Interview

.env_example

@@ -0,0 +1,11 @@
+JUPYTER_PASSWORD=password
+DJ_HOST=134.2.168.16
+DJ_USER=djuser
+DJ_PASS=keyboard-stroke-face-roll
+MINIO_ENDPOINT=endpoint
+MINIO_ACCESS_KEY=pass
+MINIO_SECRET_KEY=pass
+GITHUB_TOKEN=pass
+GITHUB_USER=gituser
+DEV_SOURCE=gituser
+BASE_IMAGE=sinzlab/pytorch:v3.8-torch1.5.0-cuda10.2-dj0.12.4

.gitignore

@@ -0,0 +1,27 @@
+secrets/
+*MNIST/
+*egg-info/
+.log
+.ftpconfig
+*.pyc
+*.pyd
+*~
+.idea
+.cache
+.ipynb_checkpoints
+*.h5
+*.env
+*.jpg
+*.gif
+.vscode
+.env
+.env*
+*.env
+.env_pipeline
+.env/
+logs/
+*.sif
+*.zip
+*.json
+*.log
+*.csv

Dockerfile

@@ -1,11 +1,30 @@
-FROM sinzlab/pytorch:latest
+FROM sinzlab/pytorch:v3.8-torch1.7.0-cuda11.0-dj0.12.7
+ENV DEBIAN_FRONTEND=noninteractive
 
-ADD . /src/nnfabrik
-WORKDIR /src
+# Install JupyterLab
+RUN python3 -m pip install --upgrade pip
+RUN python3 -m pip install jupyterlab
+RUN pip3 install numpy==1.22.0
+RUN pip3 install torch torchvision torchaudio
+RUN pip3 install git+https://github.com/KonstantinWilleke/experanto.git@statshack
+RUN pip3 install datajoint
+RUN pip3 install opencv-python==4.8.0.74
+RUN pip3 install mlflow
+RUN pip3 install -U pytorch_warmup
+RUN pip3 install transformers==4.28.0
 
-RUN pip3 install -e nnfabrik
-RUN pip3 install -e nnfabrik/ml-utils
-RUN pip3 install -e nnfabrik/nnvision/nnvision
-RUN pip3 install -e nnfabrik/mei/mei
+#RUN pip3 install -e /notebooks/mousehiera
 
-WORKDIR /notebooks
+#ADD . /src/mousehiera
+#RUN pip3 install -e /src/mousehiera
+
+ADD . /notebooks
+RUN pip3 install -e /notebooks
+
+EXPOSE 8888
+# By default start running jupyter notebook
+WORKDIR /notebooks
+
+#RUN git clone -b statshack https://github.com/KonstantinWilleke/experanto.git
+#RUN pip3 install -e /src/experanto
+ENTRYPOINT ["jupyter", "lab", "--allow-root", "--ip=0.0.0.0", "--no-browser", "--port=8888", "--NotebookApp.token='1234'", "--notebook-dir='/notebooks'"]

MANIFEST.in

@@ -0,0 +1 @@
+include v4_multihead_attention_SOTA.pth.tar

README.md

@@ -1,2 +1,5 @@
 # nnvision
 Envisioning the biological visual system with deep neural networks.
+
+## Version 0.1
+- compatible with nnfabrik and neuralpredictors post release.

