def mkdir_if_missing(output_dir):
if not fu.exists(output_dir):
try:
fu.makedirs(output_dir)
except:
logging.error(
"Something went wrong in mkdir_if_missing. "
"Probably some other process created the directory already.")
def _launcher(config_name, logdir):
args = _setup_args(config_name, logdir)
fu.makedirs(args.logdir)
if args.control.train:
_train(args)
if args.control.test:
_test(args)
def plot_trajectory(dt, hardness, orig_maps, out_dir):
out_dir = os.path.join(out_dir, FLAGS.config_name+_get_suffix_str(),
FLAGS.imset)
fu.makedirs(out_dir)
out_file = os.path.join(out_dir, 'all_locs_at_t.pkl')
dt['hardness'] = hardness
utils.save_variables(out_file, dt.values(), dt.keys(), overwrite=True)
#Plot trajectories onto the maps
plt.set_cmap('gray')
for i in range(4000):
goal_loc = dt['all_goal_locs'][i, :, :]
locs = np.concatenate((dt['all_locs'][i,:,:],
dt['all_locs'][i,:,:]), axis=0)
xymin = np.minimum(np.min(goal_loc, axis=0), np.min(locs, axis=0))
xymax = np.maximum(np.max(goal_loc, axis=0), np.max(locs, axis=0))
xy1 = (xymax+xymin)/2. - 1.*np.maximum(np.max(xymax-xymin), 24)
xy2 = (xymax+xymin)/2. + 1.*np.maximum(np.max(xymax-xymin), 24)
fig, ax = utils.tight_imshow_figure(plt, figsize=(6,6))
ax.set_axis_on()
ax.patch.set_facecolor((0.333, 0.333, 0.333))
ax.set_xticks([])
ax.set_yticks([])
all_locs = dt['all_locs'][i,:,:]*1
uniq = np.where(np.any(all_locs[1:,:] != all_locs[:-1,:], axis=1))[0]+1
uniq = np.sort(uniq).tolist()
uniq.insert(0,0)
uniq = np.array(uniq)
all_locs = all_locs[uniq, :]
ax.plot(dt['all_locs'][i, 0, 0],
dt['all_locs'][i, 0, 1], 'b.', markersize=24)
ax.plot(dt['all_goal_locs'][i, 0, 0],
dt['all_goal_locs'][i, 0, 1], 'g*', markersize=19)
ax.plot(all_locs[:,0], all_locs[:,1], 'r', alpha=0.4, linewidth=2)
ax.scatter(all_locs[:,0], all_locs[:,1],
c=5+np.arange(all_locs.shape[0])*1./all_locs.shape[0],
cmap='Reds', s=30, linewidth=0)
ax.imshow(orig_maps, origin='lower', vmin=-1.0, vmax=2.0, aspect='equal')
ax.set_xlim([xy1[0], xy2[0]])
ax.set_ylim([xy1[1], xy2[1]])
file_name = os.path.join(out_dir, 'trajectory_{:04d}.png'.format(i))
print(file_name)
with fu.fopen(file_name, 'w') as f:
plt.savefig(f)
plt.close(fig)
def plot_trajectory(dt, hardness, orig_maps, out_dir):
out_dir = os.path.join(out_dir, FLAGS.config_name+_get_suffix_str(),
FLAGS.imset)
fu.makedirs(out_dir)
out_file = os.path.join(out_dir, 'all_locs_at_t.pkl')
dt['hardness'] = hardness
utils.save_variables(out_file, dt.values(), dt.keys(), overwrite=True)
#Plot trajectories onto the maps
plt.set_cmap('gray')
for i in range(4000):
goal_loc = dt['all_goal_locs'][i, :, :]
locs = np.concatenate((dt['all_locs'][i,:,:],
dt['all_locs'][i,:,:]), axis=0)
xymin = np.minimum(np.min(goal_loc, axis=0), np.min(locs, axis=0))
