def search(self):
game = self.game
parent = self.root
while not parent.is_leaf_node():
best_puct = None
for child in parent.children:
curr_puct = self.PUCT(parent, child)
if (best_puct == None or curr_puct >= best_puct):
best_child = child
best_puct = curr_puct
self.level += 1
parent = best_child
self.game.execute_move(best_child.last_action)
raw_pred = self.agent.predict(np.array([game.get_board()]))
result = loss.softmax(np.array(game.get_legal_NN_output()),
raw_pred[0])
if not self.game.is_final():
valid_moves = game.get_moves()
for move in valid_moves:
Node(game, parent, move, result[0][move])
self.back_propagate(parent, raw_pred[1][0][0])
self.level = 0
else:
self.back_propagate(parent, raw_pred[1][0][0])
self.level = 0
def _evaluate(self, result, epsilon=0.000001):
# return 0, {str(act): 1 / len(self.game.get_moves()) for num, act in enumerate(self.game.get_moves())}
# return random.uniform(-1, 1), {str(act): random.random() for num, act in enumerate(self.game.get_moves())}
# state = state.reshape(self.NN_input_dim)
policy, value = result
policy = policy.flatten()
legal_moves = np.array(self.game.get_legal_NN_output())
num_legal_moves = np.sum(legal_moves)
policy_norm = loss.softmax(legal_moves, policy)
policy_norm = (policy_norm + epsilon) * legal_moves
outp = self.NN_output_to_moves_func(policy_norm)
policy_norm = policy_norm[policy_norm > 0]
if len(self.state_visits) == 0 and self.dirichlet_noise:
noise = np.random.dirichlet(np.array([self.alpha for _ in range(num_legal_moves)]), (1))
noise = noise.reshape(noise.shape[1])
# print("Adding noise", policy_norm)
return value, {str(act): (1 - self.epsilon) * policy_norm[num] + self.epsilon * noise[num] for num, act in
enumerate(outp)}
else:
# print("No noise", policy_norm)
return value, {str(act): policy_norm[num] for num, act in enumerate(outp)}
def perplexity(input_data):
from loss import softmax
N, T, V = input_data.shape
p = 1
for i in range(N):
p *= softmax(input_data[i])
return pow(1 / p, 1 / N) / N
def train(game,
config,
num_filters,
num_res_blocks,
num_sim=125,
epochs=50,
games_each_epoch=10,
batch_size=32,
num_train_epochs=10):
h, w, d = config.board_dims[1:]
agent = ResNet.ResNet.build(h,
w,
d,
num_filters,
config.policy_output_dim,
num_res_blocks=num_res_blocks)
agent.compile(
loss=[softmax_cross_entropy_with_logits, 'mean_squared_error'],
optimizer=SGD(lr=0.001, momentum=0.9))
agent.summary()
for epoch in range(epochs):
x, y_pol, y_val = generate_data(game,
agent,
config,
num_sim=num_sim,
games=games_each_epoch)
print("Epoch")
print(x.shape)
raw = agent.predict(x)
for num in range(len(x)):
print("targets-predictions")
print(y_pol[num], y_val[num])
print(softmax(y_pol[num], raw[0][num]), raw[1][num])
agent.fit(x=x,
y=[y_pol, y_val],
batch_size=min(batch_size, len(x)),
epochs=num_train_epochs,
callbacks=[])
agent.save_weights("Models/" + Config.name + "/" + str(epoch) + ".h5")
return agent
def get_prior_probabilities(self, board_state):
pred = self.agent.predict(board_state)
return loss.softmax(np.array(self.game.get_legal_NN_output()),
pred[0]), pred[1]
def main():
"""Create the model and start the evaluation process."""
