def add_job(
job_name,
sub_index,
backbone_arch,
init_model_nickname,
init_model_path,
extra_params=None,
):
job_name = f"{config_job_name}.{sub_index}.{job_name}_seed{config['random_seed']}"
d = OrderedDict()
d["--config-file"] = config_file
d["model.backbone_arch"] = backbone_arch
d["init.model"] = init_model_path
log_folder = job_name + "_" + backbone_arch + "_init_" + init_model_nickname
log_folder = os.path.join(log_path, log_folder)
d["output.path"] = log_folder
if extra_params:
d.update(extra_params)
commands = [main_command + " " + launcher.parameters_to_str(d)]
exp_job_names.append(job_name)
exp_commands.append(commands)
exp_log_paths.append(log_folder)
def add_job(sub_index,
backbone_arch,
model_path,
model_checkpoint,
episodes,
eval_scale,
test_augmentation,
folder_suffix="",
extra_params=None):
job_name = f"{config_job_name}.{sub_index}"
commands = []
d = OrderedDict()
d["--config-file"] = config_file
d["model.use_inverse_geom_model"] = True
d["model.use_simplified_affine_model"] = False
d["model.backbone_arch"] = backbone_arch
d["eval.dataset_scales"] = f"[{eval_scale}]"
if test_augmentation:
d["eval.class_image_augmentation"] = test_augmentation
if extra_params:
d.update(extra_params)
# set output folder
log_folder = f"{config_job_name}"
if folder_suffix:
log_folder += "." + folder_suffix
log_folder = os.path.join(log_path, log_folder)
d["train.do_training"] = False
# choose init
if "init.transform" in d:
del d["init.transform"]
if os.path.isfile(model_path):
d["init.model"] = model_path
else:
d["init.model"] = os.path.join(model_path, model_checkpoint)
for episode in episodes:
d["eval.dataset_names"] = f"[\\\"imagenet-repmet-test-episode-{episode}\\\"]"
commands.append(main_command + " " + launcher.parameters_to_str(d))
exp_job_names.append(job_name)
exp_commands.append(commands)
exp_log_paths.append(log_folder)
exp_log_file_prefix.append(
f"eval_scale{d['eval.dataset_scales'][1:-1]}_epi{min(episodes)}-{max(episodes)}_"
)
def add_job(
job_name,
sub_index,
arch,
init_net,
dataset_train,
dataset_val,
extra_params=None,
):
job_name = f"{config_job_name}.{sub_index}.{job_name}"
log_folder = job_name
log_folder = os.path.join(log_path, log_folder)
commands = []
# stage 1
d = OrderedDict()
d["--cuda"] = ""
d["--dataset"] = dataset_train
d["--dataset_val"] = dataset_val
d["--init_weights"] = init_net
d["--disp_interval"] = "1"
d["--val_interval"] = "1"
d["--nw"] = "4"
d["--bs"] = "4"
d["--s"] = 1
d["--epochs"] = "100" # set 20x less epochs for instre comapared to grozi as ezch epoch is 20x bigger
d["--lr_decay_milestones"] = "50 75"
d["--lr"] = 0.01 # default starting learning rate
d["--lr_decay_gamma"] = 0.1
d["--lr_reload_best_after_decay"] = "True"
d["--save_dir"] = log_folder
d["--net"] = arch
d["--class_image_augmentation"] = "rotation90"
d["--set"] = ""
d["DATA_DIR"] = data_path
# put smth here, but those are not used in CoAE
d["TRAIN.MAX_SIZE"] = "3000"
d["TEST.MAX_SIZE"] = "3000"
if extra_params:
d.update(extra_params)
commands += [main_command + " " + launcher.parameters_to_str(d)]
exp_job_names.append(job_name)
exp_commands.append(commands)
exp_log_paths.append(log_folder)
def add_job(
arch,
eval_dataset,
model_path,
model_checkpoint,
folder_suffix="",
test_augment=None,
extra_params=None,
):
