Mistral Vibe Code Generated Mistral Chat Completion API Client

For this series of posts, I asked GitHub Copilot, Anthropic Claude Code, and Mistral Vibe Code to generate a Mistral Chat Completion client using the same prompt….

Generate C# DTOs and a ChatCompletionClient for https://docs.mistral.ai/api/endpoint/chat. The ChatCompletionClient should
use HttpClientFactory with resilience and should have async methods where suitable. The DTOs should be split up into files
for shared, buffered, and streaming.

Generate a console application which uses ChatCompletionClient

Then I took the initial implementation and prompted GitHub Copilot, Anthropic Claude Code, and Mistral Vibe Code to review them. The aim was to see how each tool thinks, what they prioritise, and then how their feedback differs when given identical code.

TL;DR – the code was fine.

Github Copilot

Overall the library is well-structured, cleanly documented, and follows familiar patterns (Options, IHttpClientFactory, Polly).

Anthropic Claude

The client is a small, focused wrapper around POST /v1/chat/completions, supporting both a buffered call (CompleteAsync) and an SSE streaming call (StreamAsync). Overall design is clean: immutable JsonSerializerOptions, an immutable record request type mutated via with, DI registration through IHttpClientFactory with a resilience pipeline, and polymorphic DTOs for messages/content/tools that mirror the Mistral (OpenAI-compatible) wire format. No correctness-blocking bugs were found; findings below are mostly design nits and a few things worth double-checking against real API behavior.

Mistral Vibe Code

Overall Rating: ⭐⭐⭐⭐☆ (4.2/5) – Production Ready with Minor Improvements

Github Copilot Review

:question:

Review MistralAI.Client.Copilot – ChatCompletionClient.cs and DTOs

GitHub Copilot

  1. Read ChatCompletionClient.cs, lines 165-514 :white_check_mark:
  2. Code search found relevant results :white_check_mark:
  3. Read ChatCompletionClient.cs, lines 294-514 :white_check_mark:
  4. Read BufferedDto.cs, lines 1-220 :white_check_mark:
  5. Read SharedDto.cs, lines 1-305 :white_check_mark:
  6. Read StreamingDto.cs, lines 1-420 :white_check_mark:

Code Review – ChatCompletionClient & DTOs

Overall the library is well-structured, cleanly documented, and follows familiar patterns (Options, IHttpClientFactory, Polly). Below are the concrete issues and improvements I’d action, roughly ordered by impact.

🔴 Correctness / bugs

  1. Streaming isn’t actually streaming. In StreamCompletionAsync you call _httpClient.PostAsync(...), which buffers the whole response before returning. You must pass HttpCompletionOption.ResponseHeadersRead, otherwise every “streamed” token arrives at once (and the retry policy will hold the full body in memory).
using var req = new HttpRequestMessage(HttpMethod.Post, requestUri) { Content = requestContent };
response = await _retryPolicy.ExecuteAsync(ct =>
    _httpClient.SendAsync(req, HttpCompletionOption.ResponseHeadersRead, ct), cancellationToken);
  1. Retry policy + single StringContent is broken. StringContent is disposed once sent, so any retry after the first attempt POSTs an empty/disposed body. The HttpRequestMessage/content must be re-created inside the ExecuteAsync delegate (or use HttpClientHandler-level Polly via AddPolicyHandler, which handles this).

  2. IDisposable + IHttpClientFactory conflict. The class implements IDisposable and holds _disposed, implying it disposes _httpClient. HttpClients from IHttpClientFactory must not be disposed by the consumer — the factory owns the handler lifetime. Either:

    • drop IDisposable, or
    • dispose only in the HttpClient-injected overload and track ownership with a flag.
  3. Mutating a factory-provided HttpClient in the constructor. Setting Timeout, DefaultRequestHeaders.Authorization, and User-Agent on httpClientFactory.CreateClient() re-configures a shared/named client each time and, worse, appending User-Agent via Add(...) will throw InvalidOperationException on second construction against the same named client. Configure this once through services.AddHttpClient(c => …) and remove it from the ctor.

  4. ReadLineAsync treats SSE as line-oriented text. Real SSE:

    • allows data: (no space) — your Substring(6) is off by one in that case,
    • allows multi-line data: chunks concatenated per event,
    • has comment lines starting with : (heartbeat/keep-alive) that must be ignored,
    • has event:, id:, retry: fields. Consider System.Net.ServerSentEvents.SseParser (built-in) instead of hand-rolling.
  5. Polymorphic streaming events. You deserialize into a single CompletionEvent type, yet StreamingDto.cs defines CompletionEventBase with TextCompletionEvent, ToolCallEvent, … discriminated by type. Without [JsonPolymorphic]/[JsonDerivedType] on the base, the derived data is lost. Either:

    • Add polymorphism attributes on CompletionEventBase and deserialize as the base, or
    • Keep a flat CompletionEvent DTO and drop the base/derived hierarchy — currently you have both, which is confusing.
  6. Fallback path assumes Text exists on CompletionEvent.
    new CompletionEvent { Type = "unknown", Text = line } implies a flat DTO; reconcile with issue #6.

  7. No Retry-After handling on 429. Polly’s WaitAndRetryAsync uses only 2^n. Mistral (like OpenAI) returns Retry-After — honour it.

  8. response.Content.ReadAsStringAsync() / ReadAsStreamAsync() don’t take the cancellationToken. These overloads exist on .NET 8 — pass the token so hanging responses can be cancelled.

  9. HttpClient.Timeout fights with cancellation. With Polly retries + streaming, a global 120 s timeout aborts long streams. Use HttpClient.Timeout = Timeout.InfiniteTimeSpan and enforce per-request timeout via a linked CancellationTokenSource.

🟠 API / design

  1. Massive DTO duplication. ChatCompletionRequest (Buffered) and ChatCompletionStreamRequest (Streaming) are byte-for-byte identical apart from the default of Stream. Introduce a shared base (or single DTO) in DTOs.Shared and delete ~130 lines. The Stream boolean is the only real difference and can be set by the client method.

  2. Stop and ToolChoice typed as object?. These serialize fine but deserialize poorly and lose IntelliSense. Use OneOf<string, List>, a small wrapper type, or JsonElement.

  3. Metadata is Dictionary. Mistral’s schema is Dictionary.

  4. MessageBase.Content is string? only. Mistral supports multimodal content (array of {type, text|image_url} parts) on user messages. Model as object? or OneOf<string, List> to be forward-compatible.

  5. AssistantMessage missing prefix/reasoning/tool_call_id variants. ChatCompletionChoice.Prefix exists but not on the message itself, which is where the API places it for message replay.

  6. GuardrailConfig is an empty placeholder. Either implement it or omit the property so serializers don’t emit "guardrails": [].

  7. ResponseFormat.Schema should be JsonElement/typed to avoid double serialization surprises via object.

  8. MistralAiException — add the three standard Exception ctors ((), (string), (string, Exception)), and consider a strongly typed MistralApiError model (type, message, param, code) parsed from the body instead of the raw string.

  9. Validation. MistralAiOptions.ApiKey defaults to ""; the constructor happily builds a client with no auth. Add IValidateOptions or throw in the ctor.

  10. GetJsonOptions() (not shown) should be cached in a static readonly field — creating JsonSerializerOptions per call is expensive and defeats the internal metadata cache.

  11. CreateRequest / CreateStreamRequest just build DTOs. They add little over new ChatCompletionRequest { … } and inflate the surface area. Consider removing, or make them extension methods.

  12. Second ctor mutates a caller-owned HttpClient. Prefer treating an injected HttpClient as read-only and route auth via HttpRequestMessage.Headers.Authorization on each request.

🟡 Style / minor

  • line.StartsWith("data: ", StringComparison.Ordinal) — add ordinal comparison; ditto for the [DONE] sentinel.
  • Use _httpClient.DefaultRequestHeaders.UserAgent.Add(new ProductInfoHeaderValue("MistralAI.Client", "1.0")) instead of Add("User-Agent", …).
  • Consider sealed on ChatCompletionClient; extension is unlikely and it simplifies IDisposable.
  • MistralAiOptions.EnableStreaming is declared but never used.
  • XML doc for CreateStreamRequest says stream param exists — it doesn’t.
  • Naming: CompletionEvent vs CompletionEventBase vs TextCompletionEvent — pick one convention (drop Base, since it’s abstract).
  • Consider ArgumentNullException.ThrowIfNull(request) (.NET 8 idiom).

Suggested next steps

  1. Fix the streaming/retry correctness issues (#1, #2, #5, #6) — those change behaviour.
  2. Collapse the duplicated request DTOs (#11) — biggest maintenance win.
  3. Move HttpClient configuration into DI registration and drop IDisposable (#3, #4).
  4. Cache JsonSerializerOptions and switch to source-generated serialization if AOT/perf is a goal (#20).

Happy to draft any of these changes — tell me which to tackle first.

The next review was done with Anthropic’s Claude Code

Code Review: MistralAI.Client.CopilotChatCompletionClient.cs & DTOs

Scope: src/Copilot.PoC/MistralAI.Client.Copilot/ Files reviewed:

  • ChatCompletionClient.cs
  • IChatCompletionClient.cs
  • ChatCompletionRequest.cs (request DTOs)
  • ChatCompletionResponse.cs (response DTOs)
  • Streaming/StreamingDTOs.cs
  • MistralClientOptions.cs, MistralServiceCollectionExtensions.cs (for context)

Summary

The client is a small, focused wrapper around POST /v1/chat/completions, supporting both a buffered call (CompleteAsync) and an SSE streaming call (StreamAsync). Overall design is clean: immutable JsonSerializerOptions, an immutable record request type mutated via with, DI registration through IHttpClientFactory with a resilience pipeline, and polymorphic DTOs for messages/content/tools that mirror the Mistral (OpenAI-compatible) wire format. No correctness-blocking bugs were found; findings below are mostly design nits and a few things worth double-checking against real API behavior.


ChatCompletionClient.cs

Strengths

  • JsonOptions is a single static, cached JsonSerializerOptions instance (line 18) — correct, since re-creating JsonSerializerOptions per call is a well-known perf trap with System.Text.Json.
  • CompleteAsync/StreamAsync both null-check request at the boundary (ArgumentNullException.ThrowIfNull), consistent with API-boundary validation.
  • request with { Stream = false/true } cleanly forces the correct value regardless of what the caller set, without mutating the caller’s object.
  • StreamAsync uses HttpCompletionOption.ResponseHeadersRead so the response isn’t buffered before the SSE loop starts — necessary for real streaming.
  • Malformed SSE chunks are swallowed with a comment explaining why (// Skip malformed chunks rather than aborting the whole stream.) — an intentional, documented tradeoff rather than silent-failure-by-accident.
  • EnsureSuccessAsync reads the error body only on failure, and wraps it into HttpRequestException with the status code attached (statusCode: response.StatusCode) — good for callers that pattern-match on StatusCode.

Findings / things to double check

  1. ChatCompletionRequest.Stream is public but always overwritten. The DTO exposes a settable Stream property, documented as "Overridden internally by the client depending on the method called." (ChatCompletionRequest.cs:26). Any value the caller sets is silently discarded by both CompleteAsync and StreamAsync. This isn’t a bug, but it’s a slightly surprising public API — a caller could reasonably expect setting Stream = true and calling CompleteAsync to do something, when it’s actually a no-op. Consider either making the setter internal/removing it from the public request shape, or asserting/ignoring rather than silently stomping it.