args = get_arguments()
gpu0 = args.gpu
h, w = map(int, args.input_size.split(','))
input_size = (h, w)
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
save_part_dir = os.path.join(args.save_dir, 'part_map')
if not os.path.exists(save_part_dir):
os.makedirs(save_part_dir)
save_overlay_dir = os.path.join(args.save_dir, 'part_overlay')
if not os.path.exists(save_overlay_dir):
os.makedirs(save_overlay_dir)
save_part_dcrf_dir = os.path.join(args.save_dir, 'part_map_dcrf')
if not os.path.exists(save_part_dcrf_dir):
os.makedirs(save_part_dcrf_dir)
save_dcrf_overlay_dir = os.path.join(args.save_dir, 'part_dcrf_overlay')
if not os.path.exists(save_dcrf_overlay_dir):
os.makedirs(save_dcrf_overlay_dir)
save_lm_dir = os.path.join(args.save_dir, 'landmarks')
if not os.path.exists(save_lm_dir):
os.makedirs(save_lm_dir)
save_seg_dir = os.path.join(args.save_dir, 'seg')
if not os.path.exists(save_seg_dir):
os.makedirs(save_seg_dir)
save_prob_dir = os.path.join(args.save_dir, 'prob')
if not os.path.exists(save_prob_dir):
os.makedirs(save_prob_dir)
save_dcrf_prob_dir = os.path.join(args.save_dir, 'dcrf_prob')
if not os.path.exists(save_dcrf_prob_dir):
os.makedirs(save_dcrf_prob_dir)
# create network
model = model_generator(args)
#model.load_state_dict(torch.load("snapshots_CelebA/SCOPS_K8_retrain/model_100000.pth"))
model.load_state_dict(torch.load(args.restore_from))
model.eval()
model.cuda(gpu0)
if args.dataset == 'CelebAWild':
from dataset.celeba_wild_dataset import CelebAWildDataset
dataset = CelebAWildDataset
testloader = data.DataLoader(dataset(args.data_dir,
args.data_list,
crop_size=input_size,
scale=False,
mirror=False,
mean=IMG_MEAN,
center_crop=False,
ignore_saliency_fg=False,
iou_threshold=0.3),
batch_size=1,
shuffle=False,
pin_memory=True)
elif args.dataset == 'cub':
from dataset import cub
args.batch_size = 1
testloader = cub.data_loader(args)
elif args.dataset == 'ImageNet':
from dataset import imagenet as imnet_data
args.batch_size = 1
testloader = imnet_data.imnet_dataloader(args)
elif args.dataset == 'p3d':
from dataset import p3d as p3d_data
args.batch_size = 1
testloader = p3d_data.p3d_dataloader(args)
else:
print('Not supported dataset {}'.format(args.dataset))
interp = nn.Upsample(size=input_size, mode='bilinear', align_corners=True)
colorize = utils.Colorize(args.num_parts + 1)
N = len(testloader)
landmarks = np.zeros((N, args.num_parts, 2))
#landmarks_gt = np.zeros((N,args.lm_count,2))
with torch.no_grad():
for index, batch in enumerate(testloader):
"""
import pickle
with open("/xtli-correspondence/Data/trim/bboxes.npy", 'rb') as f:
bboxes = pickle.load(f)
bbox = np.squeeze(bboxes[0])
fpath = "/xtli-correspondence/Data/trim/output_00001.jpg"
fpath = "/xtli-correspondence/CUB_200_2011/images/001.Black_footed_Albatross/Black_Footed_Albatross_0039_796132.jpg"
bbox = np.array([122, 128, 481, 379])
img = read_frame(fpath, bbox)
batch = {}
batch['img'] = img
"""
if index % 100 == 0:
path_split = args.save_dir.split('/')
print('{} processd: {}/{}'.format(index, path_split[-4],
path_split[-3]))
image = batch['img']
#label = batch['mask']
img_path = batch['img_path']
