Spaces:
Running on Zero
Running on Zero
Commit History
12 days ago
-
Own study first, complete study in its own section; preview follows the run Running
- Left column order: upload your own clip/study, example clips, then the complete 50-clip study as a separate section with its own button - The clip and reference card are outputs of the run itself, not a separate click event: clicking another example while a run was in progress switched the preview while the earlier report kept running, so clip, reference and report could disagree - An uploaded MP4 is shown in the preview while its report is generated Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>fc64276 -
Account for ZeroGPU's 1.5x duration factor in declared GPU durations
ZeroGPU scales the declared duration (spaces.zero.client duration_factor) for both the time limit and the quota pre-check: a 50-clip report declared at 100 s requested 150 s. Declared values are now the measured limits divided by 1.5, so a single clip requests 75 s and a full study ~100 s, both within an anonymous visitor's 2-minute quota. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>74b493a -
Redesigned UI, reference checks, fixed models, colour-correct samples
- Examples list (complete study + one clip per view) that generates on click; each clip shows its reference (view, study values) and a Reference check tab scores every run against it - Report streams section by section; measurements flagged with EchoPrime ROC thresholds and cross-checked against the report's own numbers - Stop on <eos> and <end_of_turn>: the report no longer runs on into a second copy - Macedonian translation with the LoRA adapter disabled (it was repeating the English report), used-terms glossary and fixed section headings; runs as its own GPU call - EchoPrime predict_metrics: each phenotype now reads its own section's retrieval - Samples rebuilt from the original DICOMs with YBR->RGB conversion (old ones were tinted and 4 of 6 got the wrong view) - RAM: skip the 1.2M report decode, label arrays instead of dicts, chunked encoding, so the full study fits; DICOM preprocessing in a small thread pool (108 s -> 23 s) - GPU durations sized from measured runs (was …bc8e1e0
20 days ago
-
Sample buttons populate upload box; elapsed-time timer in every stage + total in header
Verified signature835af76
21 days ago
-
Reword launch comment
Verified signaturedd404a7 -
Add translate_to_macedonian
Verified signaturedca66ed -
Streamed progress, Macedonian output, SSR back on (ZeroGPU SDK needs it)
Verified signature9a3c9eb -
ZeroGPU fix: force CPU during EchoPrime load (CUDA init only in GPU tasks)
Verified signatureb0d1e13 -
16GB RAM fix: free candidate_reports, drop retrieval panel
Verified signatureb8530f0 -
Per-clip ground descriptions under each sample video
Verified signaturefa09c6c -
Fix: chdir before echo_prime import (import-time assets read)
Verified signaturef64bce1 -
Fix docstring literal
Verified signature9590b94 -
Swap av for imageio-ffmpeg
Verified signatureacf8b31 -
Background loading with progress panel, imageio decode (av fix), readiness guard
Verified signaturedce3c3b -
SSR off + visible errors: fix silently-dead Analyse clicks
Verified signaturede1b944 -
All 6 sample videos playable in-page + direct one-click Analyse buttons
Verified signature977bf32
22 days ago
-
Expected-answer panel: EchoPrime retrieval report + static answer key
Verified signature161cd02 -
Free EchoPrime retrieval corpus at startup (ZeroGPU RAM fix)
Verified signatureb10f9c4 -
Video preview + How-it-works accordion
Verified signaturee18e102 -
Report display cleanup: SEP breaks + trim duplicated pass
Verified signaturebcfd6ba -
bf16 from_config build: avoid fp32 startup RAM peak (ZeroGPU OOM)
Verified signature6ab11ea -
from_config instead of from_pretrained: no mirror weight download, fast ZeroGPU startup
Verified signature37cf9c1 -
OOM fixes: expandable segments, EchoPrime on CPU, bin_size=25 + cache clear
Verified signature9d28415 -
OOM fix: encoder bin 50->10, expandable segments
Verified signature51a21e2 -
Use ungated mirror for tokenizer/config; weights still from echogemma.pt
Verified signatureebb7ae3
24 days ago
-
Add PyAV for video decoding
Verified signature5f36c89 -
Add MP4/AVI video support + one-click sample loaders
Verified signaturefe4151a -
Add public sample video 5 (from EchoPrime example study)
Verified signaturef6cd9a0 -
Add public sample video 4 (from EchoPrime example study)
Verified signature07dfad0 -
Add public sample video 3 (from EchoPrime example study)
Verified signaturef2611c9 -
Add public sample video 2 (from EchoPrime example study)
Verified signature90fcdcf -
Add public sample video 1 (from EchoPrime example study)
Verified signatured6eb130 -
Add public sample video 0 (from EchoPrime example study)
Verified signature8803a92 -
Add EchoPrime repo dirs to sys.path
Verified signature6f768ac -
Fix EchoGemma import via submodule
Verified signature5375251 -
Space card
Verified signatureea506f9 -
Pinned dependencies for ZeroGPU
Verified signature19c1366 -
EchoGemma model class adapted for ZeroGPU (bf16, lazy CUDA)
Verified signature43cb33d -
Vendored EchoGemma package init
Verified signature9a39357 -
Space app: EchoGemma report + EchoPrime QA
Verified signature5e0df8e -
initial commit
Verified signature8313fe0