Dockerfile fix + volume dep + update INSTALL.md

INSTALL.md

@@ -54,3 +54,15 @@ docker compose up
 docker compose run -i --entrypoint /bin/bash igpt
 
 ```
+
+## A note on dependencies:
+
+CUDA, torch, and opencv are notorious for version incompatibilities (mostly due to binaries being built against specific versions), so here's a quick checklist:
+
+1. `opencv` libraries should always have specified version numbers, and there should only ever be one set on the system
+
+2. `cuda` libraries should always have specified version numbers, but can exist independently using conda
+
+3. `torch` libraries should always have specified version numbers, especially for torchvision [here](https://pypi.org/project/torchvision/) and torchaudio [here](https://pytorch.org/audio/main/installation.html); always check `pytorch` and `python` versions before building
+
+If deploying as Swarm or via Kubernetes, note that the build directory should have access to tty, and individual clients should be paired with individual gpus or gpu partitions. See [here](https://docs.nvidia.com/datacenter/tesla/mig-user-guide/index.html).

docker/Dockerfile

@@ -2,21 +2,25 @@ FROM continuumio/anaconda3
 
 RUN git clone https://github.com/OpenGVLab/InternGPT.git
 WORKDIR InternGPT
+
 SHELL ["/bin/bash", "-c"]
 RUN conda init bash
 RUN source "/opt/conda/bin/activate"
 RUN conda create -n igpt python=3.8
 RUN source activate igpt
 
-RUN apt-get clean && apt-get update && apt-get -y install --no-install-recommends gcc g++ libjpeg-dev libpng-dev zlib1g-dev build-essential
+RUN apt-get clean && apt-get update && apt-get -y install --no-install-recommends gcc g++ libjpeg-dev libpng-dev zlib1g-dev build-essential libgl1-mesa-glx libxml2
 
 RUN cp /opt/conda/envs/igpt/lib/libstdc++.so.6 /usr/lib/x86_64-linux-gnu
-RUN conda install pytorch pytorch-cuda=11.7 cudatoolkit -c pytorch -c nvidia
+
+RUN conda update -n base -c defaults conda
+RUN conda install pytorch pytorch-cuda=11.7 cudatoolkit=11.7 -c pytorch -c nvidia
+RUN conda install -c conda-forge cudatoolkit-dev=11.7
+RUN conda install -c numba llvmlite numba
 
 RUN pip install -r requirements.txt
 RUN pip install git+https://github.com/facebookresearch/detectron2.git
 RUN pip install mediapipe
-RUN pip install imageio-ffmpeg
 
 RUN pip uninstall -y opencv-python  
 RUN pip uninstall -y opencv-python-headless
@@ -25,12 +29,9 @@ RUN pip uninstall -y opencv-contrib-python-headless
 RUN pip install opencv-python-headless==4.6.0.66
 RUN pip install opencv-contrib-python==4.6.0.66
 
-RUN apt-get -y install libgl1-mesa-glx
-RUN apt-get -y install libxml2
+RUN conda install torchvision=0.14.0 torchaudio=0.13.0 -c pytorch
 
-RUN conda install -c conda-forge cudatoolkit-dev
-RUN conda install --channel=numba llvmlite
-RUN conda install -c numba numba
+RUN git pull
 
 EXPOSE 7862
 

docker/docker-compose.yml

@@ -1,27 +1,27 @@
 services:
   igpt:
-    build: .
     image: "igpt"
     container_name: "igpt"
     restart: "unless-stopped"
     ports:
       - "7862:7862"
     volumes:
-      - /path/to/model_zoo:/InternGPT/model_zoo
-      - /path/to/certificate:/InternGPT/certificate
+      - /data/InternGPT/model_zoo:/InternGPT/model_zoo
+      - /data/InternGPT/certificate:/InternGPT/certificate
+      - /data/InternGPT/checkpoints:/InternGPT/checkpoints
+      - /data/InternGPT/.EasyOCR:/root/.EasyOCR/model
+      - /data/stable-diffusion-webui-docker/data/image_output:/InternGPT/image
     entrypoint: "python"
     deploy:
       resources:
         reservations:
           devices:
             - driver: nvidia
-              count: 1
+              count: all
               capabilities: [gpu]
     command:
       - "-u"
       - "app.py"
       - "--load"
-      - "StyleGAN_cuda:0"
-      - "--tab"
-      - "DragGAN"
+      - "HuskyVQA_cuda:0,SegmentAnything_cuda:0,Anything2Image_cuda:0"
       - "--https"

iGPT/models/image.py

@@ -670,7 +670,7 @@ def inference_by_mask(self, inputs):
         mask_img.save(filaname, "PNG")
         return filaname
 
-    def segment_by_mask(self, mask, features):
+    def segment_by_mask(self, mask, features=None):
         random.seed(GLOBAL_SEED)
         idxs = np.nonzero(mask)
         num_points = min(max(1, int(len(idxs[0]) * 0.01)), 16)
@@ -681,12 +681,19 @@ def segment_by_mask(self, mask, features):
         points = np.array(new_mask).reshape(2, -1).transpose(1, 0)[:, ::-1]
         labels = np.array([1] * num_points)
 
-        res_masks, scores, _ = self.predictor.predict(
-            features=features,
-            point_coords=points,
-            point_labels=labels,
-            multimask_output=True,
-        )
+        if features is None:
+            res_masks, scores, _ = self.predictor.predict(
+                point_coords=points,
+                point_labels=labels,
+                multimask_output=True,
+            )
+        else:
+            res_masks, scores, _ = self.predictor.predict(
+                features=features,
+                point_coords=points,
+                point_labels=labels,
+                multimask_output=True,
+            )
 
         return res_masks[np.argmax(scores), :, :]