tmp = img_path[0].split("/")
img_folder = tmp[4]
img_nm = tmp[5].replace(".jpg", ".png")
size = input_size
output = model(image.cuda(gpu0))
output_raw = interp(output[2])
#permute = [0, 1, 4, 3, 2]
#import pdb; pdb.set_trace()
#output_raw = output_raw[:, permute, :, :]
lms = Batch_Get_Centers(output_raw)
landmarks[index, :, :] = lms
if args.save_viz:
mean_tensor = torch.tensor(IMG_MEAN).float().expand(
int(size[1]), int(size[0]), 3).transpose(0, 2)
imgs_viz = torch.clamp(image + mean_tensor, 0.0, 255.0)
#landmark visualization
lms_viz = Batch_Draw_GT_Landmarks(imgs_viz, output_raw, lms)
output = softmax(output_raw)
# normalize part
output /= output.max(dim=3,
keepdim=True)[0].max(dim=2,
keepdim=True)[0]
output[:, 0, :, :] = 0.1
output = output.cpu().data[0].numpy()
#output = output[:,:size[0],:size[1]]
#gt = np.asarray(label[0].numpy()[:size[0],:size[1]], dtype=np.int)
output_np = output.transpose(1, 2, 0)
output_np = np.asarray(np.argmax(output_np, axis=2),
dtype=np.int)
filename = os.path.join(
save_prob_dir, '{}/{}'.format(img_folder,
img_nm.replace("png",
"pth")))
file_dir = os.path.dirname(filename)
if not os.path.exists(file_dir):
os.makedirs(file_dir)
#save probility map before softmax
torch.save(output_raw.cpu(), filename)
#Image.fromarray(output_np, 'P').save(filename)
filename = os.path.join(save_seg_dir,
'{}/{}'.format(img_folder, img_nm))
file_dir = os.path.dirname(filename)
if not os.path.exists(file_dir):
os.makedirs(file_dir)
pil_image = Image.fromarray(output_np.astype(dtype=np.uint8))
pil_image.save(filename, 'PNG')
seg_viz = colorize(output_np)
filename = os.path.join(save_part_dir,
'{}/{}'.format(img_folder, img_nm))
file_dir = os.path.dirname(filename)
if not os.path.exists(file_dir):
os.makedirs(file_dir)
Image.fromarray(seg_viz.squeeze().transpose(1, 2, 0),
'RGB').save(filename)
seg_overlay_viz = (imgs_viz.numpy() * 0.8 +
seg_viz * 0.7).clip(0,
255.0).astype(np.uint8)
filename = os.path.join(save_overlay_dir,
'{}/{}'.format(img_folder, img_nm))
file_dir = os.path.dirname(filename)
if not os.path.exists(file_dir):
os.makedirs(file_dir)
Image.fromarray(seg_overlay_viz.squeeze().transpose(1, 2, 0),
'RGB').save(filename)
if args.crf:
output_dcrf_prob = utils.denseCRF(
imgs_viz.numpy().squeeze().transpose(1, 2, 0).astype(
np.uint8).copy(), output)
filename = os.path.join(
save_dcrf_prob_dir, '{}/{}'.format(
img_folder,
img_nm.replace(".png",
".npy").replace(".JPEG", ".npy")))
file_dir = os.path.dirname(filename)
if not os.path.exists(file_dir):
os.makedirs(file_dir)
np.save(filename, output_dcrf_prob)
output_dcrf = np.asarray(np.argmax(output_dcrf_prob,
axis=2),
dtype=np.int)
seg_dcrf_viz = colorize(output_dcrf)
filename = os.path.join(save_part_dcrf_dir,
'{}/{}'.format(img_folder, img_nm))
file_dir = os.path.dirname(filename)
if not os.path.exists(file_dir):
os.makedirs(file_dir)
Image.fromarray(seg_dcrf_viz.squeeze().transpose(1, 2, 0),
'RGB').save(filename)
seg_dcrf_overlay_viz = (imgs_viz.numpy() * 0.8 +
seg_dcrf_viz * 0.7).clip(
0, 255.0).astype(np.uint8)
filename = os.path.join(save_dcrf_overlay_dir,
'{}/{}'.format(img_folder, img_nm))
