def load_atomic_data(opt, LoaderPath=""):
# Hacky workaround, you may have to change this
# if your models use different pad lengths for e1, e2, r
###################SET MAXE1, MAXE2 HERE ASWELL
###############################################################################
#print(opt)#DEBUG
if opt.data.get("maxe1", None) is None:
opt.data.maxe1 = 50 ###DEFAULT FOR COMET PRETRAINED MODELS: MAXE1=17 , MAXE2 = 35 MAXR=1
opt.data.maxe2 = 35
opt.data.maxr = 1
path = "data/atomic/processed/generation/{}.pickle".format(
utils.make_name_string(opt.data))
#data loader is made directly from data in atomic folder
data_loader = data.make_data_loader(opt, opt.data.categories)
#loaded = data_loader.load_data(path)###OLD####
#ADRIAN ADDED##############################CHANGE FOR DIFF LOADER
if (LoaderPath == ""):
print('No loader path provided, using defalut loader')
loaded = data_loader.load_data(path) ###OLD####
else:
print("Loader Path Provided:", LoaderPath)
#FPathEng = 'MULTI_COMET_DATA/It50k_MaxE50/English/Eng_Loader_It50k_maxE50.pickle'
#FPathSlo = 'MULTI_COMET_DATA/It50k_MaxE50/Slovene/Slo_Loader_It50k_maxE50.pickle'
print("IS VALID FILE: ", os.path.isfile(LoaderPath))
loaded = data_loader.load_data(LoaderPath)
##########################
#print(data_loader.vocab_encoder)#DEBUG
return data_loader
def load_conceptnet_data(opt):
# Hacky workaround, you may have to change this
# if your models use different pad lengths for r
if opt.data.get("maxr", None) is None:
if opt.data.rel == "language":
opt.data.maxr = 5
else:
opt.data.maxr = 1
path = "data/conceptnet/processed/generation/{}.pickle".format(
utils.make_name_string(opt.data))
data_loader = data.make_data_loader(opt)
loaded = data_loader.load_data(path)
return data_loader
def load_atomic_data(opt):
# Hacky workaround, you may have to change this
# if your models use different pad lengths for e1, e2, r
if opt.data.get("maxe1", None) is None:
opt.data.maxe1 = 17
opt.data.maxe2 = 35
opt.data.maxr = 1
path = "data/atomic/processed/generation/{}.pickle".format(
utils.make_name_string(opt.data))
data_loader = data.make_data_loader(opt, opt.data.categories)
loaded = data_loader.load_data(path)
return data_loader
def main(num):
# Generate configuration files depending on experiment being run
utils.generate_config_files("atomic", num)
# Loads the correct configuration file
config_file = "config/atomic/config_{}.json".format(num)
print(config_file)
# Read config file to option
config = cfg.read_config(cfg.load_config(config_file))
opt, meta = cfg.get_parameters(config)
# Set the random seeds
torch.manual_seed(opt.train.static.seed)
random.seed(opt.train.static.seed)
if config.gpu_mode:
torch.cuda.manual_seed_all(opt.train.static.seed)
# Where to find the data
splits = ["train", "dev", "test"]
opt.train.dynamic.epoch = 0
print("Loading Data")
categories = opt.data.categories
path = "data/atomic/processed/{}/{}.pickle".format(
opt.exp, utils.make_name_string(opt.data))
data_loader = data.make_data_loader(opt, categories)
loaded = data_loader.load_data(path)
print(data_loader.sequences["train"]["total"].size(0))
data_loader.opt = opt
data_loader.batch_size = opt.train.dynamic.bs
print("Done.")
