def main(args):
# set up the model. different environments need different params
tt_params = {
"BipedalWalker-v2": (24, 4, 2, 10),
"CarRacing-v0": (16**2, 3, 4, 3)
}
tt = AlphaSegmenter(*tt_params[args.env])
# reference trajectory used to plot latents over time
ref_traj = next(get_datasets(folder=f"data/{args.env}")).tensors
optimizer = optim.Adam(tt.parameters(), lr=1e-3)
losses = []
alphas = []
try:
for epoch_idx in range(3):
for i, d in enumerate(get_datasets(folder=f"data/{args.env}")):
new_losses, new_alphas = train(tt,
d,
optimizer,
ref_traj=ref_traj)
losses += new_losses
alphas += new_alphas
print(f"Epoch {epoch_idx}:\t{losses[-1]}")
except KeyboardInterrupt:
print("Stopping Early!")
fig, (ax1, ax2) = plt.subplots(nrows=2, sharex=True)
ax1.plot(np.array(losses),
label=["loss", "prior", "recon", "dynamics", "reinforce"])
ax1.legend()
ax2.imshow(np.array(alphas).T)
plt.show()
def run(run_name, config_update):
print(f'\nstart: {run_name}')
pwd = os.path.dirname(os.path.abspath(__file__))
with open(f'{pwd}/config.yml', 'r') as yml:
config = yaml.safe_load(yml)
# init_exp
dir_save, config = util.init_exp(config, config_update, run_name)
model_params = config['model']['params']
split_params = config['split']
# datast
df, target = util.get_datasets()
X_train, X_valid, y_train, y_valid = train_test_split(df,
target,
test_size=0.3,
**split_params)
# fit
print(model_params)
model = DecisionTreeClassifier(**model_params)
model.fit(X_train, y_train)
# pred
X_valid_pred = model.predict(X_valid)
# eval
acc_valid = accuracy_score(y_valid, X_valid_pred)
print(f'acc_valid: {acc_valid:.5f}')
def map_eval(eval_file, token_length, tokenizer, device, model, label_list):
model.eval()
datasets, labels = get_datasets(eval_file)
total_batches = 0
total_avp = 0.0
total_mrr = 0.0
# scores, labels = [], []
for k, dataset in tqdm(datasets.items(), desc="Eval datasets"):
examples = []
for i, data in enumerate(dataset):
examples.append(InputExample(i, data[0], data[1], '0'))
eval_features = convert_examples_to_features(examples, label_list,
token_length, tokenizer)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long).to(device)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long).to(device)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long).to(device)
# all_label_ids = torch.tensor(
# [f.label_id for f in eval_features], dtype=torch.long).to(device)
x_input_ids = torch.tensor([f.input_ids_x for f in eval_features],
dtype=torch.long).to(device)
x_input_mask = torch.tensor([f.input_mask_x for f in eval_features],
dtype=torch.long).to(device)
x_segment_ids = torch.tensor([f.segment_ids_x for f in eval_features],
dtype=torch.long).to(device)
y_input_ids = torch.tensor([f.input_ids_y for f in eval_features],
dtype=torch.long).to(device)
y_input_mask = torch.tensor([f.input_mask_y for f in eval_features],
dtype=torch.long).to(device)
y_segment_ids = torch.tensor([f.segment_ids_y for f in eval_features],
dtype=torch.long).to(device)
with torch.no_grad():
logits = model(x_input_ids, x_input_mask, x_segment_ids,
y_input_ids, y_input_mask, y_segment_ids,
all_input_ids, all_segment_ids, all_input_mask)
score = F.softmax(logits, dim=1)[:, 1].cpu().numpy()
label = np.array(list(map(int, labels[k])))
# print(score, label)
# scores.append(score)
# labels.append(label)
total_avp += mean_average_precision(label, score)
total_mrr += mean_reciprocal_rank(label, score)
total_batches += 1
mAP = total_avp / total_batches
mRR = total_mrr / total_batches
logger.info("map is : {}, mrr is : {}".format(mAP, mRR))
data = {'map': mAP, 'mrr': mRR}
with open('./result.json', 'w', encoding='utf-8') as f:
json.dump(data, f)
def main():
#ARGUMENT PARSING:
known_metrics = ['cr64', 'cg64', 'cb64','ccol',\
'dd2', 'da3', 'dd3', 'dd2_64', 'da3_64', 'dd3_64',\
'dd2_128', 'dd3_128', 'da3_128']
used_metrics = []
parser = argparse.ArgumentParser(prog='Histalyzer')
for m in known_metrics:
parser.add_argument('--%s' % m, type=int)
parser.add_argument('--nn', type=int, default=5)
parser.add_argument('--wcolor', type=int, default=1)
parser.add_argument('--wdepth', type=int, default=1)
parser.add_argument('--quiet', action='store_true')
parser.add_argument('--debug', action='store_true')
parser.add_argument('--no_files', action='store_true')
parser.add_argument('--scores', action='store_true')
parser.add_argument('categories', nargs='+')
parser.add_argument('--every', type=int, default=25)
parsed = parser.parse_args()
SET_NOFILES = parsed.no_files
SET_QUIET = parsed.quiet
SET_DEBUG = parsed.debug
SET_EVERY = parsed.every
SET_SCORES = parsed.scores
if SET_QUIET:
logging.basicConfig(level=logging.WARN)
elif SET_DEBUG:
logging.basicConfig(level=logging.DEBUG)
else:
logging.basicConfig(level=logging.INFO)
categories, SET_PARTIAL = parse_categories(parsed.categories)
weight_color = parsed.wcolor
weights_color = []
weight_depth = parsed.wdepth
weights_depth = []
neighbors = parsed.nn
for k, v in vars(parsed).iteritems():
if not v or not k in known_metrics:
continue
if k[0] == 'c':
weights_color.append((k, v))
used_metrics.append(k)
elif k[0] == 'd':
weights_depth.append((k, v))
used_metrics.append(k)
else:
raise ValueError('unknown metric type: "%s"' % k[0])
if not weights_color:
weight_color = 0
if not weights_depth:
weight_depth = 0
weight_dict = {}
weight_dict['color'] = ( weight_color , weights_color )
weight_dict['depth'] = ( weight_depth , weights_depth )
#PARAMETER SETUP:
#frameset = defs.EVERY_5TH
#frameset = defs.EVERY_25TH
frameset = defs.get_frameset(SET_EVERY)
all_individuals = parse_data(used_metrics)
if not SET_NOFILES:
dir_top, dir_raw = create_directory_structure(frameset, used_metrics, neighbors,
partial=SET_PARTIAL, weights_color=weights_color, weights_depth=weights_depth)
#OVERALL VARIABLES:
overall_tested = 0
overall_correct = 0
classifier = KNNClassifier(weight_dict, neighbors, defs.ALL_CATEGORIES)
for category in categories:
#CATEGORY VARIABLES:
category_tested = 0
category_correct = 0
for instance in all_individuals[category]:
traindata, testdata = util.get_datasets(category, instance, all_individuals, frameset[0])
#result, instance_tested, instance_correct = nn.nearest_neighbor(traindata, testdata, neighbors)
result, instance_tested, instance_correct = classifier.perform_classification(traindata, testdata)
category_tested += instance_tested
overall_tested += instance_tested
category_correct += instance_correct
overall_correct += instance_correct
if not SET_NOFILES:
f = open("%s/category_%s.csv" % (dir_raw, category), "a")
f.write('%s %s,%s\n' % (category, instance, result))
f.close()
average_aggregated = float(category_correct) / category_tested * 100
if not SET_NOFILES:
f = open("%s/category_%s.csv"% (dir_raw, category), "a")
f.write('%s average,%s\n' % (category, average_aggregated))
f.close()
if not SET_NOFILES:
classifier.print_confusion_matrix(dir_top + '/confusion.csv')
if SET_SCORES:
scrs = classifier.get_overall_scores()
print "{},{},{}".format(scrs[0], scrs[1], scrs[2])
overall_percentage = float(overall_correct)/overall_tested * 100
logging.info("Overall %% %f", overall_percentage)
pass
tf.enable_v2_behavior()
import tensorflow_hub as hub
# from tensorflow import keras
import numpy as np
import pathlib
from util import get_datasets, calculate_accuracy
from processing import preprocess, postprocess
from tflite_helper import TFLiteConvertor
##################
