Audit PyTorch workshop visual candidates

Author: alexeygrigorev

cohorts/2026/08-deep-learning/pytorch/README.md

@@ -3,6 +3,7 @@
 
 * [Video](https://www.youtube.com/watch?v=Ne25VujHRLA)
 * [Notebook](https://colab.research.google.com/drive/1nCA4Q0f8DVFiLpfXdXvZtYUh-yYDy5i_?usp=sharing)
+* [Visual audit](visual-audit-2026-09-09.md)
 
 
 This workshop introduces deep learning for image classification using PyTorch. It is based on the [ML Zoomcamp Deep Learning module (08-deep-learning)](..) but adapted to use PyTorch instead of TensorFlow/Keras.

cohorts/2026/08-deep-learning/pytorch/visual-audit-2026-09-09.md

@@ -0,0 +1,51 @@
+# PyTorch workshop visual audit
+
+Date: 2026-09-09
+Scope: [PyTorch workshop README](README.md) and source video `Ne25VujHRLA` only
+
+## Result
+
+No image is published. The README had no image references before this audit and
+still has none. Six transcript-backed moments were worth checking, but no frame
+could be acquired for visual review, so none can pass the rubric's readability,
+fidelity, or caption hard gates. No imagegen regeneration was run without an
+original frame and native crop.
+
+## Source and temporary provenance
+
+- Source: `https://www.youtube.com/watch?v=Ne25VujHRLA`
+- Transcript: `/home/alexey/.cache/youtube_transcripts/Ne25VujHRLA.txt`
+- Temporary source/candidate workspace: `/home/alexey/git/.tmp/mlzoomcamp-pytorch-audit/`
+- Local source video: not acquired; no source video or candidate frame was copied into the repository.
+- Route A (`yt-dlp` through the documented sticky Oxylabs session): unavailable with the documented `407 Proxy Authentication Required` quota response.
+- Route B (the documented Piped/Invidious mirror rotation): no usable stream metadata; Piped returned bot-block/error responses and the available Invidious endpoints were disabled, blocked, or challenge pages. The browser embed also returned `LOGIN_REQUIRED`.
+
+The strict `0/12` scores below mean “not reviewable and therefore rejected,”
+not that the transcript-backed teaching point is inherently worthless. Crop
+coordinates are `—` because no source frame exists to crop.
+
+## Candidate ledger
+
+| Timestamp | Workshop section | Teaching point | Score | Crop coordinates | Disposition |
+| --- | --- | --- | ---: | --- | --- |
+| 18:30 | 3. Pre-trained Models | A pretrained ImageNet model produces ranked class predictions for a clothing image before task-specific training. | 0/12 | — | Reject: no frame to verify the output or readability; exact output would require native source rendering. |
+| 31:16 | 5. Transfer Learning | Freeze the pretrained feature extractor, pool its 1,280 features, and replace the original head with 10 clothing outputs. | 0/12 | — | Reject: no frame to inspect; if recovered, this conceptual relationship is eligible for crop plus imagegen regeneration. |
+| 41:18 | 5. Transfer Learning | The manual PyTorch loop reports train/validation loss and accuracy across epochs. | 0/12 | — | Reject: no frame to verify the measured result; exact logs are source data, not imagegen material. |
+| 45:42 | 6. Tuning the Learning Rate | The reading-speed analogy explains why an overly high or low learning rate can hurt training. | 0/12 | — | Reject: no frame to inspect; if the drawn analogy is present, it is eligible for crop plus imagegen regeneration. |
+| 67:45 | 10. Data Augmentation | Rotation, cropping, flipping, zooming, and related variants create new training views from one image. | 0/12 | — | Reject: no frame to inspect; if the transformation visual is present, it is eligible for crop plus imagegen regeneration. |
+| 79:24 | 11. Using the Trained Model | The final model maps a preprocessed image to clothing-class scores, with pants as the winning class. | 0/12 | — | Reject: no frame to verify the exact scores; exact values must remain native/deterministic. |
+
+These are the minimum plausible candidates from the transcript. Setup screens,
+browser/Colab navigation, repeated code cells, checkpoint filenames, and the
+ONNX export cell were excluded because they are already expressible in the
+README or fail the rubric's complementarity hard gate.
+
+## Reopening this audit
+
+When a local source becomes available, extract only the six timestamps above
+plus nearby `±3 s` candidates under `.tmp`, inspect every frame at lesson size,
+record real source dimensions and crop coordinates here, and publish only a
+candidate that scores at least 7/12 without a hard-gate violation. Preserve
+exact code, UI, plots, numbers, and outputs deterministically; use imagegen
+only for a bounded conceptual diagram after retaining both the original frame
+and native crop.