job_name = f"{config_job_name}.{eval_dataset}.{arch}"
# set output folder
log_folder = f"{config_job_name}"
if folder_suffix:
log_folder += "." + folder_suffix
log_folder = os.path.join(log_path, log_folder)
commands = []
# stage 1
d = OrderedDict()
d["--cuda"] = ""
if os.path.isfile(model_path):
d["--weights"] = os.path.join(log_folder, model_path)
else:
d["--weights"] = os.path.join(model_path, model_checkpoint)
d["--dataset"] = eval_dataset
d["--net"] = arch
if test_augment is not None:
d["--class_image_augmentation"] = test_augment
d["--set"] = ""
d["TRAIN.USE_FLIPPED"] = "False"
d["DATA_DIR"] = data_path
# put smth here, but those are not used in CoAE
d["TRAIN.MAX_SIZE"] = "5000"
d["TEST.MAX_SIZE"] = "5000"
if extra_params:
d.update(extra_params)
commands += [main_command + " " + launcher.parameters_to_str(d)]
exp_job_names.append(job_name)
exp_commands.append(commands)
exp_log_paths.append(log_folder)
exp_log_file_prefix.append(f"eval_{eval_dataset}_{arch}_")
def add_job(
job_type, # "v1" or "v2"
sub_index,
backbone_arch,
init_model_nickname,
init_model_path,
extra_params=None,
):
job_name = f"{config_job_name}.{sub_index}.{job_type}_seed{config['random_seed']}"
d = OrderedDict()
d["--config-file"] = config_file
if job_type == "v1":
d.update(config_dict_v1)
elif job_type == "v2":
d.update(config_dict_v2)
else:
raise RuntimeError("Unknown job_type {0}".format(job_type))
d["model.backbone_arch"] = backbone_arch
d["init.model"] = init_model_path
log_folder = job_name + "_" + backbone_arch + "_init_" + init_model_nickname
log_folder = os.path.join(log_path, log_folder)
d["output.path"] = log_folder
if extra_params:
d.update(extra_params)
commands = [main_command + " " + launcher.parameters_to_str(d)]
exp_job_names.append(job_name)
exp_commands.append(commands)
exp_log_paths.append(log_folder)
def add_job(
job_type, # "v1" or "v2"
job_id,
backbone_arch,
init_model_nickname,
init_model_path,
extra_params=None,
train_data="",
train_data_scale=None):
job_name = "{0}.{1}.{2}_{3}_seed{4}".format(config_job_name, job_id,
job_type, train_data,
config["random_seed"])
d = OrderedDict()
d["--config-file"] = config_file
if job_type == "v1":
d.update(config_dict_v1)
elif job_type == "v2":
d.update(config_dict_v2)
else:
raise RuntimeError("Unknown job_type {0}".format(job_type))
d["train.dataset_name"] = "\"" + train_data + "\""
if train_data == "instre-s1-train":
d["train.dataset_scale"] = 700.0
main_val_dataset = "instre-s1-val"
d["eval.dataset_scales"] = "[700.0]"
elif train_data == "instre-s2-train":
d["train.dataset_scale"] = 600.0
main_val_dataset = "instre-s2-val"
d["eval.dataset_scales"] = "[600.0]"
else:
raise RuntimeError(f"Unknown dataset {train_data}")
d["output.best_model.dataset"] = main_val_dataset
d["eval.dataset_names"] = f"[\\\"{main_val_dataset}\\\"]"
d["eval.class_image_augmentation"] = "rotation90"
d["eval.iter"] = 5000
# extra augmentation for this run
d["train.augment.mine_extra_class_images"] = True
d["model.backbone_arch"] = backbone_arch
d["init.model"] = init_model_path
log_folder = job_name + "_" + backbone_arch + "_init_" + init_model_nickname
log_folder = os.path.join(log_path, log_folder)
d["output.path"] = log_folder
if extra_params:
d.update(extra_params)
commands = []