# Initialize text_encoder
text_encoder = TextEncoder(config.encoder_path, config.bpe_path)
special = [data.start_token, data.end_token]
special += ["<{}>".format(cat) for cat in categories]
special += [data.blank_token]
text_encoder.encoder = data_loader.vocab_encoder
text_encoder.decoder = data_loader.vocab_decoder
opt.data.maxe1 = data_loader.max_event
opt.data.maxe2 = data_loader.max_effect
opt.data.maxr = data.atomic_data.num_delimiter_tokens["category"]
n_special = len(special)
n_ctx = opt.data.maxe1 + opt.data.maxe2
n_vocab = len(text_encoder.encoder) + n_ctx
print(data_loader.__dict__.keys())
opt.net.vSize = n_vocab
print("Building Model")
model = models.make_model(opt,
n_vocab,
n_ctx,
n_special,
load=(opt.net.init == "pt"))
print("Done.")
print("Files will be logged at: {}".format(
utils.make_name(opt, prefix="results/losses/", is_dir=True,
eval_=True)))
data_loader.reset_offsets("train")
# Get number of examples
data.set_max_sizes(data_loader)
if config.gpu_mode:
print("Pushing to GPU: {}".format(config.gpu_index))
cfg.device = config.gpu_index
cfg.do_gpu = True
torch.cuda.set_device(cfg.device)
if config.multigpu:
model = models.multi_gpu(model, config.gpu_indices).cuda()
else:
model.cuda(cfg.device)
print("Done.")
print("Training")
optimizer = OpenAIAdam(model.parameters(),
lr=opt.train.dynamic.lr,
schedule=opt.train.static.lrsched,
warmup=opt.train.static.lrwarm,
t_total=meta.iterations,
b1=opt.train.static.b1,
b2=opt.train.static.b2,
e=opt.train.static.e,
l2=opt.train.static.l2,
vector_l2=opt.train.static.vl2,
max_grad_norm=opt.train.static.clip)
scorers = ["bleu", "rouge", "cider"]
trainer = train.make_trainer(opt, meta, data_loader, model, optimizer)
trainer.set_evaluator(opt, model, data_loader)
trainer.run()
opt.exp = "generation"
data_params = get_data_params(gens_file)
categories = data_params[
"categories"] #sorted(["oReact", "oEffect", "oWant", "xAttr", "xEffect", "xIntent", "xNeed", "xReact", "xWant"])
opt.data.categories = data_params["categories"]
if "maxe1" in data_params:
opt.data.maxe1 = data_params["maxe1"]
opt.data.maxe2 = data_params["maxe2"]
opt.data.maxr = data_params["maxr"]
path = "data/atomic/processed/generation/{}.pickle".format(
utils.make_name_string(opt.data))
data_loader = data.make_data_loader(opt, categories)
loaded = data_loader.load_data(path)
refs = {}
for i in range(data_loader.sequences[split]["total"].size(0)):
sequence = data_loader.sequences[split]["total"][i]
tmp = sequence[:data_loader.max_event + 1]
init = "".join([
data_loader.vocab_decoder[i].replace('</w>',
' ').replace("<blank>", "___ ")
for i in tmp[:-1].squeeze().tolist() if i
])
attr = data_loader.vocab_decoder[tmp[-1].item()].strip("<>")
Ref = sequence[data_loader.max_event + 1:]
# Set the random seeds
torch.manual_seed(args.seed)
random.seed(args.seed)
if config.gpu_mode:
torch.cuda.manual_seed_all(args.seed)
opt.train.dynamic.epoch = 0
print("Loading Data")
f = open("atomic_generate.txt", "w")
f.close()
categories = opt.data.categories
x = "data/atomic/processed/generation/categories_oEffect#oReact#oWant#xAttr#xEffect#xIntent#xNeed#xReact#xWant-maxe1_17-maxe2_36-maxr_1.pickle"
path = x.format(utils.make_name_string(opt.data))
data_loader = data.make_data_loader(opt, categories)
loaded = data_loader.load_data(path)
data_loader.batch_size = opt.train.dynamic.bs
print("Done.")
