) + 1), 2) coords_lvl = (centroid_lvl + delta_lvl) corr = bilinear_sampler(corr, coords_lvl) corr = corr.view(batch, h1, w1, (- 1)) out_pyramid.append(corr) out = torch.cat(out_pyramid, dim=(- 1)) return out.permute(0, 3, 1, 2).contiguous().float() @staticmethod def corr(fmap1, fmap2): (batch, dim, ht, wd) = fmap1.shape fmap1 = fmap1.view(batch, dim, (ht * wd)) fmap2 = fmap2.view(batch, dim, (ht * wd)) corr = torch.matmul(fmap1.transpose(1, 2), fmap2) corr = corr.view(batch, ht, wd, 1, ht, wd) return (corr / torch.sqrt(torch.tensor(dim).float()))
class CorrLayer(torch.autograd.Function): @staticmethod def forward(ctx, fmap1, fmap2, coords, r): fmap1 = fmap1.contiguous() fmap2 = fmap2.contiguous() coords = coords.contiguous() ctx.save_for_backward(fmap1, fmap2, coords) ctx.r = r (corr,) = correlation_cudaz.forward(fmap1, fmap2, coords, ctx.r) return corr @staticmethod def backward(ctx, grad_corr): (fmap1, fmap2, coords) = ctx.saved_tensors grad_corr = grad_corr.contiguous() (fmap1_grad, fmap2_grad, coords_grad) = correlation_cudaz.backward(fmap1, fmap2, coords, grad_corr, ctx.r) return (fmap1_grad, fmap2_grad, coords_grad, None)
class AlternateCorrBlock(): def __init__(self, fmap1, fmap2, num_levels=4, radius=4): self.num_levels = num_levels self.radius = radius self.pyramid = [(fmap1, fmap2)] for i in range(self.num_levels): fmap1 = F.avg_pool2d(fmap1, 2, stride=2) fmap2 = F.avg_pool2d(fmap2, 2, stride=2) self.pyramid.append((fmap1, fmap2)) def __call__(self, coords): coords = coords.permute(0, 2, 3, 1) (B, H, W, _) = coords.shape corr_list = [] for i in range(self.num_levels): r = self.radius fmap1_i = self.pyramid[0][0].permute(0, 2, 3, 1) fmap2_i = self.pyramid[i][1].permute(0, 2, 3, 1) coords_i = (coords / (2 ** i)).reshape(B, 1, H, W, 2).contiguous() corr = alt_cuda_corr(fmap1_i, fmap2_i, coords_i, r) corr_list.append(corr.squeeze(1)) corr = torch.stack(corr_list, dim=1) corr = corr.reshape(B, (- 1), H, W) return (corr / 16.0)
class ResidualBlock(nn.Module): def __init__(self, in_planes, planes, norm_fn='group', stride=1): super(ResidualBlock, self).__init__() self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, padding=1, stride=stride) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1) self.relu = nn.ReLU(inplace=True) num_groups = (planes // 8) if (norm_fn == 'group'): self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=planes) self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=planes) if (not (stride == 1)): self.norm3 = nn.GroupNorm(num_groups=num_groups, num_channels=planes) elif (norm_fn == 'batch'): self.norm1 = nn.BatchNorm2d(planes) self.norm2 = nn.BatchNorm2d(planes) if (not (stride == 1)): self.norm3 = nn.BatchNorm2d(planes) elif (norm_fn == 'instance'): self.norm1 = nn.InstanceNorm2d(planes) self.norm2 = nn.InstanceNorm2d(planes) if (not (stride == 1)): self.norm3 = nn.InstanceNorm2d(planes) elif (norm_fn == 'none'): self.norm1 = nn.Sequential() self.norm2 = nn.Sequential() if (not (stride == 1)): self.norm3 = nn.Sequential() if (stride == 1): self.downsample = None else: self.downsample = nn.Sequential(nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride), self.norm3) def forward(self, x): y = x y = self.relu(self.norm1(self.conv1(y))) y = self.relu(self.norm2(self.conv2(y))) if (self.downsample is not None): x = self.downsample(x) return self.relu((x + y))
