278 lines
17 KiB
Plaintext
278 lines
17 KiB
Plaintext
{
|
|
"cells": [
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 1,
|
|
"id": "3f754c96",
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"import torch\n",
|
|
"from xautodl import spaces\n",
|
|
"from xautodl.xlayers import super_core\n",
|
|
"\n",
|
|
"def _create_stel(input_dim, output_dim, order):\n",
|
|
" return super_core.SuperSequential(\n",
|
|
" super_core.SuperLinear(input_dim, output_dim),\n",
|
|
" super_core.SuperTransformerEncoderLayer(\n",
|
|
" output_dim,\n",
|
|
" num_heads=spaces.Categorical(2, 4, 6),\n",
|
|
" mlp_hidden_multiplier=spaces.Categorical(1, 2, 4),\n",
|
|
" order=order,\n",
|
|
" ),\n",
|
|
" )"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 2,
|
|
"id": "81d42f4b",
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"batch, seq_dim, input_dim = 1, 4, 6\n",
|
|
"order = super_core.LayerOrder.PreNorm"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 3,
|
|
"id": "8056b37c",
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"SuperSequential(\n",
|
|
" (0): SuperSequential(\n",
|
|
" (0): SuperLinear(in_features=6, out_features=Categorical(candidates=[12, 24, 36], default_index=None), bias=True)\n",
|
|
" (1): SuperTransformerEncoderLayer(\n",
|
|
" (norm1): SuperLayerNorm1D(shape=Categorical(candidates=[12, 24, 36], default_index=None), eps=1e-06, elementwise_affine=True)\n",
|
|
" (mha): SuperSelfAttention(\n",
|
|
" input_dim=Categorical(candidates=[12, 24, 36], default_index=None), proj_dim=Categorical(candidates=[12, 24, 36], default_index=None), num_heads=Categorical(candidates=[2, 4, 6], default_index=None), mask=False, infinity=1000000000.0\n",
|
|
" (q_fc): SuperLinear(in_features=Categorical(candidates=[12, 24, 36], default_index=None), out_features=Categorical(candidates=[12, 24, 36], default_index=None), bias=False)\n",
|
|
" (k_fc): SuperLinear(in_features=Categorical(candidates=[12, 24, 36], default_index=None), out_features=Categorical(candidates=[12, 24, 36], default_index=None), bias=False)\n",
|
|
" (v_fc): SuperLinear(in_features=Categorical(candidates=[12, 24, 36], default_index=None), out_features=Categorical(candidates=[12, 24, 36], default_index=None), bias=False)\n",
|
|
" (attn_drop): SuperDrop(p=0.0, dims=[-1, -1, -1, -1], recover=True)\n",
|
|
" )\n",
|
|
" (drop): Dropout(p=0.0, inplace=False)\n",
|
|
" (norm2): SuperLayerNorm1D(shape=Categorical(candidates=[12, 24, 36], default_index=None), eps=1e-06, elementwise_affine=True)\n",
|
|
" (mlp): SuperMLPv2(\n",
|
|
" in_features=Categorical(candidates=[12, 24, 36], default_index=None), hidden_multiplier=Categorical(candidates=[1, 2, 4], default_index=None), out_features=Categorical(candidates=[12, 24, 36], default_index=None), drop=None, fc1 -> act -> drop -> fc2 -> drop,\n",
|
|
" (_params): ParameterDict(\n",
|
|
" (fc1_super_weight): Parameter containing: [torch.FloatTensor of size 144x36]\n",
|
|
" (fc1_super_bias): Parameter containing: [torch.FloatTensor of size 144]\n",
|
|
" (fc2_super_weight): Parameter containing: [torch.FloatTensor of size 36x144]\n",
|
|
" (fc2_super_bias): Parameter containing: [torch.FloatTensor of size 36]\n",
|
|
" )\n",
|
|
" (act): GELU()\n",
|
|
" (drop): Dropout(p=0.0, inplace=False)\n",
|
|
" )\n",
|
|
" )\n",
|
|
" )\n",
|
|
" (1): SuperSequential(\n",
|
|
" (0): SuperLinear(in_features=Categorical(candidates=[12, 24, 36], default_index=None), out_features=Categorical(candidates=[24, 36, 48], default_index=None), bias=True)\n",
|
|
" (1): SuperTransformerEncoderLayer(\n",
|
|
