Spaces:
Running on Zero
Running on Zero
Commit ·
86072ea
1
Parent(s): 4ce5a27
clean up
Browse files- _utils/attn_utils_new.py +0 -12
- _utils/load_track_data.py +0 -6
- _utils/track_args.py +0 -62
- app.py +13 -18
- counting.py +7 -11
- models/seg_post_model/models.py +1 -1
- models/seg_post_model/{vit_sam.py → vit.py} +0 -0
- models/tra_post_model/data.py +1 -0
- models/tra_post_model/model.py +2 -47
- models/tra_post_model/tracking/__init__.py +0 -2
- models/tra_post_model/tracking/ilp.py +2 -0
- models/tra_post_model/tracking/tracking.py +2 -0
- models/tra_post_model/tracking/utils.py +2 -0
- models/tra_post_model/utils.py +3 -2
- tracking_one.py +19 -81
_utils/attn_utils_new.py
CHANGED
|
@@ -37,12 +37,6 @@ class CountingCrossAttnProcessor1:
|
|
| 37 |
context = encoder_hidden_states if is_cross else hidden_states
|
| 38 |
k = attn_layer.to_k(context)
|
| 39 |
v = attn_layer.to_v(context)
|
| 40 |
-
# q = attn_layer.reshape_heads_to_batch_dim(q)
|
| 41 |
-
# k = attn_layer.reshape_heads_to_batch_dim(k)
|
| 42 |
-
# v = attn_layer.reshape_heads_to_batch_dim(v)
|
| 43 |
-
# q = attn_layer.head_to_batch_dim(q)
|
| 44 |
-
# k = attn_layer.head_to_batch_dim(k)
|
| 45 |
-
# v = attn_layer.head_to_batch_dim(v)
|
| 46 |
q = self.head_to_batch_dim(q, h)
|
| 47 |
k = self.head_to_batch_dim(k, h)
|
| 48 |
v = self.head_to_batch_dim(v, h)
|
|
@@ -57,11 +51,8 @@ class CountingCrossAttnProcessor1:
|
|
| 57 |
|
| 58 |
# attention, what we cannot get enough of
|
| 59 |
attn_ = sim.softmax(dim=-1).clone()
|
| 60 |
-
# softmax = nn.Softmax(dim=-1)
|
| 61 |
-
# attn_ = softmax(sim)
|
| 62 |
self.attnstore(attn_, is_cross, self.place_in_unet)
|
| 63 |
out = torch.einsum("b i j, b j d -> b i d", attn_, v)
|
| 64 |
-
# out = attn_layer.batch_to_head_dim(out)
|
| 65 |
out = self.batch_to_head_dim(out, h)
|
| 66 |
|
| 67 |
if type(attn_layer.to_out) is torch.nn.modules.container.ModuleList:
|
|
@@ -112,9 +103,6 @@ def register_attention_control(model, controller):
|
|
| 112 |
continue
|
| 113 |
|
| 114 |
cross_att_count += 1
|
| 115 |
-
# attn_procs[name] = AttendExciteCrossAttnProcessor(
|
| 116 |
-
# attnstore=controller, place_in_unet=place_in_unet
|
| 117 |
-
# )
|
| 118 |
attn_procs[name] = CountingCrossAttnProcessor1(
|
| 119 |
attnstore=controller, place_in_unet=place_in_unet
|
| 120 |
)
|
|
|
|
| 37 |
context = encoder_hidden_states if is_cross else hidden_states
|
| 38 |
k = attn_layer.to_k(context)
|
| 39 |
v = attn_layer.to_v(context)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
q = self.head_to_batch_dim(q, h)
|
| 41 |
k = self.head_to_batch_dim(k, h)
|
| 42 |
v = self.head_to_batch_dim(v, h)
|
|
|
|
| 51 |
|
| 52 |
# attention, what we cannot get enough of
|
| 53 |
attn_ = sim.softmax(dim=-1).clone()
|
|
|
|
|
|
|
| 54 |
self.attnstore(attn_, is_cross, self.place_in_unet)
|
| 55 |
out = torch.einsum("b i j, b j d -> b i d", attn_, v)
|
|
|
|
| 56 |
out = self.batch_to_head_dim(out, h)
|
| 57 |
|
| 58 |
if type(attn_layer.to_out) is torch.nn.modules.container.ModuleList:
|
|
|
|
| 103 |
continue
|
| 104 |
|
| 105 |
cross_att_count += 1
|
|
|
|
|
|
|
|
|
|
| 106 |
attn_procs[name] = CountingCrossAttnProcessor1(
|
| 107 |
attnstore=controller, place_in_unet=place_in_unet
|
| 108 |
)
|
_utils/load_track_data.py
CHANGED
|
@@ -49,9 +49,7 @@ def _load_tiffs(folder: Path, dtype=None):
|
|
| 49 |
|
| 50 |
def load_track_images(file_dir):
|
| 51 |
|
| 52 |
-
# suffix_ = [".png", ".tif", ".tiff", ".jpg"]
|
| 53 |
def find_tif_dir(root_dir):
|
| 54 |
-
"""递归查找.tif 文件"""
|
| 55 |
tif_files = []
|
| 56 |
for dirpath, _, filenames in os.walk(root_dir):
|
| 57 |
if '__MACOSX' in dirpath:
|
|
@@ -112,7 +110,3 @@ def load_track_images(file_dir):
|
|
| 112 |
|
| 113 |
return imgs, imgs_raw, images_stable, imgs_, imgs_01, height, width
|
| 114 |
|
| 115 |
-
if __name__ == "__main__":
|
| 116 |
-
file_dir = "data/2D+Time/DIC-C2DH-HeLa/train/DIC-C2DH-HeLa/02"
|
| 117 |
-
imgs, imgs_raw, images_stable, imgs_, imgs_01, height, width = load_track_images(file_dir)
|
| 118 |
-
print(imgs.shape, imgs_raw.shape, images_stable.shape, imgs_.shape, imgs_01.shape, height, width)
|
|
|
|
| 49 |
|
| 50 |
def load_track_images(file_dir):
|
| 51 |
|
|
|
|
| 52 |
def find_tif_dir(root_dir):
|
|
|
|
| 53 |
tif_files = []
|
| 54 |
for dirpath, _, filenames in os.walk(root_dir):
|
| 55 |
if '__MACOSX' in dirpath:
|
|
|
|
| 110 |
|
| 111 |
return imgs, imgs_raw, images_stable, imgs_, imgs_01, height, width
|
| 112 |
|
|
|
|
|
|
|
|
|
|
|
|
_utils/track_args.py
DELETED
|
@@ -1,62 +0,0 @@
|
|
| 1 |
-
import configargparse
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
def parse_train_args():
|
| 5 |
-
parser = configargparse.ArgumentParser(
|
| 6 |
-
formatter_class=configargparse.ArgumentDefaultsHelpFormatter,