class BottleneckBlock(nn.Module): def __init__(self, in_planes, planes, norm_fn='group', stride=1): super(BottleneckBlock, self).__init__() self.conv1 = nn.Conv2d(in_planes, (planes // 4), kernel_size=1, padding=0) self.conv2 = nn.Conv2d((planes // 4), (planes // 4), kernel_size=3, padding=1, stride=stride) self.conv3 = nn.Conv2d((planes // 4), planes, kernel_size=1, padding=0) self.relu = nn.ReLU(inplace=True) num_groups = (planes // 8) if (norm_fn == 'group'): self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=(planes // 4)) self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=(planes // 4)) self.norm3 = nn.GroupNorm(num_groups=num_groups, num_channels=planes) if (not (stride == 1)): self.norm4 = nn.GroupNorm(num_groups=num_groups, num_channels=planes) elif (norm_fn == 'batch'): self.norm1 = nn.BatchNorm2d((planes // 4)) self.norm2 = nn.BatchNorm2d((planes // 4)) self.norm3 = nn.BatchNorm2d(planes) if (not (stride == 1)): self.norm4 = nn.BatchNorm2d(planes) elif (norm_fn == 'instance'): self.norm1 = nn.InstanceNorm2d((planes // 4)) self.norm2 = nn.InstanceNorm2d((planes // 4)) self.norm3 = nn.InstanceNorm2d(planes) if (not (stride == 1)): self.norm4 = nn.InstanceNorm2d(planes) elif (norm_fn == 'none'): self.norm1 = nn.Sequential() self.norm2 = nn.Sequential() self.norm3 = nn.Sequential() if (not (stride == 1)): self.norm4 = nn.Sequential() if (stride == 1): self.downsample = None else: self.downsample = nn.Sequential(nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride), self.norm4) def forward(self, x): y = x y = self.relu(self.norm1(self.conv1(y))) y = self.relu(self.norm2(self.conv2(y))) y = self.relu(self.norm3(self.conv3(y))) if (self.downsample is not None): x = self.downsample(x) return self.relu((x + y))
class BasicEncoder(nn.Module): def __init__(self, output_dim=128, norm_fn='batch', dropout=0.0): super(BasicEncoder, self).__init__() self.norm_fn = norm_fn if (self.norm_fn == 'group'): self.norm1 = nn.GroupNorm(num_groups=8, num_channels=64) elif (self.norm_fn == 'batch'): self.norm1 = nn.BatchNorm2d(64) elif (self.norm_fn == 'instance'): self.norm1 = nn.InstanceNorm2d(64) elif (self.norm_fn == 'none'): self.norm1 = nn.Sequential() self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3) self.relu1 = nn.ReLU(inplace=True) self.in_planes = 64 self.layer1 = self._make_layer(64, stride=1) self.layer2 = self._make_layer(96, stride=2) self.layer3 = self._make_layer(128, stride=2) self.conv2 = nn.Conv2d(128, output_dim, kernel_size=1) self.dropout = None if (dropout > 0): self.dropout = nn.Dropout2d(p=dropout) for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') elif isinstance(m, (nn.BatchNorm2d, nn.InstanceNorm2d, nn.GroupNorm)): if (m.weight is not None): nn.init.constant_(m.weight, 1) if (m.bias is not None): nn.init.constant_(m.bias, 0) def _make_layer(self, dim, stride=1): layer1 = ResidualBlock(self.in_planes, dim, self.norm_fn, stride=stride) layer2 = ResidualBlock(dim, dim, self.norm_fn, stride=1) layers = (layer1, layer2) self.in_planes = dim return nn.Sequential(*layers) def forward(self, x): is_list = (isinstance(x, tuple) or isinstance(x, list)) if is_list: batch_dim = x[0].shape[0] x = torch.cat(x, dim=0) x = self.conv1(x) x = self.norm1(x) x = self.relu1(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.conv2(x) if (self.training and (self.dropout is not None)): x = self.dropout(x) if is_list: x = torch.split(x, [batch_dim, batch_dim], dim=0) return x
class SmallEncoder(nn.Module): def __init__(self, output_dim=128, norm_fn='batch', dropout=0.0): super(SmallEncoder, self).__init__() self.norm_fn = norm_fn if (self.norm_fn == 'group'): self.norm1 = nn.GroupNorm(num_groups=8, num_channels=32) elif (self.norm_fn == 'batch'): self.norm1 = nn.BatchNorm2d(32) elif (self.norm_fn == 'instance'): self.norm1 = nn.InstanceNorm2d(32) elif (self.norm_fn == 'none'): self.norm1 = nn.Sequential() self.conv1 = nn.Conv2d(3, 32, kernel_size=7, stride=2, padding=3) self.relu1 = nn.ReLU(inplace=True) self.in_planes = 32 self.layer1 = self._make_layer(32, stride=1) self.layer2 = self._make_layer(64, stride=2) self.layer3 = self._make_layer(96, stride=2) self.dropout = None if (dropout > 0): self.dropout = nn.Dropout2d(p=dropout) self.conv2 = nn.Conv2d(96, output_dim, kernel_size=1) for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') elif isinstance(m, (nn.BatchNorm2d, nn.InstanceNorm2d, nn.GroupNorm)): if (m.weight is not None): nn.init.constant_(m.weight, 1) if (m.bias is not None): nn.init.constant_(m.bias, 0) def _make_layer(self, dim, stride=1): layer1 = BottleneckBlock(self.in_planes, dim, self.norm_fn, stride=stride) layer2 = BottleneckBlock(dim, dim, self.norm_fn, stride=1) layers = (layer1, layer2) self.in_planes = dim return nn.Sequential(*layers) def forward(self, x): is_list = (isinstance(x, tuple) or isinstance(x, list)) if is_list: batch_dim = x[0].shape[0] x = torch.cat(x, dim=0) x = self.conv1(x) x = self.norm1(x) x = self.relu1(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.conv2(x) if (self.training and (self.dropout is not None)): x = self.dropout(x) if is_list: x = torch.split(x, [batch_dim, batch_dim], dim=0) return x
class RAFT(nn.Module): def __init__(self, args): super(RAFT, self).__init__() self.args = args if args.small: self.hidden_dim = hdim = 96 self.context_dim = cdim = 64 args.corr_levels = 4 args.corr_radius = 3 else: self.hidden_dim = hdim = 128 self.context_dim = cdim = 128 args.corr_levels = 4 args.corr_radius = 4 if ('dropout' not in args._get_kwargs()): args.dropout = 0 if ('alternate_corr' not in args._get_kwargs()): args.alternate_corr = False if args.small: self.fnet = SmallEncoder(output_dim=128, norm_fn='instance', dropout=args.dropout) self.cnet = SmallEncoder(output_dim=(hdim + cdim), norm_fn='none', dropout=args.dropout) self.update_block = SmallUpdateBlock(self.args, hidden_dim=hdim) else: self.fnet = BasicEncoder(output_dim=256, norm_fn='instance', dropout=args.dropout) self.cnet = BasicEncoder(output_dim=(hdim + cdim), norm_fn='batch', dropout=args.dropout) self.update_block = BasicUpdateBlock(self.args, hidden_dim=hdim) def freeze_bn(self): for m in self.modules(): if isinstance(m, nn.BatchNorm2d): m.eval() def initialize_flow(self, img): 'Flow is represented as difference between two coordinate grids flow = coords1 - coords0' (N, C, H, W) = img.shape coords0 = coords_grid(N, (H // 8), (W // 8)).to(img.device) coords1 = coords_grid(N, (H // 8), (W // 8)).to(img.device) return (coords0, coords1) def upsample_flow(self, flow, mask): 'Upsample flow field [H/8, W/8, 2] -> [H, W, 2] using convex combination' (N, _, H, W) = flow.shape mask = mask.view(N, 1, 9, 8, 8, H, W) mask = torch.softmax(mask, dim=2) up_flow = F.unfold((8 * flow), [3, 3], padding=1) up_flow = up_flow.view(N, 2, 9, 1, 1, H, W) up_flow = torch.sum((mask * up_flow), dim=2) up_flow = up_flow.permute(0, 1, 4, 2, 5, 3) return up_flow.reshape(N, 2, (8 * H), (8 * W)) def forward(self, image1, image2, iters=20, flow_init=None, upsample=True, test_mode=True, is_normalized=False): 'Estimate optical flow between pair of frames' if (not is_normalized): image1 = ((2 * (image1 / 255.0)) - 1.0) image2 = ((2 * (image2 / 255.0)) - 1.0) image1 = image1.contiguous() image2 = image2.contiguous() hdim = self.hidden_dim cdim = self.context_dim with autocast(enabled=self.args.mixed_precision): (fmap1, fmap2) = self.fnet([image1, image2]) fmap1 = fmap1.float() fmap2 = fmap2.float() if self.args.alternate_corr: corr_fn = CorrBlockAlternate(fmap1, fmap2, radius=self.args.corr_radius) else: corr_fn = CorrBlock(fmap1, fmap2, radius=self.args.corr_radius) with autocast(enabled=self.args.mixed_precision): cnet = self.cnet(image1) (net, inp) = torch.split(cnet, [hdim, cdim], dim=1) net = torch.tanh(net) inp = torch.relu(inp) (coords0, coords1) = self.initialize_flow(image1) if (flow_init is not None): coords1 = (coords1 + flow_init) flow_predictions = [] for itr in range(iters): coords1 = coords1.detach() corr = corr_fn(coords1) flow = (coords1 - coords0) with autocast(enabled=self.args.mixed_precision): (net, up_mask, delta_flow) = self.update_block(net, inp, corr, flow) coords1 = (coords1 + delta_flow) if (up_mask is None): flow_up = upflow8((coords1 - coords0)) else: flow_up = self.upsample_flow((coords1 - coords0), up_mask) flow_predictions.append(flow_up) if test_mode: return ((coords1 - coords0), flow_up) return flow_predictions