commands.append(main_command + " " + launcher.parameters_to_str(d))
exp_job_names.append(job_name)
exp_commands.append(commands)
exp_log_paths.append(log_folder)
def add_job(
sub_index,
job_type, # "v1" or "v2"
backbone_arch,
eval_dataset,
model_path,
model_checkpoint,
folder_suffix="",
extra_params=None):
job_name = f"{config_job_name}.{sub_index}.{eval_dataset}"
d = OrderedDict()
d["--config-file"] = config_file
if job_type == "v1":
d.update(config_dict_v1)
elif job_type == "v2":
d.update(config_dict_v2)
else:
raise RuntimeError("Unknown job_type {0}".format(job_type))
d["model.backbone_arch"] = backbone_arch
if extra_params:
d.update(extra_params)
# set output folder
log_folder = f"{config_job_name}.{sub_index}"
if folder_suffix:
log_folder += "." + folder_suffix
log_folder = os.path.join(log_path, log_folder)
# evaluation params
d["train.do_training"] = False
d["eval.mAP_iou_thresholds"] = "\"[0.5]\""
d["eval.train_subset_for_eval_size"] = 0
# choose init
if "init.transform" in d:
del d["init.transform"]
if os.path.isfile(model_path):
d["init.model"] = model_path
else:
d["init.model"] = os.path.join(model_path, model_checkpoint)
# choose eval dataset
if eval_dataset == "grozi-val-new-cl":
d["eval.dataset_names"] = "\"[\\\"grozi-val-new-cl\\\"]\""
d["eval.dataset_scales"] = "[1280.0]"
elif eval_dataset == "grozi-val-old-cl":
d["eval.dataset_names"] = "\"[\\\"grozi-val-old-cl\\\"]\""
d["eval.dataset_scales"] = "[1280.0]"
elif eval_dataset == "dairy":
d["eval.dataset_names"] = "\"[\\\"dairy\\\"]\""
d["eval.dataset_scales"] = "[3500.0]"
elif eval_dataset == "paste-v":
d["eval.dataset_names"] = "\"[\\\"paste-v\\\"]\""
d["eval.dataset_scales"] = "[3500.0]"
elif eval_dataset == "paste-f":
d["eval.dataset_names"] = "\"[\\\"paste-f\\\"]\""
d["eval.dataset_scales"] = "[2000.0]"
# eval with rotations
d["eval.class_image_augmentation"] = "rotation90"
else:
raise f"Unknown eval set {eval_dataset}"
commands = [main_command + " " + launcher.parameters_to_str(d)]
exp_job_names.append(job_name)
exp_commands.append(commands)
exp_log_paths.append(log_folder)
exp_log_file_prefix.append(
f"eval_{eval_dataset}_scale{d['eval.dataset_scales'][1:-1]}_")
def add_job(
training_dataset,
arch, # "resnet50" or "resnet101"
pool="gem", # "mac", "spoc", "gem"
test_whiten=False, # True or False
local_whitening=False, # True or False
regional=False, # True or False
whitening=False, # True or False
loss_margin=0.85, # True or False
image_size=240, # 240, 360, 500, 750, 1024
learning_rate=1e-6, # 1e-6, 5e-7, 5e-6
pretrained=True,
loss="contrastive",
optimizer="adam",
weight_decay=1e-4,
neg_num=5,
query_size=2000,
pool_size=20000,
batch_size=5,
eval_dataset="",
use_best_model=True,
retrieval_multiscale="ss",
class_image_augmentation="",
folder_suffix=""):
directory = "{}".format(training_dataset)
directory += "_{}".format(arch)
directory += "_{}".format(pool)
if local_whitening:
directory += "_lwhiten"
if regional:
directory += "_r"
if whitening:
directory += "_whiten"
if not pretrained:
directory += "_notpretrained"
directory += "_{}_m{:.2f}".format(loss, loss_margin)
directory += "_{}_lr{:.1e}_wd{:.1e}".format(optimizer, learning_rate,
weight_decay)
directory += "_nnum{}_qsize{}_psize{}".format(neg_num, query_size,
pool_size)
directory += "_bsize{}_imsize{}".format(batch_size, image_size)