xymax = np.maximum(np.max(goal_loc, axis=0), np.max(locs, axis=0))
xy1 = (xymax+xymin)/2. - 1.*np.maximum(np.max(xymax-xymin), 24)
xy2 = (xymax+xymin)/2. + 1.*np.maximum(np.max(xymax-xymin), 24)
fig, ax = utils.tight_imshow_figure(plt, figsize=(6,6))
ax.set_axis_on()
ax.patch.set_facecolor((0.333, 0.333, 0.333))
ax.set_xticks([])
ax.set_yticks([])
all_locs = dt['all_locs'][i,:,:]*1
uniq = np.where(np.any(all_locs[1:,:] != all_locs[:-1,:], axis=1))[0]+1
uniq = np.sort(uniq).tolist()
uniq.insert(0,0)
uniq = np.array(uniq)
all_locs = all_locs[uniq, :]
ax.plot(dt['all_locs'][i, 0, 0],
dt['all_locs'][i, 0, 1], 'b.', markersize=24)
ax.plot(dt['all_goal_locs'][i, 0, 0],
dt['all_goal_locs'][i, 0, 1], 'g*', markersize=19)
ax.plot(all_locs[:,0], all_locs[:,1], 'r', alpha=0.4, linewidth=2)
ax.scatter(all_locs[:,0], all_locs[:,1],
c=5+np.arange(all_locs.shape[0])*1./all_locs.shape[0],
cmap='Reds', s=30, linewidth=0)
ax.imshow(orig_maps, origin='lower', vmin=-1.0, vmax=2.0, aspect='equal')
ax.set_xlim([xy1[0], xy2[0]])
ax.set_ylim([xy1[1], xy2[1]])
file_name = os.path.join(out_dir, 'trajectory_{:04d}.png'.format(i))
print(file_name)
with fu.fopen(file_name, 'w') as f:
plt.savefig(f)
plt.close(fig)
def mkdir_if_missing(output_dir):
if not fu.exists(output_dir):
fu.makedirs(output_dir)
def plot_trajectory_first_person(dt, orig_maps, out_dir):
out_dir = os.path.join(out_dir, FLAGS.config_name+_get_suffix_str(),
FLAGS.imset)
fu.makedirs(out_dir)
# Load the model so that we can render.
plt.set_cmap('gray')
samples_per_action = 8; wait_at_action = 0;
Writer = animation.writers['mencoder']
writer = Writer(fps=3*(samples_per_action+wait_at_action),
metadata=dict(artist='anonymous'), bitrate=1800)
args = sna.get_args_for_config(FLAGS.config_name + '+bench_'+FLAGS.imset)
args.navtask.logdir = None
navtask_ = copy.deepcopy(args.navtask)
navtask_.camera_param.modalities = ['rgb']
navtask_.task_params.modalities = ['rgb']
sz = 512
navtask_.camera_param.height = sz
navtask_.camera_param.width = sz
navtask_.task_params.img_height = sz
navtask_.task_params.img_width = sz
R = lambda: nav_env.get_multiplexer_class(navtask_, 0)
R = R()
b = R.buildings[0]
f = [0 for _ in range(wait_at_action)] + \
[float(_)/samples_per_action for _ in range(samples_per_action)];
# Generate things for it to render.
inds_to_do = []
inds_to_do += [1, 4, 10] #1291, 1268, 1273, 1289, 1302, 1426, 1413, 1449, 1399, 1390]
for i in inds_to_do:
fig = plt.figure(figsize=(10,8))
gs = GridSpec(3,4)
gs.update(wspace=0.05, hspace=0.05, left=0.0, top=0.97, right=1.0, bottom=0.)
ax = fig.add_subplot(gs[:,:-1])
ax1 = fig.add_subplot(gs[0,-1])
ax2 = fig.add_subplot(gs[1,-1])
ax3 = fig.add_subplot(gs[2,-1])
axes = [ax, ax1, ax2, ax3]
# ax = fig.add_subplot(gs[:,:])
# axes = [ax]
for ax in axes:
ax.set_axis_off()
