def main():
trainset, valset, testset = cath_dataset(
1800, jsonl_file=sys.argv[1]) # batch size = 1800 residues
optimizer = tf.keras.optimizers.Adam()
model = make_model()
model_id = int(datetime.timestamp(datetime.now()))
NUM_EPOCHS = 100
loop_func = util.loop
best_epoch, best_val = 0, np.inf
for epoch in range(NUM_EPOCHS):
loss, acc, confusion = loop_func(trainset,
model,
train=True,
optimizer=optimizer)
util.save_checkpoint(model, optimizer, model_id, epoch)
print('EPOCH {} TRAIN {:.4f} {:.4f}'.format(epoch, loss, acc))
util.save_confusion(confusion)
loss, acc, confusion = loop_func(valset, model, train=False)
if loss < best_val:
best_epoch, best_val = epoch, loss
print('EPOCH {} VAL {:.4f} {:.4f}'.format(epoch, loss, acc))
util.save_confusion(confusion)
# Test with best validation loss
path = util.models_dir.format(str(model_id).zfill(3), str(epoch).zfill(3))
util.load_checkpoint(model, optimizer, path)
loss, acc, confusion = loop_func(testset, model, train=False)
print('EPOCH TEST {:.4f} {:.4f}'.format(loss, acc))
util.save_confusion(confusion)
def main():
trainset, valset, testset = rocklin_dataset(
32) # batch size = 1800 residues
optimizer = tf.keras.optimizers.Adam()
model = make_model()
model_id = int(datetime.timestamp(datetime.now()))
NUM_EPOCHS = 50
loop_func = loop
best_epoch, best_val = 0, np.inf
for epoch in range(NUM_EPOCHS):
loss = loop_func(trainset, model, train=True, optimizer=optimizer)
print('EPOCH {} training loss: {}'.format(epoch, loss))
save_checkpoint(model, optimizer, model_id, epoch)
print('EPOCH {} TRAIN {:.4f}'.format(epoch, loss))
#util.save_confusion(confusion)
loss = loop_func(valset, model, train=False, val=False)
print(' EPOCH {} validation loss: {}'.format(epoch, loss))
if loss < best_val:
#Could play with this parameter here. Instead of saving best NN based on loss
#we could save it based on precision/auc/recall/etc.
best_epoch, best_val = epoch, loss
print('EPOCH {} VAL {:.4f}'.format(epoch, loss))
#util.save_confusion(confusion)
# Test with best validation loss
path = models_dir.format(str(model_id).zfill(3), str(epoch).zfill(3))
load_checkpoint(model, optimizer, path)
loss, tp, fp, tn, fn, acc, prec, recall, auc, y_pred, y_true = loop_func(
testset, model, train=False, val=True)
print('EPOCH TEST {:.4f} {:.4f}'.format(loss, acc))
#util.save_confusion(confusion)
return loss, tp, fp, tn, fn, acc, prec, recall, auc, y_pred, y_true
def test(self):
load_checkpoint(self.netG_A2B,
'./results/cyclegan/checkpoints/netG_A2B_final.pth',
self.device)
load_checkpoint(self.netG_B2A,
'./results/cyclegan/checkpoints/netG_B2A_final.pth',
self.device)
os.makedirs('./results/cyclegan/evaluation/', exist_ok=True)
self.netG_A2B.eval()
self.netG_B2A.eval()
print("=====Test Start======")
with torch.no_grad():
for iter, (real_A, real_B) in enumerate(self.testloader):
fake_A = self.netG_B2A(real_B.to(self.device))
save_image(
denorm(fake_A),
os.path.join('./results/cyclegan/evaluation',
'fake_image-{:05d}.png'.format(iter + 1)))
# Compute the Inception score
dataset = FolderDataset(
folder=os.path.join('./results/cyclegan/evaluation'))
Inception = Inception_Score(dataset)
score = Inception.compute_score(splits=1)
print('Inception Score : ', score)
def test(self):
load_checkpoint(
self.netG,
os.path.join('./results', self.type, 'checkpoints',
'netG_final.pth'), self.device)
self.netG.eval()
os.makedirs(os.path.join('./results/', self.type, 'evaluation'),
exist_ok=True)
print("=====Test Start======")
