new file: .DS_Store

	new file:   basic_utils.py
	new file:   diffusemp/.DS_Store
	new file:   diffusemp/__init__.py
	new file:   diffusemp/__pycache__/__init__.cpython-38.pyc
	new file:   diffusemp/__pycache__/gaussian_diffusion.cpython-38.pyc
	new file:   diffusemp/__pycache__/rounding.cpython-38.pyc
	new file:   diffusemp/__pycache__/step_sample.cpython-38.pyc
	new file:   diffusemp/__pycache__/text_datasets.cpython-38.pyc
	new file:   diffusemp/__pycache__/transformer_model.cpython-38.pyc
	new file:   diffusemp/config.json
	new file:   diffusemp/encode_token_with_control.py
	new file:   diffusemp/gaussian_diffusion.py
	new file:   diffusemp/make_mask.py
	new file:   diffusemp/rounding.py
	new file:   diffusemp/step_sample.py
	new file:   diffusemp/text_datasets.py
	new file:   diffusemp/transformer_model.py
	new file:   diffusemp/utils/.DS_Store
	new file:   diffusemp/utils/__init__.py
	new file:   diffusemp/utils/__pycache__/__init__.cpython-38.pyc
	new file:   diffusemp/utils/__pycache__/dist_util.cpython-38.pyc
	new file:   diffusemp/utils/__pycache__/fp16_util.cpython-38.pyc
	new file:   diffusemp/utils/__pycache__/logger.cpython-38.pyc
	new file:   diffusemp/utils/__pycache__/losses.cpython-38.pyc
	new file:   diffusemp/utils/__pycache__/nn.cpython-38.pyc
	new file:   diffusemp/utils/dist_util.py
	new file:   diffusemp/utils/fp16_util.py
	new file:   diffusemp/utils/logger.py
	new file:   diffusemp/utils/losses.py
	new file:   diffusemp/utils/nn.py
	new file:   epitome/.DS_Store
	new file:   epitome/__init__.py
	new file:   epitome/__pycache__/__init__.cpython-38.pyc
	new file:   epitome/__pycache__/empathy_classifier.cpython-38.pyc
	new file:   epitome/empathy_classifier.py
	new file:   epitome/evaluation_utils.py
	new file:   epitome/models/.DS_Store
	new file:   epitome/models/__init__.py
	new file:   epitome/models/__pycache__/__init__.cpython-38.pyc
	new file:   epitome/models/__pycache__/activations.cpython-38.pyc
	new file:   epitome/models/__pycache__/configuration_bert.cpython-38.pyc
	new file:   epitome/models/__pycache__/configuration_roberta.cpython-38.pyc
	new file:   epitome/models/__pycache__/configuration_utils.cpython-38.pyc
	new file:   epitome/models/__pycache__/file_utils.cpython-38.pyc
	new file:   epitome/models/__pycache__/modeling_bert.cpython-38.pyc
	new file:   epitome/models/__pycache__/modeling_utils.cpython-38.pyc
	new file:   epitome/models/__pycache__/models.cpython-38.pyc
	new file:   epitome/models/__pycache__/roberta.cpython-38.pyc
	new file:   epitome/models/activations.py
	new file:   epitome/models/configuration_bert.py
	new file:   epitome/models/configuration_roberta.py
	new file:   epitome/models/configuration_utils.py
	new file:   epitome/models/file_utils.py
	new file:   epitome/models/modeling_bert.py
	new file:   epitome/models/modeling_utils.py
	new file:   epitome/models/models.py
	new file:   epitome/models/roberta.py
	new file:   epitome/process_data.py
	new file:   epitome/test.py
	new file:   epitome/test_empatheticdialogue.py
	new file:   epitome/train.py
	new file:   eval_util.py
	new file:   requirements.txt
