def save_img(img, save_path):
image_numpy = util.tensor2im(img)
util.save_image(image_numpy, save_path, create_dir=True)
return image_numpy
if __name__ == '__main__':
opt = TestOptions().parse()
data_info = data.dataset_info()
datanum = data_info.get_dataset(opt)[0]
folderlevel = data_info.folder_level[datanum]
dataloaders = data.create_dataloader_test(opt)
visualizer = Visualizer(opt)
iter_counter = IterationCounter(opt, len(dataloaders[0]) * opt.render_thread)
# create a webpage that summarizes the all results
testing_queue = Queue(10)
ngpus = opt.device_count
render_gpu_ids = list(range(ngpus - opt.render_thread, ngpus))
render_layer_list = []
for gpu in render_gpu_ids:
opt.gpu_ids = gpu
render_layer = TestRender(opt)
render_layer_list.append(render_layer)
NB_STACKS, SIGMA))
sys.stdout.flush()
# define and build the model
with tf.variable_scope("model", reuse=False):
model_train = StackedHourglass(nb_stacks=NB_STACKS, sigma=SIGMA)
all_heatmaps_pred_train, p2d_pred_train = model_train(im_train, True)
# sys.exit(0)
with tf.variable_scope("model", reuse=True):
model_valid = StackedHourglass(nb_stacks=NB_STACKS, sigma=SIGMA)
all_heatmaps_pred_valid, p2d_pred_valid = model_valid(im_valid, False)
# test data
im_test = create_dataloader_test(
data_root=DATA_PATH) # load test data with batch_size=1
with tf.variable_scope("model", reuse=True):
model_test = StackedHourglass(nb_stacks=NB_STACKS, sigma=SIGMA)
all_heatmaps_pred_test, p2d_pred_test = model_test(im_test, False)
# sys.exit(0)
# compute loss
print("Loss...")
sys.stdout.flush()
loss_train = model_train.compute_loss(p2d_gt_train,
all_heatmaps_pred_train)
loss_valid = model_valid.compute_loss(p2d_gt_valid,
all_heatmaps_pred_valid)
# sys.exit(0)
def _inference(args):
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = "0"
with tf.Session(config=config) as sess:
dataloader_train, dataloader_valid, max_row, max_col, iterator_output, rcstrs_output, iterator_output_valid, rcstrs_output_valid = create_dataloader_train(valid_ratio=args.valid_ratio, batch_size=args.batch_size)
#dataloader_test, row_col_prediction, rcstrs = create_dataloader_test(batch_size=args.batch_size)
iterator_test, row_col_prediction, rcstrs = create_dataloader_test(batch_size=args.batch_size)
sess.run([iterator_test.initializer,
#iterator_output.initializer,
iterator_output_valid.initializer])
dataloader_test = iterator_test.get_next()
#dataloader_output = iterator_output.get_next()
dataloader_output_valid = iterator_output_valid.get_next()
#row_col_train, label_train = dataloader_train
#row_col_valid, label_valid = dataloader_valid
#row_col_output, label_output = dataloader_output
row_col_output_valid, label_output_valid = dataloader_output_valid
row_col_test, label_test = dataloader_test
if args.model == "NeuCF2":
#with tf.variable_scope("model", reuse=False):
# model_train = NeuCF2(row_col_train, label_train, max_row, max_col, args)
# sess.run(tf.global_variables_initializer())
#with tf.variable_scope("model", reuse=True):
# model_valid = NeuCF2(row_col_valid, label_valid, max_row, max_col, args)
# sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=False):
model_test = NeuCF2(row_col_test, label_test, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
#with tf.variable_scope("model", reuse=True):
# model_output = NeuCF2(row_col_output, label_output, max_row, max_col, args)
# sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_output_valid = NeuCF2(row_col_output_valid, label_output_valid, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
if args.external_embedding:
model_test.init_embedding(sess, args)
#model_valid.init_embedding(sess, args)
#model_test.init_embedding(sess, args)
elif args.model == "NeuCF3":
#with tf.variable_scope("model", reuse=False):
# model_train = NeuCF3(row_col_train, label_train, max_row, max_col, args)
# sess.run(tf.global_variables_initializer())
#with tf.variable_scope("model", reuse=True):
# model_valid = NeuCF3(row_col_valid, label_valid, max_row, max_col, args)
# sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=False):
model_test = NeuCF3(row_col_test, label_test, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