file_dir = os.path.dirname(filename)
if not os.path.exists(file_dir):
os.makedirs(file_dir)
Image.fromarray(
seg_dcrf_overlay_viz.squeeze().transpose(1, 2, 0),
'RGB').save(filename)
filename_lm = os.path.join(save_lm_dir,
'{}/{}'.format(img_folder, img_nm))
file_dir = os.path.dirname(filename_lm)
if not os.path.exists(file_dir):
os.makedirs(file_dir)
Image.fromarray(lms_viz[0, :, :, :].transpose(1, 2, 0),
'RGB').save(filename_lm)
def predict(self, X):
X = self.forward(X)
return np.argmax(softmax(X), axis=1)
# print(img_load) # print(label_load) #---------------------------------------------------------------- w_out = np.ones((784, 10)) z_1 = np.dot(layer_img.T, w_1).T + b_1 a_1 = activation.relu(z_1) z_2 = np.dot(a_1.T, w_2).T + b_2 a_2 = activation.relu(z_2) z_3 = np.dot(a_2.T, w_3).T + b_3 a_3 = activation.relu(z_3) z_out = np.dot(a_3.T, w_out).T output = loss.softmax(z_out) t = loss.crossEntropy(output, label) dw_3 = np.dot(a_2, (a_3 - label).T) print(dw_3) # # show image debugging # img_load = np.reshape(unpack(len(bin_img)*'B',bin_img),(28,28)) # label_load = int.from_bytes(bin_label,byteorder='big',signed=False) # plt.imshow(img_load,cmap=cm.binary) # plt.show() # print(img_load) # print(label_load)
def main():
"""Create the model and start the evaluation process."""
args = get_arguments()
gpu0 = args.gpu
h, w = map(int, args.input_size.split(','))
input_size = (h, w)
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
save_part_dir = os.path.join(args.save_dir, 'part_map')
if not os.path.exists(save_part_dir):
os.makedirs(save_part_dir)
save_overlay_dir = os.path.join(args.save_dir, 'part_overlay')
if not os.path.exists(save_overlay_dir):
os.makedirs(save_overlay_dir)
save_part_dcrf_dir = os.path.join(args.save_dir, 'part_map_dcrf')
if not os.path.exists(save_part_dcrf_dir):
os.makedirs(save_part_dcrf_dir)
save_dcrf_overlay_dir = os.path.join(args.save_dir, 'part_dcrf_overlay')
if not os.path.exists(save_dcrf_overlay_dir):
os.makedirs(save_dcrf_overlay_dir)
save_lm_dir = os.path.join(args.save_dir, 'landmarks')
if not os.path.exists(save_lm_dir):
os.makedirs(save_lm_dir)
save_seg_dir = os.path.join(args.save_dir, 'seg')
if not os.path.exists(save_seg_dir):
os.makedirs(save_seg_dir)
# create network
model = model_generator(args)
model.eval()
model.cuda(gpu0)
if args.dataset == 'CelebAWild':
from dataset.celeba_wild_dataset import CelebAWildDataset
dataset = CelebAWildDataset
testloader = data.DataLoader(dataset(args.data_dir,
args.data_list,
crop_size=input_size,
scale=False,
mirror=False,
mean=IMG_MEAN,
center_crop=False,
ignore_saliency_fg=False,
iou_threshold=0.3),
batch_size=1,
shuffle=False,
pin_memory=True)
else:
print('Not supported dataset {}'.format(args.dataset))
interp = nn.Upsample(size=input_size, mode='bilinear', align_corners=True)
colorize = utils.Colorize(args.num_parts + 1)
N = len(testloader)
landmarks = np.zeros((N, args.num_parts, 2))
landmarks_gt = np.zeros((N, args.lm_count, 2))
with torch.no_grad():
for index, batch in enumerate(testloader):
if index % 100 == 0:
path_split = args.save_dir.split('/')