node_ids = dt['all_node_ids'][i, :, 0]*1
# Prune so that last node is not repeated more than 3 times?
if np.all(node_ids[-4:] == node_ids[-1]):
while node_ids[-4] == node_ids[-1]:
node_ids = node_ids[:-1]
num_steps = np.minimum(FLAGS.num_steps, len(node_ids))
xyt = b.to_actual_xyt_vec(b.task.nodes[node_ids])
xyt_diff = xyt[1:,:] - xyt[:-1:,:]
xyt_diff[:,2] = np.mod(xyt_diff[:,2], 4)
ind = np.where(xyt_diff[:,2] == 3)[0]
xyt_diff[ind, 2] = -1
xyt_diff = np.expand_dims(xyt_diff, axis=1)
to_cat = [xyt_diff*_ for _ in f]
perturbs_all = np.concatenate(to_cat, axis=1)
perturbs_all = np.concatenate([perturbs_all, np.zeros_like(perturbs_all[:,:,:1])], axis=2)
node_ids_all = np.expand_dims(node_ids, axis=1)*1
node_ids_all = np.concatenate([node_ids_all for _ in f], axis=1)
node_ids_all = np.reshape(node_ids_all[:-1,:], -1)
perturbs_all = np.reshape(perturbs_all, [-1, 4])
imgs = b.render_nodes(b.task.nodes[node_ids_all,:], perturb=perturbs_all)
# Get action at each node.
actions = []
_, action_to_nodes = b.get_feasible_actions(node_ids)
for j in range(num_steps-1):
action_to_node = action_to_nodes[j]
node_to_action = dict(zip(action_to_node.values(), action_to_node.keys()))
actions.append(node_to_action[node_ids[j+1]])
def init_fn():
return fig,
gt_dist_to_goal = []
# Render trajectories.
def worker(j):
# Plot the image.
step_number = j/(samples_per_action + wait_at_action)
img = imgs[j]; ax = axes[0]; ax.clear(); ax.set_axis_off();
img = img.astype(np.uint8); ax.imshow(img);
tt = ax.set_title(
"First Person View\n" +
"Top corners show diagnostics (distance, agents' action) not input to agent.",
fontsize=12)
plt.setp(tt, color='white')
# Distance to goal.
t = 'Dist to Goal:\n{:2d} steps'.format(int(dt['all_d_at_t'][i, step_number]))
t = ax.text(0.01, 0.99, t,
horizontalalignment='left',
verticalalignment='top',
fontsize=20, color='red',
transform=ax.transAxes, alpha=1.0)
t.set_bbox(dict(color='white', alpha=0.85, pad=-0.1))
# Action to take.
action_latex = ['$\odot$ ', '$\curvearrowright$ ', '$\curvearrowleft$ ', r'$\Uparrow$ ']
t = ax.text(0.99, 0.99, action_latex[actions[step_number]],
horizontalalignment='right',
verticalalignment='top',
fontsize=40, color='green',
transform=ax.transAxes, alpha=1.0)
t.set_bbox(dict(color='white', alpha=0.85, pad=-0.1))
# Plot the map top view.
ax = axes[-1]
if j == 0:
# Plot the map
locs = dt['all_locs'][i,:num_steps,:]
goal_loc = dt['all_goal_locs'][i,:,:]
xymin = np.minimum(np.min(goal_loc, axis=0), np.min(locs, axis=0))
xymax = np.maximum(np.max(goal_loc, axis=0), np.max(locs, axis=0))
xy1 = (xymax+xymin)/2. - 0.7*np.maximum(np.max(xymax-xymin), 24)
xy2 = (xymax+xymin)/2. + 0.7*np.maximum(np.max(xymax-xymin), 24)
ax.set_axis_on()
ax.patch.set_facecolor((0.333, 0.333, 0.333))
ax.set_xticks([]); ax.set_yticks([]);
ax.imshow(orig_maps, origin='lower', vmin=-1.0, vmax=2.0)
ax.plot(goal_loc[:,0], goal_loc[:,1], 'g*', markersize=12)
locs = dt['all_locs'][i,:1,:]
ax.plot(locs[:,0], locs[:,1], 'b.', markersize=12)