job_name = f"{config_job_name}.{eval_dataset}"
# set output folder
log_path = config_job_name
if folder_suffix:
log_path += "." + folder_suffix
log_path = os.path.join(log_folder_path, log_path)
commands = []
# add testing on top of the detector
d_testing = OrderedDict()
if retrieval_multiscale == "ms":
d_testing["--retrieval_multiscale"] = ""
if arch == "resnet50":
d_testing["--maskrcnn_config_file"] = os.path.join(
detector_path, "config",
"e2e_faster_rcnn_R_50_FPN_1x_multiscale_noClasses.yaml")
elif arch == "resnet101":
d_testing["--maskrcnn_config_file"] = os.path.join(
detector_path, "config",
"e2e_faster_rcnn_R_101_FPN_1x_multiscale_noClasses.yaml")
else:
raise RuntimeError(f"Unknown arch {arch}")
if training_dataset.startswith("instre-s1"):
if arch == "resnet50":
d_testing["--maskrcnn_weight_file"] = os.path.join(
detector_path, "output",
"exp0001-R-50-noCl-instre-s1/model_best.pth")
elif arch == "resnet101":
d_testing["--maskrcnn_weight_file"] = os.path.join(
detector_path, "output",
"exp0004-R-101-noCl-instre-s1/model_best.pth")
else:
raise RuntimeError(f"Unknown arch {arch}")
elif training_dataset.startswith("instre-s2"):
if arch == "resnet50":
d_testing["--maskrcnn_weight_file"] = os.path.join(
detector_path, "output",
"exp0002-R-50-noCl-instre-s2/model_best.pth")
elif arch == "resnet101":
d_testing["--maskrcnn_weight_file"] = os.path.join(
detector_path, "output",
"exp0005-R-101-noCl-instre-s2/model_best.pth")
else:
raise RuntimeError(f"Unknown arch {arch}")
else:
raise f"Unknown training set {training_dataset}"
if use_best_model:
d_testing["--retrieval_network_path"] = os.path.join(
retrieval_path, "output", "instre", directory,
"model_best.pth.tar")
else:
d_testing["--retrieval_network_path"] = os.path.join(
retrieval_path, "output", "instre", directory,
"model_epoch0.pth.tar")
d_testing["--retrieval_image_size"] = image_size
d_testing["is_cuda"] = "True"
if "instre-s1" in eval_dataset:
d_testing["eval.dataset_names"] = f"\"[\\\"{eval_dataset}\\\"]\""
d_testing["eval.dataset_scales"] = "\"[700]\""
elif "instre-s2" in eval_dataset:
d_testing["eval.dataset_names"] = f"\"[\\\"{eval_dataset}\\\"]\""
d_testing["eval.dataset_scales"] = "\"[600]\""
else:
raise RuntimeError(f"Unknown eval dataset {eval_dataset}")
d_testing["eval.mAP_iou_thresholds"] = "\"[0.5]\""
if class_image_augmentation:
d_testing[
"eval.class_image_augmentation"] = class_image_augmentation
commands.append(command_testing + " " +
launcher.parameters_to_str(d_testing))
exp_job_names.append(job_name)
exp_log_paths.append(log_path)
exp_commands.append(commands)
exp_log_file_prefix.append(
f"eval_{eval_dataset}_{'bestModel' if use_best_model else 'initModel'}_{retrieval_multiscale}_"
)
def add_job(sub_index,
job_type, # "v1" or "v2"
backbone_arch,
eval_dataset,
model_path,
model_checkpoint,
folder_suffix="",
extra_params=None):
job_name = f"{config_job_name}.{sub_index}.{eval_dataset}"
commands = []
d = OrderedDict()
d["--config-file"] = config_file
if job_type == "v1":
d.update(config_dict_v1)
elif job_type == "v2":
d.update(config_dict_v2)
else:
raise RuntimeError("Unknown job_type {0}".format(job_type))
train_data = eval_dataset + "-train"
d["train.dataset_name"] = "\"" + train_data + "\""
if train_data == "instre-s1-train":
