v9: download 100K COYO images (5 partitions), 4 processes, 256 threads

job_ml.sh

@@ -154,7 +154,7 @@ for i in range(torch.cuda.device_count()):
 "
 
 ###############################################################################
-# 5. DATA
+# 5. DATA - Download COYO (100K images)
 ###############################################################################
 log "=== [5/7] Data ==="
 SHARD_DIR="$DATA0/datasets/processed/flux_train/shards"
@@ -163,11 +163,12 @@ mkdir -p "$SHARD_DIR"
 SHARD_COUNT=$(find "$SHARD_DIR" -name "*.tar" 2>/dev/null | wc -l | tr -d '[:space:]')
 SHARD_COUNT=${SHARD_COUNT:-0}
 
-if [ "$SHARD_COUNT" -ge 2 ]; then
+# Need at least 20 shards (~20K images) to train properly
+if [ "$SHARD_COUNT" -ge 20 ]; then
     log "  Shards ready ($SHARD_COUNT)"
 else
-    # Try pull from HF first
-    log "  Pulling shards from HF..."
+    # Pull existing shards from HF first
+    log "  Pulling existing shards from HF..."
     python3 -u << PYEOF || true
 from huggingface_hub import snapshot_download
 try:
@@ -185,43 +186,48 @@ PYEOF
     SHARD_COUNT=$(find "$SHARD_DIR" -name "*.tar" 2>/dev/null | wc -l | tr -d '[:space:]')
     SHARD_COUNT=${SHARD_COUNT:-0}
 
-    if [ "$SHARD_COUNT" -lt 2 ]; then
-        # Download from COYO
-        log "  Downloading from COYO (1 partition, top 20K images)..."
+    if [ "$SHARD_COUNT" -lt 20 ]; then
+        # Download from COYO - 5 partitions, filter top 100K images
+        log "  Downloading COYO metadata (5 partitions)..."
         COYO_RAW="$DATA0/datasets/raw/coyo"
         COYO_FILTERED="$DATA0/datasets/raw/coyo_filtered"
 
-        # Download 1 parquet only
         python3 -u << PYEOF
 from huggingface_hub import snapshot_download
 snapshot_download(
     repo_id='kakaobrain/coyo-700m',
     repo_type='dataset',
     local_dir='$COYO_RAW',
-    allow_patterns=['data/part-00000-*.parquet'],
-    max_workers=4,
+    allow_patterns=[
+        'data/part-00000-*.parquet',
+        'data/part-00001-*.parquet',
+        'data/part-00002-*.parquet',
+        'data/part-00003-*.parquet',
+        'data/part-00004-*.parquet',
+    ],
+    max_workers=8,
 )
 print('    Metadata OK')
 PYEOF
 
-        # Filter
+        # Filter - top 100K aesthetic images
         mkdir -p "$COYO_FILTERED"
         python3 -u "$PROJECT_DIR/scripts/data_collection/filter_coyo.py" \
             --input-dir "$COYO_RAW/data" \
             --output-dir "$COYO_FILTERED" \
-            --min-aesthetic 6.0 \
-            --min-clip 0.25 \
-            --max-watermark 0.5 \
-            --max-records 20000
+            --min-aesthetic 5.5 \
+            --min-clip 0.2 \
+            --max-watermark 0.7 \
+            --max-records 100000
 
-        # img2dataset
+        # img2dataset - download images
         python3 -u -c "
 import pandas as pd
 df = pd.read_parquet('$COYO_FILTERED/coyo_aesthetic.parquet')
 df[['url','text']].rename(columns={'url':'URL','text':'TEXT'}).to_parquet('$SHARD_DIR/_urls.parquet', index=False)
-print(f'    {len(df)} URLs ready')
+print(f'    {len(df)} URLs ready for download')
 "
-        log "  Running img2dataset..."
+        log "  Running img2dataset (100K images, ~30-60 min)..."
         img2dataset \
             --url_list "$SHARD_DIR/_urls.parquet" \
             --input_format parquet \
@@ -229,24 +235,25 @@ print(f'    {len(df)} URLs ready')
             --caption_col TEXT \
             --output_format webdataset \
             --output_folder "$SHARD_DIR" \
-            --processes_count 2 \
-            --thread_count 128 \
+            --processes_count 4 \
+            --thread_count 256 \
             --image_size 1024 \
             --resize_mode center_crop \
             --resize_only_if_bigger True \
             --number_sample_per_shard 1000 \
-            --retries 3 2>&1 | tail -10
+            --retries 3 2>&1 | tail -20
 
         SHARD_COUNT=$(find "$SHARD_DIR" -name "*.tar" 2>/dev/null | wc -l | tr -d '[:space:]')
         SHARD_COUNT=${SHARD_COUNT:-0}
+        log "  Download complete: $SHARD_COUNT shards"
 
-        if [ "$SHARD_COUNT" -eq 0 ]; then
-            log "  FATAL: No shards"
+        if [ "$SHARD_COUNT" -lt 5 ]; then
+            log "  FATAL: Too few shards ($SHARD_COUNT)"
             exit 1
         fi
 
-        # Upload shards to HF so next run is instant
-        log "  Uploading shards to HF..."
+        # Upload shards to HF for next run
+        log "  Uploading shards to HF (for faster resume)..."
         python3 -u << PYEOF || true
 from huggingface_hub import HfApi, upload_folder
 api = HfApi()