#with tf.variable_scope("model", reuse=True):
# model_output = NeuCF3(row_col_output, label_output, max_row, max_col, args)
# sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_output_valid = NeuCF3(row_col_output_valid, label_output_valid, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
if args.external_embedding:
model_test.init_embedding(sess, args)
elif args.model == "NeuCF4":
#with tf.variable_scope("model", reuse=False):
# model_train = NeuCF4(row_col_train, label_train, max_row, max_col, args)
# sess.run(tf.global_variables_initializer())
#with tf.variable_scope("model", reuse=True):
# model_valid = NeuCF4(row_col_valid, label_valid, max_row, max_col, args)
# sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=False):
model_test = NeuCF4(row_col_test, label_test, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
#with tf.variable_scope("model", reuse=True):
# model_output = NeuCF4(row_col_output, label_output, max_row, max_col, args)
# sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_output_valid = NeuCF4(row_col_output_valid, label_output_valid, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
if args.external_embedding:
model_test.init_embedding(sess, args)
elif args.model == "NeuCF":
#with tf.variable_scope("model", reuse=False):
# model_train = NeuCF(row_col_train, label_train, max_row, max_col, args)
# sess.run(tf.global_variables_initializer())
#with tf.variable_scope("model", reuse=True):
# model_valid = NeuCF(row_col_valid, label_valid, max_row, max_col, args)
# sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=False):
model_test = NeuCF(row_col_test, label_test, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
#with tf.variable_scope("model", reuse=True):
# model_output = NeuCF(row_col_output, label_output, max_row, max_col, args)
# sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_output_valid = NeuCF(row_col_output_valid, label_output_valid, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
vars = [v for v in tf.global_variables() if v.name.startswith("model/NeuCF")]
saver = tf.train.Saver(vars)
saver.restore(sess, args.model_path)
output_valid_prediction = None
for j in range( math.ceil( len(rcstrs_output_valid) / args.batch_size) ):
predict = model_output_valid.step(sess, isTesting=True, dropout_keep_prob=1)
if j == 0:
output_valid_prediction = predict
else:
output_valid_prediction = np.concatenate( [output_valid_prediction, predict] , axis=0 )
output_valid_prediction = np.reshape(output_valid_prediction, (output_valid_prediction.shape[0],))
df = pd.DataFrame( {'Id': rcstrs_output_valid,'Prediction': output_valid_prediction} )
df.to_csv(args.output_valid_path,index=False)
test_prediction = None
for j in range(math.ceil(row_col_prediction.shape[0] / args.batch_size)):
predict = model_test.step(sess, isTesting=True, dropout_keep_prob=1)
if j == 0:
test_prediction = predict
else:
test_prediction = np.concatenate([test_prediction, predict], axis=0)
test_prediction = np.reshape(test_prediction, (test_prediction.shape[0],))
# data frame is reconstructed since the direct modification is too slow
df = pd.DataFrame({'Id': rcstrs,'Prediction': test_prediction})
df.to_csv(args.test_path,index=False)
def _train(args):
os.makedirs(os.path.join(args.log_path), exist_ok=True)
logfile = open(os.path.join(args.log_path, 'log.txt'), 'w', buffering=1)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = "0"
with tf.Session(config=config) as sess:
dataloader_train, dataloader_valid, max_row, max_col, iterator_output, rcstrs_output, iterator_output_valid, rcstrs_output_valid = create_dataloader_train(valid_ratio=args.valid_ratio, batch_size=args.batch_size)
#dataloader_test, row_col_prediction, rcstrs = create_dataloader_test(batch_size=args.batch_size)
iterator_test, row_col_prediction, rcstrs = create_dataloader_test(batch_size=args.batch_size)
sess.run([iterator_test.initializer,
iterator_output.initializer,
iterator_output_valid.initializer])