	new file:   sample_seq2seq.py
	new file:   scripts/.DS_Store
	new file:   scripts/eval_seq2seq.py
	new file:   scripts/run_decode.py
	new file:   scripts/run_decode.sh
	new file:   scripts/run_train.py
	new file:   scripts/train.sh
	new file:   scripts/utils/emotion_classifier.py
	new file:   scripts/utils/empathy_classifer.py
	new file:   train.py
	new file:   train_util.py

diffusemp/config.json

@@ -0,0 +1,40 @@
+{
+ "lr": 0.0001,
+ "batch_size": 2048,
+ "microbatch": 64,
+ "learning_steps": 320000,
+ "log_interval": 20,
+ "save_interval": 40000,
+ "eval_interval": 1000,
+ "ema_rate": "0.9999",
+ "resume_checkpoint": "none",
+ "schedule_sampler": "lossaware",
+ "diffusion_steps": 2000,
+ "noise_schedule": "sqrt",
+ "timestep_respacing": "",
+ "vocab": "bert",
+ "use_plm_init": "no",
+ "vocab_size": 0,
+ "config_name": "huggingface-config",
+ "notes": "folder-notes",
+ "data_dir": "data-dir",
+ "dataset": "dataset-name",
+ "checkpoint_path": "checkpoint-path",
+ "seq_len": 128,
+ "hidden_t_dim": 128,
+ "hidden_dim": 128,
+ "dropout": 0.1,
+ "use_fp16": false,
+ "fp16_scale_growth": 0.001,
+ "seed": 102,
+ "gradient_clipping": -1.0,
+ "weight_decay": 0.0,
+ "learn_sigma": false,
+ "use_kl": false,
+ "predict_xstart": true,
+ "rescale_timesteps": true,
+ "rescale_learned_sigmas": false,
+ "sigma_small": false,
+ "emb_scale_factor": 1.0,
+ "use_visible_mask": false
+}

diffusemp/encode_token_with_control.py

@@ -0,0 +1,350 @@
+# this is helpful. thank you-> this, is, help, -ful, . , thank ,you
+# by response-level
+import copy, itertools
+from transformers import BertTokenizer
+tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+import nltk
+# context_ids = [34, 46, 42, 78] # dialogue context + emotion
+data = {"src": "My friend's boyfriend recently made a pass at me. I'm married, and I also am really good friend's with his girlfriend so I felt really bad. I made sure he knew I was loyal and honest to both!Thats horrible. Did you tell on him?I struggled with it, but my friend was in a really bad spot, with no income and raising her grandchild, he was the financial supporter. So, I didn't tell her, as I knew it would really hurt her in several ways, plus she really loves him!", "trg": ["If your best friend is truly a best friend of yours, they won't leave you over this.","It isn't right to tell his girlfriend in my opinion, but it is good if you speak to him about the situation."], "er": "emotional reaction", "ex": "", "ip": "", "sf": "Required_event Personal_relationship Required_event Personal_relationship pronoun _ pronoun _ Departing pronoun _ _ Correctness Telling _ pronoun Opinion pronoun Desirability Statement", "it": ["neutral","acknowledging"]}
+# context_ids = data["src"]
+# target = data["trg"]
+# # alltarget = data['alltrg']
+# ert = "emotional_reaction"
+# ext = data["ex"]
+# ipt = data["ip"]
+# sf = "_ _ Required_event Personal_relationship _ _ _ Required_event Personal_relationship _ _ _ _ _ _ Departing _ Proportional_quantity _ _ _ _ _ Correctness _ Telling _ _ _ _ Opinion _ _ _ _ Desirability _ _ Statement _ _ _ _ _ _"
+# it = data["it"]
+
+target = ['helpfulness', 'thank']
