Bokeh chart updates. Fixed class colors. Filter classes. Added controls to adjust dots size. New option to assign tags or create labeling jobs.

.devcontainer/Dockerfile

@@ -0,0 +1,28 @@
+FROM nvidia/cuda:12.1.0-devel-ubuntu22.04
+ENV DEBIAN_FRONTEND=noninteractive
+RUN apt update && apt install python3-pip -y
+RUN apt-get install -y git
+RUN apt-get update && apt-get install ffmpeg libsm6 libxext6  -y
+
+RUN pip3 install torch==1.11.0 torchvision==0.12.0
+
+RUN pip3 install transformers==4.33.2 timm==0.9.5 scikit-learn==1.3.1 umap-learn==0.5.4
+
+# Download metaclip base model
+RUN python3 -c 'import transformers; transformers.AutoModel.from_pretrained("facebook/metaclip-b16-fullcc2.5b")'
+
+RUN pip3 install supervisely==6.73.258
+RUN pip3 install fastapi==0.109.0 bokeh==3.1.1
+LABEL python_sdk_version=6.73.258
+
+RUN pip3 install ruamel.yaml==0.17.21
+
+RUN apt-get update
+
+RUN apt-get -y install curl
+
+RUN apt -y install wireguard iproute2
+RUN apt-get -y install wget
+RUN apt-get install nano
+
+LABEL "role"="development"

.devcontainer/devcontainer.json

@@ -0,0 +1,20 @@
+{
+    "name": "EmbeddingsDevContainer",
+    "build": {
+        "dockerfile": "Dockerfile"
+    },
+    "customizations": {
+        "vscode": {
+            "extensions": [
+                "ms-python.python",
+                "ms-python.black-formatter"
+            ]
+        }
+    },
+    "runArgs": [
+        "--ipc=host",
+        "--net=host",
+        "--cap-add",
+        "NET_ADMIN"
+    ]
+}

config.json

@@ -4,7 +4,7 @@
   "name": "Explore data with embeddings",
   "description": "Calculate and visualize embeddings",
   "entrypoint": "python -m uvicorn src.main:app --host 0.0.0.0 --port 8000",
-  "docker_image": "supervisely/embeddings_app:1.0.3",
+  "docker_image": "supervisely/embeddings_app:1.0.4",
   "port": 8000,
   "icon": "https://user-images.githubusercontent.com/115161827/211832834-eef9325d-9e00-4499-ae42-2e7ee3d97f1a.png",
   "icon_cover": true,
@@ -26,7 +26,7 @@
     "target": ["images_project", "images_dataset"],
     "context_category": "Metric Learning"
   },
-  "instance_version": "6.12.12",
+  "instance_version": "6.12.17",
   "need_gpu": false,
   "gpu": "preferred",
   "community_agent": false

dev_requirements.txt

@@ -1,4 +1,4 @@
-supervisely==6.73.258
+supervisely==6.73.266
 transformers==4.33.2
 timm==0.9.5
 torch

docker/Dockerfile

@@ -11,7 +11,7 @@ RUN pip3 install transformers==4.33.2 timm==0.9.5 scikit-learn==1.3.1 umap-learn
 # Download metaclip base model
 RUN python -c 'import transformers; transformers.AutoModel.from_pretrained("facebook/metaclip-b16-fullcc2.5b")'
 
-RUN pip3 install supervisely==6.73.258
+RUN pip3 install supervisely==6.73.266
 RUN pip3 install fastapi==0.109.0 bokeh==3.1.1
 
-LABEL python_sdk_version=6.73.258
+LABEL python_sdk_version=6.73.266

src/main.py

@@ -7,9 +7,8 @@
 from supervisely.app.content import StateJson, DataJson
 from dotenv import load_dotenv
 import torch
-import re
+from typing import List, Union, Tuple
 from supervisely.app.widgets import (
-    ScatterChart,
     Container,
     Card,
     LabeledImage,
@@ -23,19 +22,20 @@
     Button,
     Field,
     Progress,
-    SelectDataset,
     IFrame,
     Bokeh,
     SelectDatasetTree,
     NotificationBox,
+    Empty,
+    Flexbox,
 )
 