" (norm1): SuperLayerNorm1D(shape=Categorical(candidates=[24, 36, 48], default_index=None), eps=1e-06, elementwise_affine=True)\n",
|
|
" (mha): SuperSelfAttention(\n",
|
|
" input_dim=Categorical(candidates=[24, 36, 48], default_index=None), proj_dim=Categorical(candidates=[24, 36, 48], default_index=None), num_heads=Categorical(candidates=[2, 4, 6], default_index=None), mask=False, infinity=1000000000.0\n",
|
|
" (q_fc): SuperLinear(in_features=Categorical(candidates=[24, 36, 48], default_index=None), out_features=Categorical(candidates=[24, 36, 48], default_index=None), bias=False)\n",
|
|
" (k_fc): SuperLinear(in_features=Categorical(candidates=[24, 36, 48], default_index=None), out_features=Categorical(candidates=[24, 36, 48], default_index=None), bias=False)\n",
|
|
" (v_fc): SuperLinear(in_features=Categorical(candidates=[24, 36, 48], default_index=None), out_features=Categorical(candidates=[24, 36, 48], default_index=None), bias=False)\n",
|
|
" (attn_drop): SuperDrop(p=0.0, dims=[-1, -1, -1, -1], recover=True)\n",
|
|
" )\n",
|
|
" (drop): Dropout(p=0.0, inplace=False)\n",
|
|
" (norm2): SuperLayerNorm1D(shape=Categorical(candidates=[24, 36, 48], default_index=None), eps=1e-06, elementwise_affine=True)\n",
|
|
" (mlp): SuperMLPv2(\n",
|
|
" in_features=Categorical(candidates=[24, 36, 48], default_index=None), hidden_multiplier=Categorical(candidates=[1, 2, 4], default_index=None), out_features=Categorical(candidates=[24, 36, 48], default_index=None), drop=None, fc1 -> act -> drop -> fc2 -> drop,\n",
|
|
" (_params): ParameterDict(\n",
|
|
" (fc1_super_weight): Parameter containing: [torch.FloatTensor of size 192x48]\n",
|
|
" (fc1_super_bias): Parameter containing: [torch.FloatTensor of size 192]\n",
|
|
" (fc2_super_weight): Parameter containing: [torch.FloatTensor of size 48x192]\n",
|
|
" (fc2_super_bias): Parameter containing: [torch.FloatTensor of size 48]\n",
|
|
" )\n",
|
|
" (act): GELU()\n",
|
|
" (drop): Dropout(p=0.0, inplace=False)\n",
|
|
" )\n",
|
|
" )\n",
|
|
" )\n",
|
|
" (2): SuperSequential(\n",
|
|
" (0): SuperLinear(in_features=Categorical(candidates=[24, 36, 48], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=True)\n",
|
|
" (1): SuperTransformerEncoderLayer(\n",
|
|
" (norm1): SuperLayerNorm1D(shape=Categorical(candidates=[36, 72, 100], default_index=None), eps=1e-06, elementwise_affine=True)\n",
|
|
" (mha): SuperSelfAttention(\n",
|
|
" input_dim=Categorical(candidates=[36, 72, 100], default_index=None), proj_dim=Categorical(candidates=[36, 72, 100], default_index=None), num_heads=Categorical(candidates=[2, 4, 6], default_index=None), mask=False, infinity=1000000000.0\n",
|
|
" (q_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
|
|
" (k_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
|
|
" (v_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
|
|
" (attn_drop): SuperDrop(p=0.0, dims=[-1, -1, -1, -1], recover=True)\n",
|
|
" )\n",
|
|
" (drop): Dropout(p=0.0, inplace=False)\n",
|
|
" (norm2): SuperLayerNorm1D(shape=Categorical(candidates=[36, 72, 100], default_index=None), eps=1e-06, elementwise_affine=True)\n",
|
|
" (mlp): SuperMLPv2(\n",
|
|
" in_features=Categorical(candidates=[36, 72, 100], default_index=None), hidden_multiplier=Categorical(candidates=[1, 2, 4], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), drop=None, fc1 -> act -> drop -> fc2 -> drop,\n",
|
|
" (_params): ParameterDict(\n",
|
|
" (fc1_super_weight): Parameter containing: [torch.FloatTensor of size 400x100]\n",
|
|
" (fc1_super_bias): Parameter containing: [torch.FloatTensor of size 400]\n",
|
|