with torch.no_grad():
for iter in range(1000):
z = torch.randn(1, 256).to(self.device)
fake_img = self.netG(z)
save_image(
denorm(fake_img),
os.path.join('./results/', self.type, 'evaluation',
'fake_image-{:05d}.png'.format(iter + 1)))
# Compute the Inception score
dataset = FolderDataset(
folder=os.path.join('./results/', self.type, 'evaluation'))
Inception = Inception_Score(dataset)
score = Inception.compute_score(splits=1)
print('Inception Score : ', score)
def main():
print('this is alpha {}'.format(args.alpha))
os.makedirs(f'{args.save_dir}', exist_ok=True)
os.makedirs(f'{args.save_dir}/{args.out_oldweight_folder}', exist_ok=True)
os.makedirs(f'{args.save_dir}/{args.out_pruned_folder}', exist_ok=True)
os.makedirs(f'{args.save_dir}/{args.out_pruned_re_folder}', exist_ok=True)
os.makedirs(f'{args.save_dir}/{args.out_quantized_folder}', exist_ok=True)
os.makedirs(f'{args.save_dir}/{args.out_quantized_re_folder}',
exist_ok=True)
util.log(f"{args.save_dir}/{args.log}",
"--------------------------configure----------------------")
util.log(f"{args.save_dir}/{args.log}", f"{args}\n")
if args.train_mode == 1:
# Define model
model = vgg_mask.VGG16('VGG16', args.partition, mask_flag=True).cuda()
model = initial_process(model)
elif args.train_mode == 2:
if os.path.isfile(f"{args.load_model}"):
print("-------load " + f"{args.load_model} ----")
model = vgg_mask.VGG16('VGG16', args.partition,
mask_flag=True).cuda()
model = util.load_checkpoint(model, f"{args.load_model}", args)
model = pruning_process(model)
else:
print("---not found " + f"{args.load_model} ----")
elif args.train_mode == 3:
if os.path.isfile(f"{args.load_model}"):
print("-------load " + f"{args.load_model} ----")
model = vgg_mask.VGG16('VGG16', args.partition,
mask_flag=True).cuda()
model = util.load_checkpoint(model, f"{args.load_model}", args)
model = quantize_process(model)
else:
print("---not found " + f"{args.load_model} ----")
elif args.train_mode == 4: # initial train/ prune cnn/ qauntize
model = vgg_mask.VGG16('VGG16', args.partition, mask_flag=True).cuda()
model = initial_process(model)
model = pruning_process(model)
quantize_process(model)
elif args.train_mode == 5: # initial train/ qauntize
model = vgg_mask.VGG16('VGG16', args.partition, mask_flag=True).cuda()
st = time.time()
model = initial_process(model)
print('time init:', time.time() - st)
model = quantize_process(model)
print('time init+qauntize:', time.time() - st)
elif args.train_mode == 6: #load base model, prune and quantization
if os.path.isfile(f"{args.load_model}"):
print("-------load " + f"{args.load_model} ----")
model = vgg_mask.VGG16('VGG16', args.partition,
mask_flag=True).cuda()
model = util.load_checkpoint(model, f"{args.load_model}", args)
model = pruning_process(model)
model = quantize_process(model)
else:
print("---not found " + f"{args.load_model} ----")
def main():
args = parse_args('test')
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
transform_test = transforms.Compose([
transforms.Resize(224),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
test_set = DataLoader(args.test_list, transform=transform_test)
test_loader = torch.utils.data.DataLoader(test_set,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers)
# Model
print("==> creating model")
num_classes = args.num_classes
model = create_model(num_classes, args).to(device)
criterion = nn.CrossEntropyLoss()
# testing
expname = args.expname.split('/')[-1]
print('\n' + expname + ': TESTING!')