ax.set_xlim([xy1[0], xy2[0]])
ax.set_ylim([xy1[1], xy2[1]])
locs = dt['all_locs'][i,step_number,:]
locs = np.expand_dims(locs, axis=0)
ax.plot(locs[:,0], locs[:,1], 'r.', alpha=1.0, linewidth=0, markersize=4)
tt = ax.set_title('Trajectory in topview', fontsize=14)
plt.setp(tt, color='white')
return fig,
line_ani = animation.FuncAnimation(fig, worker,
(num_steps-1)*(wait_at_action+samples_per_action),
interval=500, blit=True, init_func=init_fn)
tmp_file_name = 'tmp.mp4'
line_ani.save(tmp_file_name, writer=writer, savefig_kwargs={'facecolor':'black'})
out_file_name = os.path.join(out_dir, 'vis_{:04d}.mp4'.format(i))
print(out_file_name)
if fu.exists(out_file_name):
gfile.Remove(out_file_name)
gfile.Copy(tmp_file_name, out_file_name)
gfile.Remove(tmp_file_name)
plt.close(fig)
def main(_):
args = config_distill.get_args_for_config(FLAGS.config_name)
args.logdir = FLAGS.logdir
args.solver.num_workers = FLAGS.num_workers
args.solver.task = FLAGS.task
args.solver.ps_tasks = FLAGS.ps_tasks
args.solver.master = FLAGS.master
args.buildinger.env_class = nav_env.MeshMapper
fu.makedirs(args.logdir)
args.buildinger.logdir = args.logdir
R = nav_env.get_multiplexor_class(args.buildinger, args.solver.task)
if False:
pr = cProfile.Profile()
pr.enable()
rng = np.random.RandomState(0)
for i in range(1):
b, instances_perturbs = R.sample_building(rng)
inputs = b.worker(*(instances_perturbs))
for j in range(inputs['imgs'].shape[0]):
p = os.path.join('tmp', '{:d}.png'.format(j))
img = inputs['imgs'][j, 0, :, :, :3] * 1
img = (img).astype(np.uint8)
fu.write_image(p, img)
print(inputs['imgs'].shape)
inputs = R.pre(inputs)
pr.disable()
pr.print_stats(2)
if args.control.train:
if not gfile.Exists(args.logdir):
gfile.MakeDirs(args.logdir)
m = utils.Foo()
m.tf_graph = tf.Graph()
config = tf.ConfigProto()
config.device_count['GPU'] = 1
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = 0.8
with m.tf_graph.as_default():
with tf.device(tf.train.replica_device_setter(
args.solver.ps_tasks)):
m = distill.setup_to_run(m,
args,
is_training=True,
batch_norm_is_training=True)
train_step_kwargs = distill.setup_train_step_kwargs_mesh(
m,
R,
os.path.join(args.logdir, 'train_bkp'),
rng_seed=args.solver.task,
is_chief=args.solver.task == 0,
iters=1,
train_display_interval=args.summary.display_interval)
final_loss = slim.learning.train(
train_op=m.train_op,
logdir=args.logdir,
master=args.solver.master,
is_chief=args.solver.task == 0,
number_of_steps=args.solver.max_steps,
train_step_fn=tf_utils.train_step_custom,
train_step_kwargs=train_step_kwargs,
global_step=m.global_step_op,
init_op=m.init_op,
init_fn=m.init_fn,
sync_optimizer=m.sync_optimizer,
saver=m.saver_op,
summary_op=None,
session_config=config)
if args.control.test:
m = utils.Foo()
m.tf_graph = tf.Graph()
checkpoint_dir = os.path.join(format(args.logdir))
with m.tf_graph.as_default():
m = distill.setup_to_run(m,
args,
is_training=False,
batch_norm_is_training=args.control.
