def get_resnet18_inst_and_depth(which_encode_to_depth=9,
depth_dataset_list=['pbrnet', 'scenenet']):
ret_cfg = {
"BATCH_SIZE": 8,
"QUEUE_CAP": 2560,
"network_list": ["encode", "up_projection", "depth", "category"],
"imagenet_order": ["encode", "category"]
}
for depth_dataset in depth_dataset_list:
order_name = '%s_order' % depth_dataset
ret_cfg[order_name] = ["encode", "up_projection", "depth"]
# Add encode
ret_cfg["encode"] = shared_funcs.get_encode_resnet(
func_kwargs={'num_layers': [2, 2, 2, 2]})
num_layers_enc = len(ret_cfg["encode"].keys())
ret_cfg["encode_depth"] = num_layers_enc
# Add category here
ret_cfg["category"] = shared_funcs.get_category(
128, "encode_%i" % num_layers_enc)
ret_cfg["category_depth"] = 2
if which_encode_to_depth == 9:
ret_cfg = other_tasks.add_depth_head(
ret_cfg, 'encode_%i' % which_encode_to_depth)
elif which_encode_to_depth == 7:
ret_cfg = other_tasks.add_depth_head(ret_cfg,
'encode_%i' %
which_encode_to_depth,
first_upproj_filter=128,
no_upproj_layers=3)
else:
raise NotImplementedError()
return ret_cfg
def get_resnet_18_fx_encdX(no_layers_fixed=3):
ret_cfg = {
"BATCH_SIZE": 8,
"QUEUE_CAP": 2560,
"network_list": ["inst_encode", "inst_category"],
"imagenet_order": ["inst_encode", "inst_category"],
}
# Add inst_encode here
ret_cfg["inst_encode"] = shared_funcs.get_encode_resnet(
func_kwargs={'num_layers': [2, 2, 2, 2]})
num_layers_enc = len(ret_cfg["inst_encode"].keys())
ret_cfg["inst_encode_depth"] = num_layers_enc
# Add inst_category here
ret_cfg["inst_category"] = shared_funcs.get_category(
128, "inst_encode_%i" % num_layers_enc)
ret_cfg["inst_category_depth"] = 2
assert no_layers_fixed >= 1
ret_cfg["inst_encode"][1]["conv"]["trainable"] = 0
ret_cfg["inst_encode"][1]["conv"]["bn_trainable"] = 0
for idx_layer in range(2, no_layers_fixed + 1):
ret_cfg["inst_encode"][idx_layer]["ResBlock_trainable"] = 0
ret_cfg["inst_encode"][idx_layer]["ResBlock_bn_trainable"] = 0
return ret_cfg
def get_rp_inst_resnet_18(dataset_prefix='rp_imagenet', rp_layer_offset=0):
ret_cfg = {
"BATCH_SIZE": 8,
"QUEUE_CAP": 2560,
"network_list": ["encode", "encode_head", "rp_category", "category"],
"%s_order" % dataset_prefix: ["encode", "encode_head"],
"imagenet_order": ["encode", "category"],
}
# Add encode
ret_cfg["encode"] = shared_funcs.get_encode_resnet(
func_kwargs={'num_layers': [2, 2, 2, 2]})
num_layers_enc = len(ret_cfg["encode"].keys())
ret_cfg["encode_depth"] = num_layers_enc
# Add encode_head
ret_cfg["encode_head"] = other_tasks.get_rp_resnet_18_encode_head(
"encode_%i" % (num_layers_enc - rp_layer_offset))
ret_cfg["encode_head_depth"] = 2
# Add rp_category
ret_cfg["rp_category"] = other_tasks.get_rp_resnet_18_rp_category()
ret_cfg["rp_category_depth"] = 4
# Add category here
ret_cfg["category"] = shared_funcs.get_category(
128, "encode_%i" % num_layers_enc)
ret_cfg["category_depth"] = 2
return ret_cfg
def get_cpc_resnet_18(dataset_prefix='imagenet'):
ret_cfg = {
"BATCH_SIZE": 8,
"QUEUE_CAP": 2560,