train_set = os.path.basename(args.train_list).split('.')[0]
val_set = os.path.basename(args.val_list).split('.')[0]
test_set = os.path.basename(args.test_list).split('.')[0]
if args.average:
record = pd.read_csv(os.path.join(args.output_csv_dir, 'output.csv'), index_col=0)
sorted_r = record.sort_values('f1', ascending=False)
model_list = list(sorted_r['epoch_num'].astype(int))[:5]
df = pd.DataFrame(columns=['exp', 'train', 'val', 'test', 'test_loss', 'test_acc', 'f1'])
for idx, epoch in enumerate(model_list):
model = load_checkpoint(model, os.path.join(args.save_model_dir, f'epoch{epoch}_checkpoint.pth.tar'))
test_loss, test_acc, f1, _ = test(test_loader, model, criterion, device, args, epoch=idx+1)
df.loc[len(df)] = [expname, train_set, val_set, test_set, test_loss, test_acc, f1]
output_csv_file = os.path.join(args.output_csv_dir, 'test_acc.csv')
df.to_csv(output_csv_file, index=False)
else:
df = pd.DataFrame(columns=['exp', 'train', 'val', 'test', 'test_loss', 'test_acc', 'f1'])
model = load_checkpoint(model, os.path.join(save_model_dir, 'model_best_f1.pth.tar'))
test_loss, test_acc, f1, _ = test(test_loader, model, criterion, device, args)
df.loc[len(df)] = [expname, train_set, val_set, test_set, test_loss, test_acc, f1]
output_csv_file = os.path.join(output_csv_dir, 'test_f1.csv')
df.to_csv(output_csv_file, index=False)
def initialize(self, ctx):
self.manifest = ctx.manifest
# set config, path
properties = ctx.system_properties
model_dir = properties.get('model_dir')
logger.info("model_dir: %s", model_dir)
opt = Opt(model_dir)
with open(opt.config, 'r', encoding='utf-8') as f:
self.config = json.load(f)
logger.info("%s", self.config)
if opt.num_threads > 0: torch.set_num_threads(opt.num_threads)
self.config['opt'] = opt
logger.info("opt.device: %s", opt.device)
logger.info("opt.num_threads: %s", opt.num_threads)
logger.info("opt.enable_dqm: %s", opt.enable_dqm)
# load pytorch model checkpoint
checkpoint = load_checkpoint(opt.model_path, device=opt.device)
# prepare model and load parameters
self.model = self.load_model(checkpoint)
self.model.eval()
# enable to use dynamic quantized model (pytorch>=1.3.0)
if opt.enable_dqm and opt.device == 'cpu':
self.model = torch.quantization.quantize_dynamic(self.model,
{torch.nn.Linear},
dtype=torch.qint8)
print(self.model)
# prepare tokenizer
self.tokenizer = self.prepare_tokenizer()
self.initialized = True
def main():
args = parse_args('predict')
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
transform = transforms.Compose([
transforms.Resize(224),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
test_set = DataLoader(args.ext_test, transform=transform)
test_loader = torch.utils.data.DataLoader(test_set, batch_size=args.batch_size, shuffle=False)
print("==> creating model")
num_classes = args.num_classes
model = create_model(num_classes, args).to(device)
if args.average:
record = pd.read_csv(os.path.join(args.output_csv_dir, 'output.csv'), index_col=0)
sorted_r = record.sort_values('f1', ascending=False)
model_list = list(sorted_r['epoch_num'].astype(int))[:5]
for idx, epoch in enumerate(model_list):
print('\nPredict with ' f'epoch{epoch} checkpoint')
model = load_checkpoint(model, os.path.join(args.save_model_dir, f'epoch{epoch}_checkpoint.pth.tar')).to(device)
predict(test_loader, model, device, args, idx)
else:
predict(test_loader, model, device, args)
predict(test_loader, model, device, args)
def main(rootpath=ROOT_PATH):
args = parse_args()
if args.arch == 'resnet101':
model = models.resnet101(pretrained=False)
num_ftrs = model.fc.in_features
model.fc = nn.Linear(num_ftrs, len(class_names))
elif args.arch == 'resnet152':
model = models.resnet152(pretrained=False)
num_ftrs = model.fc.in_features
model.fc = nn.Linear(num_ftrs, len(class_names))
elif args.arch == 'denset121':
model = models.desnet121(pretrained=False)
num_ftrs = model.classifer.in_features
model.classifier = nn.Linear(num_ftrs, len(class_names))
elif args.arch == 'inception_v3':
model = models.inception_v3(pretrained=False)
num_ftrs = model.fc.in_features
model.fc = nn.Linear(num_ftrs, len(class_names))
checkpoint = load_checkpoint(resume)
start_epoch = checkpoint['epoch']
best_acc = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
#optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {}, val acc {})".format(resume, checkpoint['epoch'], best_acc))
if use_gpu:
model = model.cuda()
model.train(False)
test(model, res_dir)
def test_allkd(s_name, params):
teachers = ["resnet8", "resnet14", "resnet20", "resnet26",
"resnet32", "resnet44", "resnet56",
# "resnet34", "resnet50", "resnet101", "resnet152",
]
accs = {}
for t_name in teachers:
params_t = params.copy()
params_t["teacher_name"] = t_name
t_net, best_teacher, best_t_acc = setup_teacher(t_name, params_t)
t_net = util.load_checkpoint(t_net, best_teacher, params_t["device"])
t_net = freeze_teacher(t_net)
s_net = setup_student(s_name, params_t)
params_t["test_name"] = f"{s_name}_{t_name}"
params_t["results_dir"] = params_t["results_dir"].joinpath("allkd")
util.check_dir(params_t["results_dir"])
best_acc = test_kd(s_net, t_net, params_t)
accs[t_name] = (best_t_acc, best_acc)
best_acc = 0
best_t_acc = 0
for t_name, acc in accs.items():
if acc[0] > best_t_acc:
best_t_acc = acc[0]
if acc[1] > best_acc:
best_acc = acc[1]
print(f"Best results teacher {t_name}: {acc[0]}")
print(f"Best results for {s_name}: {acc[1]}")
return best_t_acc, best_acc
def load_gen_inf_mem(algorithm, num_particles, seed):
checkpoint_path = "checkpoints/checkpoint_{}_{}_{}.pt".format(
algorithm, num_particles, seed)
(
generative_model,
inference_network,
theta_losses,
phi_losses,
cluster_cov_distances,
test_log_ps,
test_log_ps_true,
test_kl_qps,
test_kl_pqs,
test_kl_qps_true,
test_kl_pqs_true,
train_log_ps,
train_log_ps_true,
train_kl_qps,
train_kl_pqs,
train_kl_qps_true,
train_kl_pqs_true,
train_kl_memory_ps,
train_kl_memory_ps_true,
memory,
_,
_,
) = util.load_checkpoint(checkpoint_path, torch.device("cpu"))
return generative_model, inference_network, memory
def main():
"""
Load data and train a model on it.