 from . import run_utils
 from . import calculate_embeddings
 
 
 def update_globals(new_dataset_ids):
-    global dataset_ids, project_id, workspace_id, team_id, project_info, project_meta, is_marked, tag_meta
+    global dataset_ids, project_id, workspace_id, team_id, project_info, project_meta, is_marked, tag_meta, issue_tag_meta
     dataset_ids = new_dataset_ids
     if dataset_ids:
         project_id = api.dataset.get_info_by_id(dataset_ids[0]).project_id
@@ -57,15 +57,20 @@ def update_globals(new_dataset_ids):
         is_marked = False
         tag_meta = project_meta.get_tag_meta(tag_name)
         print("tag_meta is exists:", bool(tag_meta))
+        issue_tag_meta = project_meta.get_tag_meta(issue_tag_name)
+        print("issue_tag_meta is exists:", bool(issue_tag_meta))
 
 
 ### Globals init
 available_projection_methods = ["UMAP", "PCA", "t-SNE", "PCA-UMAP", "PCA-t-SNE"]
 tag_name = "MARKED"
+issue_tag_name = "ISSUE"
+instance_mode = None
 
-load_dotenv("local.env")
-load_dotenv(os.path.expanduser("~/supervisely.env"))
-api = sly.Api()
+if sly.is_development():
+    load_dotenv("local.env")
+    load_dotenv(os.path.expanduser("~/supervisely.env"))
+api = sly.Api.from_env()
 
 # if app had started from context menu, one of this has to be set:
 project_id = sly.env.project_id(raise_not_found=False)
@@ -169,18 +174,50 @@ def update_globals(new_dataset_ids):
 content = Container([btn_run, check_force_recalculate, progress, info_run])
 card_run = Card(title="Run", content=content)
 
+### Embeddings Chart Settings
+dot_size_btn = Button("Change dots size", button_size="small")
+dot_size_num = InputNumber(min=0.01, value=0.05, step=0.01)
+dot_size = Flexbox([Container([dot_size_num, dot_size_btn], direction="horizontal")])
 
 ### Embeddings Chart
-bokeh = Bokeh(plots=[], x_axis_visible=True, y_axis_visible=True, grid_visible=True)
+bokeh = Bokeh(
+    plots=[],
+    x_axis_visible=True,
+    y_axis_visible=True,
+    grid_visible=True,
+    show_legend=True,
+    legend_location="right",
+    legend_click_policy="hide",
+)
 bokeh_iframe = IFrame()
-card_chart = Card(content=bokeh_iframe)
+card_chart = Card(content=Container([dot_size, bokeh_iframe]), title="Embeddings chart", collapsable=True)
 labeled_image = LabeledImage()
 text = Text("no object selected")
 show_all_anns = False
 cur_info = None
 btn_toggle = Button(f"Show all annotations: {show_all_anns}", "default", button_size="small")
 btn_mark = Button(f"Assign tag 'MARKED'", button_size="small")
-card_preview = Card(title="Object preview", content=Container(widgets=[labeled_image, text, btn_toggle, btn_mark]))
+preview_widgets = Container([labeled_image, text, btn_toggle, btn_mark])
+preview_widgets.hide()
+
+
+cur_infos = None
+batch_text = Text()
+issue_tag_text = Text()
+add_issue_tag = Button(f"Asign 'ISSUE' tags", button_size="small", plain=True)
+job_issue = Button(f"Create Labeling Job", button_size="small", plain=True)
+batch_tagging_field = Field(
+    Container([Flexbox([add_issue_tag, Empty(), job_issue]), issue_tag_text]),
+    "Issues",
+    "Assign 'ISSUE' tag to IMAGES or Create Labeling Job",
+)
+batch_tagging_cont = Container([batch_text, batch_tagging_field])
+batch_tagging_cont.hide()
+
+card_preview = Card(
+    title="Preview card",
+    content=Container(widgets=[preview_widgets, batch_tagging_cont]),
+)
 card_embeddings_chart = Container(widgets=[card_chart, card_preview], direction="horizontal", fractions=[3, 1])
 card_embeddings_chart.hide()
 