force_batchnorm_is_training_at_test)
train_step_kwargs = distill.setup_train_step_kwargs_mesh(
m,
R,
os.path.join(args.logdir, args.control.test_name),
rng_seed=args.solver.task + 1,
is_chief=args.solver.task == 0,
iters=args.summary.test_iters,
train_display_interval=None)
sv = slim.learning.supervisor.Supervisor(
graph=ops.get_default_graph(),
logdir=None,
init_op=m.init_op,
summary_op=None,
summary_writer=None,
global_step=None,
saver=m.saver_op)
last_checkpoint = None
while True:
last_checkpoint = slim.evaluation.wait_for_new_checkpoint(
checkpoint_dir, last_checkpoint)
checkpoint_iter = int(
os.path.basename(last_checkpoint).split('-')[1])
start = time.time()
logging.info(
'Starting evaluation at %s using checkpoint %s.',
time.strftime('%Y-%m-%d-%H:%M:%S', time.localtime()),
last_checkpoint)
config = tf.ConfigProto()
config.device_count['GPU'] = 1
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = 0.8
with sv.managed_session(args.solver.master,
config=config,
start_standard_services=False) as sess:
sess.run(m.init_op)
sv.saver.restore(sess, last_checkpoint)
sv.start_queue_runners(sess)
vals, _ = tf_utils.train_step_custom(sess,
None,
m.global_step_op,
train_step_kwargs,
mode='val')
if checkpoint_iter >= args.solver.max_steps:
break
def mkdir_if_missing(output_dir):
if not fu.exists(output_dir):
fu.makedirs(output_dir)
def plot_trajectory_first_person(dt, orig_maps, out_dir):
out_dir = os.path.join(out_dir, FLAGS.config_name+_get_suffix_str(),
FLAGS.imset)
fu.makedirs(out_dir)
# Load the model so that we can render.
plt.set_cmap('gray')
samples_per_action = 8; wait_at_action = 0;
Writer = animation.writers['mencoder']
writer = Writer(fps=3*(samples_per_action+wait_at_action),
metadata=dict(artist='anonymous'), bitrate=1800)
args = sna.get_args_for_config(FLAGS.config_name + '+bench_'+FLAGS.imset)
args.navtask.logdir = None
navtask_ = copy.deepcopy(args.navtask)
navtask_.camera_param.modalities = ['rgb']
navtask_.task_params.modalities = ['rgb']
sz = 512
navtask_.camera_param.height = sz
navtask_.camera_param.width = sz
navtask_.task_params.img_height = sz
navtask_.task_params.img_width = sz
R = lambda: nav_env.get_multiplexer_class(navtask_, 0)
R = R()
b = R.buildings[0]
f = [0 for _ in range(wait_at_action)] + \
[float(_)/samples_per_action for _ in range(samples_per_action)];
# Generate things for it to render.
inds_to_do = []
inds_to_do += [1, 4, 10] #1291, 1268, 1273, 1289, 1302, 1426, 1413, 1449, 1399, 1390]
for i in inds_to_do:
fig = plt.figure(figsize=(10,8))
gs = GridSpec(3,4)
gs.update(wspace=0.05, hspace=0.05, left=0.0, top=0.97, right=1.0, bottom=0.)
ax = fig.add_subplot(gs[:,:-1])
ax1 = fig.add_subplot(gs[0,-1])
ax2 = fig.add_subplot(gs[1,-1])
ax3 = fig.add_subplot(gs[2,-1])
axes = [ax, ax1, ax2, ax3]
# ax = fig.add_subplot(gs[:,:])
# axes = [ax]
for ax in axes:
ax.set_axis_off()
node_ids = dt['all_node_ids'][i, :, 0]*1
# Prune so that last node is not repeated more than 3 times?
if np.all(node_ids[-4:] == node_ids[-1]):
while node_ids[-4] == node_ids[-1]:
node_ids = node_ids[:-1]
num_steps = np.minimum(FLAGS.num_steps, len(node_ids))
xyt = b.to_actual_xyt_vec(b.task.nodes[node_ids])
xyt_diff = xyt[1:,:] - xyt[:-1:,:]
xyt_diff[:,2] = np.mod(xyt_diff[:,2], 4)
ind = np.where(xyt_diff[:,2] == 3)[0]
xyt_diff[ind, 2] = -1
xyt_diff = np.expand_dims(xyt_diff, axis=1)
to_cat = [xyt_diff*_ for _ in f]
perturbs_all = np.concatenate(to_cat, axis=1)
perturbs_all = np.concatenate([perturbs_all, np.zeros_like(perturbs_all[:,:,:1])], axis=2)
node_ids_all = np.expand_dims(node_ids, axis=1)*1
node_ids_all = np.concatenate([node_ids_all for _ in f], axis=1)
node_ids_all = np.reshape(node_ids_all[:-1,:], -1)
perturbs_all = np.reshape(perturbs_all, [-1, 4])
imgs = b.render_nodes(b.task.nodes[node_ids_all,:], perturb=perturbs_all)