# Prepare datasets
##################
cv_dataset, test_dataset = get_datasets()
num_images_in_cv_dataset = 100
# For 224 x 224
preprocess_cv_dataset_224 = list()
for i in range(0, num_images_in_cv_dataset):
preprocess_cv_dataset_224.append(preprocess(cv_dataset[i], 224, 224))
# For 299 x 299
preprocess_cv_dataset_299 = list()
for i in range(0, num_images_in_cv_dataset):
preprocess_cv_dataset_299.append(preprocess(cv_dataset[i], 299, 299))
def representative_data_gen_224():
for input_image in preprocess_cv_dataset_224:
# then z_t should be equal to c(s_t)
s_target = s[l+1]
s_pred = self.c[l](s[l])
loss += (torch.pow(s_target - s_pred, 2) * b[l]).sum()
# enforce hierarchy. this says that if a transition occurs at
# the level above, then one should occur at this level too
loss += (b[l+1] * (1. - b[l])).sum()
return loss
def decode(self, z):
T = z.shape[1]
out = self.out_net(z.view(-1, self.h_dim)).view(self.n_levels, T, -1)
return torch.split(out, [self.z_out, self.g_out, 1], dim=2)
if __name__ == '__main__':
fc = FCSegmenter(24, 4, 10, 6)
for i, dataset in enumerate(get_datasets()):
for j, (s, a) in enumerate(DataLoader(dataset, batch_size=50, shuffle=False)):
with torch.no_grad():
z = fc.encode(s, a)
s, a, b = fc.decode(z)
fc.likelihood(s, a, b)
if j > 2:
sys.exit(0)
def go(model_index):
global send_whatsapp
global split_num
global k_fold_number
global include_dependency_tree_feature
global include_pos_feature
global spam_max_length
global non_spam_max_length
global block_all_models_if_k_big
global use_morphological
global use_both
if not use_morphological:
include_dependency_tree_feature = False
include_pos_feature = False
if k_fold_number < 1:
print("K is invalid, must be 1 or bigger.")
return
if block_all_models_if_k_big and model_index == 0 and k_fold_number > 1:
print(
"Can't train for all models at once with k_fold_num > 1 , because of memory problems."
)
print(
"Please change k_fold_number to be 1, or choose a single model (1-5)"
)
return
start_time = time.time()
print()
print("==========")
print()
datasets = get_datasets(spam_max_length, non_spam_max_length, use_both,
use_morphological)
print("Starting cross-validation, with all the datasets:")
"""
for x1 in datasets:
print (x1)
for x2 in datasets[x1]:
print(x2)
for x3 in datasets[x1][x2]:
print(x3, datasets[x1][x2][x3])
"""
for dataset in datasets.keys():
print("Testing model %d (%s) for dataset '%s':" %
(model_index, model.model_desc(model_index), dataset))
test_driver(
model.model_desc(model_index),
model.Model,
datasets[dataset],
model_index,
split=split_num,
k_fold_num=k_fold_number,
include_dependency_tree_feature=include_dependency_tree_feature,
include_pos_feature=include_pos_feature,
dataset_name=dataset)
print()
print("@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@")
print("@@@@@@ Finished dataset %s @@@@@" % dataset)
print("@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@")
print()
end_time = time.time()
elapsed = end_time - start_time
print()
print("All time: %d seconds." % elapsed)
print()
print("FINISHED")
if send_whatsapp:
# get rid of none, to send message to whatsapp
if (spam_max_length is None):
spam_max_length = "None"
if (non_spam_max_length is None):
non_spam_max_length = "None"
# send message to whatsapp
msg = """הסתיימה הרצה בשרת.
k = %d
datasets = %s
model index = %d
model name = %s
dependency_trees_feature = %r
pos_feature = %r
spam_max_length = %s
non_spam_max_length = %s
Total time: %d seconds""" % (k_fold_number, str(list(
datasets.keys())), model_index, model.model_desc(model_index),
include_dependency_tree_feature,
include_pos_feature, str(spam_max_length),
str(non_spam_max_length), elapsed)
whatsapp.send_msg(msg)