"""
args = argument_parser().parse_args()
print(args)
random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.set_device(0)
train_set, val_set, test_set = read_dataset(DATA_DIR)
if not args.pretrained:
print('Training...')
train(train_set, test_set, args.checkpoint, **train_kwargs(args))
else:
print('Restoring from checkpoint...')
model_state, op_state, meta_iteration, cur_meta_step_size, accuracy_tracking = load_checkpoint(
args.checkpoint)
train(train_set, test_set, args.checkpoint, model_state, op_state,
**train_kwargs(args))
print('\nEvaluating...')
model_state, op_state, meta_iteration, cur_meta_step_size, accuracy_tracking = load_checkpoint(
args.checkpoint)
do_evaluation(model_state, op_state, args.checkpoint, val_set, test_set,
train_set)
def setup_valid(model, batcher, config):
valid_directory = config['valid_folder']
if not os.path.exists(valid_directory):
os.makedirs(valid_directory)
saver = tf.train.Saver(max_to_keep=3)
sess = tf.Session(config=util.get_config())
#sess = tf.Session()
_ = util.load_checkpoint(
saver=saver,
sess=sess,
log_directory=config['log_folder'],
checkpoint_directory='train') ### valid or train ???
best_valid_loss = None
while True:
valid_loss = 0
for i in range(config['valid_iter']):
batch = batcher.next_batch()
to_return = model.model_valid(batch, sess)
valid_loss += to_return['loss']
### if blabla
if best_valid_loss == None or valid_loss < best_valid_loss:
best_valid_loss = valid_loss
saver.save(sess,
config['bestmodel_save_path'],
latest_filename='checkpoint_best')
print('save checkpoint')
else:
print('not update checkpoint')
def run_benchmarks(modes, params, s_name, t_name):
results = {}
# if we test allkd we do not need to train an individual teacher
if "allkd" in modes:
best_t_acc, results["allkd"] = test_allkd(s_name, params)
modes.remove("allkd")
else:
t_net, best_teacher, best_t_acc = setup_teacher(t_name, params)
for mode in modes:
mode = mode.lower()
params_s = params.copy()
# reset the teacher
t_net = util.load_checkpoint(t_net, best_teacher, params["device"])
# load the student and create a results directory for the mode
s_net = setup_student(s_name, params)
params_s["test_name"] = s_name
params_s["results_dir"] = params_s["results_dir"].joinpath(mode)
util.check_dir(params_s["results_dir"])
# start the test
try:
run_test = globals()[f"test_{mode}"]
results[mode] = run_test(s_net, t_net, params_s)
except KeyError:
raise RuntimeError(f"Training mode {mode} not supported!")
