TensorFlow: Hotdog or Not with Toppings #bbuzz
summary
Workshop builds a Hotdog-or-Not pipeline that actually trains a CNN from JPEGs, converts them to TFRecords in HDFS for fast GPU feeding, and deploys the exported protobuf model with TensorFlow Serving via gRPC behind TLS. Data prep runs on Spark, distributed training uses GPUs on YARN with ring AllReduce for throughput (examples: ~1,100–1,800 images/sec on multi‑GPU nodes). Shows hyperparameter evolutionary search and grid search, runs a Python gRPC client (MNIST test hit ~93%), while the small Hotdog dataset yields only ~50–60% accuracy.
AI-written summary · about
chapters (10)
- 0:00Workshop introduction
- 1:00Platform overview Hops
- 11:00Setting up accounts
- 26:37Creating projects
- 35:57Data visualization with Facets
- 47:05Jupyter notebooks in Hops
- 48:08Model serving demo
- 52:37Distributed training
- 1:00:06Hotdog classifier demo
- 1:22:32Q&A
related talks
35m''Hot Dogs or Not" - At Scale with Kubernetes [I]
34mDeep Learning at Scale
42mHigh Performance Distributed Tensorflow
43mDeep learning beyond the learning
42mDeep learning beyond the learning
49mDistributed Deep Learning
36mHigh-Performance Distributed Tensorflow
44mTensorflow, Tensorflow, What is it that you see?