+sf = 'Required_event _'
+ert = "emotional_reaction"
+ext = ""
+ipt = ""
+it = data["it"]
+# CM_ids = [203]
+# IP_ids = []
+# EX_ids = []
+
+# IT_ids = [201, 202]
+# sf_ids = [[101], [102], [103], [105], [106], [108]]
+
+# input_id = [[1],[2],[3,4],[5],[6],[7]]
+# encode_token_with_control(target=input_id, ert = CM_ids, ext = EX_ids, ipt = IP_ids, it=IT_ids, sf=sf_ids)
+# target = ["", ""]
+def add_special_tokens_(tokenizer, model = None): # EDGE
+ """ Add special tokens to the tokenizer and the model if they have not already been added. """
+ # orig_num_tokens = len(tokenizer.encoder)
+ # num_added_tokens = tokenizer.add_special_tokens(ATTR_TO_SPECIAL_TOKEN) # doesn't add if they are already there
+ emotion_word_list = {
+ 'surprised': 0, 'excited': 1, 'annoyed': 2, 'proud': 3, 'angry': 4, 'sad': 5, 'grateful': 6, 'lonely': 7,
+ 'impressed': 8, 'afraid': 9, 'disgusted': 10, 'confident': 11, 'terrified': 12, 'hopeful': 13,
+ 'anxious': 14, 'disappointed': 15,
+ 'joyful': 16, 'prepared': 17, 'guilty': 18, 'furious': 19, 'nostalgic': 20, 'jealous': 21,
+ 'anticipating': 22, 'embarrassed': 23,
+ 'content': 24, 'devastated': 25, 'sentimental': 26, 'caring': 27, 'trusting': 28, 'ashamed': 29,
+ 'apprehensive': 30, 'faithful': 31}.keys()
+ emotion_word_list = list(emotion_word_list)
+ # other_edge_list = ['Yes','No','Why','When','Who','What','Where','Which','?','How','Pronoun']
+ other_edge_list = ['pronoun']
+ intent_word_list = ['agreeing','acknowledging','encouraging','consoling','sympathizing','suggesting','questioning','wishing','neutral']
+ cm_word_list = ["emotional_reaction", "exploration", "interpretation"]
+ frame_word_list = []
+ with open('/path-to-repo/DiffusEmp/datasets/EmpatheticDialogue/frames_list.txt', encoding='utf-8') as f:
+ for sent in f.readlines():
+ for word in sent.strip().split(' '):
+ frame_word_list.append(word)
+ all_special_tokens = emotion_word_list+intent_word_list+cm_word_list+frame_word_list+other_edge_list
+ num_added_toks = tokenizer.add_tokens(all_special_tokens)
+ if model is not None and num_added_toks > 0:
+ model.resize_token_embeddings(len(tokenizer))
+
+def encode_token_with_control(context, target, ert, ext, ipt, it, sf): #
+ er_ids, ex_ids, ip_ids, it_ids, sf_ids, ids = [], [], [], [], [], [] #id
+ tokenized_context_id = tokenizer(context, add_special_tokens=True)['input_ids'] 
+ tokenized_er_id = tokenizer(ert, add_special_tokens=False)['input_ids'] 
+ tokenized_ex_id = tokenizer(ext, add_special_tokens=False)['input_ids']
+ tokenized_ip_id = tokenizer(ipt, add_special_tokens=False)['input_ids']
+ concat_input_ids, split_input_ids, concat_it_ids, union_it_ids = [], [], [], []
+ for i in range(len(target)):
+ temp_target = nltk.tokenize.word_tokenize(target[i].strip())
+ input_it_ids = tokenizer(it[i], add_special_tokens=False)['input_ids'] if len(it[i])>0 else [] # 9
+ subsentence_input_ids = []
+ for j in range(len(temp_target)):
+ input_ids = tokenizer(temp_target[j], add_special_tokens=False)['input_ids']#listlist[[1],[2],[3,5]]
+ subsentence_input_ids.append(input_ids)
+ temp = [t for l in subsentence_input_ids for t in l]