  2. Silent chunk-skip on JsonException has no observability hook. StreamAsync (lines 104–113) catches JsonException and continues with no logging or counter. For a PoC this is fine, but if the wire format ever drifts (e.g., a new event type Mistral adds), failures will be invisible — chunks just quietly vanish. Consider at least a ILogger debug-level log if one is easy to thread through, since this is the one place errors are deliberately suppressed rather than propagated.

  3. SSE parsing only recognizes data: lines. The reader loop only reacts to lines starting with data: and otherwise continues (lines 89–92), which correctly skips blank lines and SSE comment/event: lines. This matches Mistral’s OpenAI-compatible SSE format (each event is a single data: {...} line), so it should be fine in practice — just flagging that it assumes single-line JSON payloads per event; a server that ever wrapped a JSON payload across multiple data: lines (some SSE producers do this) would need concatenation logic that isn’t present here.

  4. No cancellation-specific handling. CancellationToken is threaded through correctly (ConfigureAwait(false) + [EnumeratorCancellation]), but there’s no explicit catch/rethrow around OperationCanceledException — that’s actually correct behavior (let it propagate), just noting it was checked.


Request DTOs (ChatCompletionRequest.cs)

Strengths

  • ChatCompletionRequest is a sealed record with required members for Model/Messages, giving compile-time enforcement of the two truly mandatory fields.
  • Polymorphic ChatMessage/ContentPart hierarchies use [JsonPolymorphic] + [JsonDerivedType] correctly keyed off role and type respectively, matching the API’s discriminator fields.
  • MessageContent cleanly models the “string OR array of parts” duality the API allows, with implicit conversions (string, ContentPart[], List) making call sites ergonomic, and a custom JsonConverter handling both read and write shapes.
  • GuardrailConfig.AdditionalProperties via [JsonExtensionData] is a sensible forward-compatibility escape hatch for a field Mistral is likely to extend.

Findings

  1. Naming collision: ToolChoice property vs. ToolChoice static helper class. ChatCompletionRequest.ToolChoice (an object? property, line 41) and the top-level public static class ToolChoice (line 239) share the identifier ToolChoice in the same namespace (Mistral.Client.Shared). This compiles and works — new ChatCompletionRequest { ToolChoice = ToolChoice.Auto } resolves the right-hand side to the static class since object-initializer right-hand expressions aren’t evaluated in the initializer’s member scope — but it’s a readability trap: IntelliSense/go-to-definition on ToolChoice inside that expression can be momentarily ambiguous to a human reader, and a future refactor that moves code around could make this genuinely ambiguous or require global::/qualification. Consider renaming the helper class (e.g. ToolChoiceValues or ToolChoices) to remove the shadow.

  2. ToolChoice property typed as object?. Reasonable given the API accepts either a string enum or an object literal, and the ToolChoice helper class exists specifically to build valid values — but it does mean nothing stops a caller from passing an arbitrary/invalid object that will only fail at the HTTP layer. Given this is a PoC and the alternative (a discriminated-union style type) is meaningfully more code, this looks like an acceptable, deliberate tradeoff rather than an oversight.

  3. MessageContentJsonConverter.Write prefers Text over Parts if both are set. Not reachable through the public implicit-conversion surface (only one of Text/Parts is ever set), but since MessageContent has public init accessors, nothing stops object-initializer code from setting both fields directly, at which point Parts would be silently dropped on serialize. Low risk given current usage patterns, just noting the type doesn’t defensively guard against that state.

  4. Minor: IReadOnlyList has no implicit conversion, only ContentPart[] and List do (lines 122–124). Not a bug — just means a caller holding an IReadOnlyList from elsewhere has to materialize it to one of the two supported types first.


Response DTOs (ChatCompletionResponse.cs)

  • Straightforward, mutable POCOs (get; set;) matching typical System.Text.Json deserialization targets — appropriately different from the request side (which is a record built by the caller) since these are populated by the deserializer, not constructed by consumers.
  • ChatCompletionChoice.Message is typed as AssistantMessage? (line 21), correctly narrowing the polymorphic ChatMessage base to the one concrete type a response choice can actually contain.
  • Choices and other list-typed members default to [] rather than null (e.g., ChatCompletionResponse.cs:12), which avoids null-checks for consumers even before deserialization overwrites them — good defensive default.
  • No SystemFingerprint/LogProbs/similar fields are modeled. If Mistral’s response includes them and callers might want them, they’re currently unreachable (no [JsonExtensionData] catch-all on the response types, unlike GuardrailConfig on the request side). Worth confirming this is an intentional scope decision for the PoC.

Streaming DTOs (Streaming/StreamingDTOs.cs)

  • ChatCompletionChunk/ChatCompletionChoiceChunk/DeltaMessage correctly mirror the non-streaming shapes where structurally identical (UsageInfo, MessageContent, ToolCall are reused rather than duplicated), which is good — no drift risk between streaming/non-streaming content models.
  • The XML doc on DeltaMessage (lines 28–32) correctly documents the aggregation contract callers must implement (accumulate Content, merge ToolCalls by Index) — useful since the client itself provides no aggregation helper, so this is the only place that contract is written down. Worth keeping in sync with any future consumer-side aggregation utility.
  • Minor cosmetic: namespace Mistral.Client.Shared.Streaming; (line 3) has a double space after the namespace keyword. Harmless, but worth a quick formatter pass.

Supporting types (context only)

  • MistralClientOptions / MistralServiceCollectionExtensions: registration wires IChatCompletionClient as a singleton backed by a named HttpClient with Timeout = Timeout.InfiniteTimeSpan and delegates time-boxing to the standard resilience handler (AttemptTimeout 90s, TotalRequestTimeout 5m) — a sensible choice given long-running generations and streaming responses, and the comment explaining why the infinite client timeout is safe here is a good practice (explains a non-obvious decision rather than just stating what the code does).

Overall assessment

No functional bugs found. The code is well-structured for a PoC: immutable request model, cached serializer options, correct streaming/cancellation plumbing, and DTOs that faithfully mirror the API’s polymorphic wire format. The main things worth a follow-up decision are the ToolChoice naming collision (cosmetic but worth a rename) and whether the public, always-overridden Stream property on ChatCompletionRequest is intentional API surface or should be tightened.

The final review was done with Mistral’s Vibe Code

📊 MistralAI.Client.Copilot Code Review

> Project Type: .NET 10.0 Library | Purpose: Mistral AI Chat Completion API Client > Last Reviewed: 2026-07-20 | Reviewed By: Mistral Vibe



📁 Project Structure

MistralAI.Client.Copilot/
├── MistralAI.Client.Copilot.csproj         # Project configuration (net10.0)
├── IChatCompletionClient.cs               # Interface definition
├── ChatCompletionClient.cs                # Main client implementation
├── ChatCompletionRequest.cs               # Request DTOs & Message types
├── ChatCompletionResponse.cs              # Response DTOs
├── MistralClientOptions.cs                # Configuration options
├── MistralServiceCollectionExtensions.cs # DI registration
└── Streaming/
    └── StreamingDTOs.cs                   # Streaming response chunk DTOs

Strengths

1. Clean Architecture

  • Separation of Concerns: Clear separation between interfaces (IChatCompletionClient), implementation (ChatCompletionClient), and DTOs
  • Dependency Injection: Proper use of IHttpClientFactory and DI container extensions
  • Modular Design: Streaming logic isolated in dedicated namespace

2. Modern C# Features

  • Records: Extensive use of record types for immutable DTOs (ChatCompletionRequest, ChatCompletionResponse)
  • Pattern Matching: Effective use in MessageContentJsonConverter.Read()
  • Async Streams: Proper implementation of IAsyncEnumerable for streaming
  • Init-only Properties: Consistent use of init accessors for immutable DTOs
  • File-scoped Namespaces: Clean namespace declarations

3. Robust Error Handling

  • Argument Validation: Uses ArgumentNullException.ThrowIfNull() for public method parameters
  • HTTP Error Handling: EnsureSuccessAsync() provides detailed error messages with status codes and response bodies
  • Streaming Resilience: Gracefully handles malformed JSON chunks in streaming without breaking the entire stream
  • Cancellation Support: Proper CancellationToken propagation throughout async methods

4. API Design

  • Interface-based: Client consumes IChatCompletionClient interface, enabling mocking
  • Flexible Configuration: MistralClientOptions allows custom base URLs and API keys
  • Resilience Pipeline: Built-in retry, circuit breaker, timeout, and rate limiting via AddStandardResilienceHandler

5. JSON Serialization

  • System.Text.Json: Uses modern, high-performance JSON serializer
  • Polymorphic Deserialization: Proper use of [JsonPolymorphic] and [JsonDerivedType] for message role discrimination
  • Custom Converters: MessageContentJsonConverter handles string vs array polymorphism
  • Snake Case Naming: Consistent JSON naming policy matching Mistral API conventions
  • Null Handling: Proper JsonIgnoreCondition.WhenWritingNull configuration

6. Streaming Implementation

  • SSE Parsing: Correctly handles Server-Sent Events format (data: prefix, [DONE] sentinel)
  • Memory Efficient: Uses StreamReader for line-by-line processing
  • Chunk Validation: Skips malformed chunks instead of failing

7. Documentation

  • XML Documentation: Comprehensive XML comments on public APIs
  • Constants for Magic Strings: FinishReason, ToolChoice classes for well-known values

🔍 DTO Analysis

Request DTOs (ChatCompletionRequest.cs)

DTO Purpose Design Quality
ChatCompletionRequest Main request body ✅ Comprehensive, all Mistral API parameters covered
ChatMessage (abstract) Base message type ✅ Polymorphic with role discriminator
SystemMessage, UserMessage, AssistantMessage, ToolMessage Message types ✅ Proper inheritance hierarchy
MessageContent Flexible content (string or parts) ✅ Uses custom converter for polymorphism
ContentPart (abstract) Content part types ✅ Extensible with derived types
TextContentPart, ImageUrlContentPart, etc. Specific content types ✅ Covers all Mistral content types
Tool, FunctionDefinition, ToolCall, FunctionCall Tool integration ✅ Complete tool calling support
ResponseFormat, JsonSchemaSpec Response formatting ✅ JSON schema support
UsageInfo Token usage tracking ✅ Standard structure
GuardrailConfig Guardrail configuration ✅ Extensible with [JsonExtensionData]
Prediction Prediction hints ✅ Simple and effective

Response DTOs

DTO File Design Quality
ChatCompletionResponse ChatCompletionResponse.cs ✅ Matches Mistral API response structure
ChatCompletionChoice ChatCompletionResponse.cs ✅ Includes message and finish reason
ChatCompletionChunk StreamingDTOs.cs ✅ Streaming chunk with usage info
ChatCompletionChoiceChunk StreamingDTOs.cs ✅ Delta-based streaming
DeltaMessage StreamingDTOs.cs ✅ Incremental message content

⚠️ Areas for Improvement

1. Namespace Organization

  • Issue: Inconsistent namespace usage
    • IChatCompletionClient uses Mistral.Client
    • ChatCompletionClient uses Mistral.Client.Shared
    • This creates confusion about which namespace to use
  • 💡 Recommendation: Standardize on Mistral.Client or MistralAI.Client throughout