categories = opt.data.categories
text_encoder = GPT2Tokenizer.from_pretrained('gpt2')
special = [data.start_token, data.end_token]
special += ["<{}>".format(cat) for cat in categories]
special += [data.blank_token]
special_tokens = {
"cls_token": "[CLS]",
print(set_of_categories)
# Iterate over sets of categories to compute perplexities for
for eval_categories in set_of_categories:
print(eval_categories)
opt.eval.categories = eval_categories
results_name = "{}/{}.{}".format(
utils.make_name(opt,
prefix="results/{}/".format("losses"),
is_dir=True,
eval_=True), split, "pickle")
print("Will save {} losses to {}".format(split, results_name))
path = "data/atomic/processed/generation/{}.pickle".format(
utils.make_name_string(opt.data).replace(
"kr_{}".format(opt.data.get("kr", 1)), "kr_1"))
data_loader = data.make_data_loader(opt, opt.data.categories)
loaded = data_loader.load_data(path)
data_loader.batch_size = opt.train.dynamic.bs
print("Done.")
text_encoder = TextEncoder(config.encoder_path, config.bpe_path)
# Set special tokens
formatted_categories = ["<{}>".format(cat) for cat in opt.data.categories]
special = [data.start_token, data.end_token]
special += formatted_categories
special += [data.blank_token]
def main(num):
# Generate configuration files depending on experiment being run
utils.generate_config_files("conceptnet", num)
# Loads the correct configuration file
config_file = "config/conceptnet/config_{}.json".format(num)
print(config_file)
# Read config file to option
config = cfg.read_config(cfg.load_config(config_file))
opt, meta = cfg.get_parameters(config)
# config.gpu_mode = torch.cuda.is_available()
# Set the random seeds
torch.manual_seed(opt.train.static.seed)
random.seed(opt.train.static.seed)
if config.gpu_mode:
torch.cuda.manual_seed_all(opt.train.static.seed)
# Load the data
splits = ["train", "dev", "test"]
opt.train.dynamic.epoch = 0
print("Loading Data")
# Initialize path to pre-set data loader
path = "data/conceptnet/processed/{}/{}.pickle".format(
opt.exp, utils.make_name_string(opt.data))
# Make data loader
data_loader = data.make_data_loader(opt)
loaded = data_loader.load_data(path)
print(data_loader.sequences["train"]["total"].size(0))
data_loader.opt = opt
data_loader.batch_size = opt.train.dynamic.bs
print("Done.")
text_encoder = TextEncoder(config.encoder_path, config.bpe_path)
categories = data.conceptnet_data.conceptnet_relations
special = [data.start_token, data.end_token]
special += ["<{}>".format(cat) for cat in categories]
if loaded:
text_encoder.encoder = data_loader.vocab_encoder
text_encoder.decoder = data_loader.vocab_decoder
else:
for special_token in special:
text_encoder.decoder[len(encoder)] = special_token
text_encoder.encoder[special_token] = len(encoder)
data_loader.make_tensors(text_encoder, special)
# Set max size of different parts of relation
context_size_e1 = data_loader.max_e1
context_size_e2 = data_loader.max_e2
context_size_r = data_loader.max_r
opt.data.maxr = context_size_r
n_special = len(special)
n_ctx = context_size_e1 + context_size_r + context_size_e2
n_vocab = len(text_encoder.encoder) + n_ctx
print(data_loader.__dict__.keys())
opt.net.vSize = n_vocab
# Build Model
print("Building Model")
model = models.make_model(opt,
n_vocab,
n_ctx,
n_special,
load=(opt.net.init == "pt"))
print("Done.")
print("Files will be logged at: {}".format(
utils.make_name(opt, prefix="results/losses/", is_dir=True,
eval_=True)))
data_loader.reset_offsets("train", keys=["total"])
data.set_max_sizes(data_loader)
# Push to GPU
if config.gpu_mode:
print("Pushing to GPU: {}".format(config.gpu_index))
cfg.device = config.gpu_index
cfg.do_gpu = True
torch.cuda.set_device(cfg.device)
if config.multigpu:
model = models.multi_gpu(model, config.gpu_indices).cuda()
else:
model.cuda(cfg.device)
print("Done.")