|
| 7 |
-
config_file_parser_class=configargparse.YAMLConfigFileParser,
|
| 8 |
-
allow_abbrev=False,
|
| 9 |
-
)
|
| 10 |
-
parser.add_argument(
|
| 11 |
-
"-c",
|
| 12 |
-
"--config",
|
| 13 |
-
default="_utils/example_config.yaml",
|
| 14 |
-
is_config_file=True,
|
| 15 |
-
help="config file path",
|
| 16 |
-
)
|
| 17 |
-
parser.add_argument("-d", "--d_model", type=int, default=256)
|
| 18 |
-
parser.add_argument("-w", "--window", type=int, default=10)
|
| 19 |
-
parser.add_argument("--spatial_pos_cutoff", type=int, default=256)
|
| 20 |
-
parser.add_argument("--num_encoder_layers", type=int, default=6)
|
| 21 |
-
parser.add_argument("--num_decoder_layers", type=int, default=6)
|
| 22 |
-
parser.add_argument("--pos_embed_per_dim", type=int, default=32)
|
| 23 |
-
parser.add_argument("--feat_embed_per_dim", type=int, default=8)
|
| 24 |
-
parser.add_argument("--dropout", type=float, default=0.00)
|
| 25 |
-
parser.add_argument(
|
| 26 |
-
"--attn_positional_bias",
|
| 27 |
-
type=str,
|
| 28 |
-
choices=["rope", "bias", "none"],
|
| 29 |
-
default="rope",
|
| 30 |
-
)
|
| 31 |
-
parser.add_argument("--attn_positional_bias_n_spatial", type=int, default=16)
|
| 32 |
-
parser.add_argument("--attn_dist_mode", default="v0")
|
| 33 |
-
parser.add_argument(
|
| 34 |
-
"--causal_norm",
|
| 35 |
-
type=str,
|
| 36 |
-
choices=["none", "linear", "softmax", "quiet_softmax"],
|
| 37 |
-
default="quiet_softmax",
|
| 38 |
-
)
|
| 39 |
-
|
| 40 |
-
args, unknown_args = parser.parse_known_args()
|
| 41 |
-
|
| 42 |
-
# # Hack to allow for --input_test
|
| 43 |
-
# allowed_unknown = ["input_test"]
|
| 44 |
-
# if not set(a.split("=")[0].strip("-") for a in unknown_args).issubset(
|
| 45 |
-
# set(allowed_unknown)
|
| 46 |
-
# ):
|
| 47 |
-
# raise ValueError(f"Unknown args: {unknown_args}")
|
| 48 |
-
|
| 49 |
-
# pprint(vars(args))
|
| 50 |
-
|
| 51 |
-
# for backward compatibility
|
| 52 |
-
# if args.attn_positional_bias == "True":
|
| 53 |
-
# args.attn_positional_bias = "bias"
|
| 54 |
-
# elif args.attn_positional_bias == "False":
|
| 55 |
-
# args.attn_positional_bias = False
|
| 56 |
-
|
| 57 |
-
# if args.train_samples == 0:
|
| 58 |
-
# raise NotImplementedError(
|
| 59 |
-
# "--train_samples must be > 0, full dataset pass not supported."
|
| 60 |
-
# )
|
| 61 |
-
|
| 62 |
-
return args
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
app.py
CHANGED
|
@@ -937,12 +937,21 @@ with gr.Blocks(
|
|
| 937 |
) as demo:
|
| 938 |
gr.Markdown(
|
| 939 |
"""
|
| 940 |
-
# 🔬 Microscopy Image Analysis Suite
|
| 941 |
|
| 942 |
-
Supporting three key tasks:
|
| 943 |
- 🎨 **Segmentation**: Instance segmentation of microscopic objects
|
| 944 |
- 🔢 **Counting**: Counting microscopic objects based on density maps
|
| 945 |
- 🎬 **Tracking**: Tracking microscopic objects in video sequences
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 946 |
"""
|
| 947 |
)
|
| 948 |
|
|
@@ -1667,26 +1676,12 @@ with gr.Blocks(
|
|
| 1667 |
outputs=[feedback_status, feedback_status]
|
| 1668 |
)
|
| 1669 |
|
| 1670 |
-
|
| 1671 |
-
"""
|
| 1672 |
-
---
|
| 1673 |
-
### 📒 Note:
|
| 1674 |
-
|
| 1675 |
-
This project is currently available with usage limits for research trial use and feedback collection. We plan to release a free public version in the future. We are actively improving the toolkit and greatly appreciate your feedback!
|
| 1676 |
-
|
| 1677 |
-
|
| 1678 |
-
|
| 1679 |
-
### 💡 Technical Details
|
| 1680 |
-
|
| 1681 |
-
**MicroscopyMatching** - A general-purpose microscopy image analysis toolkit based on Stable Diffusion
|
| 1682 |
-
|
| 1683 |
-
"""
|
| 1684 |
-
)
|
| 1685 |
|
| 1686 |
if __name__ == "__main__":
|
| 1687 |
demo.queue().launch(
|
| 1688 |
server_name="0.0.0.0",
|
| 1689 |
-
server_port=
|
| 1690 |
share=False,
|
| 1691 |
ssr_mode=False,
|
| 1692 |
show_error=True,
|
|
|
|
| 937 |
) as demo:
|
| 938 |
gr.Markdown(
|
| 939 |
"""
|
| 940 |
+
# 🔬 MicroscopyMatching: Microscopy Image Analysis Suite
|
| 941 |
|
| 942 |
+
### Supporting three key tasks:
|
| 943 |
- 🎨 **Segmentation**: Instance segmentation of microscopic objects
|
| 944 |
- 🔢 **Counting**: Counting microscopic objects based on density maps
|
| 945 |
- 🎬 **Tracking**: Tracking microscopic objects in video sequences
|
| 946 |
+
|
| 947 |
+
### 💡 Technical Details:
|
| 948 |
+
|
| 949 |
+
**MicroscopyMatching** - A general-purpose microscopy image analysis toolkit based on Stable Diffusion
|
| 950 |
+
|
| 951 |
+
### 📒 Note:
|
| 952 |
+
|
| 953 |
+
This project is currently available with usage limits for research trial use and feedback collection. We plan to release a free public version in the future. We are actively improving the toolkit and greatly appreciate your feedback!