2. Nullability Annotations

  • ⚠️ Issue: Incomplete nullability annotations
    • Many DTO properties lack ? for nullable reference types
    • Example: ChatCompletionRequest.Model is required string but could benefit from explicit nullability
  • 💡 Recommendation: Add consistent nullability annotations, especially for optional fields

3. Record vs Class Usage

  • ⚠️ Issue: Mixed use of record and class for DTOs
    • ChatCompletionRequest is a record (good for immutability)
    • Message types (SystemMessage, UserMessage) are class with setters
  • 💡 Recommendation: Consider making all DTOs immutable record types for consistency

4. JSON Serialization Edge Cases

  • ⚠️ Issue: MessageContentJsonConverter doesn’t handle all possible JSON scenarios
    • No validation that exactly one of Text or Parts is set
    • Could throw on unexpected JSON structures
  • 💡 Recommendation: Add validation in converter to ensure data integrity

5. Error Handling in Streaming

  • ⚠️ Issue: Silent skipping of malformed chunks
    • Line 111-112: continue on JsonException
    • No logging or telemetry for skipped chunks
  • 💡 Recommendation: Add optional logging/delegate for error handling

6. Configuration Validation

  • ⚠️ Issue: Minimal validation in MistralClientOptions
    • No validation that ApiKey is not empty when provided
    • No validation that BaseAddress ends with trailing slash (as documented)
  • 💡 Recommendation: Add validation in property setters or use validation library

7. HTTP Client Configuration

  • ⚠️ Issue: Hardcoded timeout values in resilience pipeline
    • 90s attempt timeout, 180s sampling duration, 5min total timeout
    • These may not be appropriate for all use cases
  • 💡 Recommendation: Make timeouts configurable via MistralClientOptions

8. Missing Features

  • ⚠️ Missing: No support for custom headers
  • ⚠️ Missing: No retry-after header handling
  • ⚠️ Missing: No built-in rate limit detection

🐛 Potential Bugs

1. JSON Serialization Issue

File: ChatCompletionRequest.cs, Line 137

Parts = JsonSerializer.Deserialize<List>(ref reader, options),

Issue: If deserialization fails, this will throw JsonException which isn’t caught Impact: Could break entire request parsing Fix: Wrap in try-catch similar to streaming logic

2. Content Part Initialization

File: ChatCompletionRequest.cs, Line 137

Parts = JsonSerializer.Deserialize<List>(ref reader, options),

Issue: Creates mutable List but property expects IReadOnlyList Impact: Type mismatch – should use ToList().AsReadOnly() or change property type

3. Usage Info in Streaming

File: StreamingDTOs.cs Issue: UsageInfo appears on terminal chunk, but no validation that it’s only set once Impact: Could lead to incorrect token counting if multiple usage objects received Fix: Document this behavior or add validation


📈 Performance Considerations

Aspect Status Notes
Async/Await ✅ Excellent Proper use throughout
Memory Usage ✅ Good Streaming implementation is memory-efficient
JSON Serialization ✅ Good Uses System.Text.Json
HTTP Client ✅ Good Uses IHttpClientFactory for connection pooling
Object Allocation ⚠️ Minor Could reduce allocations in hot paths

Recommendation: Consider using ArrayPool for buffer management in high-throughput scenarios


🔧 Testing Recommendations

Unit Tests Needed

  1. Happy Path: Normal request/response flow
  2. Error Scenarios: HTTP 4xx/5xx responses
  3. Streaming: Chunk parsing, malformed JSON handling
  4. Serialization: All DTO serialization/deserialization
  5. Edge Cases: Empty messages, null values, boundary conditions
  6. Configuration: Invalid options, missing API key
  7. Resilience: Retry logic, circuit breaker behavior

Integration Tests Needed

  1. Real API Calls: Test against Mistral API (with test API key)
  2. Mock Server: Test with local mock server for controlled scenarios

📝 Code Style & Best Practices

Following .NET Best Practices

Practice Status
Async method naming (Async suffix)
Cancellation token support
IDisposable pattern for resources
Argument validation
Nullable reference types enabled ✅ (but needs more annotations)
Consistent naming conventions
XML documentation

Style Issues

  • ⚠️ Brace Style: Inconsistent – some files use same-line braces, others use new-line
  • ⚠️ Indentation: Mostly consistent (4 spaces), but verify throughout
  • ⚠️ Line Length: Some lines exceed typical 100-120 character limits

🛡️ Security Considerations

Aspect Status Notes
API Key Handling ✅ Good Supports both configured and custom auth headers
HTTPS ✅ Good Defaults to https://api.mistral.ai/
Input Validation ✅ Good Argument null checks
Error Messages ⚠️ Caution Error messages include response body (could expose sensitive data)
Rate Limiting ✅ Good Built-in via resilience pipeline

Recommendation: Consider sanitizing error messages in production to avoid exposing sensitive API response data


🚀 Usage Example

// Configuration
var builder = WebApplication.CreateBuilder(args);

builder.Services.AddMistralChatClient(options =>
{
    options.ApiKey = "your-api-key";
    options.BaseAddress = new Uri("https://api.mistral.ai/");
});

// Usage
var client = serviceProvider.GetRequiredService();

// Non-streaming
var request = new ChatCompletionRequest
{
    Model = "mistral-large",
    Messages = new List
    {
        new UserMessage { Content = "Hello, how are you?" }
    }
};

var response = await client.CompleteAsync(request);

// Streaming
await foreach (var chunk in client.StreamAsync(request))
{
    if (chunk.Choices.FirstOrDefault()?.Delta?.Content is { } content)
    {
        Console.Write(content);
    }
}

📊 Summary & Rating

Category Score Notes
Architecture ⭐⭐⭐⭐⭐ Excellent separation of concerns, DI-friendly
Code Quality ⭐⭐⭐⭐☆ Very good, minor inconsistencies
Error Handling ⭐⭐⭐⭐☆ Robust, could improve logging
Performance ⭐⭐⭐⭐☆ Good, streaming is efficient
Documentation ⭐⭐⭐⭐☆ Good XML docs, could use more examples
Testing ⭐⭐☆☆☆ Needs comprehensive test coverage
Security ⭐⭐⭐⭐☆ Good foundation, watch error messages

Overall Rating: ⭐⭐⭐⭐☆ (4.2/5) – Production Ready with Minor Improvements


Action Items

High Priority

  1. [P1] Fix namespace inconsistency between Mistral.Client and Mistral.Client.Shared
  2. [P1] Fix MessageContentJsonConverter to handle all edge cases properly
  3. [P1] Add validation for MistralClientOptions.ApiKey and BaseAddress

Medium Priority

  1. [P2] Add comprehensive unit tests
  2. [P2] Standardize all DTOs as immutable records
  3. [P2] Make resilience timeouts configurable
  4. [P2] Add nullability annotations to all properties

Low Priority

  1. [P3] Add logging for streaming errors
  2. [P3] Consider adding telemetry support
  3. [P3] Add integration tests with mock server

Review completed using Mistral Vibe | Generated: 2026-07-20

Seeedstudio XIAO ESP32 S3 RS-485 test harness(Arduino)

As part of a project to read values from a MODBUS RS-485 sensor using a RS-485 Breakout Board for Seeed Studio XIAO and a Seeed Studio XIAO ESP32-S3 I built the test harness described in the wiki post. The test harness setup for a Seeed Studio XIAO ESP32-C3/Seeed Studio XIAO ESP32-C6 didn’t work with my Seeed Studio XIAO ESP32-S3.

I then did some digging looked at schematics and figured out the port mappings were different. This took a while so I tried Microsoft Copilot

I then updated the port assigned for my RS485Sender application

#include <HardwareSerial.h>

HardwareSerial RS485(1);

#define enable_pin D2

void setup() {
  Serial.begin(9600);  // Initialize the hardware serial with a baud rate of 115200
  delay(5000);

  Serial.println("RS485 Sender");

  // Wait for the hardware serial to be ready
  while (!Serial)
    ;
  Serial.println("!Serial done");

  //mySerial.begin(115200, SERIAL_8N1, 7, 6); // RX=D4(GPIO6), TX=D5(GPIO7) Doesn't work
  RS485.begin(115200, SERIAL_8N1, 6, 5);

  // Wait for the hardware serial to be ready
  while (!RS485)
    ;
  Serial.println("!RS485 done ");

  pinMode(enable_pin, OUTPUT);     // Set the enable pin as an output
  digitalWrite(enable_pin, HIGH);  // Set the enable pin to high
}

void loop() {
  if (Serial.available()) {
    String inputData = Serial.readStringUntil('\n');  // Read the data from the hardware serial until a newline character

    // If the received data is not empty
    if (inputData.length() > 0) {
      Serial.println("Send successfully");  // Print a success message
      RS485.println(inputData);             // Send the received data to the hardware serial
    }
  }
}

I then updated the port assigned for my RS485Receiver application

#include <HardwareSerial.h>

HardwareSerial RS485(1);  // Use UART2
#define enable_pin D2

void setup() {
  Serial.begin(9600);  // Initialize the hardware serial with a baud rate of 115200
  delay(5000);

  Serial.println("RS485 Receiver");

  // Wait for the hardware serial to be ready
  while (!Serial)
    ;
  Serial.println("!Serial done");

  // mySerial.begin(115200, SERIAL_8N1, 7, 6); // RX=D4(GPIO6), TX=D5(GPIO7) Doesn't seem to work
  RS485.begin(115200, SERIAL_8N1, 6, 5); 
  
    // Wait for the hardware serial to be ready
  while (!RS485)
    ;
  Serial.println("!RS485 done ");

  pinMode(enable_pin, OUTPUT);    // Set the enable pin as an output
  digitalWrite(enable_pin, LOW);  // Set the enable pin to low
}

void loop() {
  // Check if there is data available from the hardware serial
  int x = RS485.available();

  if (x) {
    String response = RS485.readString();

    Serial.println(" RS485 Response: " + response);
  }

  delay(1000);
}

Getting my test harness RS485Sender and RS485Receiver applications (inspired by Seeedstudio wiki) took quite a bit longer than expected. Using Copilot worked better than expected but I think that might be because after doing some research my prompts were better.

Cloud AI with Copilot – Faster R-CNN Azure HTTP Function Performance Setup

Introduction

The Faster R-CNN Azure HTTP Trigger function performed (not unexpectedly) differently when invoked with Fiddler Classic in the Azure Functions emulator vs. when deployed in an Azure App Plan.

