diff options
Diffstat (limited to 'notebooks/04-convnext.ipynb')
-rw-r--r-- | notebooks/04-convnext.ipynb | 156 |
1 files changed, 47 insertions, 109 deletions
diff --git a/notebooks/04-convnext.ipynb b/notebooks/04-convnext.ipynb index 5ab71c8..5affe85 100644 --- a/notebooks/04-convnext.ipynb +++ b/notebooks/04-convnext.ipynb @@ -2,19 +2,12 @@ "cells": [ { "cell_type": "code", - "execution_count": 14, + "execution_count": 1, "id": "7c02ae76-b540-4b16-9492-e9210b3b9249", - "metadata": {}, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "The autoreload extension is already loaded. To reload it, use:\n", - " %reload_ext autoreload\n" - ] - } - ], + "metadata": { + "tags": [] + }, + "outputs": [], "source": [ "import os\n", "os.environ['CUDA_VISIBLE_DEVICE'] = ''\n", @@ -37,9 +30,11 @@ }, { "cell_type": "code", - "execution_count": 15, + "execution_count": 2, "id": "ccdb6dde-47e5-429a-88f2-0764fb7e259a", - "metadata": {}, + "metadata": { + "tags": [] + }, "outputs": [], "source": [ "from hydra import compose, initialize\n", @@ -49,9 +44,11 @@ }, { "cell_type": "code", - "execution_count": 16, + "execution_count": 3, "id": "3cf50475-39f2-4642-a7d1-5bcbc0a036f7", - "metadata": {}, + "metadata": { + "tags": [] + }, "outputs": [], "source": [ "path = \"../training/conf/network/convnext.yaml\"" @@ -59,9 +56,11 @@ }, { "cell_type": "code", - "execution_count": 37, + "execution_count": 64, "id": "e52ecb01-c975-4e55-925d-1182c7aea473", - "metadata": {}, + "metadata": { + "tags": [] + }, "outputs": [], "source": [ "with open(path, \"rb\") as f:\n", @@ -70,17 +69,19 @@ }, { "cell_type": "code", - "execution_count": 38, + "execution_count": 65, "id": "f939aa37-7b1d-45cc-885c-323c4540bda1", - "metadata": {}, + "metadata": { + "tags": [] + }, "outputs": [ { "data": { "text/plain": [ - "{'_target_': 'text_recognizer.network.convnext.ConvNext', 'dim': 16, 'dim_mults': [2, 4, 8], 'depths': [3, 3, 6], 'downsampling_factors': [[2, 2], [2, 2], [2, 2]], 'attn': {'_target_': 'text_recognizer.network.convnext.TransformerBlock', 'attn': {'_target_': 'text_recognizer.network.convnext.Attention', 'dim': 128, 'heads': 4, 'dim_head': 64, 'scale': 8}, 'ff': {'_target_': 'text_recognizer.network.convnext.FeedForward', 'dim': 128, 'mult': 4}}}" + "{'_target_': 'text_recognizer.network.convnext.convnext.ConvNext', 'dim': 8, 'dim_mults': [2, 8], 'depths': [2, 2], 'attn': {'_target_': 'text_recognizer.network.convnext.transformer.Transformer', 'attn': {'_target_': 'text_recognizer.network.convnext.transformer.Attention', 'dim': 64, 'heads': 4, 'dim_head': 64, 'scale': 8}, 'ff': {'_target_': 'text_recognizer.network.convnext.transformer.FeedForward', 'dim': 64, 'mult': 4}}}" ] }, - "execution_count": 38, + "execution_count": 65, "metadata": {}, "output_type": "execute_result" } @@ -91,21 +92,11 @@ }, { "cell_type": "code", - "execution_count": 19, - "id": "a2b420c1", - "metadata": {}, - "outputs": [], - "source": [ - "cfg.dim_mults = [2, 4, 8, 8]\n", - "cfg.depths = [3, 3, 6, 6]\n", - "cfg.downsampling_factors = [[2, 2], [2, 2], [2, 2], [2, 1]]" - ] - }, - { - "cell_type": "code", - "execution_count": 39, + "execution_count": 66, "id": "c9589350", - "metadata": {}, + "metadata": { + "tags": [] + }, "outputs": [], "source": [ "net = instantiate(cfg)" @@ -113,9 +104,11 @@ }, { "cell_type": "code", - "execution_count": 40, + "execution_count": 67, "id": "618b997c-e6a6-4487-b70c-9d260cb556d3", - "metadata": {}, + "metadata": { + "tags": [] + }, "outputs": [], "source": [ "from torchinfo import summary" @@ -123,11 +116,9 @@ }, { "cell_type": "code", - "execution_count": 41, + "execution_count": 68, "id": "25759b7b-8deb-4163-b75d-a1357c9fe88f", - "metadata": { - "scrolled": false - }, + "metadata": {}, "outputs": [ { "data": { @@ -135,84 +126,31 @@ "====================================================================================================\n", "Layer (type:depth-idx) Output Shape Param #\n", "====================================================================================================\n", - "ConvNext [2, 128, 72, 80] 165,408\n", - "├─TransformerBlock: 1-5 [2, 128, 72, 80] (recursive)\n", - "│ └─Attention: 2-6 [2, 128, 72, 80] (recursive)\n", - "│ │ └─LayerNorm: 3-13 [2, 128, 72, 80] (recursive)\n", - "├─Conv2d: 1-3 [2, 16, 576, 640] (recursive)\n", - "├─TransformerBlock: 1 -- --\n", - "│ └─Attention: 2 -- --\n", - "│ │ └─Conv2d: 3-15 [2, 128, 72, 80] (recursive)\n", - "│ └─FeedForward: 2-7 [2, 128, 72, 80] (recursive)\n", - "│ │ └─Residual: 3-16 [2, 128, 72, 80] (recursive)\n", - "│ │ │ └─Sequential: 4-26 [2, 128, 72, 80] (recursive)\n", - "├─Conv2d: 1-3 [2, 16, 576, 640] (recursive)\n", - "├─ModuleList: 1-4 -- --\n", - "│ └─ModuleList: 2-3 -- --\n", - "│ │ └─ConvNextBlock: 3-4 [2, 16, 576, 640] --\n", - "│ │ │ └─Conv2d: 4-2 [2, 16, 576, 640] 800\n", - "│ │ │ └─Sequential: 4-3 [2, 16, 576, 640] 9,280\n", - "│ │ │ └─Identity: 4-4 [2, 16, 576, 640] --\n", - "│ │ └─ModuleList: 3-5 -- --\n", - "│ │ │ └─ConvNextBlock: 4-5 [2, 16, 576, 640] 10,080\n", - "│ │ │ └─ConvNextBlock: 4-6 [2, 16, 576, 640] 10,080\n", - "│ │ │ └─ConvNextBlock: 4-7 [2, 16, 576, 640] 10,080\n", - "│ │ └─Downsample: 3-6 [2, 32, 288, 320] --\n", - "│ │ │ └─Sequential: 4-8 [2, 32, 288, 320] 2,080\n", - "│ └─ModuleList: 2-4 -- --\n", - "│ │ └─ConvNextBlock: 3-7 [2, 32, 288, 320] --\n", - "│ │ │ └─Conv2d: 4-9 [2, 32, 288, 320] 1,600\n", - "│ │ │ └─Sequential: 4-10 [2, 32, 288, 320] 36,992\n", - "│ │ │ └─Identity: 4-11 [2, 32, 288, 320] --\n", - "│ │ └─ModuleList: 3-8 -- --\n", - "│ │ │ └─ConvNextBlock: 4-12 [2, 32, 288, 320] 