# Dump the overall results
print(f"Best results teacher {t_name}: {best_t_acc}")
for name, acc in results.items():
print(f"Best results for {s_name} with {name} method: {acc}")
def get_model(word_to_idx, label_to_idx, resume=False, use_glove=True):
"""Resume a saved model or build a new model"""
best_acc = 0 # best test accuracy
start_epoch = 0 # start from epoch 0 or last checkpoint epoch
if resume:
# load checkpoint
checkpoint = load_checkpoint()
model = checkpoint['model']
best_acc = checkpoint['acc']
start_epoch = checkpoint['epoch']
else:
print('==> Building model {}...'.format(cfg.RUN_MODE))
if cfg.RUN_MODE in ["RNN", "LSTM", "GRU"]:
model = BatchRNN(cfg.EMBEDDING_DIM,
cfg.HIDDEN_DIM,
cfg.BATCH_SIZE,
len(word_to_idx),
len(label_to_idx),
rnn_model=cfg.RUN_MODE)
else:
model = CNN_NLP(cfg.EMBEDDING_DIM, cfg.HIDDEN_DIM, cfg.BATCH_SIZE,
len(word_to_idx), len(label_to_idx))
if use_glove:
# model.load_glove_model('GloVe-1.2/vectors.txt', word_to_idx)
model.load_glove_model(cfg.GLOVE_FILE,
word_to_idx,
regenerate=True)
return model, best_acc, start_epoch
def load(algorithm, num_particles):
things = [[] for _ in range(19)]
for seed in range(1, 11):
checkpoint_path = 'checkpoints/checkpoint_{}_{}_{}.pt'.format(
algorithm, seed, num_particles)
checkpoint = util.load_checkpoint(checkpoint_path, torch.device('cpu'))
for i, x in enumerate(checkpoint[2:-3] + checkpoint[-2:]):
if i >= 17 and (algorithm == 'rws' or algorithm == 'vimco'):
things[i].append(moving_average(x, 10))
else:
things[i].append(x)
# cut to min length
for i in range(len(things)):
if i < 2:
things[i] = [moving_average(x) for x in things[i]]
if things[i][0] is not None:
min_length = min(map(len, things[i]))
things[i] = [x[:min_length] for x in things[i]]
mid, lower, upper = [], [], []
for i, thing in enumerate(things):
if thing[0] is None:
mid.append(None)
lower.append(None)
upper.append(None)
else:
mid.append(np.quantile(np.array(thing), 0.5, axis=0))
lower.append(np.quantile(np.array(thing), 0.25, axis=0))
upper.append(np.quantile(np.array(thing), 0.75, axis=0))
# mid = [np.quantile(np.array(thing), 0.5, axis=0) for thing in things]
# lower = [np.quantile(np.array(thing), 0.25, axis=0) for thing in things]
# upper = [np.quantile(np.array(thing), 0.75, axis=0) for thing in things]
return mid, lower, upper
def setup_teacher(t_name, params):
# Teacher Model
num_classes = params["num_classes"]
t_net = create_model(t_name, num_classes, params["device"])
teacher_config = params.copy()
teacher_config["test_name"] = t_name + "_teacher"
if params["t_checkpoint"]:
# Just validate the performance
print("---------- Loading Teacher -------")
best_teacher = params["t_checkpoint"]
else:
# Teacher training
print("---------- Training Teacher -------")
teacher_trainer = BaseTrainer(t_net, config=teacher_config)
teacher_trainer.train()
best_teacher = teacher_trainer.best_model_file
# reload and get the best model
t_net = util.load_checkpoint(t_net, best_teacher)
teacher_trainer = BaseTrainer(t_net, config=teacher_config)
best_t_acc = teacher_trainer.validate()
# also save this information in a csv file for plotting
name = teacher_config["test_name"] + "_val"
acc_file_name = params["results_dir"].joinpath(f"{name}.csv")
with acc_file_name.open("w+") as acc_file:
acc_file.write("Training Loss,Validation Loss\n")
for _ in range(params["epochs"]):
acc_file.write(f"0.0,{best_t_acc}\n")
return t_net, best_teacher, best_t_acc
def build_gen_dis(config):
net_gen = generator(config.DIM_NOISE, config.DIM_IMG,
config.N_CHANNEL).to(config.DEVICE)
net_dis = discriminator(config.DIM_IMG, config.N_CHANNEL).to(config.DEVICE)
if config.INIT:
net_gen.apply(init_weight)
net_dis.apply(init_weight)
loss = None
print('initialize model successed')
else:
ext = config.PATH_MODEL[-4:]
path_model = config.PATH_IMPORT_MODEL[:-4] + '_epoch_%d' % config.IMPORT_IDX_EPOCH
print(path_model)
path_model = path_model + ext
print(path_model)
if config.DEVICE == torch.device("cpu"):
device = 'cpu'
else:
device = 'cuda:0'
net_gen, net_dis, loss = util.load_checkpoint(config.EPOCHS, net_gen,
net_dis, path_model,
device)
print("load model successed.")