|
| 954 |
+
|
| 955 |
"""
|
| 956 |
)
|
| 957 |
|
|
|
|
| 1676 |
outputs=[feedback_status, feedback_status]
|
| 1677 |
)
|
| 1678 |
|
| 1679 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1680 |
|
| 1681 |
if __name__ == "__main__":
|
| 1682 |
demo.queue().launch(
|
| 1683 |
server_name="0.0.0.0",
|
| 1684 |
+
server_port=7861,
|
| 1685 |
share=False,
|
| 1686 |
ssr_mode=False,
|
| 1687 |
show_error=True,
|
counting.py
CHANGED
|
@@ -153,9 +153,9 @@ class CountingModule(pl.LightningModule):
|
|
| 153 |
loca_feature_bf_regression = loca_out["feature_bf_regression"]
|
| 154 |
adapted_emb = self.counting_adapter.adapter(loca_feature_bf_regression, boxes) # shape [1, 768]
|
| 155 |
if task_loc_idx.shape[0] == 0:
|
| 156 |
-
encoder_hidden_states[0,2,:] = adapted_emb.squeeze()
|
| 157 |
else:
|
| 158 |
-
encoder_hidden_states[0,task_loc_idx[0, 1]+1,:] = adapted_emb.squeeze()
|
| 159 |
|
| 160 |
# Predict the noise residual
|
| 161 |
noise_pred, feature_list = self.stable.unet(noisy_latents, timesteps, encoder_hidden_states)
|
|
@@ -174,7 +174,7 @@ class CountingModule(pl.LightningModule):
|
|
| 174 |
|
| 175 |
# only use 64x64 self-attention
|
| 176 |
self_attn_aggregate = attn_utils.aggregate_attention( # [res, res, 4096]
|
| 177 |
-
prompts=[self.config.prompt],
|
| 178 |
attention_store=self.controller,
|
| 179 |
res=64,
|
| 180 |
from_where=("up", "down"),
|
|
@@ -182,7 +182,7 @@ class CountingModule(pl.LightningModule):
|
|
| 182 |
select=0
|
| 183 |
)
|
| 184 |
self_attn_aggregate32 = attn_utils.aggregate_attention( # [res, res, 4096]
|
| 185 |
-
prompts=[self.config.prompt],
|
| 186 |
attention_store=self.controller,
|
| 187 |
res=32,
|
| 188 |
from_where=("up", "down"),
|
|
@@ -190,7 +190,7 @@ class CountingModule(pl.LightningModule):
|
|
| 190 |
select=0
|
| 191 |
)
|
| 192 |
self_attn_aggregate16 = attn_utils.aggregate_attention( # [res, res, 4096]
|
| 193 |
-
prompts=[self.config.prompt],
|
| 194 |
attention_store=self.controller,
|
| 195 |
res=16,
|
| 196 |
from_where=("up", "down"),
|
|
@@ -201,7 +201,7 @@ class CountingModule(pl.LightningModule):
|
|
| 201 |
# cross attention
|
| 202 |
for res in [32, 16]:
|
| 203 |
attn_aggregate = attn_utils.aggregate_attention( # [res, res, 77]
|
| 204 |
-
prompts=[self.config.prompt],
|
| 205 |
attention_store=self.controller,
|
| 206 |
res=res,
|
| 207 |
from_where=("up", "down"),
|
|
@@ -212,7 +212,7 @@ class CountingModule(pl.LightningModule):
|
|
| 212 |
task_attn_ = attn_aggregate[:, :, 1].unsqueeze(0).unsqueeze(0) # [1, 1, res, res]
|
| 213 |
attention_maps.append(task_attn_)
|
| 214 |
if self.use_box:
|
| 215 |
-
exemplar_attns = attn_aggregate[:, :, 2].unsqueeze(0).unsqueeze(0)
|
| 216 |
exemplar_attention_maps.append(exemplar_attns)
|
| 217 |
else:
|
| 218 |
exemplar_attns1 = attn_aggregate[:, :, 2].unsqueeze(0).unsqueeze(0)
|
|
@@ -266,10 +266,6 @@ class CountingModule(pl.LightningModule):
|
|
| 266 |
attn_stack = torch.cat(attn_stack, dim=1)
|
| 267 |
|
| 268 |
if not self.use_box:
|
| 269 |
-
|
| 270 |
-
# cross_self_exe_attn_np = cross_self_exe_attn.detach().squeeze().cpu().numpy()
|
| 271 |
-
# boxes = gen_dummy_boxes(cross_self_exe_attn_np, max_boxes=1)
|
| 272 |
-
# boxes = boxes.to(self.device)
|
| 273 |
|
| 274 |
loca_out = self.loca_model.forward_before_reg(input_image, boxes)
|
| 275 |
loca_feature_bf_regression = loca_out["feature_bf_regression"]
|
|
|
|
| 153 |
loca_feature_bf_regression = loca_out["feature_bf_regression"]
|
| 154 |
adapted_emb = self.counting_adapter.adapter(loca_feature_bf_regression, boxes) # shape [1, 768]
|
| 155 |
if task_loc_idx.shape[0] == 0:
|
| 156 |
+
encoder_hidden_states[0,2,:] = adapted_emb.squeeze()
|
| 157 |
else:
|
| 158 |
+
encoder_hidden_states[0,task_loc_idx[0, 1]+1,:] = adapted_emb.squeeze()
|
| 159 |
|
| 160 |
# Predict the noise residual
|
| 161 |
noise_pred, feature_list = self.stable.unet(noisy_latents, timesteps, encoder_hidden_states)
|
|
|
|
| 174 |
|
| 175 |
# only use 64x64 self-attention
|
| 176 |
self_attn_aggregate = attn_utils.aggregate_attention( # [res, res, 4096]
|
| 177 |
+
prompts=[self.config.prompt],
|
| 178 |
attention_store=self.controller,
|
| 179 |
res=64,
|
| 180 |
from_where=("up", "down"),
|
|
|
|
| 182 |
select=0
|
| 183 |
)
|
| 184 |
self_attn_aggregate32 = attn_utils.aggregate_attention( # [res, res, 4096]
|
| 185 |
+
prompts=[self.config.prompt],
|
| 186 |
attention_store=self.controller,
|
| 187 |
res=32,
|
| 188 |
from_where=("up", "down"),
|
|
|
|
| 190 |
select=0
|
| 191 |
)
|
| 192 |
self_attn_aggregate16 = attn_utils.aggregate_attention( # [res, res, 4096]
|
| 193 |
+
prompts=[self.config.prompt],
|
| 194 |
attention_store=self.controller,
|
| 195 |
res=16,
|
| 196 |
from_where=("up", "down"),
|
|
|
|
| 201 |
# cross attention
|
| 202 |
for res in [32, 16]:
|
| 203 |
attn_aggregate = attn_utils.aggregate_attention( # [res, res, 77]
|
| 204 |
+
prompts=[self.config.prompt],
|
| 205 |
attention_store=self.controller,
|
| 206 |
res=res,
|
| 207 |
from_where=("up", "down"),
|
|
|
|
| 212 |
task_attn_ = attn_aggregate[:, :, 1].unsqueeze(0).unsqueeze(0) # [1, 1, res, res]
|
| 213 |
attention_maps.append(task_attn_)
|
| 214 |
if self.use_box:
|
| 215 |
+
exemplar_attns = attn_aggregate[:, :, 2].unsqueeze(0).unsqueeze(0)
|
| 216 |
exemplar_attention_maps.append(exemplar_attns)
|
| 217 |
else:
|
| 218 |
exemplar_attns1 = attn_aggregate[:, :, 2].unsqueeze(0).unsqueeze(0)
|
|
|
|
| 266 |
attn_stack = torch.cat(attn_stack, dim=1)
|
| 267 |
|
| 268 |
if not self.use_box:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 269 |
|
| 270 |
loca_out = self.loca_model.forward_before_reg(input_image, boxes)
|
| 271 |
loca_feature_bf_regression = loca_out["feature_bf_regression"]
|
models/seg_post_model/models.py
CHANGED
|
@@ -16,7 +16,7 @@ import logging
|
|
| 16 |
models_logger = logging.getLogger(__name__)
|
| 17 |
|
| 18 |
from . import transforms, dynamics, utils
|
| 19 |
-
from .
|
| 20 |
from .core import assign_device, run_net
|
| 21 |
|
| 22 |
# _MODEL_DIR_ENV = os.environ.get("CELLPOSE_LOCAL_MODELS_PATH")
|
|
|
|
| 16 |
models_logger = logging.getLogger(__name__)
|
| 17 |
|
| 18 |
from . import transforms, dynamics, utils
|
| 19 |
+
from .vit import Transformer
|
| 20 |
from .core import assign_device, run_net
|
| 21 |
|
| 22 |
# _MODEL_DIR_ENV = os.environ.get("CELLPOSE_LOCAL_MODELS_PATH")
|
models/seg_post_model/{vit_sam.py → vit.py}
RENAMED
|
File without changes
|
models/tra_post_model/data.py
CHANGED
|
@@ -1,5 +1,6 @@
|
|
| 1 |
"""Regionprops features and its augmentations.