"network_list": ["encode"],
"%s_order" % dataset_prefix: ["encode"],
}
# Add encode
ret_cfg["encode"] = shared_funcs.get_encode_resnet(
func_kwargs={'num_layers': [2, 2, 2, 2]})
num_layers_enc = len(ret_cfg["encode"].keys())
ret_cfg["encode_depth"] = num_layers_enc
ret_cfg["encode"]["as_output"] = 1
return ret_cfg
def get_resnet_18_fx_encd1():
ret_cfg = {
"BATCH_SIZE": 8,
"QUEUE_CAP": 2560,
"network_list": ["inst_encode", "inst_category"],
"imagenet_order": ["inst_encode", "inst_category"],
}
# Add inst_encode here
ret_cfg["inst_encode"] = shared_funcs.get_encode_resnet(
func_kwargs={'num_layers': [2, 2, 2, 2]})
num_layers_enc = len(ret_cfg["inst_encode"].keys())
ret_cfg["inst_encode_depth"] = num_layers_enc
# Add inst_category here
ret_cfg["inst_category"] = shared_funcs.get_category(
128, "inst_encode_%i" % num_layers_enc)
ret_cfg["inst_category_depth"] = 2
ret_cfg["inst_encode"][1]["conv"]["trainable"] = 0
return ret_cfg
def get_rp_resnet_18_bg(dataset_prefix='rp_imagenet'):
ret_cfg = {
"BATCH_SIZE": 8,
"QUEUE_CAP": 2560,
"network_list": ["encode", "encode_head", "rp_category"],
"%s_order" % dataset_prefix: ["encode", "encode_head"],
}
# Add encode
ret_cfg["encode"] = shared_funcs.get_encode_resnet(
func_kwargs={'num_layers': [2, 2, 2, 2]})
num_layers_enc = len(ret_cfg["encode"].keys())
ret_cfg["encode_depth"] = num_layers_enc
# Add encode_head
ret_cfg["encode_head"] = get_rp_resnet_18_encode_head_bg("encode_%i" %
num_layers_enc)
ret_cfg["encode_head_depth"] = 2
# Add rp_category
ret_cfg["rp_category"] = get_rp_resnet_18_rp_category_bg()
ret_cfg["rp_category_depth"] = 4
return ret_cfg
def get_ae_resnet_18(dataset_prefix='imagenet', head_dim=128):
ret_cfg = {
"BATCH_SIZE":
8,
"QUEUE_CAP":
2560,
"network_list": ["encode", "ae_head", "ae_decode", "ae_output"],
"%s_order" % dataset_prefix:
["encode", "ae_head", "ae_decode", "ae_output"],
}
# Add encode
ret_cfg["encode"] = shared_funcs.get_encode_resnet(
func_kwargs={'num_layers': [2, 2, 2, 2]})
num_layers_enc = len(ret_cfg["encode"].keys())
ret_cfg["encode_depth"] = num_layers_enc
ret_cfg["ae_head"] = {
"input": "encode_%i" % num_layers_enc,
1: {
"ae_head": {
"dimension": head_dim
}
},
}
ret_cfg["ae_head_depth"] = 1
ae_decode = {"input": "ae_head_1"}
for idx in range(1, num_layers_enc):
curr_encode = ret_cfg["encode"][num_layers_enc + 1 - idx]
curr_resblock = copy.deepcopy(curr_encode["ResBlock"])
curr_resblock = list(reversed(curr_resblock))
for each_conv in curr_resblock:
each_conv["upsample"] = each_conv["stride"]
curr_decode = {"ResBlock": curr_resblock}
ae_decode[idx] = curr_decode
ret_cfg["ae_decode"] = ae_decode
ret_cfg["ae_decode_depth"] = num_layers_enc - 1
ret_cfg["ae_output"] = {
"input": "ae_decode_%i" % (num_layers_enc - 1),
"as_output": 1,
1: {
"conv": {
"filter_size": 3,
"stride": 2,
"num_filters": 3,
"upsample": 2
}
},
2: {
"conv": {
"filter_size": 7,
"stride": 2,
"num_filters": 3,
"output": 1,
"upsample": 2
}
},
}
ret_cfg["ae_output_depth"] = 2
return ret_cfg