def __init__(self,
layers=[],
n_batches=10,
loss='categorical_crossentropy',
metric='accuracy_score',
optimizer='adam',
optimizer_params={},
save_weights=True,
shuffle=True,
random_seed=None):
self.layers = layers
self.n_batches = n_batches # mini-batches will be generated in the stratified manner
self.loss = loss
self.metric = metric
self.optimizer = optimizer
self.optimizer_params = optimizer_params
self.save_weights = save_weights
self.template = 'tmp/nn_{0}_{1:.4f}.json'
self.shuffle = shuffle
self.random_seed = random_seed
self.best_layers_ = None
self.best_epoch_ = None
self.best_val_score_ = 0.
self.n_layers_ = len(self.layers)
self._loss = get_metric(self.loss)
if self.loss == 'categorical_crossentropy':
self._loss_grad = lambda actual, predicted: -(actual - predicted)
self._metric = get_metric(self.metric)
self._optimizer = get_optimizer(self.optimizer,
**self.optimizer_params)
self._tts = TrainTestSplitter(shuffle=self.shuffle,
random_seed=self.random_seed)
self._initialized = False
self._training = False
super(NNClassifier, self).__init__(
_y_required=True) # TODO: split into multiple NNs later
def __init__(self,
n_neighbors=3,
metric='euclidean',
parallel=False,
chunksize=5,
n_jobs=-1):
self.n_neighbors = n_neighbors
self.parallel = parallel
self.n_jobs = n_jobs if n_jobs != -1 else os.cpu_count()
self.chunksize = chunksize
self.metric = metrics.get_metric(metric)
if not self.metric:
raise ValueError(f'Unkown metric: {self.metric}')
def evaluate(self, X, y_true, metric='accuracy_score'):
y_pred = self.predict(X)
if len(y_pred.shape) == 2 and y_pred.shape[1] > 1 and (len(
y_true.shape) == 1 or y_true.shape[1] == 1):
y_true = one_hot(y_true)
return get_metric(metric)(y_true, y_pred)
def set_metric(self) -> None:
"""Set metric"""
self.metric = get_metric(self.cfg)
dsize=(padded_size, padded_size),
interpolation=cv2.INTER_NEAREST)
resized_mask_pred = resized_mask_pred[position[0]:position[1],
position[2]:position[3]]
else:
resized_mask_pred = cv2.resize(resized_mask_pred,
dsize=(w, h),
interpolation=cv2.INTER_NEAREST)
resized_mask_pred = cv2.erode(resized_mask_pred, kernel, iterations=1)
if min_length * skip_resize_ratio < max_length or max_length < skip_max_length:
# allocate new memory
step2_gray_img = np.copy(step2_out_img)
step2_gray_img[step2_gray_img > 0] = 1
step2_fmeasure, step2_pfmeasure, step2_psnr, step2_drd = get_metric(
step2_gray_img, gt_mask)
cv2.imwrite(
'%s/step2_normal/%s/%s.png' %
(save_root_dir, dibco_year, img_name), step2_out_img)
save_fmeasure['step2_normal'][dibco_year][0].append(step2_fmeasure)
save_fmeasure['step2_normal'][dibco_year][1].append(step2_pfmeasure)
save_fmeasure['step2_normal'][dibco_year][2].append(step2_psnr)
save_fmeasure['step2_normal'][dibco_year][3].append(step2_drd)
csv_tmp = [
img_name, step2_fmeasure, step2_pfmeasure, step2_psnr, step2_drd
]
else:
step2_normal_or_img = np.bitwise_or(resized_mask_pred, step2_out_img)
step2_normal_or_img_metric = np.copy(step2_normal_or_img)
step2_normal_or_img_metric[step2_normal_or_img_metric > 0] = 1
def test(self):
torch.set_grad_enabled(False)
epoch = self.optimizer.get_last_epoch()
self.ckp.write_log('\nEvaluation:')
self.ckp.add_log(
torch.zeros(1, len(self.loader_test), len(self.scale))
)
# (self): my log
mlog = {'psnr':[], 'ssim':[], 'mse':[]}
self.model.eval()
timer_test = utility.timer()
if self.args.save_results: self.ckp.begin_background()
for idx_data, d in enumerate(self.loader_test):
for idx_scale, scale in enumerate(self.scale):
d.dataset.set_scale(idx_scale)
for lr, hr, filename in tqdm(d, ncols=80):
lr, hr = self.prepare(lr, hr)
sr = self.model(lr, idx_scale)
sr = utility.quantize(sr, self.args.rgb_range)
save_list = [sr]
self.ckp.log[-1, idx_data, idx_scale] += utility.calc_psnr(
sr, hr, scale, self.args.rgb_range, dataset=d
)
# (self): log metrics
mlog['psnr'].append(es.get_metric(hr.cpu().numpy()[0], sr.cpu().numpy()[0], 'psnr'))
mlog['ssim'].append(es.get_metric(hr.cpu().numpy()[0], sr.cpu().numpy()[0], 'ssim'))
mlog['mse'].append(es.get_metric(hr.cpu().numpy()[0], sr.cpu().numpy()[0], 'mse'))
if self.args.save_gt:
save_list.extend([lr, hr])
if self.args.save_results:
self.ckp.save_results(d, filename[0], save_list, scale)
self.ckp.log[-1, idx_data, idx_scale] /= len(d)
best = self.ckp.log.max(0)
self.ckp.write_log(
'[{} x{}]\tPSNR: {:.3f} (Best: {:.3f} @epoch {})'.format(
d.dataset.name,
scale,
self.ckp.log[-1, idx_data, idx_scale],
best[0][idx_data, idx_scale],
best[1][idx_data, idx_scale] + 1
)
)
# print metrics
print(f"max psnr: {np.max(mlog['psnr'])}, mean psnr: {np.mean(mlog['psnr'])}")
print(f"max ssim: {np.max(mlog['ssim'])}, mean ssim: {np.mean(mlog['ssim'])}")
print(f"max mse: {np.max(mlog['mse'])}, mean mse: {np.mean(mlog['mse'])}")
self.ckp.write_log('Forward: {:.2f}s\n'.format(timer_test.toc()))
self.ckp.write_log('Saving...')