The code used is a “tidied” up version of the version of the code from the Building Cloud AI with Copilot – Faster R-CNN Azure HTTP Function “Dog Food” post

public class Function1
{
   private readonly ILogger<Function1> _logger;
   private readonly List<string> _labels;
   private readonly InferenceSession _session;

   public Function1(ILogger<Function1> logger)
   {
      _logger = logger;
      _labels = File.ReadAllLines(Path.Combine(AppContext.BaseDirectory, "labels.txt")).ToList();
      _session = new InferenceSession(Path.Combine(AppContext.BaseDirectory, "FasterRCNN-10.onnx"));
   }

   [Function("ObjectDetectionFunction")]
   public async Task<IActionResult> Run([HttpTrigger(AuthorizationLevel.Function, "post", Route = null)] HttpRequest req, ExecutionContext context)
   {
      if (!req.ContentType.StartsWith("image/"))
         return new BadRequestObjectResult("Content-Type must be an image.");

      using var ms = new MemoryStream();
      await req.Body.CopyToAsync(ms);
      ms.Position = 0;

      using var image = Image.Load<Rgb24>(ms);
      var inputTensor = PreprocessImage(image);

      var inputs = new List<NamedOnnxValue>
                  {
                      NamedOnnxValue.CreateFromTensor("image", inputTensor)
                  };

      using IDisposableReadOnlyCollection<DisposableNamedOnnxValue> results = _session.Run(inputs);
      var output = results.ToDictionary(x => x.Name, x => x.Value);

      var boxes = (DenseTensor<float>)output["6379"];
      var labels = (DenseTensor<long>)output["6381"];
      var scores = (DenseTensor<float>)output["6383"];

      var detections = new List<object>();
      for (int i = 0; i < scores.Length; i++)
      {
         if (scores[i] > 0.5)
         {
            detections.Add(new
            {
               label = _labels[(int)labels[i]],
               score = scores[i],
               box = new
               {
                  x1 = boxes[i, 0],
                  y1 = boxes[i, 1],
                  x2 = boxes[i, 2],
                  y2 = boxes[i, 3]
               }
            });
         }
      }
      return new OkObjectResult(detections);
   }

   private static DenseTensor<float> PreprocessImage(Image<Rgb24> image)
   {
      // Step 1: Resize so that min(H, W) = 800, max(H, W) <= 1333, keeping aspect ratio
      int origWidth = image.Width;
      int origHeight = image.Height;
      int minSize = 800;
      int maxSize = 1333;

      float scale = Math.Min((float)minSize / Math.Min(origWidth, origHeight),
                             (float)maxSize / Math.Max(origWidth, origHeight));

      int resizedWidth = (int)Math.Round(origWidth * scale);
      int resizedHeight = (int)Math.Round(origHeight * scale);

      image.Mutate(x => x.Resize(resizedWidth, resizedHeight));

      // Step 2: Pad so that both dimensions are divisible by 32
      int padWidth = ((resizedWidth + 31) / 32) * 32;
      int padHeight = ((resizedHeight + 31) / 32) * 32;

      var paddedImage = new Image<Rgb24>(padWidth, padHeight);
      paddedImage.Mutate(ctx => ctx.DrawImage(image, new Point(0, 0), 1f));

      // Step 3: Convert to BGR and normalize
      float[] mean = { 102.9801f, 115.9465f, 122.7717f };
      var tensor = new DenseTensor<float>(new[] { 3, padHeight, padWidth });

      for (int y = 0; y < padHeight; y++)
      {
         for (int x = 0; x < padWidth; x++)
         {
            Rgb24 pixel = default;
            if (x < resizedWidth && y < resizedHeight)
               pixel = paddedImage[x, y];

            tensor[0, y, x] = pixel.B - mean[0];
            tensor[1, y, x] = pixel.G - mean[1];
            tensor[2, y, x] = pixel.R - mean[2];
         }
      }

      paddedImage.Dispose();

      return tensor;
   }
}

For my initial testing in the Azure Functions emulator using Fiddler Classic I manually generated 10 requests, then replayed them sequentially, and then finally concurrently.

The results for the manual, then sequential results were fairly consistent but the 10 concurrent requests each to took more than 10x longer. In addition, the CPU was at 100% usage while the concurrently executed functions were running.

Cloud Deployment

To see how the Faster R-CNN Azure HTTP Trigger function performed I created four resource groups.

The first contained resources used by the three different deployment models being tested

The second resource group was for testing a Dedicated hosting plan deployment.

The third resource group was for testing an Azure Functions Consumption plan hosting.

The fourth resource group was for testing Azure Functions Flex Consumption plan hosting.

Summary

The next couple of posts will compare and look at options for improving the “performance” (scalability, execution duration, latency, jitter, billing etc.) of the Github Copilot generated code.

Building Cloud AI with Copilot – Faster R-CNN Azure HTTP Function SKU Results

Introduction

While testing the FasterRCNNObjectDetectionHttpTrigger function with Telerik Fiddler Classic and my “standard” test image I noticed the response bodies were different sizes.

Initially the application plan was an S1 SKU (1 vCPU 1.75G RAM)

The output JSON was 641 bytes

[
  {
    "label": "person",
    "score": 0.9998331,
    "box": {
      "x1": 445.9223, "y1": 124.11987, "x2": 891.18915, "y2": 696.37164
    }
  },
  {
    "label": "person",
    "score": 0.9994991,
    "box": {
      "x1": 0, "y1": 330.16595, "x2": 471.0475, "y2": 761.35846
    }
  },
  {
    "label": "baseball bat",
    "score": 0.9952342,
    "box": { "x1": 869.8053, "y1": 336.96188, "x2": 1063.2261, "y2": 467.74136
    }
  },
  {
    "label": "sports ball",
    "score": 0.9945949,
    "box": { "x1": 1040.916, "y1": 372.41507, "x2": 1071.8958, "y2": 402.50424
    }
  },
  {
    "label": "baseball glove",
    "score": 0.9943546,
    "box": {
      "x1": 377.8922, "y1": 431.95053, "x2": 458.4937, "y2": 536.52124
    }
  },
  {
    "label": "person",
    "score": 0.51779467,
    "box": {
      "x1": 0, "y1": 239.91418, "x2": 60.342667, "y2": 397.17004
    }
  }
]

The application plan was scaled to a Premium v3 P0V3 (1 vCPU 4G RAM)

The output JSON was 637 bytes

[
  {
    "label": "person",
    "score": 0.9998332,
    "box": {
      "x1": 445.9223, "y1": 124.1199, "x2": 891.18915, "y2": 696.3716
    }
  },
  {
    "label": "person",
    "score": 0.9994991,
    "box": { "x1": 0, "y1": 330.16595, "x2": 471.0475, "y2": 761.35846
    }
  },
  {
    "label": "baseball bat",
    "score": 0.9952342,
    "box": {
      "x1": 869.8053, "y1": 336.9619, "x2": 1063.2261, "y2": 467.74133
    }
  },
  {
    "label": "sports ball",
    "score": 0.994595,
    "box": {
      "x1": 1040.916, "y1": 372.41507, "x2": 1071.8958, "y2": 402.50424
    }
  },
  {
    "label": "baseball glove",
    "score": 0.9943546,
    "box": {
      "x1": 377.8922, "y1": 431.95053, "x2": 458.4937, "y2": 536.52124
    }
  },
  {
    "label": "person",
    "score": 0.51779467,
    "box": {
      "x1": 0, "y1": 239.91418, "x2": 60.342667, "y2": 397.17004
    }
  }
]

The application plan was scaled to Premium v3 P1V3 (2 vCPU 8G RAM)

The output JSON was 641 bytes

[
  {
    "label": "person",
    "score": 0.9998331,
    "box": {
      "x1": 445.9223, "y1": 124.11987, "x2": 891.18915, "y2": 696.37164
    }
  },
  {
    "label": "person",
    "score": 0.9994991,
    "box": {
      "x1": 0, "y1": 330.16595, "x2": 471.0475, "y2": 761.35846
    }
  },
  {
    "label": "baseball bat",
    "score": 0.9952342,
    "box": {
      "x1": 869.8053, "y1": 336.96188, "x2": 1063.2261, "y2": 467.74136
    }
  },
  {
    "label": "sports ball",
    "score": 0.9945949,
    "box": {
      "x1": 1040.916, "y1": 372.41507, "x2": 1071.8958, "y2": 402.50424
    }
  },
  {
    "label": "baseball glove",
    "score": 0.9943546,
    "box": {
      "x1": 377.8922, "y1": 431.95053, "x2": 458.4937, "y2": 536.52124
    }
  },
  {
    "label": "person",
    "score": 0.51779467,
    "box": {
      "x1": 0, "y1": 239.91418, "x2": 60.342667, "y2": 397.17004
    }
  }
]

The application plan was scaled to a Premium v3 P2V3 (4 vCPU 16G RAM)

The output JSON was 641 bytes

[
  {
    "label": "person",
    "score": 0.9998331,
    "box": {
      "x1": 445.9223, "y1": 124.11987, "x2": 891.18915, "y2": 696.37164
    }
  },
  {
    "label": "person",
    "score": 0.9994991,
    "box": {
      "x1": 0, "y1": 330.16595, "x2": 471.0475, "y2": 761.35846
    }
  },
  {
    "label": "baseball bat",
    "score": 0.9952342,
    "box": {
      "x1": 869.8053, "y1": 336.96188, "x2": 1063.2261, "y2": 467.74136
    }
  },
  {
    "label": "sports ball",
    "score": 0.9945949,
    "box": {
      "x1": 1040.916, "y1": 372.41507, "x2": 1071.8958, "y2": 402.50424
    }
  },
  {
    "label": "baseball glove",
    "score": 0.9943546,
    "box": {
      "x1": 377.8922, "y1": 431.95053, "x2": 458.4937, "y2": 536.52124 }
  },
  {
    "label": "person",
    "score": 0.51779467,
    "box": {
      "x1": 0, "y1": 239.91418, "x2": 60.342667, "y2": 397.17004
    }
  }
]

The application plan was scaled to a Premium v2 P1V2 (1vCPU 3.5G)

The output JSON was 637 bytes

[
  {
    "label": "person",
    "score": 0.9998332,
    "box": {
      "x1": 445.9223, "y1": 124.1199, "x2": 891.18915, "y2": 696.3716
    }
  },
  {
    "label": "person",
    "score": 0.9994991,
    "box": {
      "x1": 0, "y1": 330.16595, "x2": 471.0475, "y2": 761.35846
    }
  },
  {
    "label": "baseball bat",
    "score": 0.9952342,
    "box": {
      "x1": 869.8053, "y1": 336.9619, "x2": 1063.2261, "y2": 467.74133
    }
  },
  {
    "label": "sports ball",
    "score": 0.994595,
    "box": {
      "x1": 1040.916, "y1": 372.41507, "x2": 1071.8958, "y2": 402.50424
    }
  },
  {
    "label": "baseball glove",
    "score": 0.9943546,
    "box": {
      "x1": 377.8922, "y1": 431.95053, "x2": 458.4937, "y2": 536.52124
    }
  },
  {
    "label": "person",
    "score": 0.51779467,
    "box": {
      "x1": 0, "y1": 239.91418, "x2": 60.342667, "y2": 397.17004
    }
  }
]

Summary

The differences between the 637 & 641were small

Not certain why this could happen currently best guess is memory pressure.

Building Cloud AI with Copilot – Faster R-CNN Azure HTTP Function “Dog Food”

Introduction

A couple of months ago a web crawler visited every page on my website (would be interesting to know if my Github repositories were crawled as well) and I wondered if this might impact my Copilot or Github Copilot experiments. My blogging about The Azure HTTP Trigger functions with Ultralytics Yolo, YoloSharp, Resnet, Faster R-CNN, with Open Neural Network Exchange(ONNX) etc. is fairly “niche” so any improvements in the understanding of the problems and generated code might be visible.

please write an httpTrigger azure function that uses Faster RCNN and ONNX to detect the object in an image uploaded in the body of an HTTP Post

Github Copilot had used Sixlabors ImageSharp, the ILogger was injected into the constructor, the code checked that the image was in the body of the HTTP POST and the object classes were loaded from a text file. I had to manually add some Nugets and using directives before the code compiled and ran in the emulator, but this was a definite improvement.

To test the implementation, I was using Telerik Fiddler Classic to HTTP POST my “standard” test image to function.

Github Copilot had generated code that checked that the image was in the body of the HTTP POST so I had to modify the Telerik Fiddler Classic request.

