We extend Adam Optimizer but use FP32 to store gradients and moments.
\n": "Adam Optimizer\u3092\u62e1\u5f35\u3057\u307e\u3057\u305f\u304c\u3001\u30b0\u30e9\u30c7\u30fc\u30b7\u30e7\u30f3\u3068\u30e2\u30fc\u30e1\u30f3\u30c8\u306e\u4fdd\u5b58\u306b\u306fFP32\u3092\u4f7f\u7528\u3057\u3066\u3044\u307e\u3059\u3002
\n", "We extend PyTorch gradient scaler to use FP32 gradients.
\n": "PyTorch \u30b0\u30e9\u30c7\u30fc\u30b7\u30e7\u30f3\u30b9\u30b1\u30fc\u30e9\u30fc\u3092 FP32 \u30b0\u30e9\u30c7\u30fc\u30b7\u30e7\u30f3\u3092\u4f7f\u7528\u3059\u308b\u3088\u3046\u306b\u62e1\u5f35\u3057\u307e\u3059\u3002
\n", "All the state tensors use FP32.
\n": "\u3059\u3079\u3066\u306e\u30b9\u30c6\u30fc\u30c8\u30c6\u30f3\u30bd\u30eb\u306f FP32 \u3092\u4f7f\u7528\u3057\u307e\u3059\u3002
\n", "\n": "\n", "
Calculate weight decay
\n": "\u4f53\u91cd\u6e1b\u5c11\u306e\u8a08\u7b97
\n", "Call the Adam Optimizer initializer
\n": "\n", "Exponential moving average of gradients, _^_0_^_
\n": "\u52fe\u914d\u306e\u6307\u6570\u79fb\u52d5\u5e73\u5747\u3001_^_0_^_
\n", "Exponential moving average of squared gradient values, _^_0_^_
\n": "\u4e8c\u4e57\u52fe\u914d\u5024\u306e\u6307\u6570\u79fb\u52d5\u5e73\u5747\u3001_^_0_^_
\n", "Get _^_0_^_ and _^_1_^_
\n": "_^_0_^_\u53d6\u5f97\u3057\u3066 _^_1_^_
\n", "Get the FP32 gradients if available
\n": "\u53ef\u80fd\u306a\u5834\u5408\u306f FP32 \u306e\u30b0\u30e9\u30c7\u30fc\u30b7\u30e7\u30f3\u3092\u53d6\u5f97
\n", "Get the FP32 parameters
\n": "FP32 \u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u53d6\u5f97
\n", "If we are using the _^_0_^_ optimizer set _^_1_^_ to the FP32 gradients
\n": "FP32 _^_0_^_ _^_1_^_ \u30b0\u30e9\u30c7\u30fc\u30b7\u30e7\u30f3\u306b\u8a2d\u5b9a\u3055\u308c\u305f\u30aa\u30d7\u30c6\u30a3\u30de\u30a4\u30b6\u30fc\u3092\u4f7f\u7528\u3057\u3066\u3044\u308b\u5834\u5408
\n", "Increment _^_0_^_ the number of optimizer steps
\n": "_^_0_^_\u30aa\u30d7\u30c6\u30a3\u30de\u30a4\u30b6\u30fc\u306e\u30b9\u30c6\u30c3\u30d7\u6570\u3092\u5897\u3084\u3059
\n", "Loop through parameters
\n": "\u30eb\u30fc\u30d7\u30b9\u30eb\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf
\n", "Maintain a FP32 copy of the parameters
\n": "\u30d1\u30e9\u30e1\u30fc\u30bf\u306e FP32 \u30b3\u30d4\u30fc\u3092\u7ba1\u7406
\n", "Not implemented for sparse tensors
\n": "\u30b9\u30d1\u30b9\u30c6\u30f3\u30bd\u30eb\u306b\u306f\u5b9f\u88c5\u3055\u308c\u3066\u3044\u307e\u305b\u3093
\n", "Otherwise, convert the gradients to FP32
\n": "\u305d\u308c\u4ee5\u5916\u306e\u5834\u5408\u306f\u3001\u30b0\u30e9\u30c7\u30fc\u30b7\u30e7\u30f3\u3092 FP32 \u306b\u5909\u63db\u3057\u307e\u3059\u3002
\n", "Otherwise, do not convert the gradients to FP32
\n": "\u305d\u308c\u4ee5\u5916\u306e\u5834\u5408\u306f\u3001\u30b0\u30e9\u30c7\u30fc\u30b7\u30e7\u30f3\u3092 FP32 \u306b\u5909\u63db\u3057\u306a\u3044\u3067\u304f\u3060\u3055\u3044\u3002
\n", "Parameter to store 32 bit gradients. This get populated by the _^_0_^_ defined below.
\n": "32 \u30d3\u30c3\u30c8\u306e\u30b0\u30e9\u30c7\u30fc\u30b7\u30e7\u30f3\u3092\u683c\u7d0d\u3059\u308b\u30d1\u30e9\u30e1\u30fc\u30bf\u30fc\u3002_^_0_^_\u3053\u308c\u306b\u306f\u4ee5\u4e0b\u306e\u5b9a\u7fa9\u304c\u5165\u529b\u3055\u308c\u307e\u3059
\u3002\n", "Perform Adam update
\n": "Adam \u30a2\u30c3\u30d7\u30c7\u30fc\u30c8\u3092\u5b9f\u884c
\n", "Set the parameters
\n": "\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u8a2d\u5b9a
\n", "Skip non-trainable parameters
\n": "\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u4e0d\u53ef\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u30b9\u30ad\u30c3\u30d7
\n", "This is the number of optimizer steps taken on the parameter, _^_0_^_
\n": "\u3053\u308c\u306f\u3001\u30d1\u30e9\u30e1\u30fc\u30bf\u30fc\u306b\u5bfe\u3057\u3066\u5b9f\u884c\u3055\u308c\u305f\u30aa\u30d7\u30c6\u30a3\u30de\u30a4\u30b6\u30fc\u30b9\u30c6\u30c3\u30d7\u306e\u6570\u3067\u3059\u3002_^_0_^_
\n", "Unscale all the gradients
\n": "\u3059\u3079\u3066\u306e\u30b0\u30e9\u30c7\u30fc\u30b7\u30e7\u30f3\u3092\u30b9\u30b1\u30fc\u30eb\u89e3\u9664
\n", "A simple PyTorch implementation/tutorial of Adam optimizer": "Adam \u30aa\u30d7\u30c6\u30a3\u30de\u30a4\u30b6\u30fc\u306e\u7c21\u5358\u306a PyTorch \u5b9f\u88c5/\u30c1\u30e5\u30fc\u30c8\u30ea\u30a2\u30eb", "Adam Optimizer for Half Precision Training": "\u534a\u7cbe\u5ea6\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u7528\u306e Adam \u30aa\u30d7\u30c6\u30a3\u30de\u30a4\u30b6\u30fc" }