|
| 2 |
WindowedRegionFeatures (WRFeatures) is a class that holds regionprops features for a windowed track region.
|
|
|
|
| 3 |
"""
|
| 4 |
|
| 5 |
import itertools
|
|
|
|
| 1 |
"""Regionprops features and its augmentations.
|
| 2 |
WindowedRegionFeatures (WRFeatures) is a class that holds regionprops features for a windowed track region.
|
| 3 |
+
Modified from Trackastra (https://github.com/weigertlab/trackastra)
|
| 4 |
"""
|
| 5 |
|
| 6 |
import itertools
|
models/tra_post_model/model.py
CHANGED
|
@@ -599,47 +599,14 @@ class TrackingTransformer(torch.nn.Module):
|
|
| 599 |
|
| 600 |
@classmethod
|
| 601 |
def from_folder(
|
| 602 |
-
cls, folder, map_location=None,
|
| 603 |
):
|
| 604 |
folder = Path(folder)
|
| 605 |
|
| 606 |
config = yaml.load(open(folder / "config.yaml"), Loader=yaml.FullLoader)
|
| 607 |
-
if args:
|
| 608 |
-
args = vars(args)
|
| 609 |
-
for k, v in config.items():
|
| 610 |
-
errors = []
|
| 611 |
-
if k in args:
|
| 612 |
-
if config[k] != args[k]:
|
| 613 |
-
errors.append(
|
| 614 |
-
f"Loaded model config {k}={config[k]}, but current argument"
|
| 615 |
-
f" {k}={args[k]}."
|
| 616 |
-
)
|
| 617 |
-
if errors:
|
| 618 |
-
raise ValueError("\n".join(errors))
|
| 619 |
|
| 620 |
model = cls(**config)
|
| 621 |
|
| 622 |
-
# try:
|
| 623 |
-
# # Try to load from lightning checkpoint first
|
| 624 |
-
# v_folder = sorted((folder / "tb").glob("version_*"))[version]
|
| 625 |
-
# checkpoint = sorted((v_folder / "checkpoints").glob("*epoch*.ckpt"))[0]
|
| 626 |
-
# pl_state_dict = torch.load(checkpoint, map_location=map_location)[
|
| 627 |
-
# "state_dict"
|
| 628 |
-
# ]
|
| 629 |
-
# state_dict = OrderedDict()
|
| 630 |
-
|
| 631 |
-
# # Hack
|
| 632 |
-
# for k, v in pl_state_dict.items():
|
| 633 |
-
# if k.startswith("model."):
|
| 634 |
-
# state_dict[k[6:]] = v
|
| 635 |
-
# else:
|
| 636 |
-
# raise ValueError(f"Unexpected key {k} in state_dict")
|
| 637 |
-
|
| 638 |
-
# model.load_state_dict(state_dict)
|
| 639 |
-
# logger.info(f"Loaded model from {checkpoint}")
|
| 640 |
-
# except:
|
| 641 |
-
# # Default: Load manually saved model (legacy)
|
| 642 |
-
|
| 643 |
fpath = folder / checkpoint_path
|
| 644 |
logger.info(f"Loading model state from {fpath}")
|
| 645 |
|
|
@@ -656,24 +623,12 @@ class TrackingTransformer(torch.nn.Module):
|
|
| 656 |
|
| 657 |
@classmethod
|
| 658 |
def from_cfg(
|
| 659 |
-
cls, cfg_path
|
| 660 |
):
|
| 661 |
|
| 662 |
cfg_path = Path(cfg_path)
|
| 663 |
|
| 664 |
config = yaml.load(open(cfg_path), Loader=yaml.FullLoader)
|
| 665 |
-
if args:
|
| 666 |
-
args = vars(args)
|
| 667 |
-
for k, v in config.items():
|
| 668 |
-
errors = []
|
| 669 |
-
if k in args:
|
| 670 |
-
if config[k] != args[k]:
|
| 671 |
-
errors.append(
|
| 672 |
-
f"Loaded model config {k}={config[k]}, but current argument"
|
| 673 |
-
f" {k}={args[k]}."
|
| 674 |
-
)
|
| 675 |
-
if errors:
|
| 676 |
-
raise ValueError("\n".join(errors))
|
| 677 |
|
| 678 |
model = cls(**config)
|
| 679 |
|
|
|
|
| 599 |
|
| 600 |
@classmethod
|
| 601 |
def from_folder(
|
| 602 |
+
cls, folder, map_location=None, checkpoint_path: str = "model.pt"
|
| 603 |
):
|
| 604 |
folder = Path(folder)
|
| 605 |
|
| 606 |
config = yaml.load(open(folder / "config.yaml"), Loader=yaml.FullLoader)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 607 |
|
| 608 |
model = cls(**config)
|
| 609 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 610 |
fpath = folder / checkpoint_path
|
| 611 |
logger.info(f"Loading model state from {fpath}")
|
| 612 |
|
|
|
|
| 623 |
|
| 624 |
@classmethod
|
| 625 |
def from_cfg(
|
| 626 |
+
cls, cfg_path
|
| 627 |
):
|
| 628 |
|
| 629 |
cfg_path = Path(cfg_path)
|
| 630 |
|
| 631 |
config = yaml.load(open(cfg_path), Loader=yaml.FullLoader)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 632 |
|
| 633 |
model = cls(**config)
|
| 634 |
|
models/tra_post_model/tracking/__init__.py
CHANGED
|
@@ -1,5 +1,3 @@
|
|
| 1 |
-
# ruff: noqa: F401
|
| 2 |
-
|
| 3 |
from .track_graph import TrackGraph
|
| 4 |
from .tracking import (
|
| 5 |
build_graph,
|
|
|
|
|
|
|
|
|
|
| 1 |
from .track_graph import TrackGraph
|
| 2 |
from .tracking import (
|
| 3 |
build_graph,
|
models/tra_post_model/tracking/ilp.py
CHANGED
|
@@ -1,3 +1,5 @@
|
|
|
|
|
|
|
|
| 1 |
import logging
|
| 2 |
import time
|
| 3 |
from types import SimpleNamespace
|
|
|
|
| 1 |
+
# Modified from Trackastra (https://github.com/weigertlab/trackastra)
|
| 2 |
+
|
| 3 |
import logging
|
| 4 |
import time
|
| 5 |
from types import SimpleNamespace
|
models/tra_post_model/tracking/tracking.py
CHANGED
|
@@ -1,3 +1,5 @@
|
|
|
|
|
|
|
|
| 1 |
import logging
|