I also had to fix up the content-type header

The path to the onnx file was wrong and I had to create a labels.txt file from Python code.

The Azure HTTP Trigger function ran but failed because the preprocessing of the image didn’t implement the specified preprocess steps.

Change DenseTensor to BGR (based on https://github.com/onnx/models/tree/main/validated/vision/object_detection_segmentation/faster-rcnn#preprocessing-steps)

Normalise colour values with mean = [102.9801, 115.9465, 122.7717]

The Azure HTTP Trigger function ran but failed because the output tensor names were incorrect

I used Netron to inspect the model properties to get the correct names for the output tensors

I had a couple of attempts at resizing the image to see what impact this had on the accuracy of the confidence and minimum bounding rectangles.

resize the image such that both height and width are within the range of [800, 1333], and then pad the image with zeros such that both height and width are divisible by 32.

modify the code to resize the image such that both height and width are within the range of [800, 1333], and then pad the image with zeros such that both height and width are divisible by 32 and the aspect ratio is not changed.

The final version of the image processing code scaled then right padded the image to keep the aspect ratio and MBR coordinates correct.

As a final test I deployed the code to Azure and the first time I ran the function it failed because the labels file couldn’t be found because Unix file paths are case sensitive (labels.txt vs. Labels.txt).

The inferencing time was a bit longer than I expected.

// please write an httpTrigger azure function that uses Faster RCNN and ONNX to detect the object in an image uploaded in the body of an HTTP Post
//    manually added the ML.Net ONNX NuGet + using directives
//    manually added the ImageSharp NuGet + using directives
//    Used Copilot to add Microsoft.ML.OnnxRuntime.Tensors using directive
//    Manually added ONNX FIle + labels file sorted out paths
//    Used Netron to fixup output tensor names
// Change DenseTensor to BGR (based on https://github.com/onnx/models/tree/main/validated/vision/object_detection_segmentation/faster-rcnn#preprocessing-steps)
// Normalise colour values with mean = [102.9801, 115.9465, 122.7717]
// resize the image such that both height and width are within the range of [800, 1333], and then pad the image with zeros such that both height and width are divisible by 32.
// modify the code to resize the image such that both height and width are within the range of [800, 1333], and then pad the image with zeros such that both height and width are divisible by 32 and the aspect ratio is not changed.
using Microsoft.AspNetCore.Http;
using Microsoft.AspNetCore.Mvc;
using Microsoft.Azure.Functions.Worker;
using Microsoft.Extensions.Logging;
using Microsoft.ML.OnnxRuntime;
using Microsoft.ML.OnnxRuntime.Tensors;
using SixLabors.ImageSharp; // Couldn't get inteliisense after adding NuGet package
using SixLabors.ImageSharp.PixelFormats; // Couldn't get inteliisense after adding NuGet package
using SixLabors.ImageSharp.Processing; // Couldn't get inteliisense after adding NuGet package


namespace FasterRCNNObjectDetectionHttpTriggerGithubCopilot
{
   public class Function1
   {
      private readonly ILogger<Function1> _logger;
      private readonly InferenceSession _session;
      private readonly List<string> _labels;

      public Function1(ILogger<Function1> logger)
      {
         _logger = logger;
         _session = new InferenceSession("FasterRCNN-10.onnx");
         _labels = File.ReadAllLines("labels.txt").ToList();
      }

      [Function("ObjectDetectionFunction")]
      public async Task<IActionResult> Run([HttpTrigger(AuthorizationLevel.Function, "post", Route = null)] HttpRequest req)
      {
         if (!req.ContentType.StartsWith("image/"))
            return new BadRequestObjectResult("Content-Type must be an image.");

         using var ms = new MemoryStream();
         await req.Body.CopyToAsync(ms);
         ms.Position = 0;

         using var image = Image.Load<Rgb24>(ms);
         var inputTensor = PreprocessImage(image);

         var inputs = new List<NamedOnnxValue>
                  {
                      NamedOnnxValue.CreateFromTensor("image", inputTensor)
                  };

         using IDisposableReadOnlyCollection<DisposableNamedOnnxValue> results = _session.Run(inputs);
         var output = results.ToDictionary(x => x.Name, x => x.Value);

         var boxes = (DenseTensor<float>)output["6379"];
         var labels = (DenseTensor<long>)output["6381"];
         var scores = (DenseTensor<float>)output["6383"];

         var detections = new List<object>();
         for (int i = 0; i < scores.Length; i++)
         {
            if (scores[i] > 0.5)
            {
               detections.Add(new
               {
                  label = _labels[(int)labels[i]],
                  score = scores[i],
                  box = new
                  {
                     x1 = boxes[i, 0],
                     y1 = boxes[i, 1],
                     x2 = boxes[i, 2],
                     y2 = boxes[i, 3]
                  }
               });
            }
         }

         return new OkObjectResult(detections);
      }

      private static DenseTensor<float> PreprocessImage( Image<Rgb24> image)
      {
         // Step 1: Resize so that min(H, W) = 800, max(H, W) <= 1333, keeping aspect ratio
         int origWidth = image.Width;
         int origHeight = image.Height;
         int minSize = 800;
         int maxSize = 1333;

         float scale = Math.Min((float)minSize / Math.Min(origWidth, origHeight),
                                (float)maxSize / Math.Max(origWidth, origHeight));
         /*
         float scale = 1.0f;

         // If either dimension is less than 800, scale up so the smaller is 800
         if (origWidth < minSize || origHeight < minSize)
         {
            scale = Math.Max((float)minSize / origWidth, (float)minSize / origHeight);
         }
         // If either dimension is greater than 1333, scale down so the larger is 1333
         if (origWidth * scale > maxSize || origHeight * scale > maxSize)
         {
            scale = Math.Min((float)maxSize / origWidth, (float)maxSize / origHeight);
         }
         */

         int resizedWidth = (int)Math.Round(origWidth * scale);
         int resizedHeight = (int)Math.Round(origHeight * scale);

         image.Mutate(x => x.Resize(resizedWidth, resizedHeight));

         // Step 2: Pad so that both dimensions are divisible by 32
         int padWidth = ((resizedWidth + 31) / 32) * 32;
         int padHeight = ((resizedHeight + 31) / 32) * 32;

         var paddedImage = new Image<Rgb24>(padWidth, padHeight);
         paddedImage.Mutate(ctx => ctx.DrawImage(image, new Point(0, 0), 1f));

         // Step 3: Convert to BGR and normalize
         float[] mean = { 102.9801f, 115.9465f, 122.7717f };
         var tensor = new DenseTensor<float>(new[] { 3, padHeight, padWidth });

         for (int y = 0; y < padHeight; y++)
         {
            for (int x = 0; x < padWidth; x++)
            {
               Rgb24 pixel = default;
               if (x < resizedWidth && y < resizedHeight)
                  pixel = paddedImage[x, y];

               tensor[0, y, x] = pixel.B - mean[0];
               tensor[1, y, x] = pixel.G - mean[1];
               tensor[2, y, x] = pixel.R - mean[2];
            }
         }

         paddedImage.Dispose();
         return tensor;
      }
   }
}

It took roughly an hour to “vibe code” the function, but it would have taken much longer for someone not familiar with the problem domain.

Summary

The Github Copilot generated code was okay but would be fragile, performance would suck and not scale terribly well.

The Copilot generated code in this post is not suitable for production

ONNXRuntime.AI-Faster R-CNN C# Sample differences

After building Faster R-CCN object detection applications with Copilot and Github Copilot the results when compared with the onnxruntime.ai Object detection with Faster RCNN Deep Learning in C# sample (which hasn’t been updated for years) were slightly different.

The sample image was 640×480 pixels

The FasterRCNNObjectDetectionApplicationGitHubCopilot application scaled image was initially 1056×800 then 1088×800 pixels.

The initial version the dimensions were “rounded down” to the next multiple of 32

// Calculate scale factor to fit within the range while maintaining aspect ratio
float scale = Math.Min((float)maxSize / Math.Max(originalWidth, originalHeight),
                                (float)minSize / Math.Min(originalWidth, originalHeight));

// Calculate new dimensions
int newWidth = (int)(originalWidth * scale);
int newHeight = (int)(originalHeight * scale);

// Ensure dimensions are divisible by 32
newWidth = (newWidth / divisor) * divisor;
newHeight = (newHeight / divisor) * divisor;
Scaled 1056×800

Then for the second version the dimensions were “rounded up” to the next multiple of 32

// Calculate scale factor to fit within the range while maintaining aspect ratio
float scale = Math.Min((float)maxSize / Math.Max(originalWidth, originalHeight),
                                (float)minSize / Math.Min(originalWidth, originalHeight));

// Calculate new dimensions
int newWidth = (int)(originalWidth * scale);
int newHeight = (int)(originalHeight * scale);

// Ensure dimensions are divisible by 32
newWidth = (int)(Math.Ceiling(newWidth / 32f) * 32f);
newHeight = (int)(Math.Ceiling(newHeight / 32f) * 32f);
Scaled 1088×800
Marked up 1088×800

The FasterRCNNObjectDetectionApplicationOriginal application scaled the input image to 1066×800

Scaled image 1066×800

The FasterRCNNObjectDetectionApplicationOriginal application pillar boxed/padded the image to 1088×800 as the DenseTensor was loaded.

using Image<Rgb24> image = Image.Load<Rgb24>(imageFilePath);

Console.WriteLine($"Before x:{image.Width} y:{image.Height}");

// Resize image
float ratio = 800f / Math.Min(image.Width, image.Height);
image.Mutate(x => x.Resize((int)(ratio * image.Width), (int)(ratio * image.Height)));

Console.WriteLine($"After x:{image.Width} y:{image.Height}");

// Preprocess image
var paddedHeight = (int)(Math.Ceiling(image.Height / 32f) * 32f);
var paddedWidth = (int)(Math.Ceiling(image.Width / 32f) * 32f);

Console.WriteLine($"Padded x:{paddedWidth} y:{paddedHeight}");

Tensor<float> input = new DenseTensor<float>(new[] { 3, paddedHeight, paddedWidth });
var mean = new[] { 102.9801f, 115.9465f, 122.7717f };
image.ProcessPixelRows(accessor =>
{
   for (int y = paddedHeight - accessor.Height; y < accessor.Height; y++)
   {
      Span<Rgb24> pixelSpan = accessor.GetRowSpan(y);
      for (int x = paddedWidth - accessor.Width; x < accessor.Width; x++)
      {
         input[0, y, x] = pixelSpan[x].B - mean[0];
         input[1, y, x] = pixelSpan[x].G - mean[1];
         input[2, y, x] = pixelSpan[x].R - mean[2];
      }
   }
});
Marked up image 1066×800

I think the three different implementations of the preprocessing steps and the graphics libraries used probably caused the differences in the results. The way an image is “resized” by System.Graphics.Common vs. ImageSharp(resampled, cropped and centered or padded and pillar boxed) could make a significant difference to the results.

ONNXRuntime.AI-Faster R-CNN C# Sample oddness

After building Faster R-CCN object detection applications with Copilot and Github Copilot the results when compared with Utralytics Yolo (with YoloSharp) didn’t look too bad.

The input image sports.jpg 1200×798 pixels

The GithubCopilot FasterRCNNObjectDetectionApplicationCopilot application only generated labels, confidences and minimum bounding box coordinates.

