Restore working ECG-FM server from commit 58e60a2 - Fix JSON serialization error

Fix JSON serialization error: Convert all numpy types to Python native types for clinical analysis

🔧 Fix predict endpoint - Use proper fairseq_signals syntax for model inference

🔧 Fix JSON serialization issue - Convert numpy int64 to Python int for JSON compatibility

🚀 Deploy ECG-FM v2.1.0 - Physiological Parameter Extraction Now Working! - Added comprehensive physiological parameter extraction (HR, QRS, QT, PR, Axis) using ECG-FM features - Implemented statistical pattern recognition algorithms - Added clinical range validation and confidence scoring - Created comprehensive test script for real ECG samples - Updated documentation and status reports - All endpoints now provide actual measurements instead of null values

FIX: Clinical analysis output structure and proper physiological parameter extraction from ECG-FM features

CRITICAL FIXES: Remove all random values, arbitrary formulas, and hardcoded defaults - replace with proper clinical defaults and warnings

CRITICAL FIX: Remove flawed physiological parameter extraction and fix duplicate returns in clinical analysis

Fix clinical analysis to use 'out' key from finetuned model (actual working output)

Fix clinical analysis to handle classifier model output format (direct logits)

Add debug logging to diagnose finetuned model output issues

Fix model_config global scope and tensor dimension issues

Deploy ECG-FM Dual Model API v2.0.0

Quick fix: Remove duplicate return_features parameter

Fix ECG-FM model inference: Use proper keyword arguments for Wav2Vec2CMSCModel and add production clinical endpoints

CRITICAL FIX: Upgrade PyTorch to 2.1.0 to resolve weight_norm compatibility with ECG-FM checkpoints

Final fix: Pin omegaconf to 2.1.2 to resolve is_primitive_type compatibility with ECG-FM checkpoints

Fix transformers version compatibility: Pin to transformers==4.21.0 to resolve GenerationMixin import error in fairseq_signals

Fix shell command syntax: Break down fairseq_signals installation into separate RUN commands to avoid shell command failures

Final fix: Skip C++ extensions to resolve NumPy API compatibility - fairseq_signals C++ extensions use deprecated NumPy 1.7 API removed in NumPy 1.26.4

Fix fairseq_signals import path: Use proper installation method and verify import at build time

Fix NumPy version conflict: Force reinstall NumPy 1.26.4 after fairseq-signals to prevent overwrite by transformers dependency

Implement Direct HF Loading Strategy: Load ECG-FM model directly from wanglab/ecg-fm repository to work within 1GB limit

Stable deployment: Skip C++ extensions, use NumPy 1.26.4 for dependency compatibility

Fix HF Spaces build: Pin NumPy to 1.24.3 for fairseq-signals C++ extension compatibility

Fix HF Spaces deployment: Update server.py import logic and requirements_hf_spaces.txt

AGGRESSIVE CACHE INVALIDATION: Force HF Spaces rebuild - Updated timestamps to 08:50 UTC - Added unique ENV BUILD_ID to Dockerfile - Enhanced server.py BUILD VERSION comment - Multiple trigger mechanisms to bypass HF Spaces aggressive caching - Build trigger attempt #3

FORCE REBUILD ATTEMPT #2: Enhanced cache invalidation - Updated timestamps to 08:45 UTC - Added BUILD VERSION comment to server.py - Enhanced Dockerfile comments - Multiple trigger mechanisms to ensure build starts

FORCE REBUILD: Cache invalidation for updated server.py - Updated README timestamp to 08:40 UTC - Added comment to Dockerfile to invalidate COPY layer cache - Ensures updated import logic is deployed with fresh build

CLEANUP: Remove unnecessary files and embedded git repos - Removed server_fallback.py (duplicate) - Removed fairseq-signals/ from git tracking (cloned during build) - Clean repository structure for HF Spaces deployment

FIX: Prioritize fairseq_signals import for ECG-FM functionality - Updated import logic to use fairseq_signals.models first - Fixed the 'No module named fairseq' error - Maintains backward compatibility with fallbacks - Ready for full ECG-FM inference capabilities

FORCE REBUILD: Recreated Dockerfile and requirements with NumPy 1.x fix - Added timestamp to README to trigger HF Spaces build - NumPy pinned to >=1.21.3,<2.0.0 for ECG-FM compatibility

CRITICAL FIX: NumPy 1.x compatibility for ECG-FM checkpoints - Pinned NumPy to >=1.21.3,<2.0.0 - Added NumPy version checking in server.py - Fixed Dockerfile to install NumPy 1.x first - Resolves NumPy 2.x crash with ECG-FM checkpoints

FIXED: Dockerfile and requirements for HF Spaces compatibility - Official fairseq-signals installation method - Clean dependencies without corruption - Ready for HF Spaces deployment

COMPLETE FIX: Add ALL fairseq-signals dependencies + fix NumPy version conflict - All researched and tested locally

RESEARCHED FIX: Pin exact compatible versions - omegaconf==2.1.2 + hydra-core==1.1.0 for ECG-FM compatibility

CRITICAL FIX: Add hydra dependencies for fairseq-signals compatibility

CRITICAL FIX: Recreate clean requirements_hf_spaces.txt without corruption

Trigger HF Spaces build - Update README with timestamp

CRITICAL FIX: Remove fairseq==0.12.0 to prevent HF Spaces build failure

Add missing requirements_hf_spaces.txt for HF Spaces deployment

Merge branch 'main' of https://huggingface.co/spaces/mystic-cbk/ecg-fm-api into main

Fix ECG-FM deployment: Update Dockerfile, server.py, requirements.txt with fairseq-signals integration

CRITICAL FIX: Install fairseq-signals with --no-deps to prevent OmegaConf version conflicts

CRITICAL FIX: Force OmegaConf 2.0.0 after fairseq-signals to preserve is_primitive_type

Resolve merge conflicts - keep corrected versions with OmegaConf 2.0.0

Complete ECG-FM setup with OmegaConf 2.0.0 fix and all dependencies

CRITICAL FIX: Update OmegaConf to 2.0.0 (has is_primitive_type) for fairseq-signals compatibility

SYSTEMATIC FIX: Remove non-existent numpy-base, use proven dependency management

CRITICAL FIX: Force NumPy 1.24.3 and fix omegaconf version - Resolve both NumPy 2.x and is_primitive_type issues