return net_gen, net_dis, loss
def decode(self):
"""Decode examples until data is exhausted (if FLAGS.single_pass) and return, or decode indefinitely, loading latest checkpoint at regular intervals"""
t0 = time.time()
counter = 0
step = 0
while True:
step += 1
batch = self._batcher.next_batch() # 1 example repeated across batch
if batch is None: # finished decoding dataset in single_pass mode
assert self._model.single_pass, "Dataset exhausted, but we are not in single_pass mode"
tf.logging.info("Decoder has finished reading dataset for single_pass.")
tf.logging.info("Output has been saved in %s and %s. Now starting ROUGE eval...", self._rouge_ref_dir, self._rouge_dec_dir)
results_dict = rouge_eval(self._rouge_ref_dir, self._rouge_dec_dir)
rouge_log(results_dict, self._decode_dir)
return
t_start_decode = time.time()
original_article = batch.original_articles[0] # string
original_abstract = batch.original_abstracts[0] # string
original_abstract_sents = batch.original_abstracts_sents[0] # list of strings
article_withunks = data.show_art_oovs(original_article, self._vocab) # string
abstract_withunks = data.show_abs_oovs(original_abstract, self._vocab, (batch.art_oovs[0] if self._model.pointer_gen else None)) # string
# Run beam search to get best Hypothesis
best_hyp = self._model.model_decode(batch, self._sess, self._vocab)
print(best_hyp.tokens)
#best_hyp = None
# Extract the output ids from the hypothesis and convert back to words
output_ids = [int(t) for t in best_hyp.tokens[1:]]
decoded_words = data.outputids2words(output_ids, self._vocab, (batch.art_oovs[0] if self._model.pointer_gen else None))
# Remove the [STOP] token from decoded_words, if necessary
try:
fst_stop_idx = decoded_words.index(data.STOP_DECODING) # index of the (first) [STOP] symbol
decoded_words = decoded_words[:fst_stop_idx]
except ValueError:
decoded_words = decoded_words
decoded_output = ' '.join(decoded_words) # single string
if self._model.single_pass:
self.write_for_rouge(original_abstract_sents, decoded_words, counter) # write ref summary and decoded summary to file, to eval with pyrouge later
counter += 1 # this is how many examples we've decoded
else:
print_results(article_withunks, abstract_withunks, decoded_output) # log output to screen
self.write_for_attnvis(article_withunks, abstract_withunks, decoded_words, best_hyp.attn_dists, best_hyp.p_gens) # write info to .json file for visualization tool
# Check if SECS_UNTIL_NEW_CKPT has elapsed; if so return so we can load a new checkpoint
t1 = time.time()
if t1-t0 > SECS_UNTIL_NEW_CKPT:
print('We\'ve been decoding with same checkpoint for %i seconds. Time to load new checkpoint', t1-t0)
_ = util.load_checkpoint(self._saver, self._sess)
t0 = time.time()
t_end_decode = time.time()
print('decode {}-th batch requires {} seconds'.format(step, int(t_end_decode - t_start_decode)))
def main():
parser = argparse.ArgumentParser(description='Predict a file.')
parser.add_argument('input_img',
type=str,
help='Path of the input image',
default="./flowers/test/2/image_05100.jpg")
parser.add_argument('checkpoint',
type=str,
help='Path of the checkpoint',
default="./checkpoint.pth")
parser.add_argument('--top_k', help='Top k', type=int, default=5)
parser.add_argument('--gpu',
action='store_true',
help='Use GPU for inference if GPU is available')
parser.add_argument('--category_names',
action="store",
default='cat_to_name.json')
args, _ = parser.parse_known_args()
input_img = args.input_img
checkpoint = args.checkpoint
category_names = 'cat_to_name.json'
if args.category_names:
category_names = args.category_names
top_k = 5
if args.top_k:
top_k = args.top_k
cuda = False
if args.gpu:
if torch.cuda.is_available():
cuda = True
else:
print("GPU flag was set but no GPU is available in this machine.")
loaded_model = util.load_checkpoint(checkpoint, cuda)
with open('cat_to_name.json', 'r') as json_file:
cat_to_name = json.load(json_file)
probabilities, classes = util.predict(input_img, loaded_model, top_k)
labels = [cat_to_name[str(int(index) + 1)] for index in classes]
probability = np.array(probabilities)
i = 0
while i < top_k:
print("{} with a probability of {:.2f}%".format(
labels[i], probability[i] * 100))
i += 1
print("Prediction Done.")
