Remove cells for unnecessary figures

yukw777 · Mar 4, 2024 · dfb7004 · dfb7004
1 parent 02fef29
commit dfb7004
Showing 1 changed file with 0 additions and 91 deletions.
diff --git a/figures/icl_eval_figures.ipynb b/figures/icl_eval_figures.ipynb
@@ -179,97 +179,6 @@
     "draw_graphs(models, metrics, \"held-out-eval.pdf\", shots, 3)"
    ]
   },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Held-Out Evaluation (Ego4D) With VideoMAE Baselines"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "models = {\n",
-    "    \"EILEV BLIP-2 OPT-2.7B\": {\n",
-    "        \"STS-CE\": [0.2098, 0.4754, 0.4897, 0.5569, 0.612, 0.6312, 0.6363],\n",
-    "        \"STS-BE\": [0.3278, 0.5495, 0.571, 0.6284, 0.6735, 0.6898, 0.6936],\n",
-    "        \"BERTScore-F1\": [0.5234, 0.6305, 0.6399, 0.6463, 0.6543, 0.6539, 0.6529],\n",
-    "        \"ROUGE-L\": [0.2315, 0.5013, 0.5396, 0.5785, 0.6102, 0.6249, 0.6296],\n",
-    "        \"BLEU\": [0.008795, 0.1376, 0.2015, 0.2443, 0.2741, 0.2968, 0.3049],\n",
-    "        \"meta\": {\"shots\": shots, \"linestyle\": \"-\"},\n",
-    "    },\n",
-    "    \"EILEV BLIP-2 Flan-T5-xl\": {\n",
-    "        \"STS-CE\": [0.3552, 0.5039, 0.5176, 0.5539, 0.6089, 0.6276, 0.6349],\n",
-    "        \"STS-BE\": [0.426, 0.5697, 0.5812, 0.613, 0.6689, 0.6886, 0.6948],\n",
-    "        \"BERTScore-F1\": [-1.84, 0.6291, 0.6394, 0.6477, 0.6527, 0.6561, 0.6572],\n",
-    "        \"ROUGE-L\": [0.3129, 0.5032, 0.5322, 0.5648, 0.607, 0.6203, 0.623],\n",
-    "        \"BLEU\": [0.06718, 0.1507, 0.1992, 0.2373, 0.2834, 0.2931, 0.2913],\n",
-    "        \"meta\": {\"shots\": shots, \"linestyle\": \"-\"},\n",
-    "    },\n",
-    "    \"BLIP-2 OPT-2.7B\": {\n",
-    "        \"STS-CE\": [0.1505, 0.3586, 0.4012, 0.432, 0.4575, 0.4469, 0.3809],\n",
-    "        \"STS-BE\": [0.3094, 0.4233, 0.4862, 0.5222, 0.5486, 0.5422, 0.4755],\n",
-    "        \"BERTScore-F1\": [-1.716, -5.206, -0.9048, 0.3596, 0.5919, 0.5613, -0.315],\n",
-    "        \"ROUGE-L\": [0.09565, 0.3769, 0.4612, 0.5006, 0.5204, 0.5019, 0.4271],\n",
-    "        \"BLEU\": [0.003134, 0.1139, 0.1639, 0.1828, 0.1926, 0.1766, 0.1203],\n",
-    "        \"meta\": {\"shots\": shots, \"linestyle\": \"--\"},\n",
-    "    },\n",
-    "    \"BLIP-2 Flan-T5-xl\": {\n",
-    "        \"STS-CE\": [0.1327, 0.4551, 0.4371, 0.4532, 0.476, 0.4812, 0.5006],\n",
-    "        \"STS-BE\": [0.2861, 0.5235, 0.4877, 0.502, 0.5315, 0.545, 0.5681],\n",
-    "        \"BERTScore-F1\": [0.3864, 0.576, 0.5802, 0.5833, 0.5829, 0.5792, 0.5765],\n",
-    "        \"ROUGE-L\": [0.0598, 0.4606, 0.4291, 0.4403, 0.4684, 0.4846, 0.5056],\n",
-    "        \"BLEU\": [0.0005683, 0.1405, 0.1188, 0.1258, 0.1609, 0.1771, 0.2052],\n",
-    "        \"meta\": {\"shots\": shots, \"linestyle\": \"--\"},\n",
-    "    },\n",
-    "    \"Kosmos-2\": {\n",
-    "        \"STS-CE\": [0.3261, 0.3052, 0.17],\n",
-    "        \"STS-BE\": [0.4014, 0.3732, 0.2999],\n",
-    "        \"BERTScore-F1\": [0.4971, 0.364, 0.4123],\n",
-    "        \"ROUGE-L\": [0.3258, 0.176, 0.1376],\n",
-    "        \"BLEU\": [0.03476, 0.005267, 0.005217],\n",
-    "        \"meta\": {\"shots\": [0, 1, 2], \"linestyle\": \"--\"},\n",
-    "    },\n",
-    "    \"Otter\": {\n",
-    "        \"STS-CE\": [0.3514, 0.324, 0.2897, 0.271, 0.2836, 0.2912, 0.2961],\n",
-    "        \"STS-BE\": [0.4074, 0.417, 0.3905, 0.3758, 0.371, 0.3682, 0.364],\n",
-    "        \"BERTScore-F1\": [0.3534, 0.3165, 0.3044, 0.2895, 0.2884, 0.289, 0.2808],\n",
-    "        \"ROUGE-L\": [0.2555, 0.2337, 0.2234, 0.2038, 0.1906, 0.1902, 0.1846],\n",
-    "        \"BLEU\": [\n",
-    "            0.003195,\n",
-    "            0.001461,\n",
-    "            0.001018,\n",
-    "            0.001041,\n",
-    "            0.0008764,\n",
-    "            0.0004788,\n",
-    "            0.0005822,\n",
-    "        ],\n",
-    "        \"meta\": {\"shots\": shots, \"linestyle\": \"--\"},\n",
-    "    },\n",
-    "    \"VideoMAE-base\": {\n",
-    "        \"STS-CE\": [0.3799],\n",
-    "        \"STS-BE\": [0.479],\n",
-    "        \"BERTScore-F1\": [0.6516],\n",
-    "        \"ROUGE-L\": [0.4375],\n",
-    "        \"BLEU\": [0.0987],\n",
-    "        \"meta\": {\"shots\": [0], \"linestyle\": \"--\"},\n",
-    "    },\n",
-    "    \"VideoMAE-huge\": {\n",
-    "        \"STS-CE\": [0.4405],\n",
-    "        \"STS-BE\": [0.5282],\n",
-    "        \"BERTScore-F1\": [0.645],\n",
-    "        \"ROUGE-L\": [0.4509],\n",
-    "        \"BLEU\": [0.1071],\n",
-    "        \"meta\": {\"shots\": [0], \"linestyle\": \"--\"},\n",
-    "    },\n",
-    "}\n",
-    "\n",
-    "draw_graphs(models, metrics, \"held-out-eval-videomae-baselines.pdf\", shots, 3)"
-   ]
-  },
   {
    "cell_type": "markdown",
    "metadata": {},