dataloader_test = iterator_test.get_next()
dataloader_output = iterator_output.get_next()
dataloader_output_valid = iterator_output_valid.get_next()
row_col_train, label_train = dataloader_train
row_col_valid, label_valid = dataloader_valid
row_col_output, label_output = dataloader_output
row_col_output_valid, label_output_valid = dataloader_output_valid
row_col_test, label_test = dataloader_test
if args.model == "NeuCF2":
with tf.variable_scope("model", reuse=False):
model_train = NeuCF2(row_col_train, label_train, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_valid = NeuCF2(row_col_valid, label_valid, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_test = NeuCF2(row_col_test, label_test, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_output = NeuCF2(row_col_output, label_output, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_output_valid = NeuCF2(row_col_output_valid, label_output_valid, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
if args.external_embedding:
model_train.init_embedding(sess, args)
#model_valid.init_embedding(sess, args)
#model_test.init_embedding(sess, args)
elif args.model == "NeuCF3":
with tf.variable_scope("model", reuse=False):
model_train = NeuCF3(row_col_train, label_train, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_valid = NeuCF3(row_col_valid, label_valid, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_test = NeuCF3(row_col_test, label_test, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_output = NeuCF3(row_col_output, label_output, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_output_valid = NeuCF3(row_col_output_valid, label_output_valid, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
if args.external_embedding:
model_train.init_embedding(sess, args)
elif args.model == "NeuCF4":
with tf.variable_scope("model", reuse=False):
model_train = NeuCF4(row_col_train, label_train, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_valid = NeuCF4(row_col_valid, label_valid, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_test = NeuCF4(row_col_test, label_test, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_output = NeuCF4(row_col_output, label_output, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_output_valid = NeuCF4(row_col_output_valid, label_output_valid, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
if args.external_embedding:
model_train.init_embedding(sess, args)
elif args.model == "NeuCF":
with tf.variable_scope("model", reuse=False):
model_train = NeuCF(row_col_train, label_train, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_valid = NeuCF(row_col_valid, label_valid, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_test = NeuCF(row_col_test, label_test, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_output = NeuCF(row_col_output, label_output, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
with tf.variable_scope("model", reuse=True):
model_output_valid = NeuCF(row_col_output_valid, label_output_valid, max_row, max_col, args)
sess.run(tf.global_variables_initializer())
summary_writer = tf.summary.FileWriter(args.log_path, sess.graph)
vars = [v for v in tf.global_variables() if v.name.startswith("model/NeuCF")]
saver = tf.train.Saver(vars, max_to_keep=200)
should_stop = False
stopping_step = 0
cur_best_pre_0 = 1.0
for i in range(args.epochs):
epoch_loss = 0
for j in range(args.epoch_iter):
batch_loss = model_train.step(sess, isTraining=True, dropout_keep_prob=0.5)
epoch_loss += batch_loss / args.epoch_iter
valid_loss = 0
valid_sse = 0
for j in range(args.valid_iter):
batch_loss, batch_sse = model_valid.step(sess, isValidating=True, dropout_keep_prob=1)
valid_loss += batch_loss / args.valid_iter
valid_sse += batch_sse
valid_sse = valid_sse / (args.valid_iter * args.batch_size)
valid_rmse = np.sqrt(valid_sse)
print( '--Avg. Train Loss ='+str(epoch_loss)[:6] + ' --Avg. Valid Loss ='+str(valid_loss)[:6]+ ' --Valid RMSE = '+str(valid_rmse)[:6]+'\n' )
logfile.write( '--Avg. Train Loss ='+str(epoch_loss)[:6] + ' --Avg. Valid Loss ='+str(valid_loss)[:6]+ ' --Valid RMSE = '+str(valid_rmse)[:6]+'\n' )