class FlowHead(nn.Module): def __init__(self, input_dim=128, hidden_dim=256): super(FlowHead, self).__init__() self.conv1 = nn.Conv2d(input_dim, hidden_dim, 3, padding=1) self.conv2 = nn.Conv2d(hidden_dim, 2, 3, padding=1) self.relu = nn.ReLU(inplace=True) def forward(self, x): return self.conv2(self.relu(self.conv1(x)))
class ConvGRU(nn.Module): def __init__(self, hidden_dim=128, input_dim=(192 + 128)): super(ConvGRU, self).__init__() self.convz = nn.Conv2d((hidden_dim + input_dim), hidden_dim, 3, padding=1) self.convr = nn.Conv2d((hidden_dim + input_dim), hidden_dim, 3, padding=1) self.convq = nn.Conv2d((hidden_dim + input_dim), hidden_dim, 3, padding=1) def forward(self, h, x): hx = torch.cat([h, x], dim=1) z = torch.sigmoid(self.convz(hx)) r = torch.sigmoid(self.convr(hx)) q = torch.tanh(self.convq(torch.cat([(r * h), x], dim=1))) h = (((1 - z) * h) + (z * q)) return h
class SepConvGRU(nn.Module): def __init__(self, hidden_dim=128, input_dim=(192 + 128)): super(SepConvGRU, self).__init__() self.convz1 = nn.Conv2d((hidden_dim + input_dim), hidden_dim, (1, 5), padding=(0, 2)) self.convr1 = nn.Conv2d((hidden_dim + input_dim), hidden_dim, (1, 5), padding=(0, 2)) self.convq1 = nn.Conv2d((hidden_dim + input_dim), hidden_dim, (1, 5), padding=(0, 2)) self.convz2 = nn.Conv2d((hidden_dim + input_dim), hidden_dim, (5, 1), padding=(2, 0)) self.convr2 = nn.Conv2d((hidden_dim + input_dim), hidden_dim, (5, 1), padding=(2, 0)) self.convq2 = nn.Conv2d((hidden_dim + input_dim), hidden_dim, (5, 1), padding=(2, 0)) def forward(self, h, x): hx = torch.cat([h, x], dim=1) z = torch.sigmoid(self.convz1(hx)) r = torch.sigmoid(self.convr1(hx)) q = torch.tanh(self.convq1(torch.cat([(r * h), x], dim=1))) h = (((1 - z) * h) + (z * q)) hx = torch.cat([h, x], dim=1) z = torch.sigmoid(self.convz2(hx)) r = torch.sigmoid(self.convr2(hx)) q = torch.tanh(self.convq2(torch.cat([(r * h), x], dim=1))) h = (((1 - z) * h) + (z * q)) return h
class SmallMotionEncoder(nn.Module): def __init__(self, args): super(SmallMotionEncoder, self).__init__() cor_planes = (args.corr_levels * (((2 * args.corr_radius) + 1) ** 2)) self.convc1 = nn.Conv2d(cor_planes, 96, 1, padding=0) self.convf1 = nn.Conv2d(2, 64, 7, padding=3) self.convf2 = nn.Conv2d(64, 32, 3, padding=1) self.conv = nn.Conv2d(128, 80, 3, padding=1) def forward(self, flow, corr): cor = F.relu(self.convc1(corr)) flo = F.relu(self.convf1(flow)) flo = F.relu(self.convf2(flo)) cor_flo = torch.cat([cor, flo], dim=1) out = F.relu(self.conv(cor_flo)) return torch.cat([out, flow], dim=1)
class BasicMotionEncoder(nn.Module): def __init__(self, args): super(BasicMotionEncoder, self).__init__() cor_planes = (args.corr_levels * (((2 * args.corr_radius) + 1) ** 2)) self.convc1 = nn.Conv2d(cor_planes, 256, 1, padding=0) self.convc2 = nn.Conv2d(256, 192, 3, padding=1) self.convf1 = nn.Conv2d(2, 128, 7, padding=3) self.convf2 = nn.Conv2d(128, 64, 3, padding=1) self.conv = nn.Conv2d((64 + 192), (128 - 2), 3, padding=1) def forward(self, flow, corr): cor = F.relu(self.convc1(corr)) cor = F.relu(self.convc2(cor)) flo = F.relu(self.convf1(flow)) flo = F.relu(self.convf2(flo)) cor_flo = torch.cat([cor, flo], dim=1) out = F.relu(self.conv(cor_flo)) return torch.cat([out, flow], dim=1)