| 2 |
from itertools import chain
|
| 3 |
|
|
|
|
| 1 |
+
# Modified from Trackastra (https://github.com/weigertlab/trackastra)
|
| 2 |
+
|
| 3 |
import logging
|
| 4 |
from itertools import chain
|
| 5 |
|
models/tra_post_model/tracking/utils.py
CHANGED
|
@@ -1,3 +1,5 @@
|
|
|
|
|
|
|
|
| 1 |
import logging
|
| 2 |
from collections import deque
|
| 3 |
from pathlib import Path
|
|
|
|
| 1 |
+
# Modified from Trackastra (https://github.com/weigertlab/trackastra)
|
| 2 |
+
|
| 3 |
import logging
|
| 4 |
from collections import deque
|
| 5 |
from pathlib import Path
|
models/tra_post_model/utils.py
CHANGED
|
@@ -1,3 +1,5 @@
|
|
|
|
|
|
|
|
| 1 |
import logging
|
| 2 |
|
| 3 |
import dask.array as da
|
|
@@ -41,7 +43,6 @@ def blockwise_sum(
|
|
| 41 |
return B
|
| 42 |
|
| 43 |
|
| 44 |
-
# TODO allow for batch dimension. Should be faster than looping
|
| 45 |
def blockwise_causal_norm(
|
| 46 |
A: torch.Tensor,
|
| 47 |
timepoints: torch.Tensor,
|
|
@@ -70,7 +71,7 @@ def blockwise_causal_norm(
|
|
| 70 |
if mode in ("softmax", "quiet_softmax"):
|
| 71 |
# Subtract max for numerical stability
|
| 72 |
# https://stats.stackexchange.com/questions/338285/how-does-the-subtraction-of-the-logit-maximum-improve-learning
|
| 73 |
-
|
| 74 |
|
| 75 |
if mask_invalid is not None:
|
| 76 |
assert mask_invalid.shape == A.shape
|
|
|
|
| 1 |
+
# Modified from Trackastra (https://github.com/weigertlab/trackastra)
|
| 2 |
+
|
| 3 |
import logging
|
| 4 |
|
| 5 |
import dask.array as da
|
|
|
|
| 43 |
return B
|
| 44 |
|
| 45 |
|
|
|
|
| 46 |
def blockwise_causal_norm(
|
| 47 |
A: torch.Tensor,
|
| 48 |
timepoints: torch.Tensor,
|
|
|
|
| 71 |
if mode in ("softmax", "quiet_softmax"):
|
| 72 |
# Subtract max for numerical stability
|
| 73 |
# https://stats.stackexchange.com/questions/338285/how-does-the-subtraction-of-the-logit-maximum-improve-learning
|
| 74 |
+
|
| 75 |
|
| 76 |
if mask_invalid is not None:
|
| 77 |
assert mask_invalid.shape == A.shape
|
tracking_one.py
CHANGED
|
@@ -1,16 +1,11 @@
|
|
| 1 |
import os
|
| 2 |
-
import pprint
|
| 3 |
from typing import Any, List, Optional
|
| 4 |
-
import argparse
|
| 5 |
from huggingface_hub import hf_hub_download
|
| 6 |
-
import pyrallis
|
| 7 |
from pytorch_lightning.utilities.types import STEP_OUTPUT
|
| 8 |
import torch
|
| 9 |
-
import os
|
| 10 |
from PIL import Image
|
| 11 |
import numpy as np
|
| 12 |
import tifffile
|
| 13 |
-
import skimage.io as io
|
| 14 |
from config import RunConfig
|
| 15 |
from _utils import attn_utils_new as attn_utils
|
| 16 |
from _utils.attn_utils_new import AttentionStore
|
|
@@ -18,7 +13,6 @@ from _utils.misc_helper import *
|
|
| 18 |
import torch.nn.functional as F
|
| 19 |
from tqdm import tqdm
|
| 20 |
import torch.nn as nn
|
| 21 |
-
import matplotlib.pyplot as plt
|
| 22 |
import cv2
|
| 23 |
import warnings
|
| 24 |
warnings.filterwarnings("ignore", category=UserWarning)
|
|
@@ -33,7 +27,6 @@ from models.tra_post_model.utils import (
|
|
| 33 |
)
|
| 34 |
from models.tra_post_model.data import build_windows_sd, get_features
|
| 35 |
from models.tra_post_model.tracking import TrackGraph, build_graph, track_greedy
|
| 36 |
-
from _utils.track_args import parse_train_args as get_track_args
|
| 37 |
import torchvision.transforms as T
|
| 38 |
from pathlib import Path
|
| 39 |
import dask.array as da
|
|
@@ -41,7 +34,6 @@ from typing import Dict, List, Optional, Union, Literal
|
|
| 41 |
from scipy.sparse import SparseEfficiencyWarning, csr_array
|
| 42 |
import tracemalloc
|
| 43 |
import gc
|
| 44 |
-
# from memory_profiler import profile
|
| 45 |
from _utils.load_track_data import load_track_images
|
| 46 |
|
| 47 |
SCALE = 1
|
|
@@ -82,15 +74,8 @@ class TrackingModule(pl.LightningModule):
|
|
| 82 |
|
| 83 |
# load loca model
|
| 84 |
self.loca_model = build_loca_model()
|
| 85 |
-
# weights = torch.load("ckpt/loca_few_shot.pt")["model"]
|
| 86 |
-
# weights = {k.replace("module","") : v for k, v in weights.items()}
|
| 87 |
-
# self.loca_model.load_state_dict(weights, strict=False)
|
| 88 |
-
# del weights
|
| 89 |
|
| 90 |
self.counting_adapter = Counting(scale_factor=SCALE)
|
| 91 |
-
# if os.path.isfile(self.args.adapter_weight):
|
| 92 |
-
# adapter_weight = torch.load(self.args.adapter_weight,map_location=torch.device('cpu'))
|
| 93 |
-
# self.counting_adapter.load_state_dict(adapter_weight, strict=False)
|
| 94 |
|
| 95 |
### load stable diffusion and its controller
|
| 96 |
self.stable = load_stable_diffusion_model(config=self.config)
|
|
@@ -110,7 +95,6 @@ class TrackingModule(pl.LightningModule):
|
|
| 110 |
" `placeholder_token` that is not already in the tokenizer."
