Image Classification
LiteRT
LiteRT
ONNX
English
mobilenetv5
mobilenetv4
gemma-3
edge-ai
quantization
w4a8
int8
npu
qualcomm-snapdragon
google-tensor
geology
Instructions to use GeoStratum/quantification-MobileNetV5-V4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LiteRT
How to use GeoStratum/quantification-MobileNetV5-V4 with LiteRT:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
File size: 8,395 Bytes
2248c41 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 | # ==============================================================================
# R Script: Generation of Benchmark Plots for Native Qualcomm Hexagon NPU Binaries
# Standardized on 100 Runs per Model Variant (600 Total Inferences)
# Models: MobileNetV4, MobileNetV5, ScaleDetection_MobileNetV4
# Target Platform: Machine 2 - Qualcomm Snapdragon X Plus (Surface Pro 11 ARM64)
# ==============================================================================
results_base <- Sys.getenv("LITHO_RESULTS_DIR", unset = "results")
if (!dir.exists(results_base) && dir.exists("E:/fine tuning/rock identifier/lithotheque-edge-models/results")) {
results_base <- "E:/fine tuning/rock identifier/lithotheque-edge-models/results"
}
v4_csv <- file.path(results_base, "MobileNetV4/machine_2/benchmark_machine_2_v4_native_100_runs.csv")
v5_csv <- file.path(results_base, "MobileNetV5/machine_2/benchmark_machine_2_v5_native_100_runs.csv")
scale_csv <- file.path(results_base, "ScaleDetection/machine_2/benchmark_machine_2_scale_native_100_runs.csv")
all_csv <- file.path(results_base, "consolidated/csv/benchmark_machine_2_native_binaries_100_runs.csv")
color_npu_w4 <- "#2A9D8F" # Teal / W4A8 INT4 NPU
color_npu_int8 <- "#1D3557" # Navy / W8A8 INT8 NPU
grid_color <- "#EAEAEA"
cat("[INFO] Running R Native Snapdragon NPU Plot Generator (100 Runs)...\n")
cat("[INFO] Results Directory:", results_base, "\n")
# ------------------------------------------------------------------------------
# 1. Comparative Latency for All Models on Hexagon NPU (100 Runs AOT .bin)
# ------------------------------------------------------------------------------
if (file.exists(all_csv)) {
df_all <- read.csv(all_csv, stringsAsFactors = FALSE)
df_all$Latency_ms <- as.numeric(df_all$Latency_ms)
out_png_all <- file.path(results_base, "consolidated/images/machine_2_native_binaries_comparison.png")
dir.create(dirname(out_png_all), recursive = TRUE, showWarnings = FALSE)
png(filename = out_png_all, width = 1200, height = 700, res = 130)
par(mar = c(6, 6, 4, 2), bg = "#FAFAFA")
agg <- aggregate(Latency_ms ~ Model + Format, data = df_all, FUN = mean)
agg$Label <- paste0(agg$Model, "\n(", agg$Format, ")")
cols <- ifelse(agg$Format == "W4A8", color_npu_w4, color_npu_int8)
bp <- barplot(agg$Latency_ms, names.arg = agg$Label, col = cols, border = NA,
ylim = c(0, max(agg$Latency_ms) * 1.25),
main = "Machine 2 (Hexagon NPU): Standalone Native AOT Invocation Turnaround (100 Runs)",
ylab = "Turnaround Latency (ms / cold invocation - includes CLI & Disk I/O)",
font.main = 2, col.lab = "#333333", col.axis = "#555555", las = 1)
grid(col = grid_color, lty = "dotted")
barplot(agg$Latency_ms, names.arg = NA, col = cols, border = NA,
ylim = c(0, max(agg$Latency_ms) * 1.25), yaxt = "n", xaxt = "n", add = TRUE)
fps <- round(1000 / agg$Latency_ms, 1)
text(bp, agg$Latency_ms + (max(agg$Latency_ms) * 0.03),
labels = paste0(round(agg$Latency_ms, 1), " ms\n(", fps, " FPS)"),
font = 2, cex = 0.85, col = "#222222")
legend("topleft",
legend = c("Hexagon HTP W4A8 AOT (.bin)", "Hexagon HTP W8A8 / INT8 AOT (.bin)"),
fill = c(color_npu_w4, color_npu_int8), border = NA, bg = "white", box.col = "#DDDDDD")
dev.off()
cat("[OK] Generated:", out_png_all, "\n")
}
# ------------------------------------------------------------------------------
# 2. MobileNetV4 Native Stability Plot (100 Consecutive Passes on Hexagon NPU)
# ------------------------------------------------------------------------------
if (file.exists(v4_csv)) {
df_v4 <- read.csv(v4_csv, stringsAsFactors = FALSE)
df_v4$Latency_ms <- as.numeric(df_v4$Latency_ms)
out_v4_png <- file.path(results_base, "MobileNetV4/machine_2/mobilenetv4_machine_2_native_stability.png")