class SmallUpdateBlock(nn.Module): def __init__(self, args, hidden_dim=96): super(SmallUpdateBlock, self).__init__() self.encoder = SmallMotionEncoder(args) self.gru = ConvGRU(hidden_dim=hidden_dim, input_dim=(82 + 64)) self.flow_head = FlowHead(hidden_dim, hidden_dim=128) def forward(self, net, inp, corr, flow): motion_features = self.encoder(flow, corr) inp = torch.cat([inp, motion_features], dim=1) net = self.gru(net, inp) delta_flow = self.flow_head(net) return (net, None, delta_flow)
class BasicUpdateBlock(nn.Module): def __init__(self, args, hidden_dim=128, input_dim=128): super(BasicUpdateBlock, self).__init__() self.args = args self.encoder = BasicMotionEncoder(args) self.gru = SepConvGRU(hidden_dim=hidden_dim, input_dim=(128 + hidden_dim)) self.flow_head = FlowHead(hidden_dim, hidden_dim=256) self.mask = nn.Sequential(nn.Conv2d(128, 256, 3, padding=1), nn.ReLU(inplace=True), nn.Conv2d(256, (64 * 9), 1, padding=0)) def forward(self, net, inp, corr, flow, upsample=True): motion_features = self.encoder(flow, corr) inp = torch.cat([inp, motion_features], dim=1) net = self.gru(net, inp) delta_flow = self.flow_head(net) mask = (0.25 * self.mask(net)) return (net, mask, delta_flow)
def conv_bn(inp, oup, stride): return nn.Sequential(nn.Conv2d(inp, oup, 3, stride, 1, bias=False), BatchNorm2d(oup), nn.ReLU6(inplace=True))
def conv_1x1_bn(inp, oup): return nn.Sequential(nn.Conv2d(inp, oup, 1, 1, 0, bias=False), BatchNorm2d(oup), nn.ReLU6(inplace=True))
class InvertedResidual(nn.Module): def __init__(self, inp, oup, stride, expand_ratio): super(InvertedResidual, self).__init__() self.stride = stride assert (stride in [1, 2]) hidden_dim = round((inp * expand_ratio)) self.use_res_connect = ((self.stride == 1) and (inp == oup)) if (expand_ratio == 1): self.conv = nn.Sequential(nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False), BatchNorm2d(hidden_dim), nn.ReLU6(inplace=True), nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False), BatchNorm2d(oup)) else: self.conv = nn.Sequential(nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False), BatchNorm2d(hidden_dim), nn.ReLU6(inplace=True), nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False), BatchNorm2d(hidden_dim), nn.ReLU6(inplace=True), nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False), BatchNorm2d(oup)) def forward(self, x): if self.use_res_connect: return (x + self.conv(x)) else: return self.conv(x)
class MobileNetV2(nn.Module): def __init__(self, n_class=1000, input_size=224, width_mult=1.0): super(MobileNetV2, self).__init__() block = InvertedResidual input_channel = 32 last_channel = 1280 interverted_residual_setting = [[1, 16, 1, 1], [6, 24, 2, 2], [6, 32, 3, 2], [6, 64, 4, 2], [6, 96, 3, 1], [6, 160, 3, 2], [6, 320, 1, 1]] assert ((input_size % 32) == 0) input_channel = int((input_channel * width_mult)) self.last_channel = (int((last_channel * width_mult)) if (width_mult > 1.0) else last_channel) self.features = [conv_bn(3, input_channel, 2)] for (t, c, n, s) in interverted_residual_setting: output_channel = int((c * width_mult)) for i in range(n): if (i == 0): self.features.append(block(input_channel, output_channel, s, expand_ratio=t)) else: self.features.append(block(input_channel, output_channel, 1, expand_ratio=t)) input_channel = output_channel self.features.append(conv_1x1_bn(input_channel, self.last_channel)) self.features = nn.Sequential(*self.features) self.classifier = nn.Sequential(nn.Dropout(0.2), nn.Linear(self.last_channel, n_class)) self._initialize_weights() def forward(self, x): x = self.features(x) x = x.mean(3).mean(2) x = self.classifier(x) return x def _initialize_weights(self): for m in self.modules(): if isinstance(m, nn.Conv2d): n = ((m.kernel_size[0] * m.kernel_size[1]) * m.out_channels) m.weight.data.normal_(0, math.sqrt((2.0 / n))) if (m.bias is not None): m.bias.data.zero_() elif isinstance(m, BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.Linear): n = m.weight.size(1) m.weight.data.normal_(0, 0.01) m.bias.data.zero_()