" (fc2_super_weight): Parameter containing: [torch.FloatTensor of size 100x400]\n",
|
|
" (fc2_super_bias): Parameter containing: [torch.FloatTensor of size 100]\n",
|
|
" )\n",
|
|
" (act): GELU()\n",
|
|
" (drop): Dropout(p=0.0, inplace=False)\n",
|
|
" )\n",
|
|
" )\n",
|
|
" )\n",
|
|
")\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"out1_dim = spaces.Categorical(12, 24, 36)\n",
|
|
"out2_dim = spaces.Categorical(24, 36, 48)\n",
|
|
"out3_dim = spaces.Categorical(36, 72, 100)\n",
|
|
"layer1 = _create_stel(input_dim, out1_dim, order)\n",
|
|
"layer2 = _create_stel(out1_dim, out2_dim, order)\n",
|
|
"layer3 = _create_stel(out2_dim, out3_dim, order)\n",
|
|
"model = super_core.SuperSequential(layer1, layer2, layer3)\n",
|
|
"print(model)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"id": "4fd53a7c",
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"> \u001b[0;32m/Users/xuanyidong/anaconda3/lib/python3.8/site-packages/xautodl-0.9.9-py3.8.egg/xautodl/xlayers/super_transformer.py\u001b[0m(116)\u001b[0;36mforward_raw\u001b[0;34m()\u001b[0m\n",
|
|
"\u001b[0;32m 114 \u001b[0;31m \u001b[0;32mimport\u001b[0m \u001b[0mpdb\u001b[0m\u001b[0;34m;\u001b[0m \u001b[0mpdb\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mset_trace\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
|
|
"\u001b[0m\u001b[0;32m 115 \u001b[0;31m \u001b[0;31m# feed-forward layer -- MLP\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
|
|
"\u001b[0m\u001b[0;32m--> 116 \u001b[0;31m \u001b[0my\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mnorm2\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mx\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
|
|
"\u001b[0m\u001b[0;32m 117 \u001b[0;31m \u001b[0mouts\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mx\u001b[0m \u001b[0;34m+\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mmlp\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0my\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
|
|
"\u001b[0m\u001b[0;32m 118 \u001b[0;31m \u001b[0;32melif\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_order\u001b[0m \u001b[0;32mis\u001b[0m \u001b[0mLayerOrder\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mPostNorm\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
|
|
"\u001b[0m\n",
|
|
"ipdb> print(self)\n",
|
|
"SuperTransformerEncoderLayer(\n",
|
|
" (norm1): SuperLayerNorm1D(shape=Categorical(candidates=[36, 72, 100], default_index=None), eps=1e-06, elementwise_affine=True)\n",
|
|
" (mha): SuperSelfAttention(\n",
|
|
" input_dim=Categorical(candidates=[36, 72, 100], default_index=None), proj_dim=Categorical(candidates=[36, 72, 100], default_index=None), num_heads=Categorical(candidates=[2, 4, 6], default_index=None), mask=False, infinity=1000000000.0\n",
|
|
" (q_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
|
|
" (k_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
|
|
" (v_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
|
|
" (attn_drop): SuperDrop(p=0.0, dims=[-1, -1, -1, -1], recover=True)\n",
|
|
" )\n",
|
|
" (drop): Dropout(p=0.0, inplace=False)\n",
|
|
" (norm2): SuperLayerNorm1D(shape=Categorical(candidates=[36, 72, 100], default_index=None), eps=1e-06, elementwise_affine=True)\n",
|
|
" (mlp): SuperMLPv2(\n",
|
|
" in_features=Categorical(candidates=[36, 72, 100], default_index=None), hidden_multiplier=Categorical(candidates=[1, 2, 4], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), drop=None, fc1 -> act -> drop -> fc2 -> drop,\n",
|
|
" (_params): ParameterDict(\n",
|
|
" (fc1_super_weight): Parameter containing: [torch.FloatTensor of size 400x100]\n",