38,592\n", - "│ │ │ └─ConvNextBlock: 4-13 [2, 32, 288, 320] 38,592\n", - "│ │ │ └─ConvNextBlock: 4-14 [2, 32, 288, 320] 38,592\n", - "│ │ └─Downsample: 3-9 [2, 64, 144, 160] --\n", - "│ │ │ └─Sequential: 4-15 [2, 64, 144, 160] 8,256\n", - "│ └─ModuleList: 2-5 -- --\n", - "│ │ └─ConvNextBlock: 3-10 [2, 64, 144, 160] --\n", - "│ │ │ └─Conv2d: 4-16 [2, 64, 144, 160] 3,200\n", - "│ │ │ └─Sequential: 4-17 [2, 64, 144, 160] 147,712\n", - "│ │ │ └─Identity: 4-18 [2, 64, 144, 160] --\n", - "│ │ └─ModuleList: 3-11 -- --\n", - "│ │ │ └─ConvNextBlock: 4-19 [2, 64, 144, 160] 150,912\n", - "│ │ │ └─ConvNextBlock: 4-20 [2, 64, 144, 160] 150,912\n", - "│ │ │ └─ConvNextBlock: 4-21 [2, 64, 144, 160] 150,912\n", - "│ │ │ └─ConvNextBlock: 4-22 [2, 64, 144, 160] 150,912\n", - "│ │ │ └─ConvNextBlock: 4-23 [2, 64, 144, 160] 150,912\n", - "│ │ │ └─ConvNextBlock: 4-24 [2, 64, 144, 160] 150,912\n", - "│ │ └─Downsample: 3-12 [2, 128, 72, 80] --\n", - "│ │ │ └─Sequential: 4-25 [2, 128, 72, 80] 32,896\n", - "├─TransformerBlock: 1-5 [2, 128, 72, 80] (recursive)\n", - "│ └─Attention: 2-6 [2, 128, 72, 80] (recursive)\n", - "│ │ └─LayerNorm: 3-13 [2, 128, 72, 80] (recursive)\n", - "│ │ └─Conv2d: 3-14 [2, 768, 72, 80] 98,304\n", - "│ │ └─Conv2d: 3-15 [2, 128, 72, 80] (recursive)\n", - "│ └─FeedForward: 2-7 [2, 128, 72, 80] (recursive)\n", - "│ │ └─Residual: 3-16 [2, 128, 72, 80] (recursive)\n", - "│ │ │ └─Sequential: 4-26 [2, 128, 72, 80] (recursive)\n", - "├─LayerNorm: 1-6 [2, 128, 72, 80] 128\n", + "ConvNext [2, 64, 14, 256] --\n", + "├─Conv2d: 1-1 [2, 8, 56, 1024] 400\n", + "├─ModuleList: 1-2 -- 30,512\n", + "├─Transformer: 1-3 [2, 64, 14, 256] 98,688\n", + "├─LayerNorm: 1-4 [2, 64, 14, 256] 64\n", "====================================================================================================\n", - "Total params: 1,558,144\n", - "Trainable params: 1,558,144\n", + "Total params: 129,664\n", + "Trainable params: 129,664\n", "Non-trainable params: 0\n", - "Total mult-adds (G): 114.00\n", + "Total mult-adds (G): 2.00\n", "====================================================================================================\n", - "Input size (MB): 2.95\n", - "Forward/backward pass size (MB): 3822.06\n", - "Params size (MB): 5.57\n", - "Estimated Total Size (MB): 3830.58\n", + "Input size (MB): 0.46\n", + "Forward/backward pass size (MB): 260.57\n", + "Params size (MB): 0.52\n", + "Estimated Total Size (MB): 261.55\n", "====================================================================================================" ] }, - "execution_count": 41, + "execution_count": 68, "metadata": {}, "output_type": "execute_result" } ], "source": [ - "summary(net, (2, 1, 576, 640), device=\"cpu\", depth=4)" + "summary(net, (2, 1, 56, 1024), device=\"cpu\", depth=1)" ] }, { |