|
| 111 |
)
|
| 112 |
try:
|
| 113 |
-
# task_embed_from_pretrain = torch.load("pretrained/task_embed.pth")
|
| 114 |
task_embed_from_pretrain = hf_hub_download(
|
| 115 |
repo_id="phoebe777777/111",
|
| 116 |
filename="task_embed.pth",
|
|
@@ -144,30 +128,17 @@ class TrackingModule(pl.LightningModule):
|
|
| 144 |
self.placeholder_token_id = placeholder_token_id
|
| 145 |
|
| 146 |
fpath = Path("_utils/config.yaml")
|
| 147 |
-
args_ = get_track_args()
|
| 148 |
|
| 149 |
model = TrackingTransformer.from_cfg(
|
| 150 |
cfg_path=fpath,
|
| 151 |
-
args=args_,
|
| 152 |
)
|
| 153 |
-
# model = TrackingTransformer.from_folder(
|
| 154 |
-
# Path(*fpath.parts[:-1]),
|
| 155 |
-
# args=args_,
|
| 156 |
-
# checkpoint_path=Path(*fpath.parts[-1:]),
|
| 157 |
-
# )
|
| 158 |
-
|
| 159 |
|
| 160 |
self.track_model = model
|
| 161 |
-
self.track_args = args_
|
| 162 |
|
| 163 |
|
| 164 |
def move_to_device(self, device):
|
| 165 |
self.stable.to(device)
|
| 166 |
-
# if self.loca_model is not None and self.counting_adapter is not None:
|
| 167 |
-
# self.loca_model.to(device)
|
| 168 |
-
# self.counting_adapter.to(device)
|
| 169 |
self.counting_adapter.to(device)
|
| 170 |
-
# self.dino.to(device)
|
| 171 |
self.loca_model.to(device)
|
| 172 |
self.track_model.to(device)
|
| 173 |
|
|
@@ -221,9 +192,9 @@ class TrackingModule(pl.LightningModule):
|
|
| 221 |
adapted_emb = self.counting_adapter.adapter(loca_feature_bf_regression, boxes) # shape [1, 768]
|
| 222 |
|
| 223 |
if task_loc_idx.shape[0] == 0:
|
| 224 |
-
encoder_hidden_states[0,2,:] = adapted_emb.squeeze()
|
| 225 |
else:
|
| 226 |
-
encoder_hidden_states[:,task_loc_idx[0, 1]+1,:] = adapted_emb.squeeze()
|
| 227 |
|
| 228 |
# Predict the noise residual
|
| 229 |
noise_pred, feature_list = self.stable.unet(noisy_latents, timesteps, encoder_hidden_states)
|
|
@@ -242,7 +213,7 @@ class TrackingModule(pl.LightningModule):
|
|
| 242 |
|
| 243 |
# only use 64x64 self-attention
|
| 244 |
self_attn_aggregate = attn_utils.aggregate_attention( # [res, res, 4096]
|
| 245 |
-
prompts=[self.config.prompt for i in range(bsz)],
|
| 246 |
attention_store=self.controller,
|
| 247 |
res=64,
|
| 248 |
from_where=("up", "down"),
|
|
@@ -250,7 +221,7 @@ class TrackingModule(pl.LightningModule):
|
|
| 250 |
select=0
|
| 251 |
)
|
| 252 |
self_attn_aggregate32 = attn_utils.aggregate_attention( # [res, res, 4096]
|
| 253 |
-
prompts=[self.config.prompt for i in range(bsz)],
|
| 254 |
attention_store=self.controller,
|
| 255 |
res=32,
|
| 256 |
from_where=("up", "down"),
|
|
@@ -258,7 +229,7 @@ class TrackingModule(pl.LightningModule):
|
|
| 258 |
select=0
|
| 259 |
)
|
| 260 |
self_attn_aggregate16 = attn_utils.aggregate_attention( # [res, res, 4096]
|
| 261 |
-
prompts=[self.config.prompt for i in range(bsz)],
|
| 262 |
attention_store=self.controller,
|
| 263 |
res=16,
|
| 264 |
from_where=("up", "down"),
|
|
@@ -269,7 +240,7 @@ class TrackingModule(pl.LightningModule):
|
|
| 269 |
# cross attention
|
| 270 |
for res in [32, 16]:
|
| 271 |
attn_aggregate = attn_utils.aggregate_attention( # [res, res, 77]
|
| 272 |
-
prompts=[self.config.prompt for i in range(bsz)],
|
| 273 |
attention_store=self.controller,
|
| 274 |
res=res,
|
| 275 |
from_where=("up", "down"),
|
|
@@ -279,7 +250,7 @@ class TrackingModule(pl.LightningModule):
|
|
| 279 |
|
| 280 |
task_attn_ = attn_aggregate[:, :, 1].unsqueeze(0).unsqueeze(0) # [1, 1, res, res]
|
| 281 |
attention_maps.append(task_attn_)
|
| 282 |
-
exemplar_attns = attn_aggregate[:, :, 2].unsqueeze(0).unsqueeze(0)
|
| 283 |
exemplar_attention_maps.append(exemplar_attns)
|
| 284 |
|
| 285 |
|
|
@@ -306,7 +277,7 @@ class TrackingModule(pl.LightningModule):
|
|
| 306 |
attn_stack = torch.cat(attn_stack, dim=1)
|
| 307 |
|
| 308 |
|
| 309 |
-
attn_after_new_regressor, loss = self.counting_adapter.regressor(input_image, attn_stack, feature_list, mask.cpu().numpy(), training=False)
|
| 310 |
|
| 311 |
return {
|
| 312 |
"attn_after_new_regressor":attn_after_new_regressor,
|
|
@@ -364,9 +335,9 @@ class TrackingModule(pl.LightningModule):
|
|
| 364 |
adapted_emb = self.adapt_emb.to(self.device)
|
| 365 |
task_loc_idx = torch.nonzero(input_ids == self.placeholder_token_id)
|
| 366 |
if task_loc_idx.shape[0] == 0:
|
| 367 |
-
encoder_hidden_states[0,5,:] = adapted_emb.squeeze()
|
| 368 |
else:
|
| 369 |
-
encoder_hidden_states[:,task_loc_idx[0, 1]+4,:] = adapted_emb.squeeze()
|
| 370 |
|
| 371 |
# Predict the noise residual
|
| 372 |
noise_pred, feature_list = self.stable.unet(noisy_latents, timesteps, encoder_hidden_states)
|
|
@@ -386,7 +357,7 @@ class TrackingModule(pl.LightningModule):
|
|
| 386 |
|
| 387 |
# only use 64x64 self-attention
|
| 388 |
self_attn_aggregate = attn_utils.aggregate_attention( # [res, res, 4096]
|
| 389 |
-
prompts=[self.config.prompt for i in range(bsz)],
|
| 390 |
attention_store=self.controller,
|
| 391 |
res=64,
|
| 392 |
from_where=("up", "down"),
|
|
@@ -397,7 +368,7 @@ class TrackingModule(pl.LightningModule):
|
|
| 397 |
# cross attention
|
| 398 |
for res in [32, 16]:
|
| 399 |
attn_aggregate = attn_utils.aggregate_attention( # [res, res, 77]
|
| 400 |
-
prompts=[self.config.prompt for i in range(bsz)],
|
| 401 |
attention_store=self.controller,
|
| 402 |
res=res,
|
| 403 |
from_where=("up", "down"),
|
|
@@ -408,13 +379,13 @@ class TrackingModule(pl.LightningModule):
|
|
| 408 |
task_attn_ = attn_aggregate[:, :, 1].unsqueeze(0).unsqueeze(0) # [1, 1, res, res]
|
| 409 |