def test_multikd(s_net, t_net1, params):
t_net1 = freeze_teacher(t_net1)
print("---------- Training MULTIKD -------")
kd_config = params.copy()
params["t2_name"] = "WRN22_4"
t_net2 = create_model(params["t2_name"], params["num_classes"],
params["device"])
t_net2 = util.load_checkpoint(t_net2, "pretrained/WRN22_4_cifar10.pth")
t_net2 = freeze_teacher(t_net2)
params["t3_name"] = "resnet18"
t_net3 = create_model(params["t3_name"], params["num_classes"],
params["device"])
t_net3 = util.load_checkpoint(t_net3, "pretrained/resnet18_cifar10.pth")
t_net3 = freeze_teacher(t_net3)
t_nets = [t_net1, t_net2]
kd_trainer = MultiTrainer(s_net, t_nets=t_nets, config=kd_config)
best_acc = kd_trainer.train()
return best_acc
def get_model(args):
if args.model_mode == 'c':
if args.use_model == 'alex':
args.prune_rates = [0.16, 0.62, 0.65, 0.63, 0.63]
model = AlexNet(100, mask_flag=True).to(args.device)
elif args.use_model == 'vgg16':
# args.prune_rates = [0.42, 0.78, 0.66, 0.64, 0.47, 0.76, 0.58, 0.68, 0.73, 0.66, 0.65, 0.71, 0.64] # deepC
args.prune_rates = [0.50, 0.50, 0.00, 0.00, 0.00, 0.00, 0.00, 0.50, 0.75, 0.75, 0.75, 0.75, 0.75] # PFEC
model = vgg16().to(args.device)
else:
raise Exception
else:
model = AlexNetMask(args.partition, 100, mask_flag=True).to(args.device)
if os.path.isfile(f"{args.load_model}"):
model, args.best_top1_acc = util.load_checkpoint(model, f"{args.load_model}", args)
print("-------load " + f"{args.load_model} ({args.best_top1_acc:.3f})----")
return model
def run_takd_distillation(s_net, ta_net, t_net, **params):
# Teaching Assistant training
print("---------- Training TA -------")
ta_config = params.copy()
ta_name = ta_config["ta_name"]
ta_config["test_name"] = f"{ta_name}_ta_trainer"
ta_trainer = KDTrainer(ta_net, t_net=t_net, config=ta_config)
best_ta_acc = ta_trainer.train()
best_ta = ta_trainer.best_model_file
ta_net = util.load_checkpoint(ta_net, best_ta)
# Student training
print("---------- Training TA Student -------")
s_trainer = KDTrainer(s_net, t_net=ta_net, config=params)
best_s_acc = s_trainer.train()
print(f"Best results teacher {ta_name}: {best_ta_acc}")
return best_s_acc
def get_error_rate_with_two_classes(training=False):
"""
Compute the overall error rate of the trained model.
Combine positive and negative as one class, compared to neutral.
If training is False, use test_data, otherwise training_data.
"""
model = load_checkpoint()['model']
data, word_to_idx, _ = data_loader(training)
targets = torch.LongTensor()
predicts = torch.LongTensor()
for sentences, _targets, seq_lengths in get_batch_data(data,
cfg.BATCH_SIZE,
word_to_idx,
shuffle=True):
_predicts = evaluate_batch(model, sentences, seq_lengths)
targets = torch.cat((targets, _targets), 0)
predicts = torch.cat((predicts, _predicts), 0)
# combine positive and negative as one class
targets[targets == 2] = 1
predicts[predicts == 2] = 1
error_rate = (targets != predicts).sum() / targets.size(0)
print(Counter(targets.numpy()), Counter(predicts.numpy()))
print('error rate: ', error_rate)
idx2label = {i: v for v, i in label_to_idx.items()}
labels = [idx2label[idx] for idx in sorted(idx2label.keys())]
print(
'Report:\n',
metrics.classification_report(targets.numpy(),
predicts.numpy(),
target_names=labels))
print('Confusion matrix: \n',
metrics.confusion_matrix(targets.numpy(), predicts.numpy()))
return error_rate
def main(args):
# general
if torch.cuda.is_available():
device = torch.device("cuda")
util.logging.info("Using CUDA")
else:
device = torch.device("cpu")
util.logging.info("Using CPU")
util.set_seed(args.seed)
# init
checkpoint_path = util.get_checkpoint_path(args)
if not Path(checkpoint_path).exists():
util.logging.info("Training from scratch")
model, optimizer, stats = util.init(args, device)
else:
model, optimizer, stats, args = util.load_checkpoint(checkpoint_path, device)
# train
train.train(model, optimizer, stats, run_args=args)
def main():
model = util.load_checkpoint(pa.load_dir)
with open(pa.category_name, 'r') as f:
cat_to_name = json.load(f)
# print(cat_to_name.get)
# print(len(cat_to_name))
top_p_list, top_flowers = util.predict(image_path, pa.load_dir,
cat_to_name, pa.top_k)
# labels = [cat_to_name[str(index+1)] for index in np.naray(top_p_list[1][0])
# return print(top_p_list[0])
i = 0
while i < pa.top_k:
print('{} with a probability of {}'.format(top_flowers[i],
top_p_list[i]))
i += 1
def get_error_rate(model=None, training=False, report=False):
"""
Compute the overall error rate of the trained model.