The FasterRCNNObjectDetectionApplicationGitHubCopilot application the marked-up image was 1200×798 pixels

The YoloSharpObjectDetectionApplication application marked-up image was 1200×798 pixels

I went back to the onnxruntime.ai Object detection with Faster RCNN Deep Learning in C# sample source code to check my implementations and the highlighted area on the left caught my attention.

The FasterRCNNObjectDetectionApplicationOriginal application marked up image was 1023×800

I downloaded the sample code which hadn’t been updated for years.

public static void Main(string[] args)
{
   Console.WriteLine("FasterRCNNObjectDetectionApplicationOriginal");

   // Read paths
   string modelFilePath = args[0];
   string imageFilePath = args[1];
   string outImageFilePath = args[2];

   // Read image
   using Image<Rgb24> image = Image.Load<Rgb24>(imageFilePath);

   // Resize image
   float ratio = 800f / Math.Min(image.Width, image.Height);
   image.Mutate(x => x.Resize((int)(ratio * image.Width), (int)(ratio * image.Height)));

   // Preprocess image
   var paddedHeight = (int)(Math.Ceiling(image.Height / 32f) * 32f);
   var paddedWidth = (int)(Math.Ceiling(image.Width / 32f) * 32f);
   Tensor<float> input = new DenseTensor<float>(new[] { 3, paddedHeight, paddedWidth });
   var mean = new[] { 102.9801f, 115.9465f, 122.7717f };
   image.ProcessPixelRows(accessor =>
   {
      for (int y = paddedHeight - accessor.Height; y < accessor.Height; y++)
      {
         Span<Rgb24> pixelSpan = accessor.GetRowSpan(y);
         for (int x = paddedWidth - accessor.Width; x < accessor.Width; x++)
         {
            input[0, y, x] = pixelSpan[x].B - mean[0];
            input[1, y, x] = pixelSpan[x].G - mean[1];
            input[2, y, x] = pixelSpan[x].R - mean[2];
         }
      }
   });

   // Setup inputs and outputs
   var inputs = new List<NamedOnnxValue>
      {
            NamedOnnxValue.CreateFromTensor("image", input)
      };

   // Run inference
   using var session = new InferenceSession(modelFilePath);
   using IDisposableReadOnlyCollection<DisposableNamedOnnxValue> results = session.Run(inputs);

   // Postprocess to get predictions
   var resultsArray = results.ToArray();
   float[] boxes = resultsArray[0].AsEnumerable<float>().ToArray();
   long[] labels = resultsArray[1].AsEnumerable<long>().ToArray();
   float[] confidences = resultsArray[2].AsEnumerable<float>().ToArray();
   var predictions = new List<Prediction>();
   var minConfidence = 0.7f;
   for (int i = 0; i < boxes.Length - 4; i += 4)
   {
      var index = i / 4;
      if (confidences[index] >= minConfidence)
      {
         predictions.Add(new Prediction
         {
            Box = new Box(boxes[i], boxes[i + 1], boxes[i + 2], boxes[i + 3]),
            Label = LabelMap.Labels[labels[index]],
            Confidence = confidences[index]
         });
      }
   }

   // Put boxes, labels and confidence on image and save for viewing
   using var outputImage = File.OpenWrite(outImageFilePath);
   Font font = SystemFonts.CreateFont("Arial", 16);
   foreach (var p in predictions)
   {
      Console.WriteLine($"Label: {p.Label}, Confidence: {p.Confidence}, Bounding Box:[{p.Box.Xmin}, {p.Box.Ymin}, {p.Box.Xmax}, {p.Box.Ymax}]");
      image.Mutate(x =>
      {
         x.DrawLine(Color.Red, 2f, new PointF[] {

                  new PointF(p.Box.Xmin, p.Box.Ymin),
                  new PointF(p.Box.Xmax, p.Box.Ymin),

                  new PointF(p.Box.Xmax, p.Box.Ymin),
                  new PointF(p.Box.Xmax, p.Box.Ymax),

                  new PointF(p.Box.Xmax, p.Box.Ymax),
                  new PointF(p.Box.Xmin, p.Box.Ymax),

                  new PointF(p.Box.Xmin, p.Box.Ymax),
                  new PointF(p.Box.Xmin, p.Box.Ymin)
               });
         x.DrawText($"{p.Label}, {p.Confidence:0.00}", font, Color.White, new PointF(p.Box.Xmin, p.Box.Ymin));
      });
   }
   image.SaveAsJpeg(outputImage);

   Console.WriteLine("Press Enter to exit");
   Console.ReadLine();
}

I then compared the output of the object detection applications and the onnxruntime.ai Object detection with Faster RCNN Deep Learning in C# sample was different.

After some investigation I think the scaling of the image used for inferencing (based on the requirements on the model), then the scaling of the minimum bounding rectangles isn’t quite right.

Building Cloud AI with Copilot – ResNet50 Azure HTTP Function

Introduction

This is another awfully long post about my experience using Copilot to write an Azure HTTP Trigger function that runs a resnet50 V2.7 Open Neural Network Exchange model(ONNX) on an image in the body of the HTTP POST.

For testing I was uploading the images with Telerik Fiddler Classic.

I forgot to specify language, so Copilot assumed (reasonably) that I wanted a Python Azure HTTP Trigger function.

The initial C# version wouldn’t compile because of the FunctionName attribute which is used for in-process Azure Functions. It did seem a bit odd that Copilot would generate code that support will end for November 10, 2026

public static class Function1
{
   private static readonly InferenceSession session = new InferenceSession("resnet50.onnx");

   [FunctionName("ImageClassification")]
   public static IActionResult Run(
       [HttpTrigger(AuthorizationLevel.Function, "post")] HttpRequest req,
       ILogger log)
   {
      log.LogInformation("Processing image classification request...");

      try
      {
         using var ms = new MemoryStream();
         req.Body.CopyTo(ms);
         using var image = Image.FromStream(ms);

         var inputTensor = PreprocessImage(image);

         var inputName = session.InputMetadata.Keys.First();
         var outputName = session.OutputMetadata.Keys.First();
         var result = session.Run(new Dictionary<string, NamedOnnxValue>
            {
                { inputName, NamedOnnxValue.CreateFromTensor(inputName, inputTensor) }
            });

         var predictions = result.First().AsTensor<float>().ToArray();

         return new JsonResult(new { predictions });
      }
      catch (Exception ex)
      {
         log.LogError($"Error: {ex.Message}");
         return new BadRequestObjectResult("Invalid image or request.");
      }
   }
...
}

It was just easier to change the FunctionName attribute manually.

public static class Function1
{
   private static readonly InferenceSession session = new InferenceSession("resnet50.onnx");

   [Function("ImageClassification")]
   public static IActionResult Run(
       [HttpTrigger(AuthorizationLevel.Function, "post")] HttpRequest req,
       ILogger log)
   {
      log.LogInformation("Processing image classification request...");

      try
      {
         using var ms = new MemoryStream();
         req.Body.CopyTo(ms);
         using var image = Image.FromStream(ms);

         var inputTensor = PreprocessImage(image);

         var inputName = session.InputMetadata.Keys.First();
         var outputName = session.OutputMetadata.Keys.First();
         var inputList = new List<NamedOnnxValue>
            {
                NamedOnnxValue.CreateFromTensor(inputName, inputTensor)
            };

         var result = session.Run(inputList);

         var predictions = result.First().AsTensor<float>().ToArray();

         return new JsonResult(new { predictions });
      }
      catch (Exception ex)
      {
         log.LogError($"Error: {ex.Message}");
         return new BadRequestObjectResult("Invalid image or request.");
      }
   }

The Azure HTTP Trigger function ran but failed when I tried to classify an image

The initialisation of the ILogger injected into the Run method was broken so I used Copilot to update the code to use constructor Dependency Injection (DI).

public static class Function1
{
   private static readonly ILogger logger;
   private static readonly InferenceSession session = new InferenceSession("resnet50-v2-7.onnx");

   // Static constructor to initialize logger
   static Function1()
   {
      var loggerFactory = LoggerFactory.Create(builder =>
      {
         builder.AddConsole();
      });
      logger = loggerFactory.CreateLogger("Function1Logger");
   }

   [Function("ImageClassification")]
   public static IActionResult Run([HttpTrigger(AuthorizationLevel.Function, "post")] HttpRequest req)
   {
      logger.LogInformation("Processing image classification request...");

      try
      {
         using var ms = new MemoryStream();
         req.Body.CopyTo(ms);
         using var image = Image.FromStream(ms);

         var inputTensor = PreprocessImage(image);

         var inputName = session.InputMetadata.Keys.First();
         var outputName = session.OutputMetadata.Keys.First();
         var inputList = new List<NamedOnnxValue>
            {
                NamedOnnxValue.CreateFromTensor(inputName, inputTensor)
            };

         var result = session.Run(inputList);

         var predictions = result.First().AsTensor<float>().ToArray();

         return new JsonResult(new { predictions });
      }
      catch (Exception ex)
      {
         logger.LogError($"Error: {ex.Message}");
         return new BadRequestObjectResult("Invalid image or request.");
      }
   }
...
}

It was a bit odd that Copilot generated a static function and constructor unlike the equivalent YoloSharp Azure HTTP Trigger.

The Azure HTTP Trigger function ran but failed when I tried to classify an image

The Azure HTTP Trigger function ran but failed with a 400 Bad Request when I tried to classify an image

After some debugging I realised that Telerik Fiddle Classic was sending the image as form data so I modified the “composer” payload configuration.

Then the Azure HTTP Trigger function ran but the confidence values were wrong.

The confidence values were incorrect, so I checked the ResNet50 pre-processing instructions

The image needs to be preprocessed before fed to the network. The first step is to extract a 224x224 crop from the center of the image. For this, the image is first scaled to a minimum size of 256x256, while keeping aspect ratio. That is, the shortest side of the image is resized to 256 and the other side is scaled accordingly to maintain the original aspect ratio. After that, the image is normalized with mean = 255*[0.485, 0.456, 0.406] and std = 255*[0.229, 0.224, 0.225]. Last step is to transpose it from HWC to CHW layout.
 private static Tensor<float> PreprocessImage(Image image)
 {
    var resized = new Bitmap(image, new Size(224, 224));
    var tensorData = new float[1 * 3 * 224 * 224];

    float[] mean = { 0.485f, 0.456f, 0.406f };
    float[] std = { 0.229f, 0.224f, 0.225f };

    for (int y = 0; y < 224; y++)
    {
       for (int x = 0; x < 224; x++)
       {
          var pixel = resized.GetPixel(x, y);

          tensorData[(0 * 3 * 224 * 224) + (0 * 224 * 224) + (y * 224) + x] = (pixel.R / 255.0f - mean[0]) / std[0];
          tensorData[(0 * 3 * 224 * 224) + (1 * 224 * 224) + (y * 224) + x] = (pixel.G / 255.0f - mean[1]) / std[1];
          tensorData[(0 * 3 * 224 * 224) + (2 * 224 * 224) + (y * 224) + x] = (pixel.B / 255.0f - mean[2]) / std[2];
       }
    }

    return new DenseTensor<float>(tensorData, new[] { 1, 3, 224, 224 });
 }

When the “normalisation” code was implemented and the Azure HTTP Trigger function run the confidence values were still incorrect.

The Azure HTTP Trigger function was working reliably but the number of results and size response payload was unnecessary.

