| | class SearchMood: |
| | def __init__(self, mood_prompt, prior_init): |
| | self.prior_init = prior_init |
| | self.mood_prompt = mood_prompt |
| | self.model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2') |
| | self.embeding = lambda mood_prompt, mood_state: (self.model.encode(mood_prompt, convert_to_tensor=True), self.model.encode(mood_state, convert_to_tensor=True)) |
| | self.similar = lambda similarx, similary: util.pytorch_cos_sim(similarx, similary) |
| | self.cx_sample = shelve.open('cx_sample.db')['sample'] |
| | self.database = shelve.open('database.db') |
| | SearchMood.prior_component = torch.tensor([0.,1.]) |
| | self.prior_sample = torch.normal(self.prior_component[0], self.prior_component[1], size=(5,)) |
| | self.sample_losses = None |
| | |
| | |
| |
|
| |
|
| |
|
| |
|
| |
|
| | def Hierarchical(self, data): |
| | |
| |
|
| | |
| | clusters, distances = hierarchical(data, distance='cosine', linkage='complete', return_clusters=True) |
| |
|
| | |
| | print(clusters) |
| |
|
| | def embedings(self, samplex, sampley): |
| | emb = self.embeding(samplex, sampley) |
| | embHx = self.Hierarchical(emb[0]) |
| | embHy = self.Hierarchical(emb[1]) |
| | similarity = self.similar(embHx, embHy) |
| |
|
| | return(similarity) |
| |
|
| |
|
| |
|
| | def mood_dist(self, data_sample=False, mood_prompt=False, search=True): |
| | cx_index = [] |
| | if search == True: |
| | for mood_state in self.cx_sample: |
| | index_sample = [] |
| | max_sample = 0 |
| | index_sample = 0 |
| | for index, mood_prompts in enumerate(self.database['database']): |
| | simemb = self.embedings(mood_state, mood_prompts) |
| | if max_sample < simemb: |
| | max_sample = simemb |
| | index_sample = index |
| | |
| | |
| | cx_index.append((float(index_sample))) |
| |
|
| |
|
| | else: |
| | cx_index.append(self.embedings(mood_prompt, data_sample)) |
| |
|
| | |
| | |
| | return(torch.tensor(cx_index)) |
| |
|
| |
|
| |
|
| | def loss_fn(self): |
| | |
| | for sample in self.prior_sample: |
| | sample = sample.item() |
| | data_sample = self.database['database'][round(sample)] |
| | samp_loss = self.mood_dist(data_sample, self.mood_prompt, search=False) |
| | print(samp_loss) |
| | if samp_loss.item() >= 1.: |
| | print('test') |
| | break |
| |
|
| |
|
| |
|
| | return(torch.tensor([samp_loss*-1])) |
| |
|
| |
|
| |
|
| | def search_compose(self): |
| | for d in range(100): |
| | optimizer = optim.Adagrad((self.prior_component[0], self.prior_component[1])) |
| | optimizer.step(closure=self.loss_fn) |
| | state_dict = optimizer.state_dict() |
| | params = state_dict['param_groups'][0]['params'] |
| |
|
| | self.prior_component[0] = params[0] |
| | self.prior_component[1] = params[1] |
| | self.prior_sample = torch.normal(self.prior_component[0], self.prior_component[1], size=(5,)) |
| |
|
| |
|
| |
|
| | |
| |
|
| |
|
| |
|