png(filename = out_v4_png, width = 1200, height = 600, res = 130)
par(mar = c(5, 5, 4, 2), bg = "#FAFAFA")
w4_runs <- df_v4$Latency_ms[df_v4$Format == "W4A8"]
int8_runs <- df_v4$Latency_ms[df_v4$Format == "INT8"]
plot(1:length(w4_runs), w4_runs, type = "l", col = color_npu_w4, lwd = 2,
ylim = c(min(c(w4_runs, int8_runs)) * 0.9, max(c(w4_runs, int8_runs)) * 1.15),
xlab = "Inference Run ID (100 Consecutive Field Images)", ylab = "Hardware Latency (ms)",
main = "MobileNetV4: Native HTP Stability on Machine 2 (Hexagon NPU, 100 Runs)",
las = 1, font.main = 2)
grid(col = grid_color, lty = "dotted")
lines(1:length(int8_runs), int8_runs, col = color_npu_int8, lwd = 2)
legend("topright",
legend = c(paste0("W4A8 (Mean: ", round(mean(w4_runs), 1), " ms, p95: ", round(quantile(w4_runs, 0.95), 1), " ms)"),
paste0("INT8 (Mean: ", round(mean(int8_runs), 1), " ms, p95: ", round(quantile(int8_runs, 0.95), 1), " ms)")),
col = c(color_npu_w4, color_npu_int8), lty = 1, lwd = 2, bg = "white")
dev.off()
cat("[OK] Generated:", out_v4_png, "\n")
}
# ------------------------------------------------------------------------------
# 3. MobileNetV5 Native Stability Plot (100 Consecutive Passes on Hexagon NPU)
# ------------------------------------------------------------------------------
if (file.exists(v5_csv)) {
df_v5 <- read.csv(v5_csv, stringsAsFactors = FALSE)
df_v5$Latency_ms <- as.numeric(df_v5$Latency_ms)
out_v5_png <- file.path(results_base, "MobileNetV5/machine_2/mobilenetv5_machine_2_native_stability.png")
png(filename = out_v5_png, width = 1200, height = 600, res = 130)
par(mar = c(5, 5, 4, 2), bg = "#FAFAFA")
w4_runs <- df_v5$Latency_ms[df_v5$Format == "W4A8"]
int8_runs <- df_v5$Latency_ms[df_v5$Format == "INT8"]
plot(1:length(w4_runs), w4_runs, type = "l", col = color_npu_w4, lwd = 2,
ylim = c(min(c(w4_runs, int8_runs)) * 0.95, max(c(w4_runs, int8_runs)) * 1.05),
xlab = "Inference Run ID (100 Consecutive Field Images)", ylab = "Hardware Latency (ms)",
main = "MobileNetV5: Native HTP Stability on Machine 2 (Hexagon NPU, 100 Runs)",
las = 1, font.main = 2)
grid(col = grid_color, lty = "dotted")
lines(1:length(int8_runs), int8_runs, col = color_npu_int8, lwd = 2)
legend("topright",
legend = c(paste0("W4A8 (Mean: ", round(mean(w4_runs), 1), " ms, p95: ", round(quantile(w4_runs, 0.95), 1), " ms)"),
paste0("INT8 (Mean: ", round(mean(int8_runs), 1), " ms, p95: ", round(quantile(int8_runs, 0.95), 1), " ms)")),
col = c(color_npu_w4, color_npu_int8), lty = 1, lwd = 2, bg = "white")
dev.off()
cat("[OK] Generated:", out_v5_png, "\n")
}
# ------------------------------------------------------------------------------
# 4. ScaleDetection V4 Native Plot (100 Consecutive Passes on Hexagon NPU)
# ------------------------------------------------------------------------------
if (file.exists(scale_csv)) {
df_scale <- read.csv(scale_csv, stringsAsFactors = FALSE)
df_scale$Latency_ms <- as.numeric(df_scale$Latency_ms)
out_scale_png <- file.path(results_base, "ScaleDetection/machine_2/scaledetection_machine_2_native_stability.png")
png(filename = out_scale_png, width = 1200, height = 600, res = 130)
par(mar = c(5, 5, 4, 2), bg = "#FAFAFA")
w4_runs <- df_scale$Latency_ms[df_scale$Format == "W4A8"]
int8_runs <- df_scale$Latency_ms[df_scale$Format == "INT8"]
plot(1:length(w4_runs), w4_runs, type = "l", col = color_npu_w4, lwd = 2,
ylim = c(min(c(w4_runs, int8_runs)) * 0.9, max(c(w4_runs, int8_runs)) * 1.15),
xlab = "Inference Run ID (100 Consecutive Macro Images)", ylab = "Hardware Latency (ms)",
main = "ScaleDetection V4: 1024x1024 Metrology Stability on Machine 2 (Hexagon NPU, 100 Runs)",
las = 1, font.main = 2)
grid(col = grid_color, lty = "dotted")
lines(1:length(int8_runs), int8_runs, col = color_npu_int8, lwd = 2)
legend("topright",
legend = c(paste0("W4A8 (Mean: ", round(mean(w4_runs), 1), " ms, p95: ", round(quantile(w4_runs, 0.95), 1), " ms)"),
paste0("INT8 (Mean: ", round(mean(int8_runs), 1), " ms, p95: ", round(quantile(int8_runs, 0.95), 1), " ms)")),
col = c(color_npu_w4, color_npu_int8), lty = 1, lwd = 2, bg = "white")
dev.off()
cat("[OK] Generated:", out_scale_png, "\n")
}
cat("[SUCCESS] All 100-run native Snapdragon NPU plots generated successfully in English.\n")
|