attention_maps.append(task_attn_)
|
| 410 |
# if self.boxes is not None and not self.training:
|
| 411 |
-
exemplar_attns1 = attn_aggregate[:, :, 2].unsqueeze(0).unsqueeze(0)
|
| 412 |
exemplar_attention_maps1.append(exemplar_attns1)
|
| 413 |
-
exemplar_attns2 = attn_aggregate[:, :, 3].unsqueeze(0).unsqueeze(0)
|
| 414 |
exemplar_attention_maps2.append(exemplar_attns2)
|
| 415 |
-
exemplar_attns3 = attn_aggregate[:, :, 4].unsqueeze(0).unsqueeze(0)
|
| 416 |
exemplar_attention_maps3.append(exemplar_attns3)
|
| 417 |
-
exemplar_attns4 = attn_aggregate[:, :, 5].unsqueeze(0).unsqueeze(0)
|
| 418 |
exemplar_attention_maps4.append(exemplar_attns4)
|
| 419 |
|
| 420 |
|
|
@@ -540,8 +511,7 @@ class TrackingModule(pl.LightningModule):
|
|
| 540 |
|
| 541 |
for n in range(n_forward):
|
| 542 |
len_ = min(74, n_instance - n * 74)
|
| 543 |
-
encoder_hidden_states[:,(task_loc_idx[0, 1]+1):(task_loc_idx[0, 1]+1+len_),:] = adapted_emb[n*74:n*74+len_].squeeze()
|
| 544 |
-
# encoder_hidden_states: # [bsz, 77, 768], 其中第1位是task prompt的embedding, 第二位开始可以是object prompt的embedding, 最后一位应该保留原始embedding
|
| 545 |
|
| 546 |
|
| 547 |
# Predict the noise residual
|
|
@@ -556,7 +526,7 @@ class TrackingModule(pl.LightningModule):
|
|
| 556 |
# cross attention
|
| 557 |
for res in [32, 16]:
|
| 558 |
attn_aggregate = attn_utils.aggregate_attention( # [res, res, 77]
|
| 559 |
-
prompts=[self.config.prompt for i in range(bsz)],
|
| 560 |
attention_store=self.controller,
|
| 561 |
res=res,
|
| 562 |
from_where=("up", "down"),
|
|
@@ -567,7 +537,7 @@ class TrackingModule(pl.LightningModule):
|
|
| 567 |
task_attn_ = attn_aggregate[:, :, 1].unsqueeze(0).unsqueeze(0) # [1, 1, res, res]
|
| 568 |
attention_maps.append(task_attn_)
|
| 569 |
try:
|
| 570 |
-
exemplar_attns = attn_aggregate[:, :, (task_loc_idx[0, 1]+1):(task_loc_idx[0, 1]+1+len_)].unsqueeze(0)
|
| 571 |
except:
|
| 572 |
print(n_instance, len_)
|
| 573 |
exemplar_attns = torch.permute(exemplar_attns, (0, 3, 1, 2)) # [1, len_, res, res]
|
|
@@ -728,11 +698,6 @@ class TrackingModule(pl.LightningModule):
|
|
| 728 |
|
| 729 |
A = self.track_model.normalize_output(A, timepoints, coords)
|
| 730 |
|
| 731 |
-
# # Spatially far entries should not influence the causal normalization
|
| 732 |
-
# dist = torch.cdist(coords[0, :, 1:], coords[0, :, 1:])
|
| 733 |
-
# invalid = dist > model.config["spatial_pos_cutoff"]
|
| 734 |
-
# A[invalid] = -torch.inf
|
| 735 |
-
|
| 736 |
A = A.squeeze(0).detach().cpu().numpy()
|
| 737 |
|
| 738 |
del feats, coords, timepoints, batch
|
|
@@ -1020,30 +985,3 @@ class TrackingModule(pl.LightningModule):
|
|
| 1020 |
track_graph = self._track_from_predictions(predictions, mode=mode, **kwargs)
|
| 1021 |
|
| 1022 |
return track_graph, masks
|
| 1023 |
-
|
| 1024 |
-
|
| 1025 |
-
|
| 1026 |
-
# def inference(data_path, box=None):
|
| 1027 |
-
# if box is not None:
|
| 1028 |
-
# use_box = True
|
| 1029 |
-
# else:
|
| 1030 |
-
# use_box = False
|
| 1031 |
-
|
| 1032 |
-
# model = TrackingModule(use_box=use_box)
|
| 1033 |
-
# load_msg = model.load_state_dict(torch.load("pretrained/microscopy_matching_tra.pth"), strict=True)
|
| 1034 |
-
|
| 1035 |
-
# model.move_to_device(torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu'))
|
| 1036 |
-
|
| 1037 |
-
|
| 1038 |
-
# track_graph, masks = model.track(file_dir=data_path, dataname="inference_sequence")
|
| 1039 |
-
|
| 1040 |
-
# if not os.path.exists(f"tracked_ours_seg_pred3/"):
|
| 1041 |
-
# os.makedirs(f"tracked_ours_seg_pred3/")
|
| 1042 |
-
# ctc_tracks, masks_tracked = graph_to_ctc(
|
| 1043 |
-
# track_graph,
|
| 1044 |
-
# masks,
|
| 1045 |
-
# outdir=f"tracked_ours_seg_pred3/",
|
| 1046 |
-
# )
|
| 1047 |
-
|
| 1048 |
-
# if __name__ == "__main__":
|
| 1049 |
-
# inference(data_path="example_imgs/2D+Time/Fluo-N2DL-HeLa/train/Fluo-N2DL-HeLa/02")
|
|
|
|
| 1 |
import os
|
|
|
|
| 2 |
from typing import Any, List, Optional
|
|
|
|
| 3 |
from huggingface_hub import hf_hub_download
|
|
|
|
| 4 |
from pytorch_lightning.utilities.types import STEP_OUTPUT
|
| 5 |
import torch
|
|
|
|
| 6 |
from PIL import Image
|
| 7 |
import numpy as np
|
| 8 |
import tifffile
|
|
|
|
| 9 |
from config import RunConfig
|
| 10 |
from _utils import attn_utils_new as attn_utils
|
| 11 |
from _utils.attn_utils_new import AttentionStore
|
|
|
|
| 13 |
import torch.nn.functional as F
|
| 14 |
from tqdm import tqdm
|
| 15 |
import torch.nn as nn
|
|
|
|
| 16 |
import cv2
|
| 17 |
import warnings
|
| 18 |
warnings.filterwarnings("ignore", category=UserWarning)
|
|
|
|
| 27 |
)
|
| 28 |
from models.tra_post_model.data import build_windows_sd, get_features
|
| 29 |
from models.tra_post_model.tracking import TrackGraph, build_graph, track_greedy
|
|
|
|
| 30 |
import torchvision.transforms as T
|
| 31 |
from pathlib import Path
|
| 32 |
import dask.array as da
|
|
|
|
| 34 |
from scipy.sparse import SparseEfficiencyWarning, csr_array
|
| 35 |
import tracemalloc
|
| 36 |
import gc
|
|
|
|
| 37 |
from _utils.load_track_data import load_track_images
|
| 38 |
|
| 39 |
SCALE = 1
|
|
|
|
| 74 |
|
| 75 |
# load loca model
|
| 76 |
self.loca_model = build_loca_model()
|
|
|
|
|
|
|
|
|
|
|
|
|
| 77 |
|
| 78 |
self.counting_adapter = Counting(scale_factor=SCALE)
|
|
|
|
|
|
|
|
|
|
| 79 |
|
| 80 |
### load stable diffusion and its controller
|
| 81 |
self.stable = load_stable_diffusion_model(config=self.config)
|
|
|
|
| 95 |
" `placeholder_token` that is not already in the tokenizer."