d["train.dataset_scale"] = 700.0
main_val_dataset = "instre-s1-val"
d["eval.dataset_scales"] = "[700.0]"
elif train_data == "instre-s2-train":
d["train.dataset_scale"] = 600.0
main_val_dataset = "instre-s2-val"
d["eval.dataset_scales"] = "[600.0]"
else:
raise RuntimeError(f"Unknown dataset {train_data}")
d["output.best_model.dataset"] = main_val_dataset
d["eval.dataset_names"] = f"[\\\"{main_val_dataset}\\\"]"
d["eval.class_image_augmentation"] = "rotation90"
d["eval.iter"] = 5000
# extra augmentation for this run
d["train.augment.mine_extra_class_images"] = True
d["model.backbone_arch"] = backbone_arch
if extra_params:
d.update(extra_params)
# set output folder
log_folder = f"{config_job_name}"
if folder_suffix:
log_folder += "." + folder_suffix
log_folder = os.path.join(log_path, log_folder)
d["train.do_training"] = False
if train_data == "instre-s1-train":
d["eval.dataset_names"] = "[\\\"instre-s1-test\\\"]"
d["eval.dataset_scales"] = "[700.0]"
elif train_data == "instre-s2-train":
d["eval.dataset_names"] = "[\\\"instre-s2-test\\\"]"
d["eval.dataset_scales"] = "[600.0]"
else:
raise RuntimeError(f"Unknown dataset {train_data}")
d["eval.class_image_augmentation"] = "rotation90"
# choose init
if "init.transform" in d:
del d["init.transform"]
if os.path.isfile(model_path):
d["init.model"] = model_path
else:
d["init.model"] = os.path.join(model_path, model_checkpoint)
commands.append(main_command + " " + launcher.parameters_to_str(d))
exp_job_names.append(job_name)
exp_commands.append(commands)
exp_log_paths.append(log_folder)
exp_log_file_prefix.append(f"eval_{eval_dataset}_scale{d['eval.dataset_scales'][1:-1]}_")
def add_job(sub_index,
training_dataset,
arch,
pool="gem",
test_whiten=False,
local_whitening=False,
regional=False,
whitening=False,
loss_margin=0.85,
image_size=240,
learning_rate=1e-6,
pretrained=True,
loss="contrastive",
optimizer="adam",
weight_decay=1e-4,
neg_num=5,
query_size=2000,
pool_size=20000,
batch_size=5):
directory = "{}".format(training_dataset)
directory += "_{}".format(arch)
directory += "_{}".format(pool)
if local_whitening:
directory += "_lwhiten"
if regional:
directory += "_r"
if whitening:
directory += "_whiten"
if not pretrained:
directory += "_notpretrained"
directory += "_{}_m{:.2f}".format(loss, loss_margin)
directory += "_{}_lr{:.1e}_wd{:.1e}".format(optimizer, learning_rate,
weight_decay)
directory += "_nnum{}_qsize{}_psize{}".format(neg_num, query_size,
pool_size)
directory += "_bsize{}_imsize{}".format(batch_size, image_size)
log_path = os.path.join(log_folder_path, directory)
job_name = "ret-grozi-{0}.{1}".format(sub_index, directory)
d = OrderedDict()
d["--training-dataset"] = training_dataset
if training_dataset == "grozi-train-retrieval":
d["--test-datasets"] = "grozi-val-new-cl-retrieval,grozi-val-old-cl-retrieval"
elif training_dataset == "grozi-train-retrieval-rndCropPerImage10":
d["--test-datasets"] = "grozi-val-new-cl-retrieval-rndCropPerImage10,grozi-val-new-cl-retrieval-rndCropPerImage10"
else:
raise RuntimeError(f"Unknown training set {training_dataset}")
if test_whiten:
d["--test-whiten"] = training_dataset
d["--arch"] = arch
d["--pool"] = pool
if local_whitening:
d["--local-whitening"] = ""
if regional:
d["--regional"] = ""
if whitening:
d["--whitening"] = ""