print("Training")
optimizer = OpenAIAdam(model.parameters(),
lr=opt.train.dynamic.lr,
schedule=opt.train.static.lrsched,
warmup=opt.train.static.lrwarm,
t_total=meta.iterations,
b1=opt.train.static.b1,
b2=opt.train.static.b2,
e=opt.train.static.e,
l2=opt.train.static.l2,
vector_l2=opt.train.static.vl2,
max_grad_norm=opt.train.static.clip)
trainer = train.make_trainer(opt, meta, data_loader, model, optimizer)
print(data_loader.sequences["dev"]["total"].max())
trainer.set_generator(opt, model, data_loader)
trainer.set_evaluator(opt, model, data_loader)
trainer.run()
def main(num):
# Generate configuration files depending on experiment being run
#utils.generate_config_files("conceptnet", num)
# Loads the correct configuration file
config_file = "config/conceptnet/config_{}.json".format(num)
print(config_file)
# Read config file to option
config = cfg.read_config(cfg.load_config(config_file))
opt, meta = cfg.get_parameters(config)
# config.gpu_mode = torch.cuda.is_available()
# Set the random seeds
torch.manual_seed(opt.train.static.seed)
random.seed(opt.train.static.seed)
if config.gpu_mode:
torch.cuda.manual_seed_all(opt.train.static.seed)
# Load the data
splits = ["train", "dev", "test"]
opt.train.dynamic.epoch = 0
print("Loading Data")
# Initialize path to pre-set data loader
x = "data/conceptnet/processed/generation/rel_language-trainsize_100-devversion_12-maxe1_200-maxe2_200.pickle"
path = x.format(opt.exp, utils.make_name_string(opt.data))
print(path)
# Make data loader
data_loader = data.make_data_loader(opt)
loaded = data_loader.load_data(path)
#print(data_loader.sequences["train"]["total"].size(0))
data_loader.opt = opt
data_loader.batch_size = opt.train.dynamic.bs
print("Done.")
print(data_loader)
#text_encoder = TextEncoder(config.encoder_path, config.bpe_path)
text_encoder = GPT2Tokenizer.from_pretrained('gpt2')
special_tokens = {
"cls_token": "[CLS]",
"unk_token": "[UNK]"
} #, "mask": '["MASK"]',"separator": '["SEP"]', "start_of_sentence": '["SOS"]', "end_of_sentence": '["EOS"]'}
text_encoder = GPT2Tokenizer.from_pretrained("gpt2",
cls_token="[CLS]",
unk_token="[UNK]",
mask='["MASK"]',
separator='["SEP"]',
start_of_sentence='["SOS"]',
end_of_sentence='["EOS"]')
text_encoder.add_special_tokens(special_tokens)
#categories = data.conceptnet_data.conceptnet_relations
special = [data.start_token, data.end_token]
#special += ["<{}>".format(cat) for cat in categories]
if loaded:
text_encoder.encoder = data_loader.vocab_encoder
text_encoder.decoder = data_loader.vocab_decoder
else:
for special_token in special:
text_encoder.decoder[len(encoder)] = special_token
text_encoder.encoder[special_token] = len(encoder)
data_loader.make_tensors(text_encoder, special)
# Set max size of different parts of relation
context_size_i1 = data_loader.max_input1
context_size_i2 = data_loader.max_input2
context_size_i3 = data_loader.max_input3
context_size_i4 = data_loader.max_input4
context_size_o1 = data_loader.max_output1
context_size_o2 = data_loader.max_output2
context_size_o3 = data_loader.max_output3
context_size_o4 = data_loader.max_output4
#opt.data.maxr = context_size_r
n_special = len(special)
n_ctx = context_size_i1 + context_size_i2 + context_size_i3 + context_size_i4 + context_size_o1 + context_size_o2 + context_size_o3 + context_size_o4
n_vocab = len(text_encoder.encoder) + n_ctx
opt.net.vSize = n_vocab
# Build Model
print("Building Model")
print(opt.net.init == "pt")
model = models.make_model(opt, n_vocab, n_ctx, n_special)
model.resize_token_embeddings(len(text_encoder))
model_knowledge = model_knowledge_story.make_model(opt, n_vocab, n_ctx,
n_special)
model_knowledge.resize_token_embeddings(len(text_encoder))
print("Done.")