# Get action at each node.
actions = []
_, action_to_nodes = b.get_feasible_actions(node_ids)
for j in range(num_steps-1):
action_to_node = action_to_nodes[j]
node_to_action = dict(zip(action_to_node.values(), action_to_node.keys()))
actions.append(node_to_action[node_ids[j+1]])
def init_fn():
return fig,
gt_dist_to_goal = []
# Render trajectories.
def worker(j):
# Plot the image.
step_number = j/(samples_per_action + wait_at_action)
img = imgs[j]; ax = axes[0]; ax.clear(); ax.set_axis_off();
img = img.astype(np.uint8); ax.imshow(img);
tt = ax.set_title(
"First Person View\n" +
"Top corners show diagnostics (distance, agents' action) not input to agent.",
fontsize=12)
plt.setp(tt, color='white')
# Distance to goal.
t = 'Dist to Goal:\n{:2d} steps'.format(int(dt['all_d_at_t'][i, step_number]))
t = ax.text(0.01, 0.99, t,
horizontalalignment='left',
verticalalignment='top',
fontsize=20, color='red',
transform=ax.transAxes, alpha=1.0)
t.set_bbox(dict(color='white', alpha=0.85, pad=-0.1))
# Action to take.
action_latex = ['$\odot$ ', '$\curvearrowright$ ', '$\curvearrowleft$ ', r'$\Uparrow$ ']
t = ax.text(0.99, 0.99, action_latex[actions[step_number]],
horizontalalignment='right',
verticalalignment='top',
fontsize=40, color='green',
transform=ax.transAxes, alpha=1.0)
t.set_bbox(dict(color='white', alpha=0.85, pad=-0.1))
# Plot the map top view.
ax = axes[-1]
if j == 0:
# Plot the map
locs = dt['all_locs'][i,:num_steps,:]
goal_loc = dt['all_goal_locs'][i,:,:]
xymin = np.minimum(np.min(goal_loc, axis=0), np.min(locs, axis=0))
xymax = np.maximum(np.max(goal_loc, axis=0), np.max(locs, axis=0))
xy1 = (xymax+xymin)/2. - 0.7*np.maximum(np.max(xymax-xymin), 24)
xy2 = (xymax+xymin)/2. + 0.7*np.maximum(np.max(xymax-xymin), 24)
ax.set_axis_on()
ax.patch.set_facecolor((0.333, 0.333, 0.333))
ax.set_xticks([]); ax.set_yticks([]);
ax.imshow(orig_maps, origin='lower', vmin=-1.0, vmax=2.0)
ax.plot(goal_loc[:,0], goal_loc[:,1], 'g*', markersize=12)
locs = dt['all_locs'][i,:1,:]
ax.plot(locs[:,0], locs[:,1], 'b.', markersize=12)
ax.set_xlim([xy1[0], xy2[0]])
ax.set_ylim([xy1[1], xy2[1]])
locs = dt['all_locs'][i,step_number,:]
locs = np.expand_dims(locs, axis=0)
ax.plot(locs[:,0], locs[:,1], 'r.', alpha=1.0, linewidth=0, markersize=4)
tt = ax.set_title('Trajectory in topview', fontsize=14)
plt.setp(tt, color='white')
return fig,
line_ani = animation.FuncAnimation(fig, worker,
(num_steps-1)*(wait_at_action+samples_per_action),
interval=500, blit=True, init_func=init_fn)
tmp_file_name = 'tmp.mp4'
line_ani.save(tmp_file_name, writer=writer, savefig_kwargs={'facecolor':'black'})
out_file_name = os.path.join(out_dir, 'vis_{:04d}.mp4'.format(i))
print(out_file_name)
if fu.exists(out_file_name):
gfile.Remove(out_file_name)
gfile.Copy(tmp_file_name, out_file_name)
gfile.Remove(tmp_file_name)
plt.close(fig)
def main(_):
args = config_distill.get_args_for_config(FLAGS.config_name)
args.logdir = FLAGS.logdir
args.solver.num_workers = FLAGS.num_workers
args.solver.task = FLAGS.task
args.solver.ps_tasks = FLAGS.ps_tasks
args.solver.master = FLAGS.master
args.buildinger.env_class = nav_env.MeshMapper