|
| 96 |
)
|
| 97 |
try:
|
|
|
|
| 98 |
task_embed_from_pretrain = hf_hub_download(
|
| 99 |
repo_id="phoebe777777/111",
|
| 100 |
filename="task_embed.pth",
|
|
|
|
| 128 |
self.placeholder_token_id = placeholder_token_id
|
| 129 |
|
| 130 |
fpath = Path("_utils/config.yaml")
|
|
|
|
| 131 |
|
| 132 |
model = TrackingTransformer.from_cfg(
|
| 133 |
cfg_path=fpath,
|
|
|
|
| 134 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 135 |
|
| 136 |
self.track_model = model
|
|
|
|
| 137 |
|
| 138 |
|
| 139 |
def move_to_device(self, device):
|
| 140 |
self.stable.to(device)
|
|
|
|
|
|
|
|
|
|
| 141 |
self.counting_adapter.to(device)
|
|
|
|
| 142 |
self.loca_model.to(device)
|
| 143 |
self.track_model.to(device)
|
| 144 |
|
|
|
|
| 192 |
adapted_emb = self.counting_adapter.adapter(loca_feature_bf_regression, boxes) # shape [1, 768]
|
| 193 |
|
| 194 |
if task_loc_idx.shape[0] == 0:
|
| 195 |
+
encoder_hidden_states[0,2,:] = adapted_emb.squeeze()
|
| 196 |
else:
|
| 197 |
+
encoder_hidden_states[:,task_loc_idx[0, 1]+1,:] = adapted_emb.squeeze()
|
| 198 |
|
| 199 |
# Predict the noise residual
|
| 200 |
noise_pred, feature_list = self.stable.unet(noisy_latents, timesteps, encoder_hidden_states)
|
|
|
|
| 213 |
|
| 214 |
# only use 64x64 self-attention
|
| 215 |
self_attn_aggregate = attn_utils.aggregate_attention( # [res, res, 4096]
|
| 216 |
+
prompts=[self.config.prompt for i in range(bsz)],
|
| 217 |
attention_store=self.controller,
|
| 218 |
res=64,
|
| 219 |
from_where=("up", "down"),
|
|
|
|
| 221 |
select=0
|
| 222 |
)
|
| 223 |
self_attn_aggregate32 = attn_utils.aggregate_attention( # [res, res, 4096]
|
| 224 |
+
prompts=[self.config.prompt for i in range(bsz)],
|
| 225 |
attention_store=self.controller,
|
| 226 |
res=32,
|
| 227 |
from_where=("up", "down"),
|
|
|
|
| 229 |
select=0
|
| 230 |
)
|
| 231 |
self_attn_aggregate16 = attn_utils.aggregate_attention( # [res, res, 4096]
|
| 232 |
+
prompts=[self.config.prompt for i in range(bsz)],
|
| 233 |
attention_store=self.controller,
|
| 234 |
res=16,
|
| 235 |
from_where=("up", "down"),
|
|
|
|
| 240 |
# cross attention
|
| 241 |
for res in [32, 16]:
|
| 242 |
attn_aggregate = attn_utils.aggregate_attention( # [res, res, 77]
|
| 243 |
+
prompts=[self.config.prompt for i in range(bsz)],
|
| 244 |
attention_store=self.controller,
|
| 245 |
res=res,
|
| 246 |
from_where=("up", "down"),
|
|
|
|
| 250 |
|
| 251 |
task_attn_ = attn_aggregate[:, :, 1].unsqueeze(0).unsqueeze(0) # [1, 1, res, res]
|
| 252 |
attention_maps.append(task_attn_)
|
| 253 |
+
exemplar_attns = attn_aggregate[:, :, 2].unsqueeze(0).unsqueeze(0)
|
| 254 |
exemplar_attention_maps.append(exemplar_attns)
|
| 255 |
|
| 256 |
|
|
|
|
| 277 |
attn_stack = torch.cat(attn_stack, dim=1)
|
| 278 |
|
| 279 |
|
| 280 |
+
attn_after_new_regressor, loss = self.counting_adapter.regressor(input_image, attn_stack, feature_list, mask.cpu().numpy(), training=False)
|
| 281 |
|
| 282 |
return {
|
| 283 |
"attn_after_new_regressor":attn_after_new_regressor,
|
|
|
|
| 335 |
adapted_emb = self.adapt_emb.to(self.device)
|
| 336 |
task_loc_idx = torch.nonzero(input_ids == self.placeholder_token_id)
|
| 337 |
if task_loc_idx.shape[0] == 0:
|
| 338 |
+
encoder_hidden_states[0,5,:] = adapted_emb.squeeze()
|
| 339 |
else:
|
| 340 |
+
encoder_hidden_states[:,task_loc_idx[0, 1]+4,:] = adapted_emb.squeeze()
|
| 341 |
|
| 342 |
# Predict the noise residual
|
| 343 |
noise_pred, feature_list = self.stable.unet(noisy_latents, timesteps, encoder_hidden_states)
|
|
|
|
| 357 |
|
| 358 |
# only use 64x64 self-attention
|
| 359 |
self_attn_aggregate = attn_utils.aggregate_attention( # [res, res, 4096]
|
| 360 |
+
prompts=[self.config.prompt for i in range(bsz)],
|
| 361 |
attention_store=self.controller,
|
| 362 |
res=64,
|
| 363 |
from_where=("up", "down"),
|
|
|
|
| 368 |
# cross attention
|
| 369 |
for res in [32, 16]:
|
| 370 |
attn_aggregate = attn_utils.aggregate_attention( # [res, res, 77]
|
| 371 |
+
prompts=[self.config.prompt for i in range(bsz)],
|
| 372 |
attention_store=self.controller,
|
| 373 |
res=res,
|
| 374 |
from_where=("up", "down"),
|
|
|
|
| 379 |
task_attn_ = attn_aggregate[:, :, 1].unsqueeze(0).unsqueeze(0) # [1, 1, res, res]
|
| 380 |
attention_maps.append(task_attn_)
|
| 381 |
# if self.boxes is not None and not self.training:
|
| 382 |
+
exemplar_attns1 = attn_aggregate[:, :, 2].unsqueeze(0).unsqueeze(0)
|
| 383 |
exemplar_attention_maps1.append(exemplar_attns1)
|
| 384 |
+
exemplar_attns2 = attn_aggregate[:, :, 3].unsqueeze(0).unsqueeze(0)
|
| 385 |
exemplar_attention_maps2.append(exemplar_attns2)
|
| 386 |
+
exemplar_attns3 = attn_aggregate[:, :, 4].unsqueeze(0).unsqueeze(0)
|
| 387 |
exemplar_attention_maps3.append(exemplar_attns3)
|
| 388 |
+
exemplar_attns4 = attn_aggregate[:, :, 5].unsqueeze(0).unsqueeze(0)
|
| 389 |
exemplar_attention_maps4.append(exemplar_attns4)
|
| 390 |
|
| 391 |
|
|
|
|
| 511 |
|
| 512 |
for n in range(n_forward):
|
| 513 |
len_ = min(74, n_instance - n * 74)
|
| 514 |
+
encoder_hidden_states[:,(task_loc_idx[0, 1]+1):(task_loc_idx[0, 1]+1+len_),:] = adapted_emb[n*74:n*74+len_].squeeze()
|
|
|
|
| 515 |
|
| 516 |
|
| 517 |
# Predict the noise residual
|
|
|
|
| 526 |
# cross attention
|
| 527 |
for res in [32, 16]:
|
| 528 |
attn_aggregate = attn_utils.aggregate_attention( # [res, res, 77]
|
| 529 |
+
prompts=[self.config.prompt for i in range(bsz)],
|
| 530 |
attention_store=self.controller,
|
| 531 |
res=res,
|
| 532 |
from_where=("up", "down"),
|
|
|
|
| 537 |
task_attn_ = attn_aggregate[:, :, 1].unsqueeze(0).unsqueeze(0) # [1, 1, res, res]
|
| 538 |
attention_maps.append(task_attn_)
|
| 539 |
try:
|
| 540 |
+
exemplar_attns = attn_aggregate[:, :, (task_loc_idx[0, 1]+1):(task_loc_idx[0, 1]+1+len_)].unsqueeze(0)
|
| 541 |
except:
|
| 542 |
print(n_instance, len_)
|
| 543 |
exemplar_attns = torch.permute(exemplar_attns, (0, 3, 1, 2)) # [1, len_, res, res]
|
|
|
|
| 698 |
|
| 699 |
A = self.track_model.normalize_output(A, timepoints, coords)
|
| 700 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 701 |
A = A.squeeze(0).detach().cpu().numpy()
|
| 702 |
|
| 703 |
del feats, coords, timepoints, batch
|
|
|
|
| 985 |
track_graph = self._track_from_predictions(predictions, mode=mode, **kwargs)
|
| 986 |
|
| 987 |
return track_graph, masks
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|