print("Files will be logged at: {}".format(
utils.make_name(opt, prefix="results/losses/", is_dir=True,
eval_=True)))
data_loader.reset_offsets("train", keys=["total"])
data.set_max_sizes(data_loader)
# Push to GPU
if config.gpu_mode:
print("Pushing to GPU: {}".format(config.gpu_index))
cfg.device = config.gpu_index
cfg.do_gpu = True
torch.cuda.set_device(cfg.device)
if config.multigpu:
#print("!!! I am here !!!")
model = models.multi_gpu(model, config.gpu_indices).cuda()
#model.to(f'cuda:{model.device_ids[0]}')
model_knowledge = model_knowledge_story.multi_gpu(
model_knowledge, config.gpu_indices).cuda()
#model_knowledge.to(f'cuda:{model.device_ids[1]}')
else:
model.cuda(cfg.device)
model_knowledge.cuda(cfg.device)
print("Done.")
print("Training")
optimizer_m = OpenAIAdam(model.parameters(),
lr=opt.train.dynamic.lr,
schedule=opt.train.static.lrsched,
warmup=opt.train.static.lrwarm,
t_total=meta.iterations,
b1=opt.train.static.b1,
b2=opt.train.static.b2,
e=opt.train.static.e,
l2=opt.train.static.l2,
vector_l2=opt.train.static.vl2,
max_grad_norm=opt.train.static.clip)
optimizer_k = Knowledge_Adam(model_knowledge.parameters(),
lr=opt.train.dynamic.lr,
schedule=opt.train.static.lrsched,
warmup=opt.train.static.lrwarm,
t_total=meta.iterations,
b1=opt.train.static.b1,
b2=opt.train.static.b2,
e=opt.train.static.e,
l2=opt.train.static.l2,
vector_l2=opt.train.static.vl2,
max_grad_norm=opt.train.static.clip)
trainer = train.make_trainer(opt, meta, data_loader, model,
model_knowledge, optimizer_m, optimizer_k)
trainer.set_generator(opt, model, model_knowledge, data_loader)
trainer.set_evaluator(opt, model, model_knowledge, data_loader)
trainer.run()
]
special = [data.start_token, data.end_token]
special += ["<{}>".format(relation) for relation in relations]
encoder_path = "model/encoder_bpe_40000.json"
bpe_path = "model/vocab_40000.bpe"
text_encoder = TextEncoder(encoder_path, bpe_path)
for special_token in special:
text_encoder.decoder[len(text_encoder.encoder)] = special_token
text_encoder.encoder[special_token] = len(text_encoder.encoder)
data_loader = cdata.GenerationDataLoader(opt)
data_loader.load_data("data/conceptnet/")
data_loader.make_tensors(text_encoder, special, test=False)
opt.data.maxr = data_loader.max_r
save_path = "data/conceptnet/processed/generation"
save_name = os.path.join(save_path,
"{}.pickle".format(utils.make_name_string(opt.data)))
utils.mkpath(save_path)
print("Data Loader will be saved to {}".format(save_name))
torch.save(data_loader, save_name)
special = [data.start_token, data.end_token]#, data.sep_token]
special += ["<{}>".format(cat) for cat in categories]
special += [data.blank_token]
for special_token in special:
text_encoder.decoder[len(encoder)] = special_token
encoder[special_token] = len(encoder)
save_path = "data/atomic/processed/{}".format(opt.exp)
utils.mkpath(save_path)
#ipdb.set_trace()
if opt.pickled_data:
save_name = opt.pickled_data
else:
save_name = os.path.join(save_path, "{}.pickle".format(utils.make_name_string(opt.data)))
data_loader = data.make_data_loader(opt, categories)
data_loader.load_data("data/atomic/")
random.shuffle(data_loader.data["dev"]["total"])
data_loader.make_tensors(text_encoder, special, test=False)
data_loader.reset_offsets()
opt.data.maxe1 = data_loader.max_event
opt.data.maxe2 = data_loader.max_effect
opt.data.maxr = 1
if opt.pickled_data:
save_name = opt.pickled_data