docker-compose.yml

@@ -1,48 +1,26 @@
-version: '2.3'
+version: '12.5.1'
+
 services:
-  notebook:
+  jupyterlab:
     build: .
-    env_file: .env
-    ports:
-      - "5003:8888"
     volumes:
-      - ../data:/data
-      #- /mnt/qb/data:/data
-      - /tmp:/tmp
-      - .:/src/nnfabrik
-      - ./notebooks:/notebooks
+      - /mnt:/mnt
+      - .:/src/foundation_models
     environment:
-      - DISPLAY=$DISPLAY
-      - NVIDIA_VISIBLE_DEVICES=0
-  gpu-server_develop:
+      - JUPYTER_ENABLE_LAB=yes
+    command: >
+      /bin/bash -c "pip3 install torch torchvision torchaudio && 
+      pip3 install --upgrade jupyterlab &&
+      pip3 install -e /src/foundation_models && 
+      jupyter lab --ip=0.0.0.0 --no-browser --allow-root --NotebookApp.token='1234'"
+
+  jupyter_dev:
     build: .
-    ports:
-      - "8888:8888"
     volumes:
-      - /var/lib/nova/sinz-shared:/data
-      - ../projects:/notebooks
-    env_file: .env
+      - /var/shared:/data
+      - /mnt:/mnt
+      - .:/notebooks
     environment:
-      - DISPLAY=$DISPLAY
+      - JUPYTER_ENABLE_LAB=yes
     runtime: nvidia
-  server_develop:
-    build: .
-    ports:
-      - "8888:8888"
-    volumes:
-      - ../../../var/sinz-shared:/data
-      - ../projects:/notebooks
-    env_file: .env
-    environment:
-      - DISPLAY=$DISPLAY
-    runtime: nvidia
-  gpu_server_playground:
-    build: .
-    ports:
-      - "8888:8888"
-    volumes:
-      - /var/lib/nova/sinz-shared:/data
-    env_file: .env
-    environment:
-      - DISPLAY=$DISPLAY
-    runtime: nvidia
+    shm_size: 1536g

nnvision/analyses/mei.py

@@ -0,0 +1,77 @@
+"""
+this script will run the following analysis:
+- load a PyTorch model
+- get meis
+- for mei:
+   - get the corresponding image and augment it:
+   - get 15x15 translations in height in width
+   - get 30 rotations for each translation
+   - show all 15x15x30 images to the model and get all predictions
+
+- for each neuron, get the maximum prediction for each MEI
+- order neurons by cluster_id, and show the predictions in a large matrix
+"""
+import numpy as np
+import matplotlib.pyplot as plt
+import scipy
+from scipy.ndimage import rotate
+from scipy.ndimage import shift
+import torch
+
+from nnvision.models.trained_models.v4_data_driven import (
+    v4_multihead_attention_ensemble_model,
+)
+
+model = v4_multihead_attention_ensemble_model
+model.eval().cuda()
+
+pop_key = {
+    "model_fn": "nnvision.models.simclr.ConvNextFC",
+    "model_hash": "6510835da772e26b8ceb7013718f367f",
+    "dataset_fn": "nnvision.datasets.mei_loaders.mei_tsmincne_loader",
+    "dataset_hash": "5ea066fad56931cbfca1fceb5ac3e5ca",
+    "trainer_fn": "nnvision.training.simclr.tsimcne_trainer",
+    "trainer_hash": "83f17636d44655570809589946dd296d",
+    "seed": 1000,
+}
+
+dl = (Dataset & pop_key).get_dataloader()
+unit_ids = dl[
+    "validation"
+].dataset.units  # array of shape (n_units*n_seeds,), 11000units, with 10-20 seeds each
+unique_unit_ids = np.unique(unit_ids)  # array of shape (n_units,), 11000 units
+
+# get the MEI for each unit
+meis = dl[
+    "validation"
+].dataset.meis  # array of shape (n_units*n_seeds,1,100,100), 11000units, with 10-20 seeds each, 1 channel, 100x100 pixels
+
+# create a for loop:
+# for each mei, get the 15x15 translations in height and width, and the translations
+
+response_array = []
+for img in meis:
+    # get the 15x15 translations in height and width, and the translations
+    # get 15x15 translations in height in width
+    # get 30 rotations for each translation
+    # show all 15x15x30 images to the model and get all predictions
+    # for each neuron, get the maximum prediction for each MEI
+    # order neurons by cluster_id, and show the predictions in a large matrix
+
+    # get the 15x15 translations in height in width
+    # get 30 rotations for each translation
+    # show all 15x15x30 images to the model and get all predictions
+
+    augmented_imgs = []
+    for i in range(15):
+        for j in range(15):
+            for k in range(30):
+                img_rot = rotate(img.numpy(), k * 6, reshape=False)
+                img_shift = shift(img_rot, [i * 2, j * 2, 0])
+                augmented_imgs.append(img_shift)
+
+    augmented_imgs = torch.from_numpy(np.array(augmented_imgs))
+    with torch.no_grad():
+        outputs = model(augmented_imgs.cuda(), data_key="all_sessions")
+        outputs = outputs.cpu().numpy().max(0)
+    response_array.append(outputs)