logfile.flush()
cur_best_pre_0, stopping_step, should_stop = early_stopping(valid_rmse, cur_best_pre_0,
stopping_step, expected_order='dec',
flag_step=args.flag_step)
if should_stop == True:
break
#saver.save(sess, os.path.join(args.log_path,'model'), global_step=i, write_meta_graph=False)
if valid_rmse == cur_best_pre_0 and args.save_flag:
saver.save(sess, os.path.join(args.log_path,'model'), global_step=i+1, write_meta_graph=False)
output_prediction = None
for j in range( math.ceil( len(rcstrs_output) / args.batch_size) ):
predict = model_output.step(sess, isTesting=True, dropout_keep_prob=1)
if j == 0:
output_prediction = predict
else:
output_prediction = np.concatenate([output_prediction, predict], axis=0)
output_prediction = np.reshape(output_prediction, (output_prediction.shape[0],))
df = pd.DataFrame( {'Id': rcstrs_output,'Prediction': output_prediction} )
df.to_csv(os.path.join(args.log_path, 'output' + str(i+1)+".csv" ),index=False)
#copyfile( os.path.join(args.log_path, 'output' + str(i+1)+".csv" ) , "../" )
output_valid_prediction = None
for j in range( math.ceil( len(rcstrs_output_valid) / args.batch_size) ):
predict = model_output_valid.step(sess, isTesting=True, dropout_keep_prob=1)
if j == 0:
output_valid_prediction = predict
else:
output_valid_prediction = np.concatenate( [output_valid_prediction, predict] , axis=0 )
output_valid_prediction = np.reshape(output_valid_prediction, (output_valid_prediction.shape[0],))
df = pd.DataFrame( {'Id': rcstrs_output_valid,'Prediction': output_valid_prediction} )
df.to_csv(os.path.join(args.log_path, 'output_valid' + str(i+1)+".csv" ),index=False)
copyfile( os.path.join(args.log_path, 'output_valid' + str(i+1)+".csv" ) ,
os.path.join("../cache", args.log_path.split("/")[-2] ) )
test_prediction = None
for j in range(math.ceil(row_col_prediction.shape[0] / args.batch_size)):
predict = model_test.step(sess, isTesting=True, dropout_keep_prob=1)
if j == 0:
test_prediction = predict
else:
test_prediction = np.concatenate([test_prediction, predict], axis=0)
test_prediction = np.reshape(test_prediction, (test_prediction.shape[0],))
# data frame is reconstructed since the direct modification is too slow
df = pd.DataFrame({'Id': rcstrs,'Prediction': test_prediction})
df.to_csv(os.path.join(args.log_path, 'submission' + str(i+1)+".csv" ),index=False)
copyfile( os.path.join(args.log_path, 'submission' + str(i+1)+".csv" ) ,
os.path.join("../test", args.log_path.split("/")[-2] ) )
sess.run([iterator_test.initializer,
iterator_output.initializer,
iterator_output_valid.initializer])
dataloader_test = iterator_test.get_next()
dataloader_output = iterator_output.get_next()
dataloader_output_valid = iterator_output_valid.get_next()
row_col_test, label_test = dataloader_test
row_col_output, label_output = dataloader_output
row_col_output_valid, label_output_valid = dataloader_output_valid
BATCH_SIZE = 1 # predict one by one, allows to use drop_remainder=true which gives back a batched set with a know batch dimension (used by tf.unstack() in HG model)
# Path
list_of_files = glob.glob(os.path.join(CURR_DIR, "log_HG2D", "*"))
LOG_PATH = max(list_of_files, key=os.path.getctime) # latest created dir for latest experiment
CHECKPOINTS_PATH =os.path.join(LOG_PATH, "checkpoints") # we restore the lastest saved model from the latest experiment
CLUSTER_PATH = "/cluster/project/infk/hilliges/lectures/mp19/project2/"
LOCAL_PATH = "."
if os.path.exists(CLUSTER_PATH):
DATA_PATH = CLUSTER_PATH
else:
DATA_PATH = LOCAL_PATH
with tf.Session() as sess:
# load images
im = create_dataloader_test(data_root=DATA_PATH)
# define model
model = StackedHourglass(nb_stacks=NB_STACKS, sigma=SIGMA)
# build the model
all_heatmaps_pred, p2d_pred = model(im, training=False)
# restore weights
saver = tf.train.Saver()
saver.restore(sess,tf.train.latest_checkpoint(CHECKPOINTS_PATH))
predictions = None
with trange(math.ceil(NUM_SAMPLES)) as t: # generate predictions for all images
for i in t:
image, p2d_out_value = sess.run([im, p2d_pred])