fu.makedirs(args.logdir)
args.buildinger.logdir = args.logdir
R = nav_env.get_multiplexor_class(args.buildinger, args.solver.task)
if False:
pr = cProfile.Profile()
pr.enable()
rng = np.random.RandomState(0)
for i in range(1):
b, instances_perturbs = R.sample_building(rng)
inputs = b.worker(*(instances_perturbs))
for j in range(inputs['imgs'].shape[0]):
p = os.path.join('tmp', '{:d}.png'.format(j))
img = inputs['imgs'][j,0,:,:,:3]*1
img = (img).astype(np.uint8)
fu.write_image(p, img)
print(inputs['imgs'].shape)
inputs = R.pre(inputs)
pr.disable()
pr.print_stats(2)
if args.control.train:
if not gfile.Exists(args.logdir):
gfile.MakeDirs(args.logdir)
m = utils.Foo()
m.tf_graph = tf.Graph()
config = tf.ConfigProto()
config.device_count['GPU'] = 1
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = 0.8
with m.tf_graph.as_default():
with tf.device(tf.train.replica_device_setter(args.solver.ps_tasks)):
m = distill.setup_to_run(m, args, is_training=True,
batch_norm_is_training=True)
train_step_kwargs = distill.setup_train_step_kwargs_mesh(
m, R, os.path.join(args.logdir, 'train'),
rng_seed=args.solver.task, is_chief=args.solver.task==0, iters=1,
train_display_interval=args.summary.display_interval)
final_loss = slim.learning.train(
train_op=m.train_op,
logdir=args.logdir,
master=args.solver.master,
is_chief=args.solver.task == 0,
number_of_steps=args.solver.max_steps,
train_step_fn=tf_utils.train_step_custom,
train_step_kwargs=train_step_kwargs,
global_step=m.global_step_op,
init_op=m.init_op,
init_fn=m.init_fn,
sync_optimizer=m.sync_optimizer,
saver=m.saver_op,
summary_op=None, session_config=config)
if args.control.test:
m = utils.Foo()
m.tf_graph = tf.Graph()
checkpoint_dir = os.path.join(format(args.logdir))
with m.tf_graph.as_default():
m = distill.setup_to_run(m, args, is_training=False,
batch_norm_is_training=args.control.force_batchnorm_is_training_at_test)
train_step_kwargs = distill.setup_train_step_kwargs_mesh(
m, R, os.path.join(args.logdir, args.control.test_name),
rng_seed=args.solver.task+1, is_chief=args.solver.task==0,
iters=args.summary.test_iters, train_display_interval=None)
sv = slim.learning.supervisor.Supervisor(
graph=ops.get_default_graph(), logdir=None, init_op=m.init_op,
summary_op=None, summary_writer=None, global_step=None, saver=m.saver_op)
last_checkpoint = None
while True:
last_checkpoint = slim.evaluation.wait_for_new_checkpoint(checkpoint_dir, last_checkpoint)
checkpoint_iter = int(os.path.basename(last_checkpoint).split('-')[1])
start = time.time()
logging.info('Starting evaluation at %s using checkpoint %s.',
time.strftime('%Y-%m-%d-%H:%M:%S', time.localtime()),
last_checkpoint)
config = tf.ConfigProto()
config.device_count['GPU'] = 1
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = 0.8
with sv.managed_session(args.solver.master,config=config,
start_standard_services=False) as sess:
sess.run(m.init_op)
sv.saver.restore(sess, last_checkpoint)
sv.start_queue_runners(sess)
vals, _ = tf_utils.train_step_custom(
sess, None, m.global_step_op, train_step_kwargs, mode='val')
if checkpoint_iter >= args.solver.max_steps:
break