@@ -199,6 +236,12 @@ def update_globals(new_dataset_ids):
 )
 
 
+@dot_size_btn.click
+def change_dot_size():
+    bokeh.update_radii(dot_size_num.value)
+    bokeh_iframe.set(bokeh.html_route_with_timestamp, height="650px", width="100%")
+
+
 @btn_toggle.click
 def toggle_ann():
     global show_all_anns
@@ -209,17 +252,65 @@ def toggle_ann():
 
 
 @bokeh.value_changed
-def on_click(selected_idxs):
-    global global_idxs_mapping, all_info_list, project_meta, is_marked, tag_meta
-    if len(selected_idxs) >= 1:
-        info = all_info_list[selected_idxs[0]]
+def on_click(selected_idxs: List[Tuple[Union[int, str], List[int]]]):
+    global global_idxs_mapping, all_info_list, project_meta, is_marked, tag_meta, cur_infos
+
+    issue_tag_text.text = ""
+    batch_tagging_cont.show()
+
+    selected_ids = [global_idxs_mapping[d.plot_id][i] for d in selected_idxs for i in d.selected_ids]
+    selected_cnt = len(selected_ids)
+    if selected_cnt == 1:
+        batch_text.text = ""
+        preview_widgets.show()
+        info = all_info_list[selected_ids[0]]
+        cur_infos = [info]
         if tag_meta is not None:
             tag = read_tag(info["image_id"], info["object_id"])
             is_marked = bool(tag)
             update_marked()
         show_image(info, project_meta)
         if btn_mark.is_hidden():
             btn_mark.show()
+    elif selected_cnt > 1:
+        preview_widgets.hide()
+        cur_infos = [all_info_list[i] for i in selected_ids]
+        obj_clss = list(set([info["object_cls"] for info in cur_infos]))
+        is_objects = any([info["object_id"] is not None for info in cur_infos])
+        is_images = any([info["object_id"] is None for info in cur_infos])
+        both = is_objects and is_images
+
+        t = f"{len(cur_infos)} "
+        t += "items. " if both else "images. " if is_images else "objects. "
+        t += f"Object classes: {str(obj_clss)}. "
+        batch_text.set(t, "info")
+
+
+@job_issue.click
+def create_labeling_job():
+    global cur_infos
+    issue_tag_text.text = ""
+    if cur_infos is not None:
+        ds_id_to_img_ids = defaultdict(set)
+        for info in cur_infos:
+            ds_id_to_img_ids[info["dataset_id"]].add(info["image_id"])
+        jobs = []
+        for ds_id, img_ids in ds_id_to_img_ids.items():
+            if len(img_ids) > 0:
+                jobs.extend(
+                    api.labeling_job.create(
+                        f"Labeling job for {project_info.name} project embeddings",
+                        ds_id,
+                        [api.user.get_my_info().id],
+                        images_ids=list(img_ids),
+                        # include_images_with_tags=[issue_tag_name],
+                    )
+                )
+        if len(jobs) > 0:
+            ids = [job.id for job in jobs]
+            issue_tag_text.set(f"Labeling jobs created IDs: {ids}", "success")
+        else:
+            issue_tag_text.set("No objects to create labeling job", "warning")
 
 
 def update_marked():
@@ -230,6 +321,34 @@ def update_marked():
         btn_mark.text = "Assign tag 'MARKED'"
 
 
+@add_issue_tag.click
+def issue_tagging():
+    global project_meta, cur_infos, issue_tag_meta
+    if issue_tag_meta is None:
+        print("first marking, creating tag_meta")
+        issue_tag_meta = sly.TagMeta(issue_tag_name, sly.TagValueType.NONE)
+        project_meta, issue_tag_meta = get_or_create_tag_meta(project_id, issue_tag_meta)
+
+    ds_ids_to_img_ids = defaultdict(set)
+    for info in cur_infos:
+        ds_ids_to_img_ids[info["dataset_id"]].add(info["image_id"])
+
+    added = 0
+    for ds_id, img_ids in ds_ids_to_img_ids.items():
+        img_ids_to_mark = []
+        img_ids = list(img_ids)
+        for img_id, tag in zip(img_ids, read_img_tags(ds_id, img_ids, issue_tag_meta)):
+            if tag is None:
+                img_ids_to_mark.append(img_id)
+
+        if len(img_ids_to_mark) > 0:
+            add_img_tags(list(img_ids_to_mark), issue_tag_meta)
+            added += len(img_ids_to_mark)
+
+    if added > 0:
+        issue_tag_text.set(f"Assigned 'ISSUE' tags: {added} images", "success")
+
+
 @btn_mark.click
 def on_mark():
     global project_info, project_meta, tag_meta, cur_info, is_marked
@@ -280,7 +399,7 @@ def update_table():
 