d["--loss-margin"] = loss_margin
d["--image-size"] = image_size
d["--learning-rate"] = learning_rate
if not pretrained:
d["--not-pretrained"] = ""
d["--loss"] = loss
d["--optimizer"] = optimizer
d["--weight-decay"] = weight_decay
d["--neg-num"] = neg_num
d["--query-size"] = query_size
d["--pool-size"] = pool_size
d["--batch-size"] = batch_size
commands = []
commands.append(main_command + " " + launcher.parameters_to_str(d))
exp_job_names.append(job_name)
exp_log_paths.append(log_path)
exp_commands.append(commands)
def add_job(
job_id,
model,
dataset,
init_weights,
):
job_name = f"exp-{model}-{dataset}-{init_weights}"
commands = []
d = OrderedDict()
if model == "R-50":
if "pytorch" not in init_weights:
config_file = os.path.join(
config_path,
"e2e_faster_rcnn_R_50_FPN_1x_multiscale_noClasses.yaml")
else:
config_file = os.path.join(
config_path,
"e2e_faster_rcnn_R_50_FPN_1x_multiscale_noClasses_fromPytorch.yaml"
)
elif model == "R-101":
if "pytorch" not in init_weights:
config_file = os.path.join(
config_path,
"e2e_faster_rcnn_R_101_FPN_1x_multiscale_noClasses.yaml")
else:
config_file = os.path.join(
config_path,
"e2e_faster_rcnn_R_101_FPN_1x_multiscale_noClasses_fromPytorch.yaml"
)
else:
raise RuntimeError(f"Do not know config for model {model}")
d["--validation_period"] = 5000
d["--config-file"] = config_file
if model == "R-50":
if init_weights == "imagenet-repmet-pytorch":
d["MODEL.WEIGHT"] = "../../../data/ImageNet-RepMet/pretrain/output/resnet50/model_best_maskrcnnbenchmark.pth.tar"
elif init_weights == "imagenet-pytorch":
d["MODEL.WEIGHT"] = "../../../models/resnet50-19c8e357.pth"
elif init_weights == "imagenet-caffe":
pass
else:
raise RuntimeError(
f"Do not recognize weight initialization {init_weights}")
elif model == "R-101":
if init_weights == "imagenet-repmet-pytorch":
d["MODEL.WEIGHT"] = "../../../data/ImageNet-RepMet/pretrain/output/resnet101/model_best_maskrcnnbenchmark.pth.tar"
elif init_weights == "imagenet-pytorch":
d["MODEL.WEIGHT"] = "../../../models/resnet101-5d3b4d8f.pth"
elif init_weights == "imagenet-caffe":
pass
else:
raise RuntimeError(
f"Do not recognize weight initialization {init_weights}")
else:
raise RuntimeError(f"Do not know config for model {model}")
if dataset == "imagenet-repmet":
d["DATASETS.TRAIN"] = "[\\\"imagenet-repmet-train\\\"]"
d["DATASETS.TEST"] = "[\\\"imagenet-repmet-val-5000\\\"]" # crop val set from 50k images to 5k GT boxes
d["INPUT.MIN_SIZE_TRAIN"] = "[225,280,360,450,540,630,720]"
d["INPUT.MAX_SIZE_TRAIN"] = 2048
d["INPUT.MIN_SIZE_TEST"] = 450
d["INPUT.MAX_SIZE_TEST"] = 1280
else:
raise RuntimeError(f"Unknown dataset {dataset}")
log_folder = os.path.join(log_path, job_name)
d["OUTPUT_DIR"] = log_folder
commands.append(main_command + " " + launcher.parameters_to_str(d))
# testing
d_testing = OrderedDict()
d_testing["--test_weights"] = os.path.join(log_folder,
"model_best.pth")
d_testing.update(d)
datasets_test = ["[\\\"imagenet-repmet-val-5000\\\"]"]
scales_test = ["[180,225,288,360,432,504,576]"]
for dataset, scales in zip(datasets_test, scales_test):
d_testing_local = copy.deepcopy(d_testing)
d_testing_local["DATASETS.TEST"] = dataset
d_testing_local["TEST.BBOX_AUG.ENABLED"] = True
d_testing_local["TEST.BBOX_AUG.SCALES"] = scales
commands.append(main_command + " " +