if self.args.save_results:
self.ckp.end_background()
if not self.args.test_only:
self.ckp.save(self, epoch, is_best=(best[1][0, 0] + 1 == epoch))
self.ckp.write_log(
'Total: {:.2f}s\n'.format(timer_test.toc()), refresh=True
)
torch.set_grad_enabled(True)
axes[1, 1].plot(flux_in, throughput, color=color)
axes[1, 2].plot(flux_out, throughput, color=color)
axes[1, 0].plot(flux_in_mean, flux_out_mean, marker='o', color=color)
axes[1, 1].plot(flux_in_mean, throughput_mean, marker='o', color=color)
axes[1, 2].plot(flux_out_mean, throughput_mean, marker='o', color=color)
axes[0, 0].set_yscale('log')
axes[0, 1].set_yscale('log')
axes[0, 2].set_yscale('log')
axes[0, 3].set_yscale('log')
axes[1, 0].set_yscale('log')
axes[1, 0].set_xscale('log')
axes[1, 1].set_yscale('log')
axes[1, 1].set_xscale('log')
axes[1, 2].set_yscale('log')
axes[1, 2].set_xscale('log')
axes[0, 3].legend()
plt.show()
if __name__ == '__main__':
obs = ObservatoryMaster(params, iteration=0)
metric_name = 'pix_yield'
metric_config = metrics.get_metric(metric_name, master_cam=obs.cam)
metric_test = MetricTester(obs, metric_config)
debugging_noise(metric_test)
debugging_throughput(metric_test)
def main(args):
"""
main function
"""
model_config = json.load(open(args.model_config, 'r'))
if args.use_cuda:
paddle.set_device("gpu")
else:
paddle.set_device("cpu")
if args.is_distributed:
strategy = fleet.DistributedStrategy()
fleet.init(is_collective=args.use_cuda, strategy=strategy)
train_loader = create_dataloader(
data_dir=args.train_data,
model_config=model_config)
valid_loader = create_dataloader(
data_dir=args.valid_data,
model_config=model_config)
encoder_model = ProteinEncoderModel(model_config, name='protein')
model = ProteinModel(encoder_model, model_config)
if args.is_distributed:
model = fleet.distributed_model(model)
# Generate parameter names needed to perform weight decay.
# All bias and LayerNorm parameters are excluded.
decay_params = [
p.name for n, p in model.named_parameters()
if not any(nd in n for nd in ["bias", "norm"])
]
grad_clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=1.0)
optimizer = paddle.optimizer.AdamW(
learning_rate=1e-4,
epsilon=1e-06,
weight_decay=0.01,
parameters=model.parameters(),
apply_decay_param_fun=lambda x: x in decay_params)
if args.is_distributed:
optimizer = fleet.distributed_optimizer(optimizer)
criterion = ProteinCriterion(model_config)
metric = get_metric(model_config['task'])
if args.init_model:
print("load init_model")
# for hot_start
if args.hot_start == 'hot_start':
model.load_dict(paddle.load(args.init_model))
# for pre_train
else:
encoder_model.load_dict(paddle.load(args.init_model))
train_sum_loss = 0
valid_min_loss = 10000
steps_per_epoch = 20
cur_step = 0
while True:
model.train()
for (text, pos, label) in train_loader:
# print("text: ", text)
cur_step += 1
pred = model(text, pos)
label = label.reshape([-1, 1])
pred = pred.reshape([-1, pred.shape[-1]])
loss = criterion.cal_loss(pred, label)
print("loss: ", loss)
train_sum_loss += loss.numpy()
loss.backward()
optimizer.minimize(loss)
model.clear_gradients()
pred = pred.numpy()
label = label.numpy()
loss = loss.numpy()
metric.update(pred, label, loss)
if cur_step % 10 == 0:
print('step %d, avg loss %.5f' % (cur_step, train_sum_loss / 10))
metric.show()
train_sum_loss = 0
metric.clear()
# save best_model
if cur_step % steps_per_epoch == 0:
print("eval begin_time: ", time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
valid_cur_loss = eval(model, valid_loader, criterion, metric)
print("valid_cur_loss: ", valid_cur_loss)
print("eval end_time: ", time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
if valid_cur_loss < valid_min_loss:
print("%s Save best model step_%d." % \
(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()), cur_step))
paddle.save(encoder_model.state_dict(), 'models/epoch_best_encoder.pdparams')
paddle.save(model.state_dict(), 'models/epoch_best.pdparams')
valid_min_loss = valid_cur_loss
os.system("cp -rf models/epoch_best.pdparams models/step_%d.pdparams" % (cur_step))
os.system("cp -rf models/epoch_best_encoder.pdparams models/step_%d_encoder.pdparams" % (cur_step))
model.train()
