Sapiens2 try-it API — student / collaborator client

You don’t need any of the server code. Everything you need is one Python file.

1. Get the client

On Bhaskar (or any machine that can read /DATA):

cp /DATA/nipun.batra/sapiens2/client/student_client.py .

If you’re on your own laptop, scp it across:

scp bhaskar.iitgn.ac.in:/DATA/nipun.batra/sapiens2/client/student_client.py .

2. Install deps

pip install requests tqdm

3. Reach the server

The API is at http://10.0.62.159:8511 on the IITGN intranet. You’re good if:

you’re on IITGN-Wifi / eduroam on campus, or
you’re connected to the IITGN VPN, or

you SSH-tunnel from your laptop:

ssh -L 18511:localhost:8511 bhaskar.iitgn.ac.in
# then use http://localhost:18511 instead

Tell the script where the server is (default is the intranet URL):

export SAPIENS2_API=http://10.0.62.159:8511          # campus / VPN
export SAPIENS2_API=http://localhost:18511           # SSH tunnel

4. Run

One image, summary printed:

python student_client.py path/to/photo.jpg --task pose

A whole folder (recursive walk; resume-aware):

python student_client.py /path/to/images/ --task pose --workers 4 --out results.jsonl

Re-running the same command picks up where it stopped. The output is JSONL: one JSON object per line, with the full server response plus a src field pointing at the input. The big base64 image strings are stripped to keep the file small.

5. Tasks available

pose — 308 keypoints (body + face + hands + feet)
seg — 29-class body-part segmentation (dome29 palette)
normal — disabled in this build
pointmap — disabled in this build

6. Throughput

The server runs Sapiens2 1B on a single A5000, ~0.88 image/s. Practical estimates:

N images	Time
100	~ 2 minutes
1,000	~ 19 minutes
7,000	~ 2 h 13 min

--workers higher than 4 won’t help — the GPU is the bottleneck. If you need more throughput, ping Nipun about running a second API instance on the second GPU, or switching to the smaller 0.4B model.

7. Reading the JSONL output

import json
records = [json.loads(line) for line in open("results.jsonl")]

# Pose example
pose = next(r for r in records if r.get("task") == "pose")
for kp in pose["keypoints"]:
    if kp["score"] >= 0.5:
        print(f"{kp['name']:20s}  ({kp['x']:.0f}, {kp['y']:.0f})  score={kp['score']:.2f}")

8. Troubleshooting

can't reach <URL>: not on the IITGN network / VPN / tunnel. See step 3.
503 disabled: you asked for normal or pointmap; only pose and seg are enabled in this build.
500 with traceback: server-side bug; copy the JSON body and send it to Nipun.

# Sapiens2 try-it API — student / collaborator client You don't need any of the server code. Everything you need is one Python file. ## 1. Get the client On Bhaskar (or any machine that can read `/DATA`): ```bash cp /DATA/nipun.batra/sapiens2/client/student_client.py . ``` If you're on your own laptop, scp it across: ```bash scp bhaskar.iitgn.ac.in:/DATA/nipun.batra/sapiens2/client/student_client.py . ``` ## 2. Install deps ```bash pip install requests tqdm ``` ## 3. Reach the server The API is at `http://10.0.62.159:8511` on the IITGN intranet. You're good if: - you're on `IITGN-Wifi` / `eduroam` on campus, **or** - you're connected to the IITGN VPN, **or** - you SSH-tunnel from your laptop: ```bash ssh -L 18511:localhost:8511 bhaskar.iitgn.ac.in # then use http://localhost:18511 instead ``` Tell the script where the server is (default is the intranet URL): ```bash export SAPIENS2_API=http://10.0.62.159:8511 # campus / VPN export SAPIENS2_API=http://localhost:18511 # SSH tunnel ``` ## 4. Run One image, summary printed: ```bash python student_client.py path/to/photo.jpg --task pose ``` A whole folder (recursive walk; resume-aware): ```bash python student_client.py /path/to/images/ --task pose --workers 4 --out results.jsonl ``` Re-running the same command picks up where it stopped. The output is JSONL: one JSON object per line, with the full server response plus a `src` field pointing at the input. The big base64 image strings are stripped to keep the file small. ## 5. Tasks available - `pose` — 308 keypoints (body + face + hands + feet) - `seg` — 29-class body-part segmentation (dome29 palette) - `normal` — disabled in this build - `pointmap` — disabled in this build ## 6. Throughput The server runs Sapiens2 1B on a single A5000, ~0.88 image/s. Practical estimates: | N images | Time | |-----------:|----------------| | 100 | ~ 2 minutes | | 1,000 | ~ 19 minutes | | 7,000 | ~ 2 h 13 min | `--workers` higher than 4 won't help — the GPU is the bottleneck. If you need more throughput, ping Nipun about running a second API instance on the second GPU, or switching to the smaller 0.4B model. ## 7. Reading the JSONL output ```python import json records = [json.loads(line) for line in open("results.jsonl")] # Pose example pose = next(r for r in records if r.get("task") == "pose") for kp in pose["keypoints"]: if kp["score"] >= 0.5: print(f"{kp['name']:20s} ({kp['x']:.0f}, {kp['y']:.0f}) score={kp['score']:.2f}") ``` ## 8. Troubleshooting - `can't reach <URL>`: not on the IITGN network / VPN / tunnel. See step 3. - `503 disabled`: you asked for `normal` or `pointmap`; only `pose` and `seg` are enabled in this build. - `500` with traceback: server-side bug; copy the JSON body and send it to Nipun.