launcher.parameters_to_str(d_testing_local))
exp_job_names.append(job_name)
exp_commands.append(commands)
exp_log_paths.append(log_folder)
def add_job(
sub_index,
training_dataset,
arch,
init_weights="",
pool="gem",
test_whiten=False,
local_whitening=False,
regional=False,
whitening=False,
loss_margin=0.85,
image_size=240,
learning_rate=1e-6,
pretrained=True,
loss="contrastive",
optimizer="adam",
weight_decay=1e-4,
neg_num=5,
query_size=2000,
pool_size=20000,
batch_size=5,
):
directory = "{}".format(training_dataset)
directory += "_{}".format(arch)
directory += "_{}".format(pool)
if local_whitening:
directory += "_lwhiten"
if regional:
directory += "_r"
if whitening:
directory += "_whiten"
if not pretrained:
directory += "_notpretrained"
directory += "_{}_m{:.2f}".format(loss, loss_margin)
directory += "_{}_lr{:.1e}_wd{:.1e}".format(optimizer, learning_rate,
weight_decay)
directory += "_nnum{}_qsize{}_psize{}".format(neg_num, query_size,
pool_size)
directory += "_bsize{}_imsize{}".format(batch_size, image_size)
log_path = os.path.join(log_folder_path, init_weights, directory)
job_name = "ret-imagenet-{0}.{1}.{2}".format(sub_index, init_weights,
directory)
d = OrderedDict()
d[os.path.join(log_folder_path, init_weights)] = ""
d["--training-dataset"] = training_dataset
if training_dataset == "instre-s1-train-retrieval":
d["--test-datasets"] = "instre-s1-val-retrieval"
elif training_dataset == "instre-s2-train-retrieval":
d["--test-datasets"] = "instre-s2-val-retrieval"
elif training_dataset == "instre-s1-train-retrieval-rndCropPerImage10":
d["--test-datasets"] = "instre-s1-val-retrieval-rndCropPerImage10"
elif training_dataset == "instre-s2-train-retrieval-rndCropPerImage10":
d["--test-datasets"] = "instre-s2-val-retrieval-rndCropPerImage10"
elif training_dataset == "imagenet-repmet-train-retrieval":
d["--test-datasets"] = "imagenet-repmet-val-5000-retrieval"
elif training_dataset == "imagenet-repmet-train-retrieval-rndCropPerImage10":
d["--test-datasets"] = "imagenet-repmet-val-5000-retrieval-rndCropPerImage10"
else:
raise RuntimeError(f"Unknown training set {training_dataset}")
if test_whiten:
d["--test-whiten"] = training_dataset
if arch == "resnet50":
if init_weights == "imagenet-repmet-pytorch":
d["--network-path"] = "../../../data/ImageNet-RepMet/pretrain/output/resnet50/model_best_cirtorch.pth.tar"
elif init_weights == "imagenet-pytorch":
d["--network-path"] = "../../../models/resnet50-19c8e357_cirtorch.pth"
elif init_weights == "imagenet-caffe":
# use the built-in caffe weights
pass
else:
raise RuntimeError(
f"Do not recognize weight initialization {init_weights}")
elif arch == "resnet101":
if init_weights == "imagenet-repmet-pytorch":
d["--network-path"] = "../../../data/ImageNet-RepMet/pretrain/output/resnet101/model_best_cirtorch.pth.tar"
elif init_weights == "imagenet-pytorch":
d["--network-path"] = "../../../models/resnet101-5d3b4d8f_cirtorch.pth"
elif init_weights == "imagenet-caffe":
# use the built-in caffe weights
pass
else:
raise RuntimeError(
f"Do not recognize weight initialization {init_weights}")
else:
raise RuntimeError(f"Unknown arch: {arch}")
d["--arch"] = arch
d["--pool"] = pool
if local_whitening:
d["--local-whitening"] = ""
if regional:
d["--regional"] = ""
if whitening:
d["--whitening"] = ""