print('{} processd: {}/{}'.format(index, path_split[-4],
path_split[-3]))
image = batch['img']
label = batch['saliency']
size_org = batch['size']
name = batch['name']
landmarks_gt[index, :, :] = batch['landmarks']
size = input_size
output = model(image.cuda(gpu0))
output = interp(output[2])
lms = Batch_Get_Centers(output)
landmarks[index, :, :] = lms
if args.save_viz:
mean_tensor = torch.tensor(IMG_MEAN).float().expand(
int(size[1]), int(size[0]), 3).transpose(0, 2)
imgs_viz = torch.clamp(image + mean_tensor, 0.0, 255.0)
#landmark visualization
lms_viz = Batch_Draw_GT_Landmarks(imgs_viz, output, lms)
output = softmax(output)
# normalize part
output /= output.max(dim=3,
keepdim=True)[0].max(dim=2,
keepdim=True)[0]
output[:, 0, :, :] = 0.1
output = output.cpu().data[0].numpy()
output = output[:, :size[0], :size[1]]
gt = np.asarray(label[0].numpy()[:size[0], :size[1]],
dtype=np.int)
output_np = output.transpose(1, 2, 0)
output_np = np.asarray(np.argmax(output_np, axis=2),
dtype=np.int)
filename = os.path.join(save_seg_dir,
'{}.png'.format(name[0][:-4]))
file_dir = os.path.dirname(filename)
if not os.path.exists(file_dir):
os.makedirs(file_dir)
Image.fromarray(output_np, 'P').save(filename)
seg_viz = colorize(output_np)
filename = os.path.join(save_part_dir,
'{}.png'.format(name[0][:-4]))
file_dir = os.path.dirname(filename)
if not os.path.exists(file_dir):
os.makedirs(file_dir)
Image.fromarray(seg_viz.squeeze().transpose(1, 2, 0),
'RGB').save(filename)
seg_overlay_viz = (imgs_viz.numpy() * 0.8 +
seg_viz * 0.7).clip(0,
255.0).astype(np.uint8)
filename = os.path.join(save_overlay_dir,
'{}.png'.format(name[0][:-4]))
file_dir = os.path.dirname(filename)
if not os.path.exists(file_dir):
os.makedirs(file_dir)
Image.fromarray(seg_overlay_viz.squeeze().transpose(1, 2, 0),
'RGB').save(filename)
if args.crf:
output_dcrf = utils.denseCRF(
imgs_viz.numpy().squeeze().transpose(1, 2, 0).astype(
np.uint8).copy(), output)
output_dcrf = np.asarray(np.argmax(output_dcrf, axis=2),
dtype=np.int)
seg_dcrf_viz = colorize(output_dcrf)
filename = os.path.join(save_part_dcrf_dir,
'{}.png'.format(name[0][:-4]))
file_dir = os.path.dirname(filename)
if not os.path.exists(file_dir):
os.makedirs(file_dir)
Image.fromarray(seg_dcrf_viz.squeeze().transpose(1, 2, 0),
'RGB').save(filename)
seg_dcrf_overlay_viz = (imgs_viz.numpy() * 0.8 +
seg_dcrf_viz * 0.7).clip(
0, 255.0).astype(np.uint8)
filename = os.path.join(save_dcrf_overlay_dir,
'{}.png'.format(name[0][:-4]))
file_dir = os.path.dirname(filename)
if not os.path.exists(file_dir):
os.makedirs(file_dir)
Image.fromarray(
seg_dcrf_overlay_viz.squeeze().transpose(1, 2, 0),
'RGB').save(filename)
filename_lm = os.path.join(save_lm_dir,
'{}.png'.format(name[0][:-4]))
file_dir = os.path.dirname(filename_lm)
if not os.path.exists(file_dir):
os.makedirs(file_dir)
Image.fromarray(lms_viz[0, :, :, :].transpose(1, 2, 0),
'RGB').save(filename_lm)
np.save(os.path.join(args.save_dir, 'pred_kp.npy'), landmarks)
np.save(os.path.join(args.save_dir, 'gt_kp.npy'), landmarks_gt)
def predict_proba(self, X):
score, _ = self.forward(X, False)
return loss_func.softmax(score)