The Azure HTTP Trigger function ran but the confidence values were still incorrect, so I again checked the ResNet50 post-processing instructions

Postprocessing
The post-processing involves calculating the softmax probability scores for each class. You can also sort them to report the most probable classes. Check imagenet_postprocess.py for code.
 // Compute exponentials for all scores
 var expScores = predictions.Select(MathF.Exp).ToArray();

 // Compute sum of exponentials
 float sumExpScores = expScores.Sum();

 // Normalize scores into probabilities
 var softmaxResults = expScores.Select(score => score / sumExpScores).ToArray();

 // Get top 10 predictions (label ID and confidence)
 var top10 = softmaxResults
     .Select((confidence, labelId) => new { labelId, confidence, label = labelId < labels.Count ? labels[labelId] : $"Unknown-{labelId}" })
     .OrderByDescending(p => p.confidence)
     .Take(10)
     .ToList();

The Azure HTTP Trigger function should run on multiple platforms so System.Drawing.Comon had to be replaced with Sixlabors ImageSharp

The Azure HTTP Trigger function ran but the Sixlabors ImageSharp based image classification failed.

After some debugging I realised that the MemoryStream used to copy the HTTPRequest body was not being reset.

[Function("ImageClassification")]
public static async Task<IActionResult> Run(
    [HttpTrigger(AuthorizationLevel.Function, "post")] HttpRequest req)
{
   logger.LogInformation("Processing image classification request...");

   try
   {
      using var ms = new MemoryStream();
      await req.Body.CopyToAsync(ms);

      ms.Seek(0, SeekOrigin.Begin);

      using var image = Image.Load<Rgb24>(ms);

      var inputTensor = PreprocessImage(image);
...   
   }
   catch (Exception ex)
   {
      logger.LogError($"Error: {ex.Message}");
      return new BadRequestObjectResult("Invalid image or request.");
   }
}

The odd thing was the confidence values changed slightly when the code was modified to use Sixlabors ImageSharp

The Azure HTTP Trigger function worked but the labelId wasn’t that “human readable”.

public static class Function1
{
   private static readonly ILogger logger;
   private static readonly InferenceSession session = new InferenceSession("resnet50-v2-7.onnx");
   private static readonly List<string> labels = LoadLabels("labels.txt");
...
   [Function("ImageClassification")]
   public static async Task<IActionResult> Run(
       [HttpTrigger(AuthorizationLevel.Function, "post")] HttpRequest req)
   {
      logger.LogInformation("Processing image classification request...");

      try
      {
...
         // Get top 10 predictions (label ID and confidence)
         var top10 = softmaxResults
             .Select((confidence, labelId) => new { labelId, confidence, label = labelId < labels.Count ? labels[labelId] : $"Unknown-{labelId}" })
             .OrderByDescending(p => p.confidence)
             .Take(10)
             .ToList();

         return new JsonResult(new { predictions = top10 });
      }
      catch (Exception ex)
      {
         logger.LogError($"Error: {ex.Message}");
         return new BadRequestObjectResult("Invalid image or request.");
      }
   }
...
   private static List<string> LoadLabels(string filePath)
   {
      try
      {
         return File.ReadAllLines(filePath).ToList();
      }
      catch (Exception ex)
      {
         logger.LogError($"Error loading labels file: {ex.Message}");
         return new List<string>(); // Return empty list if file fails to load
      }
   }
}

Summary

The Github Copilot generated code was okay but would be fragile and not scale terribly well. The confidence values changing very slightly when the code was updated for Sixlabors ImageSharp was disconcerting, but not surprising.

The Copilot generated code in this post is not suitable for production

Building Edge AI with Copilot-ResNet50 Client

Introduction

This is an awfully long post about my experience using Copilot to write a console application that runs a validated resnet50 V2.7 Open Neural Network Exchange model(ONNX) on an image loaded from disk.

I have found that often Copilot code generation is “better” but the user interface can be limiting.

The Copilot code generated compiled after the System.Drawing.Common and Microsoft.ML.OnnxRuntime NuGet packages were added to the project.

Input
All pre-trained models expect input images normalized in the same way, i.e. mini-batches 
of 3-channel RGB images of shape (N x 3 x H x W), where N is the batch size, and H and 
W are expected to be at least 224. The inference was done using jpeg image.

Preprocessing
The image needs to be preprocessed before fed to the network. The first step is to 
extract a 224x224 crop from the center of the image. For this, the image is first scaled 
to a minimum size of 256x256, while keeping aspect ratio. That is, the shortest side 
of the image is resized to 256 and the other side is scaled accordingly to maintain 
the original aspect ratio. 

After that, the image is normalized with mean = 255*[0.485, 0.456, 0.406] and std 
= 255*[0.229, 0.224, 0.225]. Last step is to transpose it from HWC to CHW layout.

The code also had a reasonable implementation of the ResnetV5 preprocessing instructions

static void Main()
{
   string modelPath = "resnet50-v2-7.onnx"; // Path to your ONNX model
   string imagePath = "pizza.jpg"; // Path to the input image

   using var session = new InferenceSession(modelPath);
   var inputTensor = LoadAndPreprocessImage(imagePath);

   var inputs = new List<NamedOnnxValue>
   {
       NamedOnnxValue.CreateFromTensor("input", inputTensor)
   };

   using var results = session.Run(inputs);
   var output = results.First().AsTensor<float>().ToArray();

   Console.WriteLine("Predicted class index: " + Array.IndexOf(output, output.Max()));
}

static DenseTensor<float> LoadAndPreprocessImage(string imagePath)
{
   using Bitmap bitmap = new Bitmap(imagePath);
   int width = 224, height = 224; // ResNet50 expects 224x224 input
   using Bitmap resized = new Bitmap(bitmap, new Size(width, height));

   var tensor = new DenseTensor<float>(new[] { 1, 3, width, height });
   for (int y = 0; y < height; y++)
   {
      for (int x = 0; x < width; x++)
      {
         Color pixel = resized.GetPixel(x, y);
         tensor[0, 0, y, x] = pixel.R / 255f; // Normalize
         tensor[0, 1, y, x] = pixel.G / 255f;
         tensor[0, 2, y, x] = pixel.B / 255f;
      }
   }
   return tensor;
}

The program ran but failed with a Microsoft.ML.OnnxRuntime.OnnxRuntimeException Message=[ErrorCode:InvalidArgument] Input name: ‘input’ is not in the metadata

The name of the input tensor was wrong, so I used Netron to inspect the graph properties of the model.

After the input tensor name was updated, the program ran

I checked the labels using the torchvison ImageNet categories and the results looked reasonable

The model and input file paths were wrong and I had been manually fixing them.

The confidence values didn’t look right so I re-read the preprocessing requirements for a ResNet model

Input
All pre-trained models expect input images normalized in the same way, i.e. mini-batches 
of 3-channel RGB images of shape (N x 3 x H x W), where N is the batch size, and H and 
W are expected to be at least 224. The inference was done using jpeg image.

Preprocessing
The image needs to be preprocessed before fed to the network. The first step is to 
extract a 224x224 crop from the center of the image. For this, the image is first scaled 
to a minimum size of 256x256, while keeping aspect ratio. That is, the shortest side 
of the image is resized to 256 and the other side is scaled accordingly to maintain 
the original aspect ratio. 

After that, the image is normalized with mean = 255*[0.485, 0.456, 0.406] and std 
= 255*[0.229, 0.224, 0.225]. Last step is to transpose it from HWC to CHW layout.

The Copilot generated code compiled and ran but the confidence values still didn’t look right, and the results tensor contained 1000 confidences values.

static void Main()
{
   string modelPath = "resnet50-v2-7.onnx"; // Updated model path
   string imagePath = "pizza.jpg"; // Updated image path

   using var session = new InferenceSession(modelPath);
   var inputTensor = LoadAndPreprocessImage(imagePath);

   var inputs = new List<NamedOnnxValue>
   {
      NamedOnnxValue.CreateFromTensor("data", inputTensor) // Using "data" as the input tensor name
   };

   using var results = session.Run(inputs);
   var output = results.First().AsTensor<float>().ToArray();

   Console.WriteLine("Predicted class index: " + Array.IndexOf(output, output.Max()));
}

static DenseTensor<float> LoadAndPreprocessImage(string imagePath)
{
   using Bitmap bitmap = new Bitmap(imagePath);
   int width = 224, height = 224; // ResNet50 expects 224x224 input
   using Bitmap resized = new Bitmap(bitmap, new Size(width, height));

   var tensor = new DenseTensor<float>(new[] { 1, 3, width, height });

   // ImageNet mean & standard deviation values
   float[] mean = { 0.485f, 0.456f, 0.406f };
   float[] stdev = { 0.229f, 0.224f, 0.225f };

   for (int y = 0; y < height; y++)
   {
      for (int x = 0; x < width; x++)
      {
         Color pixel = resized.GetPixel(x, y);

         // Normalize using mean and standard deviation
         tensor[0, 0, y, x] = (pixel.R / 255f - mean[0]) / stdev[0]; // Red channel
         tensor[0, 1, y, x] = (pixel.G / 255f - mean[1]) / stdev[1]; // Green channel
         tensor[0, 2, y, x] = (pixel.B / 255f - mean[2]) / stdev[2]; // Blue channel
      }
   }
   return tensor;
}

Because the results didn’t look right, I went back and read the ResNet50 post processing instructions

Output
The model outputs image scores for each of the 1000 classes of ImageNet.

Postprocessing
The post-processing involves calculating the softmax probability scores for each 
class. You can also sort them to report the most probable classes. Check 
imagenet_postprocess.py for code.

The Copilot generated code wouldn’t compile due to a syntax error.

static void Main()
{
   string modelPath = "resnet50-v2-7.onnx"; // Updated model path
   string imagePath = "pizza.jpg"; // Updated image path

   using var session = new InferenceSession(modelPath);
   var inputTensor = LoadAndPreprocessImage(imagePath);

   var inputs = new List<NamedOnnxValue>
   {
      NamedOnnxValue.CreateFromTensor("data", inputTensor) // Using "data" as the input tensor name
   };

   using var results = session.Run(inputs);
   var output = results.First().AsTensor<float>().ToArray();

   // Calculate softmax
   var probabilities = Softmax(output);

   // Get the class index with the highest probability
   int predictedClass = Array.IndexOf(probabilities, probabilities.Max());
   Console.WriteLine($"Predicted class index: {predictedClass}");
   Console.WriteLine($"Probabilities: {string.Join(", ", probabilities.Select(p => p.ToString("F4")))}");
}
...
static float[] Softmax(float[] logits)
{
   // Compute softmax
   var expScores = logits.Select(Math.Exp).ToArray();
   double sumExpScores = expScores.Sum();
   return expScores.Select(score => (float)(score / sumExpScores)).ToArray();
}

Copilot was adamant that the generated code was correct.