+ subsentence_it_ids = input_it_ids * len(temp)
+
+ #iinten 
+ union_it_ids.append((i, len(subsentence_it_ids)))
+
+ concat_it_ids.extend(subsentence_it_ids)
+ concat_input_ids.extend(temp)
+ split_input_ids.extend(subsentence_input_ids)
+
+ if len(sf) == 0:
+ sf = ['_']*len(concat_input_ids) 
+ else:
+ sf = sf.split()
+
+ concat_sf_ids, union_sf_ids = [], []
+ input_sf_ids = tokenizer(sf, add_special_tokens=False)['input_ids']list [[6],[7],[8]]
+ for i in range(len(input_sf_ids)):
+ concat_sf_ids.extend(input_sf_ids[i] * len(split_input_ids[i]))
+ union_sf_ids.append((i, len(split_input_ids[i])))
+
+ allid = tokenized_er_id + tokenized_ex_id + tokenized_ip_id + concat_it_ids + concat_sf_ids + concat_input_ids
+
+ position2id = {i:a for a, i in zip(allid, range(len(allid)))}
+ same_sf = max([len(list(v)) for k,v in itertools.groupby(concat_sf_ids)])
+ ### Target
+ vertice = [(i,a) for i, a in zip(range(len(allid)), allid)]
+ vertice_index = 0
+ er_vertice = [(i,a) for i, a in zip(range(vertice_index, vertice_index+len(tokenized_er_id)), tokenized_er_id)]
+ vertice_index += len(er_vertice)
+ ex_vertice = [(i,a) for i, a in zip(range(vertice_index, vertice_index+len(tokenized_ex_id)), tokenized_ex_id)]
+ vertice_index += len(ex_vertice)
+ ip_vertice = [(i,a) for i, a in zip(range(vertice_index, vertice_index+len(tokenized_ip_id)), tokenized_ip_id)]
+ vertice_index += len(ip_vertice)
+ it_vertice_index = vertice_index
+ it_vertice = [(i,a) for i, a in zip(range(vertice_index, vertice_index+len(concat_it_ids)), concat_it_ids)]
+ vertice_index += len(it_vertice)
+
+ sf_vertice_index = vertice_index
+ sf_vertice = [(i,a) for i, a in zip(range(vertice_index, vertice_index+len(concat_sf_ids)), concat_sf_ids)]
+ vertice_index += len(sf_vertice)
+ target_vertice = [(i,a) for i, a in zip(range(vertice_index, vertice_index+len(concat_input_ids)), concat_input_ids)]
+ control_vertice = er_vertice + ex_vertice + ip_vertice + it_vertice + sf_vertice
+
+ union_it = []
+ it_front_len = it_vertice_index
+ for (i, l) in union_it_ids:
+ union_it.append(list(range(it_front_len, it_front_len+l)))
+ it_front_len += l
+
+ union_sf = []
+ sf_front_len = sf_vertice_index
+ for (i, l) in union_sf_ids:
+ union_sf.append(list(range(sf_front_len, sf_front_len+l)))
+ sf_front_len += l
+
+ union = union_it + union_sf
+
+
+ target_edges = {}
+ control_edges = {}
+ for j in range(len(target_vertice)):
+ control_vertice_j = tokenized_er_id + tokenized_ex_id + tokenized_ip_id + [concat_it_ids[j]] + [concat_sf_ids[j]]
+ target_edges[target_vertice[j]] = target_vertice + er_vertice + ex_vertice + ip_vertice + [it_vertice[j]] + [sf_vertice[j]]
+ for k, v in target_edges.items():
+ for vi in v:
+ if vi not in control_edges.keys():
+ control_edges[vi] = [k]
+ else:
+ control_edges[vi].extend([k])
+ for i in range(len(union)):
+ if len(union[i]) == 1: continue
+ else:
+ temp = []
+ for ui in union[i]:
+ temp.append((ui, position2id[ui]))
+ union[i] = temp
+ union_reverse_dict = {}
+ for i in range(len(union)):
+ if len(union[i]) == 1:continue
+ else:
+ a = union[i][0]
+ for bi in range(1, len(union[i])):
+ b = union[i][bi]
+ union_reverse_dict[b] = a