def mobilenetv2(pretrained=False, **kwargs): 'Constructs a MobileNet_V2 model.\n\n Args:\n pretrained (bool): If True, returns a model pre-trained on ImageNet\n ' model = MobileNetV2(n_class=1000, **kwargs) if pretrained: model.load_state_dict(load_url(model_urls['mobilenetv2']), strict=False) return model
def conv3x3(in_planes, out_planes, stride=1): '3x3 convolution with padding' return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False)
class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = BatchNorm2d(planes) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = BatchNorm2d(planes) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if (self.downsample is not None): residual = self.downsample(x) out += residual out = self.relu(out) return out
class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, downsample=None): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn1 = BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn2 = BatchNorm2d(planes) self.conv3 = nn.Conv2d(planes, (planes * 4), kernel_size=1, bias=False) self.bn3 = BatchNorm2d((planes * 4)) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) if (self.downsample is not None): residual = self.downsample(x) out += residual out = self.relu(out) return out
class ResNet(nn.Module): def __init__(self, block, layers, num_classes=1000): self.inplanes = 128 super(ResNet, self).__init__() self.conv1 = conv3x3(3, 64, stride=2) self.bn1 = BatchNorm2d(64) self.relu1 = nn.ReLU(inplace=True) self.conv2 = conv3x3(64, 64) self.bn2 = BatchNorm2d(64) self.relu2 = nn.ReLU(inplace=True) self.conv3 = conv3x3(64, 128) self.bn3 = BatchNorm2d(128) self.relu3 = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2) self.layer3 = self._make_layer(block, 256, layers[2], stride=2) self.layer4 = self._make_layer(block, 512, layers[3], stride=2) self.avgpool = nn.AvgPool2d(7, stride=1) self.fc = nn.Linear((512 * block.expansion), num_classes) for m in self.modules(): if isinstance(m, nn.Conv2d): n = ((m.kernel_size[0] * m.kernel_size[1]) * m.out_channels) m.weight.data.normal_(0, math.sqrt((2.0 / n))) elif isinstance(m, BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def _make_layer(self, block, planes, blocks, stride=1): downsample = None if ((stride != 1) or (self.inplanes != (planes * block.expansion))): downsample = nn.Sequential(nn.Conv2d(self.inplanes, (planes * block.expansion), kernel_size=1, stride=stride, bias=False), BatchNorm2d((planes * block.expansion))) layers = [] layers.append(block(self.inplanes, planes, stride, downsample)) self.inplanes = (planes * block.expansion) for i in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def forward(self, x): x = self.relu1(self.bn1(self.conv1(x))) x = self.relu2(self.bn2(self.conv2(x))) x = self.relu3(self.bn3(self.conv3(x))) x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.avgpool(x) x = x.view(x.size(0), (- 1)) x = self.fc(x) return x
def resnet50(pretrained=False, **kwargs): 'Constructs a ResNet-50 model.\n\n Args:\n pretrained (bool): If True, returns a model pre-trained on ImageNet\n ' model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs) if pretrained: model.load_state_dict(load_url(model_urls['resnet50']), strict=False) return model
def resnet18(pretrained=False, **kwargs): 'Constructs a ResNet-18 model.\n Args:\n pretrained (bool): If True, returns a model pre-trained on ImageNet\n ' model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs) if pretrained: model.load_state_dict(load_url(model_urls['resnet18'])) return model
def _sum_ft(tensor): 'sum over the first and last dimention' return tensor.sum(dim=0).sum(dim=(- 1))