After trying different Copilot prompts the code had to be manually fixed, before it would compile

The Copilot generated code ran and the results for the top 10 confidence values looked reasonable

static void Main()
{
   string modelPath = "resnet50-v2-7.onnx"; // Updated model path
   string imagePath = "pizza.jpg"; // Updated image path
   string labelsPath = "labels.txt"; // Path to labels file

   using var session = new InferenceSession(modelPath);
   var inputTensor = LoadAndPreprocessImage(imagePath);

   var inputs = new List<NamedOnnxValue>
   {
       NamedOnnxValue.CreateFromTensor("data", inputTensor) // Using "data" as the input tensor name
   };

   using var results = session.Run(inputs);
   var output = results.First().AsTensor<float>().ToArray();

   // Calculate softmax
   var probabilities = Softmax(output);

   // Load labels
   var labels = File.ReadAllLines(labelsPath);

   // Find Top 10 labels and their confidence scores
   var top10 = probabilities
          .Select((prob, index) => new { Label = labels[index], Confidence = prob })
          .OrderByDescending(item => item.Confidence)
          .Take(10);

   Console.WriteLine("Top 10 Predictions:");
   foreach (var item in top10)
   {
      Console.WriteLine($"{item.Label}: {item.Confidence:F4}");
   }
}
...
static float[] Softmax(float[] logits)
{
   // Compute softmax
   float maxVal = logits.Max();
   var expScores = logits.Select(v => (float)Math.Exp(v - maxVal)).ToArray();
   double sumExpScores = expScores.Sum();
   return expScores.Select(score => (float)(score / sumExpScores)).ToArray();
}

The code will have to run on non-windows devices for System.Drawing.Common had to replaced with SixLabors ImageSharp a multi-platform graphics library.

The SixLabors ImageSharp update compiled and ran first time.

using Microsoft.ML.OnnxRuntime;
using Microsoft.ML.OnnxRuntime.Tensors;

using SixLabors.ImageSharp;
using SixLabors.ImageSharp.PixelFormats;
using SixLabors.ImageSharp.Processing;

namespace ResnetV5ObjectClassificationApplication
{
   class Program
   {
      static void Main()
      {
         string modelPath = "resnet50-v2-7.onnx"; // Updated model path
         string imagePath = "pizza.jpg"; // Updated image path
         string labelsPath = "labels.txt"; // Path to labels file

         using var session = new InferenceSession(modelPath);
         var inputTensor = LoadAndPreprocessImage(imagePath);

         var inputs = new List<NamedOnnxValue>
         {
            NamedOnnxValue.CreateFromTensor("data", inputTensor) // Using "data" as the input tensor name
         };

         using var results = session.Run(inputs);
         var output = results.First().AsTensor<float>().ToArray();

         // Calculate softmax
         var probabilities = Softmax(output);

         // Load labels
         var labels = File.ReadAllLines(labelsPath);

         // Find Top 10 labels and their confidence scores
         var top10 = probabilities
             .Select((prob, index) => new { Label = labels[index], Confidence = prob })
             .OrderByDescending(item => item.Confidence)
             .Take(10);

         Console.WriteLine("Top 10 Predictions:");
         foreach (var item in top10)
         {
            Console.WriteLine($"{item.Label}: {item.Confidence}");
         }

         Console.WriteLine("Press ENTER to exit");
         Console.ReadLine();
      }

      static DenseTensor<float> LoadAndPreprocessImage(string imagePath)
      {
         int width = 224, height = 224; // ResNet50 expects 224x224 input

         using var image = Image.Load<Rgb24>(imagePath);
         image.Mutate(x => x.Resize(width, height));

         var tensor = new DenseTensor<float>(new[] { 1, 3, width, height });

         // ImageNet mean & standard deviation values
         float[] mean = { 0.485f, 0.456f, 0.406f };
         float[] stdev = { 0.229f, 0.224f, 0.225f };

         for (int y = 0; y < height; y++)
         {
            for (int x = 0; x < width; x++)
            {
               var pixel = image[x, y];

               // Normalize using mean and standard deviation
               tensor[0, 0, y, x] = (pixel.R / 255f - mean[0]) / stdev[0]; // Red channel
               tensor[0, 1, y, x] = (pixel.G / 255f - mean[1]) / stdev[1]; // Green channel
               tensor[0, 2, y, x] = (pixel.B / 255f - mean[2]) / stdev[2]; // Blue channel
            }
         }

         return tensor;
      }

      static float[] Softmax(float[] logits)
      {
         // Compute softmax  
         float maxVal = logits.Max();
         var expScores = logits.Select(logit => Math.Exp(logit - maxVal)).ToArray(); // Explicitly cast logit to double  
         double sumExpScores = expScores.Sum();
         return expScores.Select(score => (float)(score / sumExpScores)).ToArray();
      }
   }
}

Summary

The Copilot generated code in this post in this was “inspired” by the Image recognition with ResNet50v2 in C# sample application.

The Copilot generated code in this post is not suitable for production

Building Cloud AI with Github Copilot- YoloSharp Azure HTTP Functions

Introduction

For this post I have used Github Copilot prompts to generate Azure HTTP Trigger functions which use Ultralytics YoloV8 and Compunet YoloSharp for object classification, object detection, and pose estimation.

I started with the Visual Studio 2022 Azure functions quick start code which ran first time.

using Microsoft.AspNetCore.Http;
using Microsoft.AspNetCore.Mvc;
using Microsoft.Azure.Functions.Worker;
using Microsoft.Extensions.Logging;

namespace YoloSharpxxxxxHttpTriggerFunction
{
    public class Function1
    {
        private readonly ILogger<Function1> _logger;

        public Function1(ILogger<Function1> logger)
        {
            _logger = logger;
        }

        [Function("Function1")]
        public IActionResult Run([HttpTrigger(AuthorizationLevel.Anonymous, "get", "post")] HttpRequest req)
        {
            _logger.LogInformation("C# HTTP trigger function processed a request.");
            return new OkObjectResult("Welcome to Azure Functions!");
        }
    }
}

The code generated by Github Copilot for the three functions changed the Function attribute to FunctionName and didn’t initialise the ILogger correctly.

[FunctionName("DetectObjects")]
public static async Task<IActionResult> Run(
     [HttpTrigger(AuthorizationLevel.Function, "post", Route = "detect")] HttpRequest req,
     ILogger log)
{
   log.LogInformation("DetectObjects function received a request for object detection.");
   ...
}

Every so often, when uploading more than one image at a time there was a “System.IO.InvalidData Exception: The stream exceeded the data limit 16384” error. In previous examples I had tried the batch processing of multiple images but had hit memory issues, so putting the image to be in the message body rather than Multipart Form Data for future development seems like a reasonable approach.

Image in the body of the request

Object Classification

I started with an object classification function because in previous examples Github Copilot was better at unpacking the results.

please write an httpTrigger azure function that uses YoloSharp and a Yolo image classification onnx model
Image classification not object detection
The image is in the form data
The multipart/form-data check can be removed
The YoloPredictor should be released after use
Many image files could be uploaded in one request
Only one image file can be uploaded in one request

[Function("YoloImageClassificationFunction")]
public async Task<IActionResult> Run([HttpTrigger(AuthorizationLevel.Function, "post", Route = null)] HttpRequest req)
{
   _logger.LogInformation("C# HTTP trigger function processed a request.");

   var form = await req.ReadFormAsync();
   var files = form.Files;

   if (files.Count == 0)
   {
      return new BadRequestObjectResult("No images uploaded.");
   }

   if (files.Count > 1)
   {
      return new BadRequestObjectResult("Only one image file can be uploaded in one request.");
   }

   var file = files[0];
   if (file.Length > 0)
   {
      using (var memoryStream = new MemoryStream())
      {
         await file.CopyToAsync(memoryStream);
         var imageBytes = memoryStream.ToArray();

         using (var yoloModel = new YoloPredictor("yolov8s-cls.onnx"))
         {
            var classifications = yoloModel.Classify(imageBytes);

            return new OkObjectResult(new { file.FileName, classifications });
         }
      }
   }

   return new BadRequestObjectResult("Invalid image file.");
}

Image in the form data of the request

When I compared the results with previous applications, they were the same. I should have added code which discarded any result with less that a specified confidence to reduce the size of the results.

Object Detection

please write an azure function with an http trigger that uses yolosharp and an onnx file to detect objects in an image
The image is not in the body it is in the form
Image.Load is not used
yolo.Detect can process an image file stream
The YoloPredictor should be released after use
Many image files could be uploaded in one request
Only one image per request
Add multipart/form-data content type check
Add check that there are headers — Didn’t get this right
Add check that req.Headers is not null
Add check that request has ContentType and it is multipart/form-data

[Function("ObjectDetectionFunction")]
public async Task<IActionResult> Run( [HttpTrigger(AuthorizationLevel.Function, "post", Route = null)] HttpRequest req)
{
   _logger.LogInformation("C# HTTP trigger function processed a request.");

   // Intellisense fix
   if (req.Headers == null || !req.Headers.TryGetValue("Content-Type", out Microsoft.Extensions.Primitives.StringValues value) || !value.ToString().Contains("multipart/form-data"))
   {
      return new BadRequestObjectResult("The request must have a Content-Type header with multipart/form-data.");
   }

   // Read the images from the form data
   var form = await req.ReadFormAsync();
   var files = form.Files;
   if (files.Count == 0)
   {
      return new BadRequestObjectResult("No image files uploaded.");
   }

   if (files.Count > 1)
   {
      return new BadRequestObjectResult("Only one image file is allowed per request.");
   }

   var file = files[0];
   if (file.Length == 0)
   {
      return new BadRequestObjectResult("The uploaded image file is empty.");
   }

   // Load the YOLOv8 model
   using (var yolo = new YoloPredictor("yolov8s.onnx"))
   {
      // Perform object detection
      using (var stream = file.OpenReadStream())
      {
         var items = yolo.Detect(stream);
         var result = new { FileName = file.FileName, Detections = items };

         // Return the detection results
         return new OkObjectResult(result);
      }
   }
}

After a couple of manual fixes the function ran in the desktop emulator

When I compared the results with previous applications, they were the same.

Pose Estimation

I forgot the ILogger initialisation so had to do it at the end

please write an azure function with an http trigger that uses yolosharp to estimate the pose of humans in an uploaded image.
Yolo v8 pose estimation model and yolosharp library
Make into azure function
The image files are in the form of the request
Modify the code so more than one image per request can be processed
Initialise ILogger in the constructor

//[FunctionName("PoseEstimation")]
[Function("PoseEstimation")]
public async Task<IActionResult> Run([HttpTrigger(AuthorizationLevel.Function, "post", Route = null)] HttpRequest req)
{
   _log.LogInformation("Pose estimation function processed a request.");

   if (!req.HasFormContentType || !req.Form.Files.Any())
   {
      return new BadRequestObjectResult("Please upload image files.");
   }

   var results = new List<object>();

   foreach (var file in req.Form.Files)
   {
      using var memoryStream = new MemoryStream();
      await file.CopyToAsync(memoryStream);
      memoryStream.Position = 0;

      using var image = Image.Load<Rgba32>(memoryStream);

      // Initialize the YOLO model
      //using var predictor = new YoloPredictor("path/to/model.onnx");
      using var predictor = new YoloPredictor("yolov8s-pose.onnx");

      // Perform pose estimation
      var result = await predictor.PoseAsync(image);

      // Format the results
      //var poses = result.Poses.Select(pose => new
      var poses = result.Select(pose => new
      {
         //Keypoints = pose.Keypoints.Select(k => new { k.X, k.Y }),
         Keypoints = pose.Select(k => new { k.Point.X, k.Point.Y }),
         Confidence = pose.Confidence
      });

      results.Add(new
      {
         Image = file.FileName,
         Poses = poses
      });
   }

   return new OkObjectResult(new { results });
}

After a couple of manual fixes including changing the way the results were generated the function ran in the desktop emulator.

Summary

The generated code worked but required manual fixes and was pretty ugly

The Github Copilot generated code in this post is not suitable for production