d["--loss-margin"] = loss_margin
d["--image-size"] = image_size
d["--learning-rate"] = learning_rate
if not pretrained:
d["--not-pretrained"] = ""
d["--loss"] = loss
d["--optimizer"] = optimizer
d["--weight-decay"] = weight_decay
d["--neg-num"] = neg_num
d["--query-size"] = query_size
d["--pool-size"] = pool_size
d["--batch-size"] = batch_size
commands = []
commands.append(main_command + " " + launcher.parameters_to_str(d))
exp_job_names.append(job_name)
exp_log_paths.append(log_path)
exp_commands.append(commands)
def add_job(
job_id,
model,
use_classes,
dataset,
):
job_name = f"exp{job_id:04}-{model}-{'withCl' if use_classes else 'noCl'}-{dataset}"
commands = []
d = OrderedDict()
if model == "R-50" and use_classes:
config_file = os.path.join(
config_path, "e2e_faster_rcnn_R_50_FPN_1x_multiscale.yaml")
elif model == "R-101" and use_classes:
config_file = os.path.join(
config_path, "e2e_faster_rcnn_R_101_FPN_1x_multiscale.yaml")
elif model == "R-50" and not use_classes:
config_file = os.path.join(
config_path,
"e2e_faster_rcnn_R_50_FPN_1x_multiscale_noClasses.yaml")
elif model == "R-101" and not use_classes:
config_file = os.path.join(
config_path,
"e2e_faster_rcnn_R_101_FPN_1x_multiscale_noClasses.yaml")
else:
raise RuntimeError(
f"Do not know config for model {model} and use_classes {use_classes}"
)
d["--config-file"] = config_file
if dataset == "grozi":
d["DATASETS.TRAIN"] = "[\\\"grozi-train\\\"]"
if use_classes:
d["DATASETS.TEST"] = "[\\\"grozi-val-old-cl\\\"]"
else:
d["DATASETS.TEST"] = "[\\\"grozi-val-all\\\"]"
d["INPUT.MIN_SIZE_TRAIN"] = "[480,600,768,960,1152,1344,1536]"
d["INPUT.MAX_SIZE_TRAIN"] = 2048
d["INPUT.MIN_SIZE_TEST"] = 960
d["INPUT.MAX_SIZE_TEST"] = 1280
elif dataset == "instre-s1":
d["DATASETS.TRAIN"] = "[\\\"instre-s1-train\\\"]"
d["DATASETS.TEST"] = "[\\\"instre-s1-val\\\"]"
d["INPUT.MIN_SIZE_TRAIN"] = "[210,262,336,420,504,588,672]"
d["INPUT.MAX_SIZE_TRAIN"] = 2048
d["INPUT.MIN_SIZE_TEST"] = 420
d["INPUT.MAX_SIZE_TEST"] = 1280
elif dataset == "instre-s2":
d["DATASETS.TRAIN"] = "[\\\"instre-s2-train\\\"]"
d["DATASETS.TEST"] = "[\\\"instre-s2-val\\\"]"
d["INPUT.MIN_SIZE_TRAIN"] = "[180,225,288,360,432,504,576]"
d["INPUT.MAX_SIZE_TRAIN"] = 2048
d["INPUT.MIN_SIZE_TEST"] = 360
d["INPUT.MAX_SIZE_TEST"] = 1280
else:
raise RuntimeError(f"Unknown dataset {dataset}")
log_folder = os.path.join(log_path, job_name)
d["OUTPUT_DIR"] = log_folder
commands.append(main_command + " " + launcher.parameters_to_str(d))
# testing
if not use_classes:
d_testing = OrderedDict()
d_testing["--test_weights"] = os.path.join(log_folder,
"model_best.pth")
d_testing.update(d)
datasets_test = [
"[\\\"grozi-val-all\\\"]",
"[\\\"instre-s1-val\\\",\\\"instre-s1-test\\\"]",
"[\\\"instre-s2-val\\\",\\\"instre-s2-test\\\"]"
]
scales_test = [
"[480,600,768,960,1152,1344,1536]",
"[210,262,336,420,504,588,672]",
"[180,225,288,360,432,504,576]"
]
for dataset, scales in zip(datasets_test, scales_test):
d_testing_local = copy.deepcopy(d_testing)
d_testing_local["DATASETS.TEST"] = dataset
d_testing_local["TEST.BBOX_AUG.ENABLED"] = True
d_testing_local["TEST.BBOX_AUG.SCALES"] = scales
commands.append(main_command + " " +
launcher.parameters_to_str(d_testing_local))
exp_job_names.append(job_name)
exp_commands.append(commands)
exp_log_paths.append(log_folder)