for ch in parent_to_child[p]:
temp_df = pickle.load(
open(
pkl_dump_dir + "exclusive/" + algo + "/" + str(iteration) +
"it/" + ch + ".pkl", "rb"))
temp_df["label"] = [ch] * len(temp_df)
if df_weaksup is None:
df_weaksup = temp_df
else:
df_weaksup = pd.concat([df_weaksup, temp_df])
df = pd.concat([df, df_weaksup])
coarse_input_ids, coarse_attention_masks = basic_gpt2_tokenize(
tokenizer, df.text.values, df.label.values, pad_token_dict)
# Combine the training inputs into a TensorDataset.
dataset = TensorDataset(coarse_input_ids, coarse_attention_masks)
# Create a 90-10 train-validation split.
coarse_train_dataloader, coarse_validation_dataloader = create_data_loaders(
dataset, batch_size=4)
model = train(model, tokenizer, coarse_train_dataloader,
coarse_validation_dataloader, doc_start_ind, all_labels,
device, pad_token_dict)
test_generate(model, tokenizer, all_labels, pad_token_dict, device)
tokenizer.save_pretrained(tok_path)
torch.save(model, model_path + model_name)
pickle.dump(pad_token_dict, open(pkl_dump_dir + "pad_token_dict.pkl",
"wb"))
batch_size=1 # Trains with this batch size.
)
label_to_exclusive_dataloader[ch] = func(dataloader)
fine_posterior, fine_model = train(coarse_model,
fine_model,
coarse_tokenizer,
fine_tokenizer,
train_dataloader,
validation_dataloader,
label_to_exclusive_dataloader,
doc_start_ind,
index_to_label,
device,
secondary_device)
test_generate(fine_model, fine_tokenizer, children, pad_token_dict, device)
# true, preds, _ = test(fine_model, fine_posterior, fine_input_ids, fine_attention_masks, doc_start_ind,
# index_to_label, label_to_index, list(temp_df.label.values), device)
# all_true += true
# all_preds += preds
fine_tokenizer.save_pretrained(fine_tok_path)
torch.save(fine_model, fine_model_path + p + ".pt")
torch.save(fine_posterior, fine_label_path + "/fine_posterior.pt")
pickle.dump(index_to_label, open(fine_label_path + "/index_to_label.pkl", "wb"))
pickle.dump(label_to_index, open(fine_label_path + "/label_to_index.pkl", "wb"))
print("*" * 80)
# print(classification_report(all_true, all_preds), flush=True)
print("*" * 80, flush=True)
kl_dataloader = func(kl_dataloader)
# Combine the training inputs into a TensorDataset.
dataset = TensorDataset(input_ids, attention_masks, labels)
# Create a 90-10 train-validation split.
train_dataloader, validation_dataloader = create_data_loaders(dataset, batch_size=4)
doc_start_ind_dict = {}
for ch in children:
doc_start_ind_dict[ch] = doc_start_ind
model = train(fine_model, fine_tokenizer, train_dataloader, validation_dataloader, kl_dataloader,
index_to_label, pad_token_dict, doc_start_ind_dict, device, coarse_model, coarse_tokenizer,
doc_start_ind)
test_generate(model, fine_tokenizer, set(children), pad_token_dict, device)
fine_label_path = base_fine_path + p
os.makedirs(fine_label_path, exist_ok=True)
fine_tok_path = fine_label_path + "/tokenizer"
fine_model_path = fine_label_path + "/model/"
os.makedirs(fine_tok_path, exist_ok=True)
os.makedirs(fine_model_path, exist_ok=True)
fine_tokenizer.save_pretrained(fine_tok_path)
torch.save(fine_model, fine_model_path + p + ".pt")
pickle.dump(index_to_label, open(fine_label_path + "/index_to_label.pkl", "wb"))
pickle.dump(label_to_index, open(fine_label_path + "/label_to_index.pkl", "wb"))
pickle.dump(pad_token_dict, open(fine_label_path + "/pad_token_dict.pkl", "wb"))
print("*" * 80)