If training is False, use test_data, otherwise training_data.
If report is True, print precision, recall, F1-score, and confusion matrix.
"""
model = model or load_checkpoint()['model']
data, word_to_idx, label_to_idx = data_loader(training=training)
targets = torch.LongTensor()
predicts = torch.LongTensor()
for sentences, _targets, seq_lengths in get_batch_data(data,
cfg.BATCH_SIZE,
word_to_idx,
shuffle=True):
_predicts = evaluate_batch(model, sentences, seq_lengths)
targets = torch.cat((targets, _targets), 0)
predicts = torch.cat((predicts, _predicts), 0)
error_rate = (targets != predicts).sum() / targets.size(0)
if report:
print('targets:', Counter(targets.numpy()), 'predicts:',
Counter(predicts.numpy()))
print('error rate: ', error_rate)
idx2label = {i: v for v, i in label_to_idx.items()}
labels = [idx2label[idx] for idx in sorted(idx2label.keys())]
print(
'Report:\n',
metrics.classification_report(targets.numpy(),
predicts.numpy(),
target_names=labels))
print('Confusion matrix: \n',
metrics.confusion_matrix(targets.numpy(), predicts.numpy()))
return error_rate
def __init__(self, model, batcher, vocab):
"""
Args:
model: a Seq2SeqAttentionModel object.
batcher: a Batcher object.
vocab: Vocabulary object
"""
self._model = model
self._batcher = batcher
self._vocab = vocab
self._saver = tf.train.Saver() # load checkpoints for decoding
self._sess = tf.Session(config=util.get_config())
# Load an initial checkpoint to use for decoding
ckpt_path = util.load_checkpoint(self._saver, self._sess, checkpoint_directory = 'train')
if self._model.single_pass:
# Make a descriptive decode directory name
#ckpt_name = "ckpt-" + ckpt_path.split('-')[-1] # this is something of the form "ckpt-123456"
#self._decode_dir = os.path.join(self.model.log_folder, get_decode_dir_name(ckpt_name))
self._decode_dir = self._model.decode_folder
if os.path.exists(self._decode_dir):
print("single_pass decode directory should not already exist")
#raise Exception("single_pass decode directory %s should not already exist" % self._decode_dir)
else: # Generic decode dir name
self._decode_dir = self._model.decode_folder
# Make the decode dir if necessary
if not os.path.exists(self._decode_dir): os.mkdir(self._decode_dir)
if self._model.single_pass:
# Make the dirs to contain output written in the correct format for pyrouge
self._rouge_ref_dir = os.path.join(self._decode_dir, "reference")
if not os.path.exists(self._rouge_ref_dir): os.mkdir(self._rouge_ref_dir)
self._rouge_dec_dir = os.path.join(self._decode_dir, "decoded")
if not os.path.exists(self._rouge_dec_dir): os.mkdir(self._rouge_dec_dir)
def main():
print("==> loading model")
num_classes = args.num_classes
model = create_model(num_classes, args).to(device)
model = load_checkpoint(model, checkpoint).to(device)
transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
input_data = args.input
if os.path.isdir(input_data):
input_folder_lst = [
os.path.join(input_data, i) for i in os.listdir(input_data)
]
for input_folder in input_folder_lst:
print(os.path.basename(input_folder))
input_data = os.path.join(input_folder, 'patch.json')
filter_patches(model, input_data, transform)
else:
filter_patches(model, input_data, transform)
def main(args):
if torch.cuda.is_available():
device = torch.device("cuda")
util.logging.info("Using CUDA")
else:
device = torch.device("cpu")
util.logging.info("Using CPU")
if args.checkpoint_path is None:
checkpoint_paths = list(util.get_checkpoint_paths())
else:
checkpoint_paths = [args.checkpoint_path]
for checkpoint_path in checkpoint_paths:
try:
model, optimizer, stats, run_args = util.load_checkpoint(
checkpoint_path, device=device)
except FileNotFoundError as e:
print(e)
if "No such file or directory" in str(e):
print(e)
continue
plot_losses(f"{util.get_save_dir(run_args)}/losses.pdf", stats.losses)