+ for k,v in target_edges.items():
+ for vi in range(len(v)):
+ if v[vi] in union_reverse_dict.keys():
+ v[vi] = union_reverse_dict[v[vi]]
+ target_edges[k] = v
+ for k,v in control_edges.items():
+ if k in er_vertice: control_edges[k] += er_vertice
+ if k in ex_vertice: control_edges[k] += ex_vertice
+ if k in ip_vertice: control_edges[k] += ip_vertice
+ if k in it_vertice: control_edges[k] += it_vertice
+ if k in sf_vertice: control_edges[k] += sf_vertice
+ if k in target_edges.keys(): control_edges[k] = target_edges[k]
+ for k,v in control_edges.items(): #
+ for vi in range(len(v)):
+ if v[vi] in union_reverse_dict.keys():
+ v[vi] = union_reverse_dict[v[vi]]
+ control_edges[k] = v
+ if k in union_reverse_dict.keys():
+ control_edges[union_reverse_dict[k]]+=control_edges[k]
+ control_edges[k] = []
+ for k,v in control_edges.items():
+ control_edges[k] = list(set(control_edges[k]))
+ new_control_edges = {}
+ for k,v in control_edges.items():
+ if len(v)>0: new_control_edges[k] = v
+ old_vertices = sorted(new_control_edges.keys())
+ new_vertices = []
+ for i in range(len(old_vertices)):
+ old = old_vertices[i]
+ new_vertices.append((i, old[1]))
+ old2new = {k:v for k,v in zip(old_vertices, new_vertices)}
+
+ edge = {} # control+target
+ for k,v in control_edges.items(): #
+ if k in old2new.keys():
+ ki = old2new[k]
+ edge[ki] = []
+ for vi in range(len(v)):
+ edge[ki].append(old2new[v[vi]])
+ input_ids_ctrl_target = [v[1] for v in sorted(edge.keys())]
+ sep_between_control_target = -len(concat_input_ids)
+ input_ids_ctrl = input_ids_ctrl_target[:-len(concat_input_ids)]
+ input_ids_x = tokenized_context_id + input_ids_ctrl + [tokenizer.sep_token_id]
+ # input_ids_ctrl_sep_target = input_ids_ctrl_target.insert(sep_between_control_target, tokenizer.sep_token_id)
+ # input_ids_c =[tokenizer.cls_token_id] + [v[1] for v in sorted(edge.keys())] + [tokenizer.sep_token_id] 
+ input_ids_y = [tokenizer.cls_token_id] + concat_input_ids + [tokenizer.sep_token_id]
+ edge_index = {}
+ context_len = len(tokenized_context_id)
+ for k,v in edge.items():
+ nk = k[0] + context_len
+ nv = [vi[0] + context_len for vi in v] + list(range(context_len))
+ edge_index[nk] = sorted(nv)
+ control_target_len = len(edge_index.keys())
+ for i in range(context_len):
+ # temp = context_len + control_target_len
+ edge_index[i] = list(range(context_len + control_target_len))
+ # context_len = len(tokenized_context_id)
+ total_len = len(edge_index)
+ mask = []
+ sep_between_control_target = -len(concat_input_ids)
+ sep_between_x_y = sep_between_control_target + 1
+ cls_target = sep_between_control_target + 2
+ sep_target = -1
+ for i in range(total_len):
+ temp = [0] * total_len
+ for j in edge_index[i]:
+ temp[j] = 1
+ temp.insert(sep_between_control_target, 1)
+ temp.insert(sep_between_control_target, 1)
+ temp.insert(sep_between_control_target, 1)
+ temp.append(1)
+ mask.append(temp)
+ mask.insert(sep_between_control_target, [1]*(total_len+4))
+ mask.insert(sep_between_control_target, [1]*(total_len+4))
+ mask.insert(sep_between_control_target, [1]*(total_len+4))
+ mask.append([1]*(total_len+4))
+ input_ids_c = edge_index