 @btn_run.click
 def run():
-    global model_name, global_idxs_mapping, all_info_list  # , project_meta, dataset_ids, project_id, workspace_id, team_id
+    global model_name, global_idxs_mapping, all_info_list, instance_mode
 
     selected_datasets = set()
     for dataset_id in dataset_selector.get_selected_ids():
@@ -385,17 +504,22 @@ def run():
     print(f"n_classes = {len(obj_classes)}")
     series, pre_colors, global_idxs_mapping = run_utils.make_series(projections, all_info_list, project_meta)
 
-    series_len = len(series)
-    x_coordinates, y_coordinates, colors = [], [], []
-    for s, color in zip(series, pre_colors):
-        x_coordinates.extend([i["x"] for i in s["data"]])
-        y_coordinates.extend([i["y"] for i in s["data"]])
-        colors.extend([color] * len(s["data"]))
-
-    r = 0.15 if series_len > 1000 else 0.1
-    plot = Bokeh.Circle(x_coordinates, y_coordinates, radii=r, colors=colors)
     bokeh.clear()
-    bokeh.add_plots([plot])
+    plots = []
+    for s, color in zip(series, pre_colors):
+        x_coordinates = [i["x"] for i in s["data"]]
+        y_coordinates = [i["y"] for i in s["data"]]
+        r = 0.05
+        plot = Bokeh.Circle(
+            x_coordinates,
+            y_coordinates,
+            radii=r,
+            colors=[color] * len(s["data"]),
+            legend_label=s["name"],
+            plot_id=s["name"],
+        )
+        plots.append(plot)
+    bokeh.add_plots(plots)
     bokeh_iframe.set(bokeh.html_route_with_timestamp, height="650px", width="100%")
     card_embeddings_chart.show()
     update_table()
@@ -421,6 +545,24 @@ def get_tag_meta(project_id, name) -> sly.TagMeta:
     return project_meta.get_tag_meta(name)
 
 
+def read_img_tags(ds_id, image_ids, tag_meta):
+    tags = []
+    if len(image_ids) > 0:
+        filters = [{"field": "id", "operator": "in", "value": image_ids}]
+        image_infos = api.image.get_list(ds_id, filters=filters, force_metadata_for_links=False)
+        id_to_info = {img_info.id: img_info for img_info in image_infos}
+        for img_id in image_ids:
+            curr_tags = [tag for tag in id_to_info[img_id].tags if tag["tagId"] == tag_meta.sly_id]
+            tags.append(curr_tags[0] if len(curr_tags) == 1 else None)
+    return tags
+
+
+def read_labels_tags(object_ids, tag_meta):
+    if len(object_ids) == 0:
+        return []
+    return [read_label_tag(obj_id, tag_meta) for obj_id in object_ids]
+
+
 def read_img_tag(image_id, tag_meta):
     image_info = api.image.get_info_by_id(image_id)
     tags = [tag for tag in image_info.tags if tag["tagId"] == tag_meta.sly_id]
@@ -444,6 +586,10 @@ def read_tag(image_id, object_id):
         return read_label_tag(object_id, tag_meta)
 
 
+def add_img_tags(image_ids, tag_meta, value=None):
+    return api.image.add_tag_batch(image_ids=image_ids, tag_id=tag_meta.sly_id, value=value)
+
+
 def add_img_tag(image_id, tag_meta, value=None):
     return api.image.add_tag(image_id=image_id, tag_id=tag_meta.sly_id, value=value)