|
|
" (fc1_super_bias): Parameter containing: [torch.FloatTensor of size 400]\n",
|
|
" (fc2_super_weight): Parameter containing: [torch.FloatTensor of size 100x400]\n",
|
|
" (fc2_super_bias): Parameter containing: [torch.FloatTensor of size 100]\n",
|
|
" )\n",
|
|
" (act): GELU()\n",
|
|
" (drop): Dropout(p=0.0, inplace=False)\n",
|
|
" )\n",
|
|
")\n",
|
|
"ipdb> print(inputs.shape)\n",
|
|
"torch.Size([1, 4, 100])\n",
|
|
"ipdb> print(x.shape)\n",
|
|
"torch.Size([1, 4, 96])\n",
|
|
"ipdb> print(self.mha)\n",
|
|
"SuperSelfAttention(\n",
|
|
" input_dim=Categorical(candidates=[36, 72, 100], default_index=None), proj_dim=Categorical(candidates=[36, 72, 100], default_index=None), num_heads=Categorical(candidates=[2, 4, 6], default_index=None), mask=False, infinity=1000000000.0\n",
|
|
" (q_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
|
|
" (k_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
|
|
" (v_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
|
|
" (attn_drop): SuperDrop(p=0.0, dims=[-1, -1, -1, -1], recover=True)\n",
|
|
")\n",
|
|
"ipdb> print(self.mha.candidate)\n",
|
|
"*** AttributeError: 'SuperSelfAttention' object has no attribute 'candidate'\n",
|
|
"ipdb> print(self.mha.abstract_candidate)\n",
|
|
"*** AttributeError: 'SuperSelfAttention' object has no attribute 'abstract_candidate'\n",
|
|
"ipdb> print(self.mha._abstract_child)\n",
|
|
"None\n",
|
|
"ipdb> print(self.abstract_child)\n",
|
|
"None\n",
|
|
"ipdb> print(self.abstract_child.abstract_child)\n",
|
|
"*** AttributeError: 'NoneType' object has no attribute 'abstract_child'\n",
|
|
"ipdb> print(self.mha)\n",
|
|
"SuperSelfAttention(\n",
|
|
" input_dim=Categorical(candidates=[36, 72, 100], default_index=None), proj_dim=Categorical(candidates=[36, 72, 100], default_index=None), num_heads=Categorical(candidates=[2, 4, 6], default_index=None), mask=False, infinity=1000000000.0\n",
|
|
" (q_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
|
|
" (k_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
|
|
" (v_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
|
|
" (attn_drop): SuperDrop(p=0.0, dims=[-1, -1, -1, -1], recover=True)\n",
|
|
")\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"inputs = torch.rand(batch, seq_dim, input_dim)\n",
|
|
"outputs = model(inputs)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"id": "05332b98",
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"abstract_space = model.abstract_search_space\n",
|
|
"abstract_space.clean_last()\n",
|
|
"abstract_child = abstract_space.random(reuse_last=True)\n",
|
|
"# print(\"The abstract child program is:\\n{:}\".format(abstract_child))\n",
|
|
"model.enable_candidate()\n",
|
|
"model.set_super_run_type(super_core.SuperRunMode.Candidate)\n",
|
|
"model.apply_candidate(abstract_child)\n",
|
|
"outputs = model(inputs)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"id": "3289f938",
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"print(outputs.shape)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"id": "36951cdf",
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": []
|
|
}
|
|
],
|
|
"metadata": {
|
|
"kernelspec": {
|
|
"display_name": "Python 3",
|
|
"language": "python",
|
|
"name": "python3"
|
|
},
|
|
"language_info": {
|
|
"codemirror_mode": {
|
|
"name": "ipython",
|
|
"version": 3
|
|
},
|
|
"file_extension": ".py",
|
|
"mimetype": "text/x-python",
|
|
"name": "python",
|
|
"nbconvert_exporter": "python",
|
|
"pygments_lexer": "ipython3",
|
|
"version": "3.8.8"
|
|
}
|
|
},
|
|
"nbformat": 4,
|
|
"nbformat_minor": 5
|
|
}
|