+ # sep_between_control_target = -len(concat_input_ids)
+
+ return input_ids_x, input_ids_y, mask, len(tokenized_context_id)
+
+
+def compute_mask(context_ids, input_id, sf_ids, IT_ids, CM_ids, seg_pos):
+
+ # obtain subwords seq
+ input_ids = []
+ for id in input_id:
+ if len(id) == 1:
+ input_ids.append(id[0])
+ elif len(id) > 1:
+ for id_ in id:
+ input_ids.append(id_)
+
+ # initial mask vertor for every ids
+ mask_ids = [0] * len(context_ids + CM_ids + IT_ids + sf_ids + input_ids)
+
+ pos_CM_ids = len(context_ids)
+ pos_IT_ids = pos_CM_ids + len(CM_ids)
+ pos_sf_ids = pos_IT_ids + len(IT_ids)
+ pos_input_ids = pos_sf_ids + len(sf_ids)
+
+
+ # all context_ids and input_ids can see each other, 
+ # prepared a temp for the next step
+ tmp_context = []
+ for i in range(len(context_ids)):
+ tmp_context.append(i)
+
+ tmp_IT = []
+ for i in range(pos_IT_ids+1, pos_sf_ids):
+ tmp_IT.append(i)
+
+ tmp_sf = []
+ for i in range(pos_sf_ids+1, pos_input_ids):
+ tmp_sf.append(i)
+
+ tmp_target = []
+ for i in range(pos_input_ids+1, len(mask_ids)):
+ tmp_target.append(i)
+
+ one_hots_CM = tmp_context + [pos_CM_ids] + tmp_target
+
+ one_hots_IT = []
+ for i in range(len(IT_ids)):
+ pos = tmp_context + tmp_IT + tmp_target[seg_pos[i]:seg_pos[i+1]]
+ one_hots_IT.append(pos)
+
+ # memorize the pos of one-hots
+ one_hots_target = []
+ one_hots_sf = []
+ for i in range(len(input_id)):
+ if len(input_id[i]) == 1:
+ rst = tmp_context + [pos_CM_ids] + tmp_IT + tmp_sf + tmp_target
+ one_hots_target.append(rst)
+ one_hots_sf.append(tmp_context + tmp_sf + [tmp_target[i]])
+
+ elif len(input_id[i]) > 1:
+ for j in range(len(input_id[i])):
+ rst = tmp_context + [pos_CM_ids] + tmp_IT + tmp_sf + tmp_target
+ one_hots_target.append(rst)
+ one_hots_sf.append(tmp_context + tmp_sf + [tmp_target[i]])
+
+ attention_mask = []
+ for _ in context_ids:
+ attention_mask.append([1] * len(context_ids + CM_ids + IT_ids + sf_ids + input_ids))
+
+ for _ in CM_ids:
+ new_mask_ids = copy.deepcopy(mask_ids)
+ for pos in one_hots_CM:
+ new_mask_ids[pos] = 1
+ attention_mask.append(new_mask_ids)
+
+ for _ in IT_ids:
+ new_mask_ids1 = copy.deepcopy(mask_ids)
+ for poss in one_hots_IT:
+ for pos in poss:
+ new_mask_ids1[pos] = 1
+ attention_mask.append(new_mask_ids1)
+
+ for _ in sf_ids:
+ new_mask_ids2 = copy.deepcopy(mask_ids)
+ for poss in one_hots_sf:
+ for pos in poss:
+ new_mask_ids2[pos] = 1
+ attention_mask.append(new_mask_ids2)
+
+ for _ in input_ids:
+ new_mask_ids3 = copy.deepcopy(mask_ids)
+ for poss in one_hots_target:
+ for pos in poss:
+ new_mask_ids3[pos] = 1
+ attention_mask.append(new_mask_ids3)
+
+ return attention_mask
+
+context_ids = [34, 46, 42, 78] # dialogue context + emotion
+
+CM_ids = [203]
+IP_ids = []
+EX_ids = []
+
+IT_ids = [201, 202]
+sf_ids = [[101], [102], [103], [105], [106], [108]]
+input_id = [[3,4],[1],[2],[5],[6],[7]]
+add_special_tokens_(tokenizer=tokenizer)
+encode_token_with_control("test new", target, ert, ext, ipt, it, sf)
+
+# seg_pos = [0,4,6]
+# attention_mask = compute_mask(context_ids, input_id, sf_ids, IT_ids, CM_ids, seg_pos)
+# for mask in attention_mask:
+# print(mask)